Foresight has spent the past 40 years backing frontier science too early or too interdisciplinary for legacy institutions. Foresight has helped inspire and organize 10,000+ scientists, technologists, philanthropists, investors, students, thinkers, pioneers, and dreamers, including the early 2000s space community where Elon Musk and Steve Jurvetson discussed the importance of commercial space and reusable rockets. Foresight’s Feynman Prize went to Fraser Stoddart and David Baker 9 and 20 years before each won a Nobel. Today, Foresight is one of the main communities where neurotech pioneers gather and explore the frontiers of BCI, connectomics, simulations, and uploading.

Allison also founded the Existential Hope movement, rooted on the conviction that we must imagine flourishing futures — in the tradition of ambitious science & sci-fi — as a precondition for building them. She is one of the best people thinking deeply about the future, coordinating humanity to imagine and achieve good outcomes, and working to accelerate the technologies and adoption that will get us there.

We go deep into all of that in this episode: how neurotech could let us connect far more directly than language allows; how BCIs are meaningfully improving people’s lives today; why connectomics and whole-brain emulation can keep humans compatible with Superintelligent AI; why uploading can give us a path to exploring ourselves and living vastly richer and more fulfilling lives than a single ~80 year narrative arc allows; the amazing possibility of living billions of years; and how nanotechnology could reshape the physical world, address our sustainability and energy challenges, and build the infrastructure to lift us off planet and explore the universe.

This is a ton and goes to show how deeply and broadly Allison thinks. I’m very excited to have her on the podcast, as I think she explains things clearly and bridges huge divides with care, vision, and rigor. Hope you enjoy!

Topics covered

• 00:00:00 Introduction

• 00:01:25 What is Existential Hope?

• 00:04:06 Visions of the future: nanotech, longevity, neurotech, AI

• 00:08:04 From existential risk to building better futures

• 00:16:15 How far away is nanotechnology?

• 00:24:26 Is a superintelligent AI a tool, or a new form of life?

• 00:37:33 How neurotech could connect minds far beyond language

• 00:41:56 Whole brain emulation: from C. elegans and FlyWire to the human “holy grail”

• 00:45:16 The connectome problem: a complete wiring diagram isn’t enough

• 00:47:19 Why it’s so hard to attract talent to emulation

• 01:03:04 Longevity, cryonics, and the case for living billions of years

• 01:22:24 How close are we to never having to die?

• 01:43:13 How Foresight operates to advance frontier science

• 01:54:06 From philosophy to Foresight: how a three-month stay became 10 years in SF

• 02:03:04 Allison’s vision for an optimistic future

Episode Links

• Juan Benet Podcast on YouTube

• Juan Benet Podcast on Spotify

• Juan Benet Podcast on Apple

Links From the Podcast Episode

Guest + Organizations

• Allison Duettmann

• Foresight Institute

• Existential Hope

• Existential Hope World Gallery

• Foresight Fellowship / programs

• Feynman Prize in Nanotechnology

• Mythos Ventures

Referenced External Papers & Projects:

• Whole Brain Emulation: A Roadmap - Sandberg & Bostrom (2008)

• ARIA Manufacturing Abundance

• Less Wrong

• The Bitter Lesson - Richard Sutton (2019)

• Causal Learning: Psychology, Philosophy, and Computation - Gopnik & Schulz

• FlyWire (fruit-fly connectome)

• OpenWorm / C. elegans connectome

• Wormatlas

• Max Hodak

• Until Labs

• Adam Marblestone & Convergent Research

Videos & Demonstrations

• Funding the Commons 2022

• Woman hearing for the first time

• DBS tremor surgery (spiral drawing)

• Cystic fibrosis: Progress Through New Treatments Graph

Books & Media:

• The Nexus Trilogy - Ramez Naam (2012-2015)

• Permutation City - Greg Egan (1994)

• The Beginning of Infinity - David Deutsch (2011)

• Gaming the Future - Allison Duettmann, Mark S. Miller, Christine Peterson (2022)

• Engines of Creation: The Coming Era of Nanotechnology - K. Eric Drexler (1986)

• Age of Em - Robin Hanson (2016)

• The Edge of Sentience - Jonathan Birch (2024)

• Superintelligence: Paths, Dangers, Strategies - Nick Bostrom (2014)

• The Longevity Imperative: How to Build a Healthier and More Productive Society - Andrew J. Scott (2024)

• The Precipice - Toby Ord (2020)

• The Myth of Sisyphus - Albert Camus (1942 French, 1955 English)

• The Culture series - Iain M. Banks (1987-2012)

• Unbounding the Future: The Nanotechnology Revolution - K. Eric Drexler, Chris Peterson, Gayle Pergamit (1991)

• Surely You’re Joking, Mr. Feynman! - Richard P. Feynman with Ralph Leighton (1985)

• “Letter To Posterity” - Petrarch (1350)

Juan & Protocol Labs

• Juan Benet on X

• Protocol Labs

• PL Neuro

• Disclaimer⁠

Transcription

Juan Benet

My guest today is Allison Duettmann, the CEO of the Foresight Institute, a nonprofit backing Frontier Technology for the benefit of life since 1986. Foresight runs grants, fellowships, workshops, and other programs across AI, biolongevity, molecular nanotech, and neurotech. Allison also founded the existential hope movement. She co-authored Gaming the future, a book about civilizational scale cooperation. She studied philosophy and public policy at the London School of Economics. She’s also a venture partner at Mythos Ventures. And I believe she’s one of the best people in the world thinking deeply about the future, how to coordinate humanity to achieve good outcomes and working practically today to accelerate the technologies and adoption that will get us there. So Allison, thank you so much for being with us today.

Allison Duettmann

Yeah, thanks. This is a beautiful spot. I’m excited to be here.

Juan Benet

Great. Let’s dive in. What is existential hope?

[1:28] What is Existential Hope?

Allison Duettmann

Well, first I should say I’m glad they described it as a movement. I’m not sure if I could call it that yet, but we’re definitely picking up people along the way. And I think it’s mostly just because if you look out there, the current media landscape still is filled with doom. That’s to some extent due to media bias. You know, like it always sells to have very jarring scenes to really like get people thinking about like worst case scenarios. And to some extent that’s evolutionary too and exacerbated by the media.

But nevertheless, if you look back into our past, there used to be a time when there was very ambitious positive sci-fi or when there was any sci-fi at all really that was new and that really enable people to look further. And so I think when you query people that have actually invented amazing science and technology or any other products that we rely on in our civilization, like what got them started, it’s often something like either it is sci-fi book that they read earlier on and it’s often like very similar books and they’re old and there’s none of them like it anymore. Or it’s something like a concept that they just couldn’t get out of their mind anymore that didn’t exist yet, but they try to bring it into being. Michael Nielsen from Astera calls it a hyper entity which is basically something a concept that is so powerful that it manifests itself into existence.

And so this existential hope I guess idea is mostly that we can focus all we want on the futures that we don’t want to have and we can focus a lot on the past that we don’t want to go down on. But unless we start thinking about where we actually do want to go, it’s going to be very unlikely then we’re going to end up anywhere good because we can just avoid avoid avoid but never actually get anywhere where we do want to go. There’s a lot more to say about it. We’re not trying to be prescriptive about the future. It’s not like we’re laying out this one path and now just like that’s the road map. Let’s go.

But I think what we’re trying to do is get people to think about for themselves what a good future would look like because it’s probably very different for me than for you. I’m very eager to dive into deep different concepts of utopias. There’s various different ones from pertopia to protopia to varia. But without getting all conceptual, like I think a good future might look different to you than to me. We want to have people explore that again, think about this again, and lean into this and then hopefully get inspired to build some of these positive worlds instead of some of the worlds that we currently are running from.

Juan Benet

Why do you think that media and public narratives tend to skew so dark and dystopic?

Allison Duettmann

Well, I mean, partially it’s this evolutionary bias. You know, we really need to be wary of the futures that not just the futures, the risk that we want to avoid, the threats in our life, right? Otherwise, you wouldn’t have made it very far, I think, in our ancestral environments. And that’s fair enough. I’m not trying to say that we shouldn’t be focusing on the wrist at all. I think another one that’s exacerbated by the media is just that, you know, sex sells, but also what really sells is

[4:10] Visions of the future: nanotech, longevity, neurotech, AI

like pretty gruesome crime stories and you know, horrible imagery and kids being abducted. All of this is triggering this evolutionary part of ourselves that is like just trying to survive.

But we’re now in a civilization where we don’t only need to survive anymore. We definitely do need to still do that. But in addition, we now finally have the ability, the technology, the civilizational knowhow, ability to cooperate to kind of lift our gaze a bit and think about, well, where could we be going? What would it mean to live in a universe where we don’t kill ourselves? How could we get there? What technologies might we want? What societal structure might we want? And that type of thinking is not really evolutionary ingrained in us. So, we actually need to train it and we need to think about doing it. Otherwise, we’ll forget and we’ll just end up in the red ways of zero sumness and collectively avoid different parts of futures that others would want.

Juan Benet

To make it more concrete, what are some example visions of the future that you’re thinking about? Like give us a sense of the scope of the distribution.

Allison Duettmann

Yeah. I mean, you could think along different axes. You know, one axis is the technological vantage point where that’s also something I guess that’s more foresight aligned, but we think about what technologies might be possible and then not only individually, but also how do they interplay with each other. So for example, we were founded on this vision of advanced molecular nanotechnology. The idea is that you can build with a lot more precision in an automated way and through that build much smaller build with a lot less waste. Build things that we currently cannot build. Build new structures, build new devices with new properties that might be stronger, that might be more durable, that might be less wasteful. You could clean up the environment. You could clean up all the mess that we’ve made as 20 century and 21st centurization that we have become. And that’s just nanotechnology itself, right? There’s a lot more potential benefits even through that like including like you know nanobots in your body. What could that mean for longevity and health?

Then there is longevity biotechnology as a field in itself that has grown a lot since I started at foresight and there you could think about what would it mean to live healthy long lives. And usually when we think about longevity people think about well I would just be like my grandma or you know like my parents over 70 80 90. Unfortunately they are not who they used to be and people are not very excited about extending that but that’s not what it’s about. It’s about like actually having the verality, having the mental capability, having the lust for life that you had in the early days and trying to either restore that or preserve it from the get-go.

And then there’s various other technologies like neuro, that one that we’re going to talk about, but they’re all kind of interrelated in the sense that ultimately if you want to live a long life, you need good neurotechnology or neuroscience because it’s the brain and it’s you that we’re trying to preserve here. It’s not just your body. And nanotechnology could help with all of this. AI obviously I think the interesting kind of connection is that AI might be the engines of design by which we can really design what we want to build in the science and technology space and then you have the engines of creation which might be things like nanotechnology for example or also biotechnology in some level that can like rearrange and manipulate and better edit the physical world to take these designs and implement them into reality. So I think that there’s a long way to go in terms of positive visions on the science and technology side.

And then there’s a whole other societal spectrum, you know, that you could go down on and various different ways that we can just be more peaceful, collaborate better in more positive some ways, keep our options open. That’s the Viatopia idea of like we don’t really know where we want to go yet, but we probably want to preserve optionality, freedom of information, free speech, the ability to actually have epistemically good conversations. There’s a protopia idea from Kevin Kelly or probably want to get all the way to utopia, but we want to get to just doing it a little bit better every day. And that’s how civilization has evolved. And look where we are. It’s pretty great when it’s great. So there’s different like societal concepts too of like how can we cooperate better and do so in a way where we create a world that is better for more people by their own standards.

Juan Benet

What are some of the responses you’ve seen so far from people that come in touch with existential hope? either the

[8:09] From existential risk to building better futures

visions that they’ve articulated prompted by you or in maybe their reactions to the idea space that you think are helpful in kind of generating a trajectory. Meaning from my perspective, the existential hope movement is a really wonderful counterbalance to the darker side that comes from the existential risk landscape where it was very useful and valuable to frame existential risk so that we could identify a range of problems and avoid them. But it also had the side effect of creating too much doom narratives that then also triggered like maybe the EAC effect of accelerationist landscape as a counterpoint. But I feel like essential hope is like similar counterpoint but very different in kind of like quality and perspective and vision. And so I’m wondering what are some of the ideas that you see percolating through the community or the kind of outcomes that you hope to instill.

Allison Duettmann

Yeah, I mean I think you’re totally right. You know, once you set like a pretty radical vision out there, then there’s usually like a counter radical vision and then there’s the EAL like or more the AS50 doomer movement that was eventually birthed. Then as a response, you had the EAC movement and now you have like the DACK movement in the middle which is more defensive decentralized differential acceleration. But I think in general like if you go back to the existential risk framing, I think it’s a really important framing. That’s the basics, right? It’s just that we survive individually. Maybe that would be nice, but also as a civilization. And then the same people that were really at the forefront of making the existential risk movement big and writing very eloquently about it, Toby and a few others, they were also the ones really championing the concept of existential hope at the beginning. So it’s not like one or the other, but they were basically pointing out that the same way that we talking about an existential risk like a basically an event after which I think they call it the expected value of the universe would be much lower than an existential hope event would be one where the expected value of the universe would be much higher. So like an event after which you’ be pretty sure that the universe would be amazing. That’s a very tall order. I think to be odd had something like emergence of life is one existential hope event or like going from unicellular to multicellular organism. That’s a tall order, you know, but nevertheless, I think we can also think about smaller existential hope events in our life. And nevertheless, I think the direction is a different one, you know, it’s just like focusing also on what you do want.

And things that I’ve kind of like seen percolate out of this is theme of world building as one in itself where rather than trying to be prescriptive about one future is like the main point is mostly just trying to inspire people to think about what a good future would look like for themselves. And I do think that when we have this work building competition and have had it now for quite some time and the worlds that come out are so different. I think we have a like some a gallery of over 100 different worlds or something and they’re all incredibly different but they’re all super creative and kind of amazing and I think it’s mostly this process of engaging with the future in a positive manner that we hope to see. And then hopefully through that you also get inspired to actually like get engaged and build more of the these things.

So ideally what we want to see is not just like having people imagining these worlds but then for example one of them was pointing out a future key societal coordination framework where like you actually have automated prediction markets that run on everything in the future and like have much better odds at like what’s actually going to happen maybe some of it also AI automated and then where humans come in was like we would still vote on like the future that we would want but we’d have much better information and so not that this was inspired by that world but there is a Futarchy organization out there now etc.

So I think like ultimately we want to have some kernels of ideas and then have people actually like go out and build them like fiduciary AI is another idea that I think would be awesome like an AI that actually runs locally has your best interest at heart can go out reduce search cost coordinate with other AI agents on your behalf negotiate with them and and ultimately come to agreements like there’s a variety of different things that were nothing stopping us from building this really apart from like imagining it and then going out and doing it and so we’re just hoping to inspire a bit of that.

Juan Benet

Do you have plans for maybe turning some of those visions into more evocative narratives like movies or short film or like things like that that could like help because there’s something really really powerful about certain pieces of media and art that communicate the idea space so well to lots of people.

Allison Duettmann

Yeah, we actually have a small movie that’s coming out soon. It’s animated movie. It’s on the concept of like tool AI and like basically because sometimes especially in our communities you tend to think that we really only need AGI and then that will solve everything else but ultimately I think why we’re excited about AGI is because it can do so many things that will benefit humanity right in an ideal case right and so there’s nothing stopping us from creating some of those benefits sooner I do think that ultimately AGI would be a very powerful technology but ideally we should already start thinking about today you know how we actually use AI to improve people’s lives in a very visceral manner, right? That would I think also alleviate some of the concerns that many people in the public have about AI. And so being a lot more concrete about that would be useful. And so we made a short movie about that.

And I think in general media creation is great. It’s getting easier with AI. Thanks. And that would certainly be useful.

We had a meme prize the other day where people actually submitted AI generated content, videos, memes and the winner was a voices from the future conversation where people reflected back on what it was like to live in the present when you did not have longevity yet or at least like lev like longevity escape velocity meaning when people would still die and in that future they wouldn’t they wouldn’t really age if they didn’t want to etc. And so they were reflecting back on like how cruel the present is compared to a future in which you have health available and accessible and that won and that is very much worth watching I think. So that was all AI generated and it was really quite beautiful. So yeah I do think media is important.

Juan Benet

Yeah, talk a little bit about AI and the maybe the last 10 years. Foresight has been at the forefront of lots of frontier tech both thinking ahead you know 20 30 years and beyond and helping people scientists and technologists and policy makers reason about the future. I think foresight was super ahead on AI and the AI landscape.

How have you seen that kind of unfold over the last I don’t know 10 years or so where like these communities together made a whole bunch of predictions that are now starting to happen at scale and a lot of the world is unfolding as a lot of people predicted you know it’s very different to be living it than just you know be reasoning about it into the future what are some of the things that like the landscape ended up different and how does that make you think about like the next I don’t know 5 10 years

Allison Duettmann

yeah I mean I should say I’ve only been with Foresight for a decade and it’s now our 40th anniversary. So, I’m a newbie. But with that disclaimer out of the way, what attracted me to foresight was the early archives and it was the fact that, you know, like 40 years ago, they had discussions about the future of AI in tandem with various other emerging technologies.

And you know, back in the days, I think people thought that nano would come earlier than AI or at least like people in the broader AI nano community that existed at the time. That hasn’t panned out to be the case at all. So for now it really looks like AI is going to be there first. But the two as I had already pointed out are so deeply entwined. You know like better AI might lead to better modeling and simulation which might lead to better nanotechnology and building which might lead to better chips, better like other infrastructure that would better energy that allow you to build AI faster. So they’re all kind of very deeply intertwined but nevertheless it looks like AI will be first.

Juan Benet

I think we can agree. Yeah. Yeah.

Allison Duettmann

For so that’s something it will have to be very different for that for nanotech. Who knows? Never say never.

Juan Benet

It’s possible.

Allison Duettmann

It’s possible. I do think that ultimately though, interestingly enough, like the main bottleneck to doing things will become the physical world, right? Like very soon, you know, especially when we can automate a lot, there’ll be this AI overhang. And so thinking now about what kind of projects, what infrastructure, what coordination we can put in place to actually tackle some of the physical world so that we can once AI starts to automate a lot of the non-physical world we can actually get going on the physical world because that’s going to be the rate limiting factor to progress very soon. [laughter]

Juan Benet

Yeah.

Allison Duettmann

And nanotechnology is the best way to actually rearrange the physical world and edit it the way that will be helpful for civilization.

Juan Benet

So how far away does nanotech feel today? I

[16:20] How far away is nanotechnology?

mean, it still is the sort of thing that a lot of people talk about as being many decades out, but it could be a lot closer.

Allison Duettmann

Yeah. I mean, I wouldn’t say many decades because I’m not sure what is many decades out at this point. It gets extremely difficult to reason about, you know, 20 years out into the future. But yeah, I would say that today if you think about where the capital investment is going, I don’t think that the large pools of resources take seriously the idea that you could come up with way of manufacturing chips and beyond with you know atomically precise manufacturing techniques to then generate again all the chipsets that you need or like the data centers or like the robotics or like new materials like all of this kind of stuff. Yeah, I don’t think anyone’s really taking it seriously. Still not.

I think like you know we’re maybe perhaps still one of the only organizations out there that is like making a case for it like seriously I guess if there are others I would love to meet you [laughter] but Foresight has like a great track record of building the communities I mean like the archives also have like the period where foresight was hosting all of the spaceoriented meetups like you know 20 years ago or 20 25 years ago that like Elon and Steve Jurvetson and others used to go to and then they were talked about oh great like the thing we need to do is get reusable rocketry and then Elon’s like okay great I’ll I’ll go do that, right? And like here we are.

Juan Benet

And he did.

Allison Duettmann

I guess for Nano, let’s see. There are some projects kind of like going into the right direction either from the bottom up in the sense that like there’s progress on some of the technologies and paths that we might want to use to get there. There’s also new paths emerging like there’s the protein engineering route, there’s the DNA nanotech route, there’s still the, you know, like SDM AFM, you know, actually like placing atom by atom route. Microscopes are getting better. We might be able to do it eventually.

So that kind of scientific stack or the tech stack is emerging from the real actual bottom up and at the same time you also have people building trying to build smaller from the macro scale that’s not really going to get us quite to nano but like there’s an area program I think on manufacturing abundance that’s interesting I think they’re really at least grappling with idea of like what it mean to manufacture better there is Neil Gershenfeld at MIT is interest doing interesting work on fab labs I think he was one of the creators of Fab Labs. And so he’s trying to just like get the scale down and down and down and down and automate FabLabs a lot more so that we can like maybe get eventually from 3D printing to like printing smaller and smaller. Again, we’re not going to get all the way to printing nano that way. Maybe I should never say maybe, but I think people are trying to grapple with the material world a little bit, shifting into the other parts of AI that might I don’t know like unfold over the next 5 to 10 years.

Juan Benet

What are some of the things you’re excited and I don’t know existentially hopeful about?

Allison Duettmann

Well, I think one thing where people weren’t quite sure about at the beginning and we’re still not quite sure of how AI is going to play out, I think, is it going to be more singleton like or is it going to be more decentralized? And so those are of course like big suitcase terms. When super intelligence was published Nick Bostrom book a while ago, it laid out this kind of like recursive self-improvement world where over time you might have more economies of scale and like basically dynamics that would push very much progress towards possibly creating one super intelligent entity whether that is like one AGI system or whether it’s a system by like controlled by like one entity that would very quickly be able to out compete and just scale faster build up more of a delta and more of a lead towards other AI efforts and then kind of like outrun eventually just because recursive self-improvement and scaling might eventually lead to such big benefits even small advantage points might matter so much because you can build on them right that you can just outpace your competitors eventually so that’s one path right how AI could be going and then another one is that actually maybe that’s not going to work because maybe what we’ll find out is that local knowledge is very hard to come by that there are these longtail problems that you want actually specialized super intelligent agents that can focus on one specific task really well and then instead of scaling them all the way up to a generalized system.

You want to enshrine them in some broader economy or ecology of other specialized super intelligent agents that do one thing really well and then you have this kind of work mind or multi- aent collective intelligence cooperative layer that you know negotiates manages re resources in between. And so that’s more kind of like the civilization and the economy that we currently see but hypercharged and certainly the current trajectory of the development of AI favors this distributed approach in that well so far in that maybe we needed a mega scale data center to do the training but the way in which the inference of transformers ended up working out is that it is very difficult to create that one singleton type entity and it’s a lot easier to have millions of small pretty intelligent systems Now they are all parting from the same model but the fact you can go from one model to millions of agents and that’s the easier thing to do than to kind of harness the computing power for millions of agents and make one ultra powerful singleton is like a weird path dependent thing that comes down from like how the semiconductor industry ended up like making chips and how data centers got shaped and so on.

I’ve definitely heard like, you know, the sudden spitter lesson of just like actually just, you know, throwing comput something and like scaling works pretty well. There’s still that theory. I think I just went to a talk where the narrative was more that like test time compute and scaling works so blately well that we haven’t really like got the final benefits out of it that we can maybe scaling can go a lot further and at that point you just need a lot more compute actually how the architecture eventually looks like that might still be a different question of course but generally you need a lot of comput.

Juan Benet

Yeah.

Allison Duettmann

And so whoever is building it still needs a lot of resources.

Juan Benet

Yeah.

Allison Duettmann

So you might have like these like mega scale resourced training environments. I think it is surprising and counter to a lot of the sci-fi that you would end up with millions of separate instances that are all part from the same model as opposed to like you know the classic example from a lot of sci-fi you have like one single ultra intelligent machine that everyone sort of talks to. you know, deep think is a good example or like PAL or you know, lots of lots of examples in narratives and it just turned out that like after you spend your billions of dollars of compute, you end up with like a terabyte of parameter weights and then you can run them and that’s like a weird sure outcome.

Juan Benet

That is true. But then there’s also differences between like you know who controls the AI versus what’s the architecture of the AI, right? And so you could even have like perhaps like one very powerful company controlling the AI while the architecture itself might still be like a bunch of different agents or something, you know, like a mixture of experts like working together.

Allison Duettmann

You have millions of agents but they’re all controlled by, you know, it’s possible AI Inc.

Juan Benet

Yeah. I mean, you know, AI Inc. put your favorite AI or in there.

Allison Duettmann

So I think like those types of dynamics are really interesting to watch. You know, I don’t think we know yet where we are at. I appreciate that there’s risks and benefits on both sides. The obvious one on the decentralized being you know you by giving a lot more people a lot more capability you also proliferate risk and the ability to do cause civilization scale damage in the form of bio cyber weaponry and then the you know downside of the more centralized vision is of course like that has not historically worked super well. You have single points of failure. They are vulnerable to attack. They’re vulnerable to corruption and you know maybe even accidents. And ultimately, you know, you might have some dynamics that come with that that might be pretty autonomy restricting eventually.

So, you know, there’s ups and downs to either side, but I do think that the kind of camps have evolved. You know, back in the days, people just cared about AI and they were like cared about AI safety. But now, I think we really see different kind of streams evolving that are more along the lines of like, you know, we’re trying to chart a decentralized path and this is where we think it’s actually going to be likely and normatively better to go or versus like, no, it’s actually just going to be scaling all the way and and it’s also safer that way. So, different camps and it’s very interesting to be at the front row of all of this and see how it’s going to pay out.

Juan Benet

How do you think about AI life?

[24:31] Is a superintelligent AI a tool, or a new form of life?

Because I think in all of these camps right now, most of them are kind of united on the front of AI should only be a tool and it should definitely not be a form of life or a form of having individual agency. But to me, it seems like that’s unlikely to pan out in the long term and also kind of shortsighted too in ways that leaning into charting a path for like a very good protopic coordinated path to building digital life might be like a good trajectory. I don’t know if you know how much you think about this and yeah I do actually landscape

Allison Duettmann

I do think about it a lot and I think it’s probably prudent to start thinking about it now. I don’t think we have an amazing track record at identifying life that demands our attention and moral consideration neither on the human level nor on the like other sentient creatures around aka other non-human animal level. So I do think that you know we are just not historically great at it. So I don’t think it’s too early to think about. I think in general there are definitely camps that are trying to avoid these sentient AIs. Then there’s others that are really trying to bring them about.

Ultimately, I do think that if we have incredibly intelligent entities that develop their own goals, that pursue them, that have longer context windows, that have the ability to access more resources, that are getting more agentic, more creative over time, you know, they will have these final goals, but through that they might develop instrumental goals, like things that they need to do in order to get to the final goal. And then over time these instrumental goals you know might develop into things that you curtail and by curtailing you might actually like harm or at least very much curtail that that entity.

And I think that ultimately humans we kind of developed the same way. You know, we had some goals that we wanted. Then we had like instrumental goals along the way. Eventually these mentor goals also to some extent evolved into felt goals like love for example, right? Like love very evolutionary beneficial. But love is kind of important to us now. You know, it’s like the thing that still makes people go around. And so these kind of is like the for many people it’s the top reason for fulfillment and for existing in life entirely.

Juan Benet

Yeah. And also for misery.

Allison Duettmann

So these instrumental goals are now felt goals. We feel them. like that’s the best way to make them instrumental and useful. So I don’t think we can really rule that out. There’s nothing in my head that makes me think that this is not a possibility. And so I do think because these systems currently in the way that they’re constructed are so different to us. So for example, you know, they can copy in the way that we’re currently building them, they don’t necessarily need to rely on procreation to procreate or on sex to procreate. So they will have copy and paste them and replicate them and there they are. So they will have different felt goals very likely, right, than us.

And so it’s going to be very hard for us to intuit it what they are. And so I think we need a better science around it. And there’s amazing orgs that are working on that. IOS in the Bay Area for example, they’re trying to kind of like create a better science of human and machine consciousness. Like actually trying to like create evils based on different theories of sentience and consciousness and try to evaluate different models according to how they’re fairing. not just on one theory of consciousness evil but across a spectrum of consciousness evils because everyone has their own pet theory of how consciousness works and so rather than picking one maybe like you have a breath of them. So I do think we need to get started on it and I don’t think we have an amazing track record that should make us super optimistic that we’re just going to wing it and it’s going to be fine.

Juan Benet

Very much agree. This is a really great segue into neurotech. How do you think about these kinds of questions in terms of organic brains and so on? Like one part is trying to understand the centurions of artificial intelligent systems but we barely understand consciousness or perception or qualia in our brains and our own mind. Do you think that some of those evas that people are making might end up translating into testable predictions and inorganic brains?

Allison Duettmann

I mean hopefully right like I mean there was a for a while I think there was a hope that we can basically get better at understanding current biological systems and then through them develop good evals for AI systems and I guess we’re doing that to some extent because that’s what evolves are somewhat still based on but on the other hand we haven’t really done that for animals all that well so now the hope is the opposite of like maybe by creating good evol for AI systems we can also apply them to other biological creatures including humans and others

So I do think either way of whether it’s evils or not, we’re going to learn a lot because we have to differentiate and distinguish ourselves eventually from them. We have to ask these hard questions of like and how far are these and we already are. Society is definitely asking themselves that people that get upset when they have their access cut off to you know like an older version of open areas or latest model. They are upset you know and like they’re probably companies are probably not going to do that again [laughter] without warning people. We’re already we’re already encountering these problems in the wild. So one way or the other we’re going to like have to make up our minds about this. It’s philosophy with a deadline how Nick Bostrom calls it. And so ideally we have good theories around it before we’re going to go all by matics and intuition and you know like just yeah it’s going to be complicated either way but yeah maybe thinking about neurotech are the things in across the field that you’re very excited about or that you think have like an enormous amount of promise either short or medium term.

Juan Benet

Well I think like at Foresight at least we’re like there’s there’s a lot we find interesting from like also more of a far future perspective.

Allison Duettmann

There’s like I guess three buckets. One is the newer AI lens of like actually finding out a lot more about how the brain works to inform and inspire different ways of building AI system and vice versa. The other one is the brain computer interface space where we have this amazing capability potentially to like restore augment function in significant ways. And then the third bucket of things is the emulation layer which histoly has been perhaps like a little bit more sci-fi but I do think there’s projects and people that are making good progress on that. And there the goal is like actually like can you emulate model organisms or like potentially eventually a human. And so those are the three buckets I guess that we broadly care about. I’m happy to talk more about individual each one.

Juan Benet

Yeah. Yeah. Let’s dive into the first one.

Allison Duettmann

Well the newer AI one I think is mostly from like an AI safety lens and like we are actually not funding or pursuing all that much work in that but I still think it’s kind of interesting. There are other amazing efforts for example from the Amaranth Foundation who put out a lot on the newer AI landscape and what we can actually say about interpretability by learning from human interpretability for AI systems also not just on the safety side but can we build better models by relying more on how the brain is working. Those are all I think really interesting questions and I think we can just learn a lot more vice versa on the both interpretability side but then also on the how does this actually work side right in both domains and that’s I guess partially what like NeurIPS was doing at the beginning the conference now it’s like very very AI heavy but it was really this interconnection between neuroscience and AI

Juan Benet

yeah like the science of how neural networks will work and like the math and systems of larger and larger scale networks the science of intelligence in a sense

Allison Duettmann

Yeah. So really interesting field I think and there’s a lot of academic research also in that area that’s interesting because it’s like quite theoretical. You don’t need a ton of resources. You need some compute resources but it’s not like it doesn’t require you to build new hardware devices all that much.

Juan Benet

And the other one break interfaces.

Allison Duettmann

That’s one where we’ve just seen like an explosion of amazing projects. Many of them in the bay and many of them with various different near-term and long-term use cases. But I think in general like we have a human brain. It’s working. So, so definitely working better at the beginning of your life than at the end of your life. And it’s definitely working better for some people than for others. And so you could think about it are there actually are folks that we could like lift back to where much of humanity is through brain computer interfaces.

And they have very different shapes, forms, sizes, levels of invasiveness, but like broad kind of like the restore use cases of like hey somebody has due to a genetic disorder or some accident have lost function in part of their nervous system. So either I don’t know they lost the ability to see or hear or move speak.

Juan Benet

Yeah.

Allison Duettmann

And you know the idea here would be to restore that you know to actually allow them to do that and I think that is probably the one that people most come across on the internet and I cannot watch it without having my eyes water because I think those use cases are just so hope inspiring.

Juan Benet

It’s how do you even talk about it?

Allison Duettmann

Yeah. [laughter] It’s just so impactful like we it’s come up in a few of the episodes so far. or like this one example of a woman turning on a coar implant for the first time and hearing or some other examples of being able to go from you know somebody with Parkinson’s and shaking to then yeah suddenly holding being able to move around in your own kitchen right like so if anyone listening to this has not seen those it’s just really worth like just googling it and yeah we’ll put the links in the in the description and there’s great news there in that the last 10 years have seen incredible advances in the manufacturing and building of the devices and getting them through clinical trials. So, we gotten to talk to a few of the people at the forefront of building these.

Juan Benet

So, like these are going to be in market like now or very soon like in years.

Allison Duettmann

I know. I’m pretty grateful actually that FDA and like some of these other regulatory orcs have moved fairly fast on some of those. So that’s been cool to see which is not often the case or faster than in other cases.

Juan Benet

Yeah, not quite fast enough as we might like

Allison Duettmann

but there are also of course some serious downsides to consider but nevertheless like it has really improved people’s lives. I’m just hoping that they become a lot more accessible and like just get rolled out fast.

Juan Benet

How do you think about those devices extending into beyond kind of restorative use cases into improving lives or kind of enhancing capabilities? Like one part is restoring vision, but like the same kind of device can enhance it. Like you can go from a limited part of your visual spectrum into like a much larger like able to see like eagles or see like very far away or see night vision like I don’t know. Is that kind of stuff that you think about or foresight thinks about?

Allison Duettmann

I mean, who wouldn’t want to have that, right? I think as a kid, you’re most likely to think about this of just like how do other animals perceive the world and like how much better are they at smelling or and other factors. Like I think definitely like the ability to be able to do these types of things much better would be great.

The ability to just remember better like you know if I could remember as well as when I did my A levels that would be amazing. And now of course as you age that’s not going to get better. And so I think, you know, uplifting folks that are more on the like older spectrum, but at the same time also, why only start there? I would love to remember a lot better than I did. Like even better than that, right? Like there’s nothing that is necessarily stopping. Yes, you could like remember all of Wikipedia or like remember.

Juan Benet

Yeah. There might be some things I want to black out.

Allison Duettmann

Yes. And selective memory.

Juan Benet

Selective memory.

Allison Duettmann

I think it also serves a point, right? Like we’re just not built to have everything in our mind at all time but like but the ability to tap into it easier and the ability to grow the muscles to at least remember the things you really want to like that would be great or learn new things really fast.

Juan Benet

Yeah. Learn new things.

Allison Duettmann

I mean that kind of skill acquisition kind of comes also hand in hand with the memory part right like we really want to be a lot better at acquiring skills to some extent. People are trying to do that through things like memory palace and like better ways of learning skills and stuff but install the driving skill application or

Juan Benet

sure learn to write a bike. I think the sky is the limit here. I think play piano.

Allison Duettmann

Yeah, I would love to. I’m very I’m not musical but it’s the one thing that I would love to do where I’m totally suck at and I would just love to be able to have that module in me.

I think another one is, you know, I think we’ve seen some amazing things what things like TMS can do for depression. And it’s not just TMS, there’s other modularities as well, like transcranium magnetic stimulation. There’s other ones as well, but like TMS, I think, is only still FDA approved as like the last resort for depression after you’ve exhausted all SSRIs, etc. And now you have people I think like Tim Ferris and a few others that are just kind of going out on a limb and trying to fund more TMS accessibility and availability because it has just proven especially when paired on specific supplements that he goes into at length in some of his writing on this that it has proven to be able to do the work of like you know like years of therapy if you do it right in the span of a few months and like when done in this accelerated enhanced version maybe the span of a few sittings And that would be great. You know, I’ve seen loved ones with debilitating depression and that’s been really tough to watch. And at the same time, you can think about could this be better for cognitive performance. There’s a few people that have tried to use that just and you can also think about could it be better for like modulating other parts of your life, right?

Like sometimes I’m not the person I wish I was. I have akrasia like weakness of will. I sometimes, you know, think I want other things than I do. And so like actually modulating your mood. Being able to like perceive when something is actually like bothering you and be able to kind of switch out of it, maybe meditate, maybe you know do something else leave the room rather than engaging with that path would be good, you know, like sleeping better, sleeping easier, you know, basically understanding what your mind and body tells you. I’m usually at least totally numb to it. I’m very an unintuitive non-perceptive person. Being able to understand your own mind and your body better and listening to some of these signals would be awesome. And then there’s a lot more like being able to interface both with computers and

[37:40] How neurotech could connect minds far beyond language

Allison Duettmann

also with other humans just much more deeply like the kind of bandwidth of communication. You know we have these mics we are doing a podcast. It requires a lot to set all of this up which I’m very grateful for. And of course, you know, ideally, I think that when you know people, you can mind meld with them already a little bit because you have so much shared context, right? But in other scenarios, like, you know, having just a lot better bandwidth communication with other humans might eventually be possible, right? Like you can just share context a lot easier.

And then likewise, just being able to process and have access to the all the information that’s actually online in a computer, being able to like a like get your queries out faster, being able to process information faster, creating those feedback loops of like actually becoming a lot more operating at a much higher bandwidth than we currently can with our fingers or something.

Juan Benet

Sometimes I find a difference between what I’m thinking and what I’m trying to write down is very hard to accept. Yeah, [laughter] this is also an area where sci-fi examples where BCIS are used to expand horizons or like explore the mind more or kind of generate like new kinds of experiences like there’s you know like the Nexus trilogy where people are able to like link together and kind of have shared experiences and kind of very deep ways or there’s this lesbi and more like connectics but it like in Permutation City people are able to like mind meld in various different ways and so on. So, I don’t know what of those have you like thought about or explored or are excited about or

Allison Duettmann

Yeah, I mean now we’re talking more like real speculation, but I think that’s ultimately some of the most exciting things when I think about what could a future look like. Of course, we can improve our material world, but the mind space has no limits or as far as we can tell, right? And like one book that I love is Permutation City. And even though many people think it’s a dystopia, I actually think it’s very utopian. I think you should have to read it and decide for yourself.

Juan Benet

I agree. I strongly agree in that you took a side, right?

Allison Duettmann

So, first of all, like it’s more an emulation book, but it goes from this like premise of like at the beginning someone’s kind of emulated against their will and then without wanting to take all the story away, but like ultimately they are in this kind of like very rich emulated world where they have these shared experiences with some of the other emulated minds that come along. They can mind melt in extremely deeply ways to share a lifetime in what probably in objective time outside looks like seconds or minutes or something but they have known each other for eons and you know really are able to share a life with someone.

I think something that I sometimes struggle with in the default world is just that you know sometimes you have a deeply transformative experience and you’re trying to explain it to someone else and it’s just not cutting it. You just can’t explain what just happened but you’re really longing to. And likewise when someone people that I love explain to me things I’m like I wish I had been there you know I wish I was in your mind I wish I could really comprehend what this means to you and that sense of like deep it’s much more even than empathy but that sense of like shared experience ultimately it’s like not being lonely right it’s like being more than one like having a shared experience with someone else or many people for that matter it’s something that I’m deeply missing I guess and would be amazing to have

likewise I think if you think from a productivity layer being able to actually like collaborate with others, you know, in a very deeply meaningful almost Borgl like manner where instead of just sharing your experience with one, you can share it across different minds, right? You can actually really like and you can I should say you should be able to select what you share. I don’t want to know everything that’s going on in your mind. Neither should you know everything that’s going on in my mind. But ideally being able to kind of have access control and like you know actually share the bits that are relevant for specific endeavors, civilization scale, galaxy scale endeavors that you might want to be taking on with other minds. And then work really efficiently with each other. Being able to like communicate and just actually build collectively, that must be something, you know, quite amazing.

And that’s just in human world. there’s potential like you know other sentient creatures that we might want to understand much better you know what it feels like to be them you know like I always wanted to feel like what it was like to be my dog I don’t think I ever will not what’s like to be that dog anyways but I think the ability to actually like mindmelt across different species and also have an ability through that to develop a deeper sense of empathy for them and compassion and just you know be able to live collectively better lives with each other on planet earth will be kind of amazing.

[41:59] Whole brain emulation: from C. elegans and FlyWire to the human “holy grail”

Juan Benet

You already touched a little bit on through the permutation city story into like the third category which is kind of the emulation simulation area. What is a an emulation or simulation of a brain system? How does that work?

Allison Duettmann

Yeah, I mean there are different approaches and opinions differ on that front. But in a nutshell, I guess one could say it’s something, you know, like actually creating or recreating a copy of an individual’s brain in an ideal sense that both has structure like a similar structure representation of what your actual the original brain looks like, but then also functionally is equivalent enough at least to the original brain that you actually end up with something that is like more or less looks like and behaves like the biological version, but now it’s just on a different substrate, aka it’s possibly like a digital mind.

And so, you know, usually that is the connectomics route that you know, people are favoring and pursuing here. So, you actually like slice very thinly [laughter] a model organism’s brain and then actually like image it bit by bit and then stitch these different images together so that you can actually create a connectome like a map of what the brain looks like.

At that point, you really only have even if that all works, which is very complicated process and it’s definitely not we’re not there yet for humans. But even if that all worked, you would still have to get function out of it, right? Or like if you even if you had a map, trying to actually read anything useful out of that, trying to like predict how that organism will behave, trying to also see whether it’s remotely similar to how the original would have behaved. That’s a very tall order.

Juan Benet

Concretely, the process is you have an organic brain. You’re trying to image it and extract the wiring diagram of all of the different organic neurons and how they communicate. And if you can replicate that with high fidelity or high enough fidelity, then you should be getting same exact functional output. And so you can make a high fidelity enough copy in a digital environment and you run it, then you should have a perfectly preserved snapshot of that person.

Allison Duettmann

Yeah, that’s the theory.

Juan Benet

The theory.

Allison Duettmann

Yeah. And so far I guess you know like we’ve definitely done C elegance that has that’s we’ve gotten the snapshot of C. elegans

Juan Benet

orations are not we’ve done the snapshot

Allison Duettmann

yeah and that was my part like I guess the kind of functional piece is actually kind of difficult the emulation piece but we’ve definitely done the snapshot and that’s we’ve done decades ago then recently like we’ve made real strides on the fruitfly and like the FlyWire project amazing project I’m not sure if you’ve had anyone on there from that

Juan Benet

not yet but we will

Allison Duettmann

okay

Juan Benet

you will

Allison Duettmann

great so they can tell you a lot more about the specific details of that but like that was a real breakthrough that made cop optimism inspiring. people are trying to hand a mouse different parts of a mouse maybe like stitching different parts together that’s a much taller order and then a human is still you know it’s that would be the holy grail of course you know and like just because it’s so difficult and like time and cost intensive to actually do the slicing that might be required for that to actually do the imaging to do the storage for humans like cuz human brains are just so much bigger and more complex than a mouse let alone a fruitly that all of the difficulties of a fruitfly just grow in complexity. So it’s it’s just a very very hard problem still at the scale. At that point then

[45:20] The connectome problem: a complete wiring diagram isn’t enough

you’d still need to map it right [laughter] and the mapping is I think really where things get totally astronomically more complicated.

Juan Benet

Yeah. So we were talking about connectomics work and extracting like wiring diagrams of organic brains and it’s not clear like what kind of connectivity between neurons actually matters in carrying the signal. So we don’t know how to construct the simulations yet to like effectively work. Where do you sort of see the field and what are the needs?

Allison Duettmann

Well, I think I mean different people are trying to work on different parts of the problem that you just mentioned. You know, anything from the annotation layer from the annotation problem to bring down the cost of like the actual imaging and that process of the actual imaging, right? Maybe we can have better microscopes for doing so, etc. So I think there’s different pieces of the puzzle like I think the compute problem honestly and some of the storage and the actual part of like once you have that what do you actually do with that? How do you run it long term to have like more of those and to store the information properly is also something that we haven’t really solved yet.

But I think in general like we also need more shots on goal and we just generally more projects because there are so many different research efforts that are trying to do a part of it. There are now luckily FRO you know that are trying to like actually create better tools that other projects then can then use. Bon lab has created a bunch of tools that other projects have used. So definitely people are trying to create more tooling for the ecosystem at large.

But the kind of patient capital and the efforts with the ambition of actually trying to see a project from start to finish actually see it through and trying to create an entire emulation. We’ve seen a few of those on like a national level scale or like a collaboration between different universities, but like having more shots on goal like that and actually allowing people the time and the experimentation and hopefully coordinating between them to create these more mega scale projects that I think I don’t see as much of yet as I would like to. We have different people building different tools, different people are having different approaches. Now we finally see some tools that can be reused. We see some shots on goal, but I think we need more than that.

[47:22] Why it’s so hard to attract talent to emulation

Juan Benet

For the people listening, like if you could wave a magic wand and summon assistance from the internet, like what are the things that you would point people towards? Is it just being able to form the capital required to like do the engineering projects? Is it hey actually no, we need founder level talent to be able to construct the engineering machinery like recruit the people form the team like run the thing or is it hey you need more scientists digging into the problem space and kind of thinking about it at a concept level. What are the requirements?

Allison Duettmann

[snorts] I don’t I mean, how powerful is this magic wand? I mean, I should say that like I’m not a scientist in any of these domains, but like from a field building perspective, I do think we need a lot more talent. We need talent that also has a level of ambition to want to go there and like work on these hard problems. I think one thing that you often see in academia is it’s like excellent work that’s coming out of it. You can usually be sure like this is actually great. But I think one thing that’s a problem of the academic incentive structures, it goes by like how can I like advance this field a little bit forward? you know, how can I, you know, write a grant application for a specific domain that is like good enough that it gets likely funded where but I’m also sure enough that I can actually make some progress here where I can publish afterwards, right? And so I’m not saying everyone’s doing that. There are a few rogue scientists and researchers that are very much just having some build up enough credibility to be able to do really ambitious work. This is also not their problem like it’s a problem of the field and the center of structures that they are.

But to some extent we need to think like the opposite way around of just like where do we want to go and then [clears throat] think about what needs to happen before that what needs to happen before that what needs to happen before that. And so I think like having this kind of shot on goal mindset rather than the like let’s advance where we currently are like one step at a time is kind of a problem that like a requires a mindset shift but then b it requires a funding behind it and it’s not a pot of funding once because in order to convince some of these people to leave their positions that they’ve worked 10 20 years sometimes longer for to get they finally have tenure and then you’re trying to convince them to put it all on the line to join this rag tag team of people that are trying to have a shot on goal on emulation. That’s a crazy ask for anyone, especially for these people.

So, you need to signal that this project is sustainable enough that they can actually put their career on the line for that, right? And the scientific ecosystem for better or for worse, but also for good reasons, is like extremely unforgiving in some of these areas. And so, like these people putting their careers on the line, like this needs to have real signal. And if you want to get the best people and for this type of stuff, you need to get the best people because generally not that many people are thinking about this type of stuff generally and the ones that do have a lot to lose, right?

Juan Benet

This feels similar to the BCI space pre neural link, right? Like I think pre there was maybe one or two companies that were kind of chipping away at the problem but very far away from commercialization and the appearance of Neuralink as the first extremely serious engineering very fast-paced project with very large capital like kind of checked a bunch of like required boxes opened the Overton window drew a lot more people into the field galvanized attention shaped the first range of devices popularized the idea drew a number of people out of academia into the field and almost independent of whether or not neural link specifically with the neural link specific devices are the ones that catalyze the BCI landscape already that effort move the entire field forward massively and we kind of like need a version of that for simulation it also had the benefit of having Elon which that’s just very hard to replicate

Allison Duettmann

yeah but we probably have other people that can make that kind of approach

Juan Benet

yeah

Allison Duettmann

big benefit of having Elon was like you had the network immediately and the kind of signal if you want to do something ambitious this is where you go and because some people you know that I met through the Foresight ecosystem. I just like I did neuroscience because I knew Neurolink was a thing and I wanted to work there, you know, or like I started building devices because I knew, you know, I would eventually could maybe get into Neurolink. It’s not only like that researchers switch them out of academia there, but like it created this whole new field of people that were just like I want to do what Elon’s doing. I want to be a part of that and you know, I’ll do my best.

Juan Benet

There’s some piece of this which is having a very clear sense of how to tell the story publicly broadly and we are committed to doing this very credibly that I think Elon gets you know a number of other founders also get it but this is often not well understood in academia not well understood in corporate research labs or maybe more conservative corporations like there’s just very real and tangible catalyzing value to having singular leaders who understand the problem domain enough can spell out the vision, can commit to it with the resources, energy, talent required to form the pieces and more things can get catalyzed that way.

I mean, I think like a smartphone was a version of this. Apple just gung-ho doing like doing it and building the first major scale smartphone, releasing it into the into the ecosystem, shifting opening the frontier to say forget about like the keyboard like a small screen thing like we’re going in this other direction and then rewired the rest of the phone industry to go in that direction.

Allison Duettmann

Yeah. And I guess like I think I once saw like a lecture I think it was Elon at Stanford where he went into deep into Fermi paradox land and was like this because of the Fermi paradox we need to be a Mars civilization and because of that we need things like SpaceX because of that we need reusable rockets and like really just like started at [laughter] the like there’s the thing like the Fermi paradox we need to become interplanetary we need to go to Mars that’s why we need to do everything else that SpaceX is doing.

So I think this type of vision right and for Neurolink it was similar it was like there was the AI vision of humans if we want to be compatible with the long-term future in which AI plays a role we need to be able to uplift ourselves improve ourselves like enhance ourselves or at least have the option to do that and so for that we just need newer link and for that we need to start at the restoration piece but then like we need to move ahead and so to some extent that sounds crazy I mean even SpaceX sounded crazy but the having that vision and being able to actually get that into people’s minds in an inspiring way is something that he’s been fairly good at.

Juan Benet

Yeah. Yeah. And I think there’s a range of other founders in the space taking that [laughter] course that view too. I think Max is a good example of pushing in this direction where not only did Max help shape the beginning of Neuralink, but now he’s pushing the frontier of all of this sensory restoration and pushing into the very high bandwidth BCI type use cases.

Allison Duettmann

Yeah. I mean, he’s unapologetically like just doing that thing and is also a very clear communicator about that. You definitely get that same inspiration, I think.

Juan Benet

Yeah. And I think yeah for our progress to advance faster more successfully and so on we just need more Max and Elon type people that can form these efforts and push them forward. How do we attract or generate more of those people?

Allison Duettmann

Partially I think it comes from a deep obsession with the future. I think if I remember for Max it also started very early like I think there was rarely ever a time where he wasn’t obsessed I think with the future. So whenever you ask people what inspires them it often started pretty early. So you need on the one hand like a kernel of this deep want for the future to be a thing and to have the audacity to think that you can have an impact on that and then you need somehow it paired with a sense of like technical ability to at least figure out what’s important and neglected to do there and then like to some extent that’s not always the case but it definitely helps to have some financial freedom or at least have access to capital that and be convincing enough to attract it to be able to like push for these visions and I do hope that at least the last piece through, you know, some of the writing that we’ve recently seen out of like, you know, this third wave philanthropism, that might be coming soon, thanks to AI and thanks to some of the whether it’s secondaries, whether it’s IPOs, whatever you have in the AI world right now.

And the SpaceX IPO of course also being a potential factor there, but like people being soon in a position where actually they can put some more money behind that. And we’ve seen that in crypto as well like and the people that were interested in crypto early and the people that were interested in AI early first of all there’s there sometimes similar people or the same person same people and then the second point is like they both if you’re interested in crypto and if you were interested or like early and if you’re interested in AI early that’s usually a pretty good reference class for also being interested in other really interesting things that are very important for the future and so those types of people I have high hopes for that a few of them might like branch out and just do something and we’ve seen it for example Sam Altman you know like being an instrumental part in Retro Biosciences in funding Retro Biosciences and funding merge epistem like projects like that right and so we already see that and same with crypto right so I think that’s good and hopefully in the future some of that money will end up to these very ambitious and certainly I mean definitely the AI field because it’s so visible and working and so on has attracted substantially large number strong founders to go and tackle a lot of these problems.

We’re also seeing that in a range of hardware domains now where 10 years ago it was extremely difficult but post the Tesla and SpaceX successes that generated both a crew of people that either spent time in those companies learned how to do it and then went on to build other companies or where at least the existence proof of Tesla SpaceX enabled a set of people to form the capital to actually go do it. We now have significantly large numbers of good companies in the deep tech landscape.

But somehow I feel like we still need to draw larger numbers of those type of founder type people into the neuro tech space and certainly into uploading right like into connectomics like this is like still if the new tech space is undervalued then within that the emulation space is just I mean like I think until a few years ago when we did this emulation workshop trying to kind of look back at the original road map from 2008 from Anders Sandberg and a few others.

At that point, I think we were like 30 people in this like tiny convent like space in Oxford and it just seemed like is there even a there there you know can we talk about this stuff now or is it even at a point where it makes sense to kind of dust off some of the old road maps and where we can look at this again of course there has been a ton of progress individually in that I shouldn’t be saying the space there isn’t even like the space yet but in the space is relevant for emulation there have been a lot of progress over [laughter] the last 20 years even before that but I think the concerted effort of like thinking about like could this be a possibility like if so what would be required for this you know what are some of the bottlenecks I think only over the last maybe four or five years or something I’ve even seen people reconsidering that and so it’s fairly fresh and new you know in the sense that people think that this is now a possibility that has been partially inspired because they think that we might need it visav AI systems but also partially just because we now actually might have the tools to make some progress on this or at least like seriously consider it again and So I think it’s generally a pretty exciting time but it’s very early stages for these areas and we definitely need talent. most of the neuroscience space doesn’t really think of that area as something that they could be going into that is worthwhile exploring and so I think there’s a lot of education required incentive structures required like all the things that you need for starting a field right

Juan Benet

we talked about some of the how and kind of like bottlenecks in the field and so on and we kind of touched a little bit on the edges of this but let’s maybe get deep into the why of emulation and simulation like the amazing upside case here like what would it look like to get this right and to be able to actually simulate entire brains

Allison Duettmann

the space is just so broad we on the one hand the kind of like sub field of this could be beneficial visav AI systems getting rapidly more intelligent because you know in the emulation space you might have something like an actually human aligned AI system in the big if that you get this right and that you get this done soon you have something that is like an artificial general intelligence that is not based on an totally alien architecture but you actually have something that is like you know more or less like trying to emulate as much as possible at least how humans work. Humans are not perfectly aligned, but humans are like somewhat aligned or at least like we know the pitfalls and dangers of humans a lot better than we might know those of alien minds where we don’t really know their motivational factors etc. And then you’d still have to align like humans and find architectures of cooperation under which they can these MS could coexist. there’s the question of like how compatible or competitive are they even with artificial intelligences when would they even arrive questionable whether that could be done before AGI but even if it could be done after AGI or like in tandem with AGI in some manner you might still want them as a long-term hatch for a long-term future in which you as someone that is human originating wants to be competitive or compatible with AI systems or compatible with the future really at all. I never would like to have a future in which that is the only way in which you can exist.

I do think that we need that freedom of choice and ideally we can wall off a garden in which humans can just be humans those that don’t want to be kind of along for the ride. So that’s kind of like the sci-fi case and the AI safety case for that. I don’t know how sci-fi it actually is but it’s worth consideration.

And then but even apart from that right like the emulation cases are really strong on the one hand like if we had an actually emulated simulated mind or brain we could learn a lot about how humans work what makes us tick really like understanding humans on a on a deep deep level which might shine more light into the theory of into the problems of consciousness and sentience. It might provide us with a lot more of an interpretable standard of like what do different interventions, drugs do to the human mind? Like how can we actually improve some of the kind of failure cases that we know that are actually like deeply ingrained in the way that our brain is set up at least for some people how it’s working. Can we like intervene in that very precise level and like actually have a mind that is not like optimized for something external but like that works the way that we would want, you know, that like is a little bit more reliable at the things that we would want, right? you can actually have this surgical more than surgical precision of hopefully modulating and helping the brain like live up to some of the hopes and dreams you might have for it. Restorative but also augmentative.

And then there’s the whole part that we just explored that is like more long-term and in the sense of like do you want to merge with potential AI systems? Do you want to merge with other humans? If you emulate in a sci-fi scenario at least it is explored by like Robin Hanson for example in HFM. You could like let emulations run much faster than human brains can. Or you could have them be run slower than humans. If you want to like extend or contract your subjective time, you could allow yourself to live in incredibly rich virtual environments. And often times people are like, well, but then you would just have these worker emulations and they will all live at subsistence and these automata that produce work for you. But the thing about virtual worlds is that it’s not very expensive to make nice ones.

Juan Benet

Yeah.

Allison Duettmann

So dramatically [laughter] cheaper than a

Juan Benet

or organic.

Allison Duettmann

Yeah. I mean this table like it’s hard to move, it’s hard to create, it’s expensive. The energy requirements are like economically different.

Juan Benet

Yes. You need compute and you need some good designs.

Allison Duettmann

But apart from that like worlds or the emulation space is your oyster, right? Like you can really emulate a lot easier amazing simulated environments that you can rearrange your physical worlds to your liking. And so if we had these emulated minds, it’s not clear to me that they would lead a terrible existence because it’s just so easy to make them have a good one. And so you just have a lot more space of creating amazingly rich experiences for people, having them like really like experience what it’s like to be in someone else’s shoes and all the ways that we currently can only like conceive sometime in like heightened psychedelic states, maybe like actually into it or something. But it’s just the virtual world is crazy vast

Juan Benet

with your like existential hope perspective. like what is concrete picture of like awesome future with like digital existence

Allison Duettmann

thinking up new scenarios, new environments that you want to explore that like they could either be mirrored on existing worlds or entirely new worlds that you might want to see. So I think the exploratory side is like almost unlimited on the connectivity side. I think like actually being able to go really deep with the people that you love like actually be able to like yeah mind melt for lack of a better term

[1:03:05] Longevity, cryonics, and the case for living billions of years

Allison Duettmann

with these people. Being able to learn a lot, right? Like I remember when there was a time when I was young and I thought I could eventually read all the books that there were that I wanted and then of course you eventually realize it’s totally impossible right now. I know that is a crushing moment when you learn that you’re [snorts] like wait hold on there are more books than I can read in my lifetime.

Juan Benet

Yeah.

Allison Duettmann

And usually like the more you learn the more you know that what you do not know but also the more you know the things that you would like to know. In a world in which you have like close to infinite time, you know, you have all the time, you have the skills, you have the resources, you have also the kind of learning environments to actually comprehend reality and life and other humans, the sentient creatures at a level that is totally unimaginable to us.

This is a piece I love about permutation city also like where this is not spoiling anything. One of the characters just can experience a very large array of different lives and experiences and just schedules for himself like just experiencing all of these different types of things and like spends you know entire lifetimes learning how to play a particular instrument and becoming like amazing virtuous.

Juan Benet

Yeah. Yeah.

Allison Duettmann

And then just goes into like intricacies of like woodworking in this particular way or like mastery of this like other artistic thing or an amazing exploration of all the lives. I do think that putting my David Deutsch hat on or beginning of infinity hat on that even then we’ll still have the problems that we will make up and compose so many amazing worlds experiences and you know things to learn that we will still have the problem of like we will never get it all done possibly but that’s kind of a problem that we have to live with I guess

Juan Benet

that’s a pretty good problem to have you know

Allison Duettmann

and I do remember like as Deutsch says problems are soluble problems are soluble and then you create new ones and you try to solve those again and you create better worlds through that. And we now look at our life from a longevity perspective. There’s this kind of like deep sense of almost well sadness for sure that we kind of have to pick one lane more or less, you know?

I mean like okay, if you look at our parents’ parents, they really had to pick one lane. They really had lived one life, one possibility. Then we kind of came along and we’re like, “Oh, we’re actually going to like do things more open source and we’re collaborate a little bit over here and over there and like we’re going to move to different countries and like you know, we’re going to do a few different studies.” And so we are multiple careers in your life.

Juan Benet

Yeah, you can have multiple careers.

Allison Duettmann

So we are already living kind of like more I think but there’s still so many other careers I would have liked to have right if only I had the time to do that or just like things I would have loved to experience and live.

And so I think we have to pick like one narrative arc in our life. There might be different kind of sub stories or like chapters in it, but it’s like this one narrative arc and in a world in which a you could live a lot longer on simulated or emulated time and you could just have experiences much faster, much richer, you can live all the worlds and all the lives that you can currently make out and then some. So that I think is something that is truly inspiring.

And then I think the final enemy is that if you think about super longevity from a biological perspective, if we let’s say said that we’re not going to emulate and we are somehow going to live in our biological bodies for a very very long time. This is now going into like sci-fi land a bit, but like nevertheless it might be possible.

Juan Benet

Not that far out

Allison Duettmann

that far out or like these days people a lot of people are both making significant advances in bio longevity, but also taking it a lot more seriously. Like it’s it’s become part of the public conversation when we were talking about it 10 years ago and beyond like you couldn’t really use the word longevity publicly that much. People would be like, “Oh, this is weird. Like this is strange.” And now it’s, you know, in major podcasts like part of the public conversation. A lot of people want to extend Healthspan. There’s still a massive long road to go, but like people are a lot more aware of the problems and working on the challenges and excited for those solutions.

I do hope that that’s the way that uploading is going to go and like in 5 years or maybe even sooner we’re going to sit here and be like, “Oh, do you remember that time like that wasn’t really like a thing yet and then at that point it’s so obvious that it would be crazy to do yet another podcast about it or something.” [laughter] So, I do hope that we’ll get there. I don’t think we’re that far away from what I think we’ve recently seen of how the neuro tech space has exploded.

But nevertheless, I do think what I’m talking about here more is like super longevity, you know, like now imagine like really living indefinite 200 years, like much much longer, right? Like let’s just like put that head on like and think about it for a minute.

Juan Benet

How can the sun?

Allison Duettmann

Yeah. But like not outlast the sun, but like outlast a lot more, right? Like actually be around for like the next few billions, billions, billions of years. And then eventually you have like something maybe like the heat death of the universe. And like even if you think about hyper longevity from a truly existential hope perspective eventually the universe might come to some kind of end like various different theories of how that might be but eventually like you know things would just possibly like drift apart you know like there’s a lot less going on in the universe.

And the cool thing about permutation city is that it kind of not necessarily has a solution to that, but it’s like it has a pretty good hack for that. And I don’t want to take the ending away, but I think if you think about emulated minds and mind uploading, there’s just a lot longer in subjective time than you have even than you had in a super longevity time in the universe that we have physically because you can speed up, slow down at your will. So without taking the ending away of Permutation City I think the actual ending of it is the thing that actually gives me hope and if we could do that that would be pretty awesome.

Juan Benet

Yeah. Well that’ll be a great teaser for folks out there.

Allison Duettmann

Yeah. We’ll put the links to Permutation City on the description.

Juan Benet

Speaking about longevity you know real shift in people like now understanding it. Certainly this is also something that foresight helped pioneer and helped a lot of people think about like the early cryonics work and thinking the foresight community helped shape a lot of it. Where is the field now? And well, I guess there’s the bilingity piece and there’s like the cryonics picture in terms of how do you avoid catastrophic outcomes in between now and then. You know, how do you do medical time travel? Where is that landscape today? What are you excited about? What are you hopeful for?

Allison Duettmann

Yeah. I mean, first I should say that I actually found foresight because I was really into cryonics when I was very young just because I was very disillusion. Well, first I was very into longevity. I didn’t want to die.

Juan Benet

Then I thought why actually sorry to interrupt you but well I mean like why would you want to like

Allison Duettmann

for me it was more like the opposite like I couldn’t really imagine what I thought it was preposterous that a we had to die and then also b that it was just seen as an okay thing or something like there are valid reasons why some people might want to yeah it should always be a choice and like you have like very very serious cases where like you want to just end the experience and like that’s should be a right but sometimes people talk about it in terms of the opposite.

Imagine a world where life is just continuous and it is expected that you would live forever until you don’t or like you would live until you don’t want to anymore and then somebody introduces the concept of like no now we’re going to start killing everyone after you know 80 years and we’re going to like introduce a change into society that’s going to make your body degrade and decay goosebumps when I think about it down and like and then you’re going to turn into dust.

Juan Benet

Yeah. Oh man.

Allison Duettmann

Like that would be a bummer, right? So it’s crazy to think about that that could be a world that we’d be in, you know, and then we’d look back on the present in this manner. So whether you think we just have to take the default world as a given and then modulate our expectations based on what’s currently possible, then I think framing of like death is fine, let’s just make the most of life is fine. But if you actually think like no, we can actually restructure and modulate the world to match our expectations, then the notion of accepting death is actually like pretty painful because then you’re just like selling yourself short, right?

Instead, what you should be doing is to be like, okay, there might be a chance where we can actually extend life a little bit and like if so, what would it mean to work on that? And then for a long time, probably nothing would happen. And like it’s just kind of like psychologically very unsafe to take that step because first you have to accept death is bad and then you have to accept we don’t have a solution for it and that’s really hard and that’s kind of where most of the field is in unless they trick themselves into thinking oh no we can solve death easily in our lifetimes which also some people are in.

From my perspective like death is good has been like one of the biggest cope mechanisms in you know human history. We have invented so many narratives about it just to like help us get over the fact that we like don’t know how to deal with it or like we didn’t know how to deal with it and we’re finally in the frontier of solving it.

Juan Benet

Yeah.

Allison Duettmann

There’s nothing graceful about aging, you know, really on a on a biological matter. It just means you’re working less well. You’re less the person than you used to be or want to be. So

Juan Benet

So you were a kid, you didn’t want to die. [laughter] Then what?

Allison Duettmann

Well then then I got into this deep existential crisis of like well but I still have to because there isn’t really anything like it right now that’s a panacea against like you know we don’t have to die no matter what people tell you it’s just like we don’t have someone that has lived for 300 years right now around.

So then I was like okay what else is there and then I came across biostasis or cryonics and you know I was still in Germany in Hamburg at the time and up on it this concretely is preserving the body and putting into a near frozen state or like I guess technically it’s not freezing because it’s not vitrified you want glass you don’t want ice and you can preserve as in sci-fi books you go into like cryostasis and you freeze the processes of the body you stop and then later you can rewarm and restart and so this can be helpful to like avoid disease, avoid, you know, if we don’t know how to treat a disease right now, like we might be able to know in the future.

So if you have like some terminal disease, you can like go into the cryostasis chamber, wait, you know, 5 10 years until we know how to treat the disease or longer and then kind of like recover.

Juan Benet

Yeah.

Allison Duettmann

In theory at least, you’re not really allowed, I think, to curtail your life to choose cryonics as an option. nor I think would it be super wise to do so right now based on where the science is at because there’s a few companies that are doing the work that they can and organizations like Alcor, Tomorrow Bio and a few others and people are preserved already including many foresight people but currently like at least and again I’m not a scientific expert in any of these domains but the way that I read at least you know the level of sophistication that we have in current preservation technology is just like we might not be all the way there yet that you have good chances like now there’s Ralph Merkle who was an early foresighter.

He always said if you’re on a boat and the boat is sinking, you know, you’re either going to face shore death or someone throws you a life jacket, it’s still really far away for a rescue boat to arrive, etc. So, your chances might still be very slim that you survive, but would you take the life jacket?

Juan Benet

And usually, I mean, yes, you would.

Allison Duettmann

And unless the costs are enormous, like I think it’s still useful to sign up to Quinonex now. And the costs are not ginormous for many people. you can choose life insurance to actually pay for it. That’s how I did it. And especially if you’re you should try to do that young because the younger you are, the better terms on your life insurance you’re getting. And so and you can still at a later point decide actually that your life insurance goes somewhere else. You can change the beneficiary of your life insurance later.

But you should no matter what you do, you should get life insurance because either you might benefit your family from it or you might use it to pay you for your preservation. But either way, like we’re just not there yet that we can change though, right? There’s some new companies that but I’m saying for people that are currently I think already preserved I think we would need to wait longer to rely on more advanced nanotechnology or something like that to be able to you know to put the back together again.

Juan Benet

Yeah.

Allison Duettmann

you know, like in a way we just need like what people think in the cryonics world is just we need to wait longer for them. You know, I do think in thinking about this, I think it may turn out to be that all those mechanisms ended up working out just fine because super intelligence should be able to reconstitute a lot of these pieces. I think it’ll come down to the to questions around do you have enough data points to interpolate the human in between.

Juan Benet

Yeah. I mean like how much of a book can you destroy but still decipher the story?

Allison Duettmann

Yes.

Juan Benet

Right. Things like that.

Allison Duettmann

Yeah. And then of course this is just like what is the actually like truly preserving a human as they currently are preserved state of the science. I’m still going to go for it. But then there’s a lot of other amazing things happening in between of like can we just advance the field forward so that at a time in the future we can actually make sure that the organisms that we’re preserving we can actually prove that we can already reanimate them. And that’s what more recent companies are trying to do. They’re not necessarily trying to do it for humans right yet but like for other non-human animals or for like you know for organs.

And then of course for the organ case you have a pretty big use case already because most of the organs I think currently in the US don’t get to like from the donor to the new host.

Juan Benet

Oh yeah.

Allison Duettmann

Not because the organ is bad but because of the times that it takes to actually allocate the person that they would benefit and then get them to where they’re a tragic problem. And so like if you have a condition where you need a new organ, you know, a kidney or a liver or something like that, you get put into a wait list and then because we don’t have good organ preservation technology, you basically have to have your operation to get the organ from a donor like a number of hours after their death or their donation, which means that unless a particular compatible donor happens to die close enough to when you need it, if like the timing doesn’t work out, then you die.

Juan Benet

Yeah.

Allison Duettmann

And there’s like this tragic tragic tragic thing. So if we can get organ vitrifying to work correctly, like organ kryostasis to work correctly, that would save a gigantic number of lives.

Juan Benet

Actually, I don’t know if the organs that are getting from A to B, if they’re vitrified, cuz I thought that you just put them on dry ice or something and try to get them to

Allison Duettmann

Well, today they put them in dry ice type of environments, but I think the point is like today you can’t do organ banking like you can’t put them into stasis for

Juan Benet

I see.

Allison Duettmann

That would be much preferred. Then there’s all these other projects of like growing organs in other like animals or you know like but anyway that space of sin bio is amazing and it’s growing and it’s wonderful. It’s totally sci-fi and otherworldly and there’s going to be some really good news. I’ve seen some stuff that is not public yet bring all these stories a lot closer real soon.

Allison Duettmann

Well, you I’ll probably hear it on this podcast.

Juan Benet

I would imagine [laughter] less neuro related but maybe we’ll expand it.

Allison Duettmann

Yeah, sure. I mean ultimately you know this those are all continues to make you live and live see another day but yeah so I do think that like you know projects like until you know from Laura Deming is as a fantastic project in the space where I think if I’m getting the mission correct and you should just have her on and have her talk about it. But it’s really like this ambition of like we actually want to prove that we can reanimate like model organisms and like eventually like more complex model organisms before we do anything you know like afterwards to them.

And so how amazing would that be if we actually had like this would to me be worthy of the label of existential hope in the sense of like if an event after which my expected value of the universe would go up a lot would be if we could actually reanimate like basically put down and reanimate a complex animal maybe even a human and they would come back you know that would be and I mean to some extent we’ve done that already in the sense that sometimes you know we can store embryos for a while, right? U so that’s great when we actually declare someone dead later on, you know, like they are not quite dead yet. And that time actually from being declared dead to being able to be reanimated or not being declared dead, but like actually having your organs fail to like being able to be reanimated that also has grown. So like I think we’re making progress there as well.

But I mean in general like how amazing would that be if you actually had someone that was like out for days or you know for like a considerable amount of time and that you could then later on reanimate. That would be just mindshattering news. And this alone would save hundreds of millions of lives. Even if you discount the long-term cryostasis option if you had a limit of we are only going to do this for like max 30 years or whatever. This already is sufficient to help save the lives of hundreds of millions of people that die from diseases that become preventable over time.

The most shocking way of thinking about this is there’s this graph for different diseases where you can see how different interventions, technological interventions that were invented over time increase life expectancy. The one that I love looking at the most is the cystic fibrosis one where initially somebody with the condition would only survive a few months to a few years and then as new technologies appeared the life expectancy started jumping up into like 5 years 10 years 15 30 50 and now people you know have like a normal life or close to a normal life expectancy. And so if somebody you know depending on when you were born you’re sentenced to die at a particular time shorter and if you could just like walk into a room freeze and then reanimate when the disease is now treatable you have like access to your full life and like that’s gigantic life improvement for the world today that you know working amazing.

Juan Benet

I mean on the one hand those graphs you said like are amazing to look at on the other hand they’re so sad to look at. I mean amazing in like the shocking ways amazing that we figured out how to treat it. shocking that we aren’t solving it faster and like the potential if we could just like at least unlock medical time travel like that would be amazing.

Allison Duettmann

Yeah. And I think also the way that I think I had interpreted some of the until results recently is that like once you actually get something down to minimum temperature 30 or 50 years doesn’t really matter like basically like I think once you have something preserved it’s preserved. So you know like it’s not not necessarily like after 10 years you have to like look at the clock and be like oh we better get going. So I think at some point it just you are preserved. The technology and the solution might be exactly the same to preserve for 30 years versus 3,000. But question of like hey would you like to survive this cancer that is currently untreatable? That’s like a much easier question for most people today than like hey do you want to live 3,000 years?

Juan Benet

Yeah. Yeah. But I think the question would also be is that cancer treatable in 3 years or in 3,000 years or something like basically like

Allison Duettmann

but for that you can have contracts.

Juan Benet

That’s right.

Allison Duettmann

you know, like you don’t have to, it’s not a one-sizefits-all policy, you know, and like to some extent, I mean, yeah, it’s just so intuitive. And I think the big thing in there always when I think about it is just like right now I think civilizationally we’re so close to either really break through and become a sustainable K whatever number you want to pick civilization long term and like actually climb the cut scales and just create a future that is all the ways amazing that we try to touch on like throughout this conversation already.

I remember growing up thinking like still like reading all the sci-fi and like just being like wow one day this might happen and would I want to be a part of it but I’d never actually considered that I would and so now I think we are actually quite close to it thanks to many of the individual scientific advancement and then AI accelerating many of them and so I think now it would be extra sad to miss the cut off because it’s really so close the cut off meaning here like the transition into like actually being able to live a very long life if these technologies work out and we don’t kill ourselves through an existential risk through AI or whatever.

So, I think right now it’s especially bittersweet to think about that there might be one of the last generations around that might ever have to go, you know, and I don’t want that to be my parents. Ideally, I wouldn’t want it to be my grandparents and I don’t want it to be any people close to me, right? And so, I think that right now it feels like somehow extra urgent and I know that we’re not the only generation that thought so. Many others thought that they were like close to a takeoff or like an evidential collapse, but somehow it feels like extra passion these days where I didn’t feel like it was that urgent earlier because I just didn’t think that the upside was so immediate

[1:22:28] How close are we to never having to die?

Allison Duettmann

or something, you know.

Juan Benet

Yeah, it does feel like a lot of these pieces are coming together quite quickly and if we steer well through the next 10, 20 years, then we are in a very positive outcome scenario.

Allison Duettmann

Yeah, it will always take longer to get these technologies like fully realized and actually accessible to everyone and you know actually create like a society that can adopt them and have a regulatory system that can like wave them through when it’s time. So you know like there might be this other bittersweet moment of like we have it in theory some people already getting access to it you know to like we have a world that can accommodate these technologies. So I think that is definitely also a problem for the future that we have to deal with. But in general there shouldn’t be anything stopping us from kind of like really uplifting ourselves in the next 20 years or so if we make it through you know.

Juan Benet

Yeah on a couple things that are interesting. One might be what is the kind of policy and regulatory landscape that thinks about these agencies like the FDA and others is very preventative in terms of hey stop certain harm from happening that one person is causing on another and they’re not really well structured to think about global natural harms in the same way I tend to describe it as unfortunately we’ve put the FDA in this like really difficult position where they have to play this trolley problem of they pull a lever and run a clinical trial and end up accidentally causing the death of one, five, ten people or like a hundred people in a trial, they might get the cure faster, but then they did that versus if they don’t do anything then like millions die from that preventable disease.

And unfortunately, because of how we structured it, our society holds the FDA responsible only if they pull the lever, but not if they don’t pull the lever.

Allison Duettmann

Yeah.

Juan Benet

And so it’s this like really difficult game theoretic situation where we’re not holding our systems accountable to the totally preventable deaths that are happening from natural causes and like that is like just leading to gigantic volumes of death and harm.

Allison Duettmann

Yeah. I think you like you can’t optimize or hold people accountable before what you don’t measure and so making it very legible what the trade-off actually is and the opportunity cost that would be good. And there are a few people that are trying to do that like I think his name is Andrew Scott at the London Business School. He’s trying to basically create an extrapolation of how much economic growth, how much productivity, how much societal benefit, how much human life years are actually lost by not having aging technology sooner because it’s staggering, right? like we’re able to just increase health span of people that are currently alive but are retiring early because they can’t work.

So he’s trying to basically put numbers to this and just model out what the actual opportunity cost is of like not letting a aging interventions come to fruition sooner.

But then the additional thing would be to do that for a specific treatment, intervention, medication [laughter] to try to get that counterfactual going and then hold the FDA accountable for if you had accepted this or like if you had waved it through just like two years earlier, those are the amounts of lives we could have saved. This is the qualities like qualities adjusted life years that we could have enabled and this is the economic cost that you caused or whatever it might be, you know, like things like that. So we just need to put numbers to these things and we need to not only put numbers there but then also use them to hold the FDA accountable because otherwise we have this kind of tyranny of the minority of like these very legible horrible horrible cases where we still have the other horrible side of it that we have just accepted and it’s gigantic. [laughter]

Juan Benet

And it’s gigantic.

Allison Duettmann

Yeah. And many people are suffering for it but somehow we think it’s normal.

Juan Benet

Yeah. There’s like a tyranny of normaly that the world is always under where like the we just are somehow blind to the constant harms and terrible outcomes that are just normal and we don’t really think of them as strategies in the same way.

Allison Duettmann

Yeah. Because I do think it’s back to this like level of psychological safety, right? Like actually accepting how terrible it is and like having common knowledge that other people also know how terrible it is makes it even more terrible. And so the cope is just to pretend that it’s all fine and that there’s something graceful and wonderful about it.

And you know, so I think unless we make the types of technologies that could prevent some of these like aging states so likely and so near within our reach that it would be ludicrous not to flip over to the other side and accept the psychological cost to accept that the current state is bad because now you have this other thing that you can flip to and where those things are working and you can actually just defect from the normaly and death side to the other camp and have a strategy that works for you. I think we won’t get everyone until we are able to do that because it’s just too costly psychologically for people to accept how horrible it is that we haven’t solved this yet.

Juan Benet

Something I think you’re amazingly good at is understanding the depths of all these like incredible problems and yet you’re one of the most hopeful and bright and you’re filled with like all these amazing warmth and love and excitement. And I wonder how you individually kind of like navigate this landscape and how you both fill your own life with the excitement for the possible and such like a force around you both in the foresight community but beyond. You sort of like bring this energy of this set of problems is soluble. We’re going to figure it out. Here’s how we do it. Like let’s chip away at the problem and we’re making progress. And it’s first of all it’s great that you do that. Second, how do you navigate it and how can you like inspire others to do that too?

Allison Duettmann

Well, maybe partially is because I only have dangerous half knowledge of how complex and horrible the problems actually are. I think people in those specific fields are often just like it’s actually very tough dayto-day, you know, so I’m not grinding away in a lab every day. And so it’s easy for me, I think, to be more handwra about that or like almost like polyonically naively optimistic. I do want to acknowledge that fact.

On the other hand, I do think that many of the problems that seem insurmountable right now, I do think that in theory there are solutions to it. And so it’s really just kind of like a question of like are we going to do them or not? And there’s nothing you know like in the laws of physics that says that we can’t do them. So I like I think that you know you have to live like in the world in which it matters what you do like conceptually because if you live in the other world then you’ve already lost. So I think even if the chances are like small that individually I or collectively we have an impact the other world is just not like a non-starter from the get-go just because there isn’t really anything to do in that world right so I think once you have accepted that okay maybe the chances aren’t 100% that we’ll make it through but I’m still going to try then I think you’ve kind of dealt with the fact that it’s not going to work and then you can kind of forget about that fact and just focus on like the worlds in which it matters what you do and I think to some extent that also aligns with like yes I worry a lot about the risks.

I worry gravely about the risks just because I think we’re so close and because even if we just progressed civilizationally as we already have, we have a really great future ahead. I do ultimately think that the biggest danger to that is an existential risk or something that would just curtail kind of like marginal revolution on which we are already on because yes, we’ve talked about all these amazing, you know, like scientific advancements and how amazing they could propel us into that future. But ultimately, even if we just didn’t kill ourselves, civilization is making progress in the sense that lives are getting longer. You know, we are solving many of the diseases. Like people are generally better off along the way. We’re generally more peaceful more or less. So, I think ultimately the existential risks are so damning because we are already on a good track. Even if we don’t develop all this sooner, it might not benefit us. We don’t die if we didn’t do it. But, nevertheless, like I think we are on a good trajectory of civilization in general, especially compared to our past.

So existential risks are I think a big concern but then again I think the attitude towards them is also well we could either try to slow everything down not do anything or you know like get very bogged down in doom but there are a lot of things that you can build in this like differential technology development mind frame so even if you think that the risks are really there for specific technological areas there are often specific things that you can be advancing first for neurotechnology I hope that we don’t just advance very advanced neurotechnology but we also have a mindset for privacy and for security in mind like as we build these devices that have read and write access to our brains eventually like we need to build really good ideally like encrypted data provisions for neuroscience like maybe some kind of like real object level capability framework for the neurotech devices so they only have access to what they actually need things like that so we need to build technologies always with a like a mind to the risks I think in mind but that shouldn’t mean that we shouldn’t build them it just means that we like have to build carefully and maybe sequentially. So I think usually even if you accept that they’re risks, there’s still things you can do productively even in that world. Yeah.

And then lastly, I think it’s just generally an amazing time to be alive, at least for those that are like fortunate enough to say so. And I don’t think that was always true or like at least like I remember feeling so cut off from the future in Germany. Like I had a great life. I grew up really well. I had a wonderful childhood and very loving parents. So, there was nothing wrong in my childhood, but there was still this yearning for the future that I didn’t couldn’t really like satisfy or something or like a craving.

And then I think I’m just so grateful to be in the Bay Area and having been here for 10 years, being now much closer to where the future is being built and how it’s being built and just be having like a tiny like seat in the front row. Like, so I’m just generally quite grateful that a civilization we’re at a level now where those things are seem within reach. Like I’m always reminded of like Petrarch who like I think lived just incredibly long time ago and was like trying to write a letter to the future. He was one of the first futurists and was writing like you in the future you will see amazing things and you will actually you know develop all the science and technology that like I can just not even fathom right now and you know we from the past like you know like congratulate you for that and he was just like writing at a time when none of this was possible and writing a letter to like us future beings now. So I think to some extent there’s this level of gratitude you know to just be around right now where many of these things that will be defined the future and whether there’s a future at all will are fought out in front of our eyes and then also having a front seat at it like in the bay it’s just like I think a amazing ecosystem here that I’m just so grateful to be a part of.

So I think it’s kind of hard to be very pessimistic these days in the sense there’s just so much amazing things happening all around us by many of the people that you know you and I both know but like also in the broader just society the internet was this amazing force for diffusion so like beyond the just the Bay Area that we sort of grew up with the explosion of the internet enabling this conversation to reach everyone previously was just house in a few labs and a few organizations in a few cities I mean that’s how I found out about krenics about foresight about That’s just generally I think what gave you any hope like that you know that like people are working on these issues. It’s not just like because before that I think like mostly what I turned to was like philosophy you know I was like there’s existentialism this is how you can deal with the fact that you have high expectations for the world but then a short life you know and very impressing that the among the best the 20th century had to offer was existentialism.

Juan Benet

Yeah it’s better cope too right like it is like the biggest hope of all time.

Allison Duettmann

I have like a very strong beef against existentialism. And I think it’s like beyond coping, it causes people to stop striving for a dramatically better future.

And I have like a countervailing philosophy that I describe as like Prometheanism where you might be dealt a hand that might be very difficult, but there’s an enormous amount of progress that you can achieve and it might be risky. It might be difficult. You know, your liver might end up being eaten by an eagle along the way. But if you can harness the fire and be able to advance knowledge and science and technology, you end up uplifting everything. Maybe I can do this with Claude someday. Like I wanted to write a like the myth of Prometheus similar to like the myth of a huge counterpoint because it’s a similar kind of story, similar kind of like Greek myth with a dark part to it where like the hero doesn’t necessarily end up in a good state, but the story is completely different.

Juan Benet

Yeah. But I would say I guess you know like it depends how you interpret it like I mean first like the myth of Sisyphus which was written by Camus like it’s still drawing on the old actual Sisyphus story of like Sisyphus was punished by having to roll up the boulder up the mountain and then whenever he was up there it rolled down again he had to do it over and over again eternity and like to some extent I think what Camus was pointing out is because before that mind you or like not before that but like an alternative to existentialism is also nihilism.

Allison Duettmann

Yeah, totally. In the sense that like you could just believe that nothing matters in the face of having short brutish lives. And so the universe

Juan Benet

Yeah. And so existentialism was like a little bit of a juxtaposition to that.

Allison Duettmann

Maybe not quite chronologically, but like nevertheless, like it was this notion that like okay, like we have a life that is restricted in ways that we might not like, but nevertheless, you have to imagine Sisyphus as a happy creature or something, I think is the wording that he like, you know, he’s like putting pushing the boulder up the mountain, but he’s like doing this out of like the spirit of being human. He’s like the battle crafts his own fulfillment through the effort and the act of pushing the rock uphill and creates his own meaning and this is where I think it turns negative.

It’s like if you push that really far then the only thing that matters is the meaning that you come up with and it encourages this almostic attitude where the height of meaning is just whatever you tell yourself and you come up with and it encourages this extreme subjective perspective instead of thinking about the intersubjective that we have to deal with or like the actual objective reality of like very large planet there’s a all kinds of challenges that we have to solve the sun is going to run out of energy at some point. There’s death and disease all over the place. And you can have a huge impact today.

And if you kind of like take the super existentialist route of like, oh, I’ll craft my own meaning by escaping into literature or into the things that like I craft my own joy over and I don’t instead work on like the more pro-social things that contribute to the rest of the world or contribute objectively, you end up in like a bad state. So I think like definitely like seems like it was a needed step forward but it also is I can see today that it a lot of the existentialism that lurks in the reaches of the kind of social consciousness inhibit a lot of people from taking charge of improving the world for themselves and for people.

Juan Benet

No, I agree. Like eventually like you know you want to break out of system force in the sense that like the boulder should stay up on the mountain and then you want to build a civilization on top of it, you know. So eventually you know like you want to have an impact in the real world. [laughter] You want to go and like get the fire from the that is locked away in the gods and distribute. Maybe you burn yourself.

Allison Duettmann

Yeah. Maybe you burn yourself along the way and it’s hard but like you figure it out and like you advance. I do agree that that’s better. But it’s a spectrum, right? like of things like nihilism is like on all the way all the one end extend to hope maybe all the way on the other and then there’s like different things in between and yeah these are useful polls to pull the conversation and pull perspective and to inspire action.

Juan Benet

Yeah. Yeah. I do hope that more people are like empowered I think to actually have that mind shift because it can feel like a little bit like you know just like oh it’s so easy to say you know like there’s so much so much opportunity and optionality for us already like so of course like it’s easy to think that way but I do think that like you can do that on all different levels right and like there’s always like one level further you can advance yourself to and another person that you can help and another connection you can make you know things like that there’s like endless amount of things to do that are like inspiring and also broadly helpful, you know. Yeah.

And speaking of connections and inspiring, this is one of the things that I think foresight is incredibly good at. As an institution, foresight has been gathering hundreds and thousands of people over I guess four decades now, helping shape fields and advance identify bottlenecks and like the tech trees that we have to solve and the dependencies and causing you know kind of stimulating conversations has constructed this connective lattice across different labs in academia across different companies across venture capital and philanthropy and society. You know how do you think about it? How do you like build it? How do you what do you’ve been able to catalyze over time?

Allison Duettmann

One secret was just the fact that Foresight was around for so long and that there were these group of people that were had like a mixture of characteristics. On the one hand, they were like quite optimistic about the future and like wanted to do something. On the other hand, they were not just like philosopher. There are a lot of transhumanists around etc that are just like kind of like thinking about how the future could look. No, it was like a group of scientists and technologists that were like actively working on the future.

And then they were also like risk conscious to the extent that like people were actually like engaged with the downsides but in this like very productive way where the downsides didn’t mean you couldn’t do anything but like you just had to like work on something responsibly right so it was this kind of like cool attitude towards the future that was like not super polarized or something and then I think many of these folks are kind of like have just stuck around and are still kind of inspiring that vibe and then many new folks are I think along the way are like thinking similarly like just very pragmatic optimistic but like you know build oriented towards like the futures that we want to build and so I think those people are ultimately the ones that you want to connect with each other because whether someone is doing that in neuro or in nano or in AI they often have this other share property that they care about other fields and domains

so I think one thing that I also liked about foresight is that it wasn’t just like a one technology shop you know it had this like deep ambition for the future it was not technology agnostic it had definitely had a few bets but generally speaking like a future is not built on one day not by one or not for one technology like we need to kind of like think about and build technology in tandem with each other.

And so it’s kind of like a university in the original sense of the word where you’re bringing together all the different people from all these different fields and enabling them to create a space for them to talk and discuss these ideas. And not like traditional modern universities where as a faculty member you’re supposed to swim in a particular lane and you occupy a fixed spot and you have to teach a specific set of classes and only do research within a specific frame and you have to kind of advance your field in a particular way and you can’t step on the toes of the other people in the other fields nearby or like you can’t say things that are like too crazy or too out there yet. And so in a way like here universities who are supposed to be the places where we are able to explore ideas fully somehow have become more conservative and more shut out and then the scientists and technologists and builders have to find these other outlets to talk about these things.

Juan Benet

I mean I totally agree. I remember I guess in philosophy it was a little bit more I don’t think I had that problem that much in my university but definitely what I hear from other experiences it was pretty stark.

Allison Duettmann

I do think also that like there has just been a growing movement of orgs that are picking up the slack and you know building like focused research organizations left and right or like other fantastic nonprofits that have like come along in the past I would say I mean very much so in the past 5 years but like maybe broadly over the past like 10 years or something. So there is now a space of orgs specifically for like individual areas or something where there is now like a much more of an ecosystem that I would have liked to see maybe early when I came to foresight.

So it’s definitely I think we’re in great company these days and like one thing I think that we’re trying to do is also connect these different people organizations and projects to each other because I think sometimes you know like they have their own communities but like there are so many pieces I think that people are building in parallel with each other and so having complimentarity on the one hand is great but like having different orgs people and products be able to collaborate across VC philanthropy academia engineering open source company. Like I think that’s something that I really like and often there’s like a lot that these people want to do just because they’re like so inspired at like what the future could bring.

So it’s on the one hand I have big imposter syndrome within this community but on the other hand it’s just like a very inspiring I think group to be a part of and it’s just kind of inspiring to see just how much I mean like you have must have seen this too right like over the last few years like the amount of people that are coming out of woodwork building amazing I mean, we’ve just seen the Midjourney announcement, right? Like, wow, that’s now a thing. Companies are just like using all of their funding to like create this entirely other technology that was not in the original plan, but they’re just like, “Yeah, we now have the funding to do this and we’re just going to revolutionize science.” Like things like that. It’s amazing, you know, that this is now a thing that And of course, they didn’t have much receive funding, so that was easier to do and just to decide. But in general, I think it’s a great community, I think, right now, and it’s growing rapidly. And so I don’t know there’s just a lot to do but also there’s now a lot of people that are coming and doing it

[1:43:16] How Foresight operates to advance frontier science

Juan Benet

and concretely talk us through have the range of programs like you do workshops you have a range of prices you have a range of events that gather the community you have online seminars and network connectivity structure like how do you kind of like upkeep the community or like [snorts] what role do these pieces serve?

Allison Duettmann

Yeah, I guess you can slice Foresight like one or two ways or both ways. like one is like by the focus areas that we have and again there’s a lot of overlap between them but like there’s the secure like you know decentralized cooperative AI area there’s the neuro area and like within those areas it’s mostly really the frontier stuff that we’re interested in there’s the longevity long bio area and there’s the nano area and so often you can like initially you have people coming in through one of these buckets but then often times you know once they’re in a bucket they’re just like oh but I’m also really interested in new let me go over here and we totally allow that and try to encourage people to do so.

And then like what is happening in these domains? It’s usually like a mixture between the programs that you mentioned of like in some areas like prizes are like still a good idea to encourage a competition of ideas about how to do an ambitious scientific feat. In some areas like grants are a great idea because like you know a few approaches but like you want to actually encourage a lot more innovation in those. So there’s a lot of grant giving probably like the biggest one expenditure based and then there’s the workshops which are like mostly like field building activities to the extent that like sometimes there isn’t really a field yet and so just kind of giving prizes or grants would still not really create an ecosystem like you actually need to bring people together like over a period of many years you know in many of these areas we’ve done workshop now or like I’ve done workshop for 10 years Foresight has done workshop in some of these areas for 30 years every year so it is like a growing cohort then that people can come back to and hire hear from and like coordinate around.

And then for those that are not in the Bay Area, there is the virtual seminar platform which is mostly just like trying to create virtual shelling points around these technologies where people can present new research and just try to create more connective tissue around that. And then the vision weekends are like these big events where we try to bring it all together and like just have these festivals where we celebrate scientific progress. And like one bit I guess that we added this year because it’s our 40th anniversary these AI nodes.

So rather than just supporting folks through like fellowships or grants or something, what we’ve really focused on a lot more is like actually in-house support. So often times these people really actually want to be part of a physical community and like actually build amongst each other. And that’s becoming I think more and more important not less. And because not everyone’s in the Bay or nor should they be. We have one now in the San Francisco area and then one in Berlin. And there’s now a lot of interest on like launching these like nodes, these like collectives with funding and other support structures in other areas too. Maybe like each city for example Boston could have like more of a biotech vibe and like they could all pick different specializations.

And I think the most important part there that we have started doing which is also possibly the most difficult part is to provide in-house compute to projects. And so on the one hand that seems like a totally mistaken effort because why would you build your own clusters if you can have access very easily to many of the hyperscalers and data center providers that you have. But on the other hand I think like future proofing your efforts maybe almost not necessarily requires you to build up your own cluster but like you should at least consider it because we already see compute prices I think they have doubled or something since last year or at least like in a very very short time like

Juan Benet

Yeah, depending for specific use cases, but compute costs are rising.

Allison Duettmann

I think they’re going to continue rising now that the race is on between the main frontier labs.

Juan Benet

Eventually, they might for a while.

Allison Duettmann

Yeah. I mean, unless we have some real breakthroughs like I think for quite a once we get recursive self-improvement factory tape out of just tons of chips and yeah, data centers. But then we might still want to use a lot of scale like maybe if we actually like if like new model architectures might pan out that are massively more energy. One good existence proof is you know the brain runs on like less than 30 watts. So we should be able to figure this out.

Juan Benet

Yeah. Long term for sure. Like that would be a goal and it would be sad if we didn’t make it work right like otherwise I think all of space could be filled with data centers and like you know like we’re I mean it’s just going to be very difficult to do anything.

Allison Duettmann

It’s definitely I’m pretty excited that we started building a Dyson swarm in 2026 and you know didn’t wait until 2036.

Juan Benet

Sure. I mean, you know, this why not start now?

Allison Duettmann

And I do think many of the problems, for example, like space exploration or space manufacturing, space development, but also energy like solar energy, other forms of like lean green energy might just be accelerated through compute, the need for compute. [laughter]

But I think in general like it’s unclear to me whether independent projects and AI for science efforts will have access to the compute that they need on the long run. And I’m not just saying like you know you want to build something that the government might not approve of and it might not eventually might not get computed. I’m mostly just talking about like we’ve seen what the US government did with Fable right and so that’s a model provider but like it could do that with data centers that are based in the US. I’m not saying like half of any like intel that it’s going to happen this year. But eventually the US government will wake up to you know how critical computer is.

Juan Benet

They already arguably have.

Allison Duettmann

And so I think if you want to be like sovereign and independent and actually like have a future proof strategy, you need possibly your own computer or at least it’s a theory that we have and like it’s not that we’re trying to be competitive sand into chips yourself.

Juan Benet

Yeah. Eventually you have to start grinding.

Allison Duettmann

No, but like I think it’s also kind of fun, you know, to get your hands dirty and like figure out like new ways to do community oriented.

Juan Benet

Well, certainly we can figure out nanotech correctly, then yeah, we can mega decentralize the ability to construct all this stuff.

Allison Duettmann

Yeah. And then you have like very local energy production and energy consumption eventually, right? like those all those things kind of go hand in hand to some extent but yeah it’s it’s kind of TBD how whether that future will actually be necessary in which you have your own compute but it’s at least like a hedge

Juan Benet

yeah has for done any estimations on the cost structure of a nanotech program

Allison Duettmann

there are some and with Foresight I mean like people that we have asked like most of the people you already know like Adam etc like a few people have tried to do you know cost extrapolations but I think mostly what people I think are trying to estimate and I don’t remember the exact numbers but it was not necessarily like actually molecular manufacturing would cost or producing that like a machinery that could do that but like what is molecular 3D printer or like some kind of like some intermediate step that could then inspire more capital to flow into like the longer routes and so I do hope that eventually we’ll get that funding I’m not saying for should get the funding saying generally as a field there is enough interest and incentive again to like think about Nano because that will become the main bottleneck.

Juan Benet

But I feel like the US needs to appoint Adam as the chief roadmapper.

Allison Duettmann

That would be wonderful. I would totally vote for that if I could vote in the US. Yeah, I mean I think he has his hands in many of the areas that people care about like not just nano neuro of course and to some extent bio and so that would be wonderful not just roadmapper but also like instigator right just like froze all over the place and I guess to some extent like you know the US government is taking some hints recently like x labs and like the tech labs I guess how they call it the NSF labs like yeah that the new tech labs are FRO inspired yeah so that’s great we also have to give credit where credit’s due like that’s pretty awesome and certainly Adams thinking was shaped by ARPA and DARPA and like it’s all it’s all remixes.

Juan Benet

It’s all remixes, but I’m just glad that these are happening.

Allison Duettmann

And then of course you see amazing efforts like Arya, you know, that are yet again like remixes of this the upper model. And I just think that generally we need a lot more of those.

Juan Benet

Maybe on prices specifically like I think Foresight has had a bunch of success with prizes in both identifying and recognizing innovations and people who then went on to do like a bunch of like great work like what are some of those stories?

Allison Duettmann

Okay, so let’s see on the nano side at least we gave two Feynman prizes relatively early that then I think it was in one case nine years and in the other case multiple decades later got awarded with a Nobel prize and I mean the latest one is perhaps interesting just because it was awarded to AlphaFold and to David Baker in the same year [laughter] David Baker was one of them basically where we tried to like award progress in nanotechnology early perhaps like strange approaches you know like Baker’s case you know he’s now working on like aa or like a real computational AI approach to make progress on like protein engineering. And so we try to just be early on some of these fronts.

And in Fraser Stoddart’s case, who was the other Feynman Prize winner who later on went to win a Nobel Prize. It was like really like building these very functional molecular machines where you could actually like really like distinguish a function. They actually look like little machines, you know, [laughter] not that the ribosome doesn’t, but yeah.

So there were like some of these kind of like good early pigs like and I think now especially it’s our 40th anniversary I’m diving again into our archives and Christine Peterson who co-ounded foresight also coined the term open source in this like I think it was 1998 effort to package the idea that software should be open should be editable should be modifiable rather than should all be patented and locked away and not interoperable. try to like package that idea because at that point that was a new idea and it wasn’t the way the software was going and then now look where we are with open source right like it’s like the main way of doing things not saying that Christine had like anything to do with the way that the field turned out but I think sometimes the naming and like the early conceptualization you know really like trying to put something on the map can have an impact there

Juan Benet

our ability to describe name refer to things advances our capability of doing a thing examples of this include like mathematical notation like once to develop like the right mathematical notation to do certain kinds of operations, a whole world opens up. You know, this is like why I don’t know Leibniz and others were so influential. It’s just like figuring out that notation structure. At the end of the day, like these words or terminology are like abstractions and then through these like abstractions you can build a higher level abstractions and you know it’s like composed like knowledge in a more packageable computer can now handle it better.

Allison Duettmann

Yeah. Yeah. Totally. So I think naming things are definitely important. I think ultimately and popularizing it and all that kind of stuff. Yeah, that I think the popularizing bit is definitely I don’t know. I think foresight is always like a little bit maybe too early, you know. It’s definitely like we’re not a very popular organization. It is named foresight. It is named foright not after sight.

Juan Benet

That’s true.

Allison Duettmann

We’re not the hindsight institute.

Juan Benet

No, that’d be great to start on April Fool’s.

Allison Duettmann

There is actually that organization.

Juan Benet

That’s great.

Allison Duettmann

Yes, there is something called the hindsight institute and Christine is involved with it and it’s it’s a hilarious joke. They have business cards and they hand them out at Foresight events.

Juan Benet

That’s fantastic. [laughter]

Allison Duettmann

It’s all very self-reerential and hilarious.

Juan Benet

We’ve heard a little bit about you as a kid and being inspired by certain ideas. What shaped you? Who inspired you? How did you develop? There’s a lot of people out there who grow in lots of different

[1:54:09] From philosophy to Foresight: how a three-month stay became 10 years in SF

Juan Benet

ways and so hearing individual stories can be super helpful.

Allison Duettmann

I definitely like grew up reading a lot because also there wasn’t all that much out there yet, at least early on, right? reading a lot of philosophy like nature and then a lot of existentialism and then I was like this is not good enough what else is out there then I kind of switched into early transhumanism extropianism cosmism and some other just like early versions of less wrong what were they called Arbital and some of the early Scott Alexander writing and there was some early futurism online that was very interesting lots of sci-fi once I discovered it I actually discovered it quite late but yeah so I think like reading and then the internet was like I think it’s just a big level and that’s available to everyone really right and then I had a great philosophy instructor and teacher and then at university I had a few courses that I just really loved like I generally had a great I think I loved my school life and I loved my university life I actually had a really good experience in my educational system that seems to be very rare but it happened I mostly grew up a lot outside like I had a talk that I mostly spent time outside with.

I guess like through university I think I really dug deep into philosophy of AI. That was my like the thing that I did my master of science dissertation in and then read a lot of for example from Jonathan Birch who wrote a recent book the edge of sentience that just came out and that’s a lot of like different types of minds and how they work and like had a big neuro kind of like mind emulation or like you know different types of minds investigation angle. So there was just some really interesting classes around actually at my university. So that was awesome.

And then I discovered Foresight eventually through the internet and like those weird corners and then eventually I cold emailed them and I asked if I could volunteer for them and then they asked me to come over to the bay. I had no clue where that is. I actually thought at that point the Golden Gate Bridge was in Los Angeles and I didn’t really have a differentiation between San Francisco and Los Angeles in my head at all and was very confused when I ended up here.

But then I went to my first effect of altruism global and back in the days there was just all kinds of amazing world improvement ideas and projects there and there were so many quirky Bay Area events at that point already and so many communities like last one was already a big scene then and yeah at that point it was just a big frenzy of people that were all working on interesting science world improvement and big ideas projects and so it felt like a breath of fresh air to come here and at that point longevity and AI were like my two biggest focus areas and there were both people here working on that very much less so than now but definitely already more so than in any other place in the world. So yeah that’s I guess that’s it and since then I thought I would stay for 3 months but a decade later I guess I’m still here.

Juan Benet

So you came initially for 3 months and you just kept extending.

Allison Duettmann

Yeah I kept extending.

Juan Benet

Yeah that’s awesome. And how do you think about how the rest of the world has developed over time? I think foresight has inspired a ton of people around the world and you also help connect a lot of people around the world. Is it a lot of like talent flow to the Bay Area or have you seen a lot of activity stay in other places or flow between other locations?

Allison Duettmann

I mean, it’s definitely a lot more exciting to go to Germany now than it was before and to go to the UK. It’s just kind of awesome. We just had a vision weekend in London and it was amazing. So I do think that many of the other pockets where at least I didn’t have a community of like minds back in the day neither really in Germany nor really in the UK there are now just you know there’s things happening now you know like there’s communities now around these types of issues and also like I should say incredible talent and incredible creativity and it’s awesome I definitely think that there’s a lot more going on now than there ever was and at the same time I think maybe during co there was a move away from the bay to some extent but like many people are coming back and I’ve never seen the Bay Area as kind of popping as it is right now.

And whenever I try to go anywhere else or fly anywhere else for an event, it’s a real cost to not be here for like a few days or a single weekend because every single weekend there is a conference here or meeting here or workshop here that is like expressive of at least what I care about. So it actually takes a lot of willpower to believe these days. [laughter] And then of course what added to it is I had a child. So that makes you a lot more stable [laughter] and a lot more stationary than before. But it also inspires you a lot with hope for the next generation and for what’s to come and definitely with a lot more incentive to get the future right.

Juan Benet

So do you think a high quality conference workshop type environments? It’s just so rate limited to happen digitally until like we get uploading right or do you think we’re going to figure out some way of having that kind of experience online?

Allison Duettmann

I don’t know if you can have it online. And I know that like people are definitely trying and there’s some great I think there was a funding the comments that was online that was really fun or at least like a part of it was online. First one was online.

Juan Benet

Yeah.

Allison Duettmann

See that was really good. Had some of the best vlogs so far.

Juan Benet

That was awesome. I remember that.

Allison Duettmann

That was really good. So yeah people are doing that also you know there are many organiz so we had some excuses there. I know. So co I think actually helped us also get global because suddenly everyone was online and through that we because we did twice a day online programming during co to catch both time zones I guess both the eur European and US time zone we gathered a huge global community so I think co was really the time where we like went from a [snorts] Bay Area ORC to a global orc and now most of our staff is not here we have a node in Berlin we’re trying to have nodes elsewhere I’m not saying everyone should come to the bay it’s also super saturated and we are actually trying to like also create other dependencies to like decentralize or at least like create more connective tissue between places

and then there’s a lot of amazing other events right that are happening like you guys are hosting amazing workshops like the neuro workshop the other day was awesome I wouldn’t have wanted to be anywhere else a day than in one talk so I do think there is a lot going on as well it’s just that there’s generally a lot more happening now you know and every once in a while I do think that it would be kind of better to not live here just because you get more done because you’re not so distracted by all the social stuff that’s happening all the time you know but ultimately that’s how also how you hear about what’s happening and stuff culture is a is a big part of innovation, but I do think you can actually get most of the intel from Twitter.

Juan Benet

Yeah, there’s not very much that is said in private channels here that people don’t also say on Twitter. I think often there’s definitely group chats that are a lot more exclusive, but in general, I’m always pretty amazed by what you can learn on Twitter. What other books shape you or you love or you think like are very compelling that people should go out there and read or

Allison Duettmann

Well, there’s Beginning of Infinity. I think that’s just a book that should be required reading.

Juan Benet

Yeah, I think that is the best philosophy book. I think also the David Deutschian way of thinking of really just trying to figure out that nothing that isn’t actually physically impossible should be out of reach long term.

Allison Duettmann

Not forbidden by the laws of physics.

Juan Benet

Yeah, it’s possible. Then it’s possible.

Allison Duettmann

And you know there’s also some interesting theories of the universal constructor came out of that et like I mean he’s just a very brilliant person in general. His conception of knowledge was incredibly formative for me also.

Juan Benet

Yeah. I had to think about mimemetics and knowledge structures and the only thing preventing you from doing something is knowing how.

Allison Duettmann

Yeah. Constructing the ideas and the idea space and so on. I don’t think we fully grasped yet all the ideas that are in this book. We’re in the [laughter] of realizing like the outcomes of like this new way of thinking. And I think the interesting thing is that David D is not very AGI at all. So that’s kind of I’m still trying to come to terms with that.

Juan Benet

I think he’s actually appealed in that he does think we will construct AIS eventually, but he has been negative on the work so far.

Allison Duettmann

Yeah, I think there is like a strange aversion to digital minds being able to have the same level of creativity that organic minds, but I think he’s about to be proven wrong on that.

Juan Benet

Well, maybe one day I’ll just understand him and I know actually that he was right in some foundational sense that I can’t even grasp yet or something.

Allison Duettmann

Yeah. Yeah. Usually there’s some crux assumption somewhere that yields this outcome. We will definitely find out one way or the other soon. I guess I think another one is just like what is it? Surely you’re joking, Mr. Feynman. It’s just like such a fun read. It’s just so easy to understand and you get this kind of download of this crazy wonderful mind that existed and just like did things in the world. And so that’s just an amazing book in terms of like how you can be in the world, still be brilliant and functional because obviously brilliant, but he also just had a very fun way of going about life. He was this incredibly curiositydriven and most of the time you weren’t sure if he was joking or not, you know. So that was an interesting read.

I mean, there’s so many amazing sci-fi books as well, but I almost feel like Permutation City, there’s nothing really cut quite like it for me. So I guess I did like the Culture series. It was more space opera but it was trying to draw out how can a kind of post-scarcity utopian society live. Player of Games was about like you know what games might they play? How can they find meaning in life and stuff? So there

[2:03:05] Allison’s vision for an optimistic future

were some interesting aspects in that too. I think for me Greg Egan is just kind of untouchable.

Juan Benet

Yeah. What is your optimistic vision for the future?

Allison Duettmann

I ultimately hope that we can develop all the science and tech that is we have discussed from advanced nanotechnology to rearrange like the physical world in a way that is like both sustainable for the biosphere and the planet that we inhabit but then also to create amazing structures and devices and a physical infrastructure that could lift us off this planet and that could make us really like explore what is out there. So I think the hardware component is actually kind of important.

The other one of course that we discussed is the neuro layer of like you actually can’t go out very far unless you are a digital mind. So we do have to get that right eventually if you want to see some some parts of the universe and then of course you don’t have to go out that far into the universe. You could also stay in a very rich virtual environment right here on your local machine. So I do think that neural layer is amazing.

The kind of like base layer for all of that is I think individually like human longevity and actually like being able to have a future at all. So I think that’s what we absolutely need to get right for our individual ability to like benefit from any of this.

And then on a civilization layer like we do have to get AI right enough that it can benefit all these other fields and right enough to not you know lead us into destructive nuclear war bio- risk scenarios or cyber risk scenarios that can take out a lot of civilization or at least set it back by a lot. and we don’t want to have like a runaway AI that either can destroy the world itself or that is sentient and suffering you know things like that those are like the necessary condition of the ingredients that we need for the future and we need to kind of arrange them such that they can benefit each other but not that they do destroy humanity

and then I think generally what types of society do we want well I ultimately think that this might come from my philosophy background but like on a meta ethics level I think it’s going to be very tough to agree on one path to go I think we both agree on a But I don’t think we will actually want to live up in the same or like in exactly the same utopia. I think there are a few people like you’re probably one amongst them where I’m like okay I actually feel like more or less comfortable in handing you the stick to the future. But ultimately I think that’s definitely not true for everyone. And I think this notion that we often had for alignment of like we want to create a coherent extrapolated valition basically like figure out what humans would want if we thought longer. We’re more the people that we wish we were. We’re more rational. We’re smarter. Like I just don’t think that exists in the sense that people have different intuitions to situations. They’re either genetically enshrined in them through various different genetic varants that we have and or they are kind of enshrined us through our upbringing and so we actually see the worldly, right? Like we actually see the world differently, right? And usually we think that’s a good thing. We value diversity. And so I don’t think we’re going to agree on one future path.

And so we need to have a future that allows for pluralism that allows for diversity of viewpoints that allows for many different ways to live. Some want to have ultimate morphological freedom and create very different substrates for themselves and look very different. And some will just want to, you know, have their garden and be the crochet futurists. And so I think we need to have a future that can allow for both. And so a we need to have a future that can have peace between them. So ideally like they respect each other’s boundaries. They don’t interfere. They there’s no war. There’s no matic interference. no violence. And then on the other hand, we also want futures in which those that actually do have a more or less aligned vision on some part of the future can cooperate much easier and can actually go out and build stuff together and just kind of like align on the ways they need to create futures that are again evaluated each better than before by their own standards.

And so we don’t really actually need to agree on something. We just need to agree enough to collaborate on some specific thing and then we can go out our merry ways again. So I think ultimately hopefully we can have this like parotopian vision of the future kind of come true. And I don’t think much is stopping us. But I do think it requires like a level of physical safety, abundance, and just kind of like level of living that allows at least everyone to like flourish in their own way and actually go out and live their best lives, you know, to each by their own standards.

Juan Benet

Wonderful. Thank you very much.

Allison Duettmann

Thanks. That was very fun. And I should also say really not just thanks for the conversation, but you and Protocol Labs in general I think have been such an amazing force for good in not just No, I actually really mean it. Like I think I remember like before I actually knew you personally, I’ve always like followed some of your writing. I was like, well, they’re doing such interesting work. How’s this like company, you know, like doing so much interesting work?

And I think ultimately companies and efforts are like expressions of some of the people that like instigate them. And I think you’ve like just had a really long reaching breath for the future, like an interest in the future, like an obsession with it, like a craving for it. And I think it showed in the way that you build the critical apps, but also the entire ecosystem around it. And then the big neuro effort now is just I think great to see. I wish more companies did things like that. It’s very similar to what Midjourney did, right, of just like this is what we want to bank on in the future and that we’re just going to go ahead and do it.

And so it’s been awesome to see and the collaborations that we’ve had between PL and foresight and how you know we’ve like partnered on so many of their efforts has just been really fun and so I hope that there’s a lot more of that out there especially now that PL’s kind of more of a network structure there’s probably going to be lots of different pockets of that popping up. So it’s been great to see.

Juan Benet

Thank you very much.

Allison Duettmann

Yeah, that was very fun.

Juan Benet

Thanks. I hope you enjoyed this episode. This is a new podcast so we need your help to get the word out. Please like, rate, and subscribe on your favorite platform and share it with people you think would find it interesting. Thank you. See you next time.

Konrad argues the most promising path to understanding how the brain works is to read the brain’s wiring directly, down to the molecular detail of each connection, and to build compilers and simulations to understand the brain’s computation directly.

In this episode we go deep into how neurons work, how neurons wire together, and how organic and artificial neural networks differ. We discuss why organic neurons are doing much more; how a model of a single organic neuron can solve MNIST — computing more like a 3-layer artificial neural network; how the brain might learn by solving credit assignment with only local signals; how to approximate backprop without a global algorithm; why AI and humans are intelligent along different dimensions; why Konrad isn’t very worried about AI replacing us; economic models of intelligence and physical work; and much more.

Konrad is a brilliant, contrarian thinker who explains complex concepts very intuitively. It is a solid computational neuroscience primer. I hope you enjoy this conversation as much as I did!

Other links to this episode and references below.

Topics covered

• 00:00:00 Introduction

• 00:01:01 How organic neurons work

• 00:24:13 How the brain learns: circuits and credit assignment

• 00:45:29 Recording the brain

• 00:52:47 Why simulating brains is hard

• 01:05:00 A new approach: connectomes and compilers

• 01:21:00 Why simulate brains?

• 01:29:50 How AI and human intelligence differ

• 01:41:04 Evolution, intelligence and AI risk

• 01:52:42 Robotics, causality, and the roots of intelligence

• 02:05:53 AI for science and scientific rigor

• 02:13:05 The economics of intelligence

• 02:27:50 A hopeful future

Podcast Links

• Juan Benet Podcast on YouTube

• Juan Benet Podcast on Spotify

• Juan Benet Podcast on Apple

• Juan Benet Podcast on Amazon

Links From the Podcast Episode

Guest + Organizations

• KordingLab: kordinglab.com

• KordingLab on GitHub: https://github.com/KordingLab

• KordingLab on X: https://x.com/kordinglab

Papers directly from Konrad Kording’s lab:

• Can Single Neurons Solve MNIST? The Computational Power of Biological Dendritic Trees (2020)

• Comparing Dendritic Trees with Actual Trees (2023)

• (Artificial) Intelligence Saturation and the Future of Work (2025)

• Compiling Molecular Ultrastructure into Neural Dynamics (2026)

Referenced external papers:

• Millisecond-timescale, genetically targeted optical control of neural activity (2005)

• Dopamine Reward Prediction Error Coding — Wolfram Schultz (2016)

• Single Cortical Neurons as Deep Artificial Neural Networks — Beniaguev, Idan Segev & London (2021)

• Neural Signal Propagation Atlas of C. elegans — Randi, Sharma, Dvali & Leifer (Andrew Leifer’s lab) (2023)

Books & Media:

• Causal Learning: Psychology, Philosophy, and Computation — Alison Gopnik & Laura Schulz

Juan & Protocol Labs

• Juan Benet on X

• Protocol Labs

• PL Neuro

• Disclaimer⁠

Transcription

Juan Benet

My guest today is Professor Konrad Kording, a computational neuroscientist known for bridging neuroscience, machine learning, and motor control. He is professor at the University of Pennsylvania, spanning neuroscience, bioengineering, and physics. He’s one of the top scientists working to understand how our brains work.

He’s done tons of pioneering work, including, you know, some of the earliest Bayesian brain hypotheses and figuring out why control works. He’s a clear contrarian thinker on what neuroscience does and does not yet understand about the brain. And he consistently pushes for more rigorous, more falsifiable models in neuroscience and beyond.

Konrad, thank you for being here and let’s dive in.

Konrad Kording

Thanks so much for having me, Juan.

Juan Benet

Great. So we’re going to start from the bottom up. What is a neuron and how does it work?

Konrad Kording

Well that’s a great question. So neurons are small structures that do all the information processing in brains. They usually have a cell body that is maybe ten micrometer big, hundreds of a millimeter. And then they have a piece that takes the information from other neurons. It’s called a dendrite. And it might be a millimeter long.

There’s a large variation there. And then it has an axon, which is the wire along which it sends the signals that it produces, which may be very short, maybe just a few micrometers, or it might be all the way from the brain to the feet of the giraffe, which would be meters in scale. And the way in neuroscience, how we think about them is that they are input output devices.

We have inputs. They’re called the synapses from other neurons. It’s putting all these inputs together produces an output which is spikes. And ultimately these spikes control your body and are the basis of all the interactions that happen in your brain.

Juan Benet

And when a neuron interacts with another neuron, what are the different. How does that communication happen through the synapse.

Konrad Kording

Yeah. So in most cases neurons are what we call spiking. So there’s an electrical signal that comes out of the neuron. And that then reaches the synapse which is the place where they meet one another, where it reaches that in so-called chemical synapses. We have this phenomenon where the cell where the signal comes from throws a lot of chemicals out of the cell, and the next cell that it’s connected to.

The downstream neuron basically has receptors for these chemicals. We call those chemicals neurotransmitters. And these chemicals. Then once they bind the receptors produce an influence on the postsynaptic cell. Usually that is just the current. It happens by having ion channels just like a battery.

They open and then the ions go through it. Current produces the signal on the postsynaptic side, and that is the basis of computation in neurons and variance.

Juan Benet

You said most neurons are spiking neurons. What are the what are the non most neurons.

Konrad Kording

Yeah there are some neurons that are what we call analog neurons. That way you can say what changes over time is the voltage in it. So spike is something that is either there or it’s not there. It’s like a telegraph signal like beep, beep, beep and it arrives on the other side. Whereas whereas analog neurons, you can think about it more like wires that can go up and down and are just graded signals.

Much of computation in human brains is in terms of spiking neurons. Much of the computation in very small systems maybe once or something is graded. much of the information processing in your eyes and your retina is graded as analog. And then the outputs of. From your eye are spiking. And then basically the eye signal sends a signal to the rest of the brain.

That’s the basis of seeing.

Juan Benet

And how all these different functions get computed in the brain. In the computer science model that we have. We have a very, you know, simple abstraction where we tend to look at a neuron as just one mathematical object with a set of inputs coming through and then firing out. And they’re effectively the same everywhere.

But brain neurons aren’t like that. They’re very different. There are many different types.

Konrad Kording

Yeah, there’s lots of brain, lots of different neurons and brains. And I think their function is much more complicated where you can say so in the, in the neuron. I told you, there’s a lot of synapses that go there, roughly 10,000 depending on where you are in the brain. And what happens is every synapse produces a current on the postsynaptic side.

But it’s not as in computer science, where we say the output of the neuron is just the sum of all the input that goes. But we have local synapses and they do something non-linear. For example, we have this thing called NMDA spikes where you have a couple of synapses, and if they’re all active at the same time, they’re much more strong than if just one of them comes.

And then we have if we go from these very local pieces on the dendrite, two more global pieces, we have things that’s called calcium spikes, where you can say the neuron is very active for some period of time, maybe a 10th of a second. And then ultimately we have this spiking in the neuron. So when people do that analysis where they ask how complex do neurons appear to be?

And it’s based on some simulations. So take it with a grain of salt. But I think the best way of thinking about it is that a given neuron is a little bit like a three layer, four layer neural network that just has one output. So in that sense, you can say that a real neuron is really complicated and might be doing much, much more work than one neuron in an artificial neural network would do.

Juan Benet

That’s a very good insight of Modeling. So we could model one neuron with like a four layer artificial neural network. I think is there some kind of estimate on the parameter count that an average neuron might have? Because that implies a lot of different parameters and a lot of computation that is happening in one single organic neuron?

Konrad Kording

Yeah. I think you need a very large number of to describe the computation in one neuron. Imagine you have 10,000 synapses. It’s not that you now have 10,000 parameters, which is one weight for each of them. It’s much worse. So there are tons to be that. There’s some connections that are very fast and something happens and then it’s gone.

And there’s other connections that are much, much slower. There are excitatory inhibitory ones. There are those where two inputs, if they come immediately after one another, it’s like the second one doesn’t count. We call that depressing. And then there’s the opposite of facilitating. If just one input comes, very little happens.

But if two of them happen. One right after the other. It’s a very strong signal. So all these things, we need parameters for it. So I think there’s going to be at least ten parameters for each synapse, that number of synapses on or that number of parameters. Now we’re at maybe 100,000 per cell, but there’s a lot of extra ones coming from the local non-linear properties of dendrites.

We have these things called ion channels where that can amplify signals. They can weaken signals, they can make signals nonlinear. There might be a million of them on the cell. So the number of parameters that we need to describe a cell is undoubtedly very, very large.

Juan Benet

Across different types of organic neurons. Do they follow at least a set of patterns where you roughly like if you were to kind of model one organic neuron with a particular structure of a, of an artificial network, is that viable, or are they just so uniquely tuned that it’s kind of like learning? For each individual neuron, you would have to learn a different network from scratch.

Konrad Kording

Yeah, I think every neuron is different and they have to be, you know, like. Because what do you what do neurons do? They embody what we know about the world. So if they were all identical, they couldn’t start any information. Like there must all be different. And because they are all different, they have lots of lots of parameters.

And therefore, if we want to model them well, we need to use lots of parameters.

Juan Benet

Well, we push them back on like the artificial neural networks are all mathematically identical. They just they store the differences in like different parameter weights. The network structure of the artificial neuron is identical while the parameter is different. And so here you’re saying inorganic neuron terms.

How different are they. Is it like the more variable like a parameter setting? Or is it just like the actual wiring and structure of them is just so fundamentally different? Wonder until the next. That is very difficult to like generate a model that can kind of abstract between them.

Konrad Kording

I think they’re all very different, and yet they’re all the same. And let me highlight what I mean with it. So you could say if I describe what a neuron does with its inputs, indeed they are non-linear and they’re all different from one another. But if you look, if you zoom into it and ask how it works, they all work in the same way, which is there are ion channels on it, there’s synapses on it, there’s synapses, and ion channels produce electrical currents on that neuron.

And then that neuron uses these electrical currents that get integrated by the physics of the cable equation integrates all of that. And after integrating it, it produces an output. And that output is sent to the axon. They are like slides. It’s biology. No, evolution does whatever works. So there’s slight variations of it where you can say, sometimes a neuron tell something to another neuron that doesn’t go through this axonal output.

And that’s those are the areas where we don’t quite know how big these effects are. But those effects do exist. So for example, if you have two neurons that are just like touching one another, even if there’s no axon from one to the other, the fact that there touch, if one of them is electrical active, the other one is electrically influenced by the first one, it’s called a ephaptic coupling.

And that means that two neurons that are right next to one another will influence one another beyond just the axons.

Juan Benet

Yeah, you mentioned the cable equation. So what are the computational models that we’ve built to represent neurons today? And what is sort of like the zoo of different models that we’ve tried and what has worked and what hasn’t worked?

Konrad Kording

Yeah. So there’s a continuum of abstract to more realistic models. Let’s start with the realistic model. The realistic model basically thinks about the neuron like that where you can say the neuron has this dendritic tree. If we zoom into the physics of that, it’s a little bit not like what is the dendritic tree.

It basically is a water and salt on the inside. And we know in physics how we should think about water and salt. It Current can flow through. It has a certain amount of resistance locally, and you can say a long branch of a dendrite may just be like a lot of resistors, one after the other. There’s also around that there’s cell membrane.

Now how should we think about the cell membrane. It’s a little bit like a capacitance to the outside of that. Now different cells have different lengths of that. And they branch out. And they might be thin or thick, or they might have a thick layer around it or thin layer on it, which from a physics perspective just means the different sets of resistors, different numbers of them, maybe a tree of resistors, maybe some of them have bigger capacitance than others, but they’re all fit into like that.

Same way of thinking that a cell has basically lots of resistors, which lots of capacitance and lots of sources of currents. Now let’s talk about sources of currents, like how does a current go into it? If we zoom all the way into it, we have an ion channel. There’s a mechanism that opens it often, and when that closes it, if it’s open, it means that ions will go through it.

It’s like attaching a little battery with a resistor to it. And that’s how we model it when we try to build very realistic models. And then you can say we have the synapses. It’s the same as an ion channel, only that it also matters what comes from the presynaptic neuron. So at some level we have a very coherent way of thinking about it.

Let me highlight the places where it gets more complicated, which is within the cells. We have a lot of molecules that influence that. So an ion channel is regulated by intracellular biochemical processes. That’s a place where complexity comes in. And that’s a place where we are not that certain about how that complexity works.

There’s aspects that we understand very well. For example, if you have a spike, the most famous model basically says we just have like one piece of axon. It’s like little resistors between with capacitance to the outside only that there is also this thing which lets in more current when the voltage is very high, and then once the voltage was high for a bit, it makes the voltage low again.

And this what gives rise to the spikes where you reach the place where it starts. Then the voltage goes high and then it goes down again. And if you put this across space, it like sends a spike through the axon and transmits it to other neurons. And we understand it at that level very well. Now when it comes to the non-linear regulation phenomena that happened there, like there’s just a voltage gated sodium channel that we need for that, that is still relatively simple.

In reality, there’s these things we call them second messengers, where where various cellular phenomena affect ion channels and complex ways. And learning is just the extreme way of that, where there’s a process that, if you want, starts a molecular cascade that ultimately Changes. How strong a synapse is, but also if new synapses get started.

Juan Benet

Going back to kind of like the different models. So we have some realistic description both in physics. And then you can build a computational model of that. But now that would be kind of way too granular. And with two details like, like let’s start kind of abstracting out, like what are more abstract kind of computational structures that can represent what the neuron is doing to like a great degree.

But maybe it’s like losing some information.

Konrad Kording

Yeah. So let’s go to the abstractions we use. And like a cable equation model is expensive to simulate because you as you end up simulating every little piece, every resistor, every capacitance. And on top of it, it’s nonlinear and with inputs that change over time. So it’s very difficult, very expensive computationally to simulate that.

There’s therefore there’s various abstractions of that. There are ways, for example, where you can say we take the cable equation, which tries to model like every little piece of that of the cell, and you can say, let’s, let’s just cause grain. That way you can say, maybe I take everything that originally was 100 resistors, and I replace it with one resistor that’s approximating what the system does.

Then I have just fewer resistors, fewer capacitance. It’s easier for me to simulate. We can go to the next step of abstraction where we can say, well, let’s assume that all the dendrites don’t overly matter. And that gives us the integrate and fire neuron. The idea there is we take every synapse we see on average.

How strongly does it influence the voltage at the soma at the cell body of the neuron? We then pretend that all this complicated dendrite, with its delays and time scales and nonlinearities and so forth. It doesn’t exist. It’s just all linearly added to it. Integrate and fire. It at least produces spikes.

It’s the basis of many works in spiking neural networks, which is a branch of artificial neural networks with the promise of making computation a lot more efficient. And then we can go farther with to what we call a rate model, where we can say, yeah, there’s time, and but time is fast. And if you want like over the period of a second, we might have a spike or multiple spikes.

And therefore it doesn’t matter all that much of what all those biophysics are. It’s enough to know as a function of how many spikes come in and how many spikes come out. And so you have this whole continuum of biological realism there.

Juan Benet

What do you guess is going to be the right computational model to then start building larger and larger scale representations and networks where you can recapitulate behavior well, meaning like, can you truly abstract it all the way to leaky integrated fire type of structure? Or do you need like something much more real?

Like what do you think right now? Or like, how do you think that this might develop over time?

Konrad Kording

I think there’s the information processing in the cell actually truly matters because there’s strong non-linearities. That and nonlinearities is the basis of artificial neural networks. So the idea that you can take one very simple thing that just adds things and then thresholds it and gives you the spikes in the way we do in artificial neural networks, I think is wrong.

But what we can do is we can now use machine learning to say, well, here you have this simulation of the neuron that’s really expensive. We want to simulate it fast, and it’s fine if it costs us some time to understand that. So we call that amortized inference where we run our simulations a lot and do a lot of computation.

And others with the idea that we fast put in a lot of compute, and then later it’s cheap to do that. You can say a neuron, a realistic neuron that gets lots of inputs, lots of and has outputs that change over time. We can still use machine learning to approximate it, where you can say, what? What does anyone do?

It gets like a couple thousand inputs as a function of time. It still just computes a function that. We can use machine learning to approximate that function, and in all likelihood, we can do that pretty efficiently. And in fact, there’s some work that already does that. If I tell you that complicated neuron can maybe approximated by a three layer neural network, well, you can simulate that three layer neural network massively faster than the cable equations for a neuron.

So there is this path by which we can take realistic neurons and in principle take those neuron models and move it into the space where we can really efficiently simulate it. And I think that’s essential if we want to build big things.

Juan Benet

What are the sort of functions that you think each of these neurons is calculating? Because at least in artificial neural networks, you get to use these very simple linear algebra style functions, and you build the complexity out of the layers of the in the stack. When you want to learn and approximate some more complex function, you’re just doing it across the layers where each individual neuron is not actually very complicated, but it sounds here.

The dynamics of a single organic neuron are so complicated that it could actually be learning. One single neuron could be learning a very complex function its own.

Konrad Kording

Yeah, it’s great that you mentioned that because we’ve done some fun research there where you can say we took a neuron that was somewhat realistic, and we asked, could a single neuron solve a machine learning task? And we found that a single neuron can solve MNIST. So for people who don’t know MNIST, MNIST is basically you put in the image of a number and it tells you what number that is.

And it uses real US Postal Service data. So. So everyone writes numbers differently. I’m from Germany. Are you my the way I do the number one is different to the way you do it. And yet it recognizes all of them. And so what we did in that study, we took a neuron with all of free parameters that they have.

And it was still a somewhat approximated, somewhat simplified neuron. And then we asked, can we predict, can we train that neuron with all the free parameters so that it would recognize if something is, say, a number. Seven was this a number? One? A single neuron can do that very well, actually. Therefore, you can say that it just gives you an example that the power is higher.

But at the same time, if we take a model of that neuron and we make it very efficient, it’s still structured where you can say there is a place where there are two parts of the neuron, there are two dendritic branches meet. That still means that not all possible three layer neural networks could be implemented.

It needs to be one where information meets and then it meets other information meets other information ultimately produce the output of the neuron. And so in that sense, it might rather be that these realistic neurons make it easier to compute these complex functions, because they have a lot of structure and sparseness and structure just promises to make a lot of computations better.

Juan Benet

When you mention the 3 to 4 layers, I immediately thought of MNIST. So that’s awesome that, that actually that actually has been done. Have we pushed into the, the limits of that, like what is like the most complex things that we have gotten single neurons to try to do? And where do they sort of break down?

What kind of functions are they just not able to then figure out?

Konrad Kording

I think in principle, a single neuron could probably compute a lot of the functions that a three layer neural network can. So what we did in that area and there’s not that there’s there’s emerging literature like pushing it further. And I think you can push it a lot further. But MNIST is famous for being a somewhat old fashioned machine learning problem, because it’s just basically binary drawings of numbers.

And there are more interesting models that people use today. CIFAR-10, for example, is one of them where you have, where you have just ten different object classes. There’s fashion, MNIST, this, this. They’re still simple. They’re not don’t think of it as an LLM or something, but they’re pretty complicated tasks.

A single neuron could solve those problems pretty well as well, I think. No. Like, there’s no doubt that it won’t be able to do an LLM. And you can you can say there’s something about the number of parameters. A neuron has 10,000 parameters. We shouldn’t expect a neuron to be able to do more than A and with about 10,000 parameters.

So you can say 10,000 parameters is kind of like the machine learning year 1995, say state. But machine learning 1995 does an awful lot of like non-trivial things. So I think we are not giving neurons enough credit for all the things that they could be doing.

Juan Benet

Super powerful in the three layer that you’re describing or 3 to 4 ish. Does that come from the dendritic structure or does that come from the parameter space of the individual synapses, or.

Konrad Kording

It comes from the structure of the dendrites. You know, if you look at it, dendrite, it’s this tree like structures. Now it really looks like a tree. Look at the harmonica drawing you kind of like. It almost looks like. In fact, we had a fun little paper where we quantified how similar trees are to neurons, and we found that they’re remarkably similar.

And so. So it’s a bit of a tree and you can say in that tree, wherever two branches meet, you have the chance to combine two signals. So I think that’s where the three layers comes from.

Juan Benet

And is it as straightforward as just, you know, if you look at the if you just take a snapshot of the neuron and you trace the tree structure, you can kind of represent that with like a logical neural network following that same tree structure. Or does it get more much more complicated than that? That’s how we did it.

Konrad Kording

There’s nice research from Eden Segev, who kind of looks more holistically at that. But I think first order approximation is this. Take a neuron that has tree structure and model it by an artificial neural network that has tree structure is a pretty good approximation.

Juan Benet

So this has kind of covers, you know, single neurons to some degree. When we start kind of putting them together into larger and larger units, into small circuits, and then later, like larger and larger networks, how do those work? And what are the kind of communication pathways that we that we see? Right.

So in artificial neural networks, we have a very we have constrained ourselves to build, like most of the deep learning structures just have these single directional neural networks where things are just kind of flowing in one pass, just because that makes the learning algorithms much easier. But organic neurons are not like that.

There’s all kinds of like recurrence and channels coming back and can maybe speak to that a bit. And what is the complexity when we try to model these like circuits?

Konrad Kording

Yeah, I mean, you mentioned the first complexity already, which is in say, a transformer architecture. We have strict feedforward transmission of information. There’s one layer, it goes to, the next layer goes to the next layer, and eventually you’re finished in the brain. It looks very different if in the brain, let’s say we’re both neurons.

I talk to you and you’ll very likely talk back to me, or if not directly, kind of your team talks with my team. That’s roughly the logic in brains. So you have this massive recurrence where it goes one direction and it comes back. I think it’s not that disjunct from modern machine learning, where a lot of people in machine learning use neural networks, and I think they’re having a bit of a recurrence these days, resurgence where we want to allow systems to feed back information.

I think that’s crucial. It’s incredibly useful. But let’s first talk about what we know biologically. The brain has hundreds of different brain areas, many, many structures. Neurons look different in all of them. In general, know this pattern that if it goes from one place to another place, it usually comes back from the other place to the first place is a general pattern that by and large holds.

Juan Benet

And this sort of a different channel like information flowing down. Or is it use the exact same channel back?

Konrad Kording

Yeah. No, it projects slightly differently. So if you look at cortex, which is the area that most people study, there’s parts of cortex that send the signal up and then there’s the neurons that are up, send the signals and they come back in a different place. So they come to the superficial layers and they’re being sent forward from the deeper layers in cortex.

So there is structure there. So if you want it’s useful from a learning perspective and from a processing perspective, that kind of information is separated. And you know what comes from the bottom and you know what comes from the top. And I think that is very important.

Juan Benet

Yeah. And it’s very useful algorithmically. Like that’s the basis for machine learning. Learning rules is like separation of information is what lets you do like the feedforward and backprop learning and all that kind of stuff. Right. And so some of the tension between the two fields is that in artificial neural networks we get to do this global computation, whereas in organic neurons we don’t.

Or we have to figure out how to represent the global computation in a in a local way.

Konrad Kording

Yeah. So let’s break down these two things. So that’s the first thing which is processing. If you make a decision, if you want to say a word say information, your brain will go forward and backwards and forwards three layers and then backwards two. It kind of goes back and forth all the time until the decision is made by your brain to like, say, given what if you look at an artificial neural network, it only goes forward.

Now in the artificial neural network, information also goes backwards, but it’s only information for learning in the artificial neural network at the end, you know, was this a good thing or a bad thing? And then you kind of go back and say like, hey, I did this bad thing and you were responsible. And then the neural net was responsible.

So I was like, okay, but those guys were responsible for me being active. And then those guys kind of like, that is how the so-called credit assignment problem is solved in AI in the brain. It’s much less clear how it works. So like it’s it’s clear. You need to know. If you want to get better at learning, you somehow need to know, well, should I, as a neuron have been more active or should I have been less active?

Otherwise, how do you know if you should do it differently next time? So this credit assignment problem kind of intuitively clearly needs to be solved. I need to know how I should change an AI. We know how we do that. We have a very efficient algorithm called backpropagation that tells every neuron if it should be more or less active, and every neuron tells it to the to the weights between them, the synapses if you want in brains, we don’t know.

It’s clear that somehow, if you want to get better, it means that the neurons where it helps. If they’re more active, they must become more active, and the neurons where it helps that they are less active must become less active. We know that it can. It’s a simple mathematical proof that on average that needs to be happening.

But how that works is very unclear. And like everyone, every neuron constantly talks with lots of other neurons, and somehow out of this dynamics must emerge, something which tells the right neurons to become more active and the wrong neurons to become less active.

Juan Benet

In neural networks, we get to apply this whole learning structure on top of the entire network the feedforward and backprop and backpropagation model with like the whole, you know, whole network gradient descent works in great part because we know the bounding box of this entire network and we know what is the feedforward pass, what is a backprop pass.

And, you know, there’s a whole string of papers that try to inspect both the artificial neural networks and neural networks to try and figure out is the brain doing something similar to that or a totally different type of learning? Because recurrent neural networks in the artificial models are very hard and notoriously very hard to train because of this bounding box problem and the computation getting very messy, The brain has to be doing something locally, right?

Like not whole. Brain synchronized. The learning and the algorithms have to be at least if not in a single neuron, in some local patch of some sort. Or maybe this like a question. What? What do we know?

Konrad Kording

Now let’s talk about physics. Yeah. Imagine you’re like a sign up somewhere in the brain. What information can you possibly have? Now, the only thing that you can know locally is what did the upstream neuron do? What did the downstream neuron do? And it turns out there’s this thing called back propagating action potential, where basically if the cell that I’m on as a synapse is active, I’m going to hear it.

It’s like everyone knows that just happened because that’s where we are locally. And then we can like maybe see locally a few so-called neurotransmitters. So there’s neurotransmitters that kind of seem to be related to like this was good or this was unexpectedly bad. So there is the possibility that I have some information there, but mostly I just see what happened of the Pignon?

What happened to me.

Juan Benet

In those neurotransmitters, eh? What are those? And be. Do they happen locally or in larger areas or globally?

Konrad Kording

Yeah. So there’s a lot unknown about it, but there’s suddenly dopamine seems to tell you about expected reward. And then there’s a huge literature that shows that this approximation is a really bad approximation. But like, there’s there’s some experiments, beautiful experiments by Schultz that basically say, you give me a reward that I don’t expect, someone comes in, gives me some coffee, dopamine will be like, yep, great.

And then alternatively, I expect coffee. Someone comes. I’m like, sorry, no coffee right now. And the dopamine minions will go like, ooh,

That’s sad. And so there is this way of representing those things, but reality is much more complicated. And that’s because I think like signals, not like an AI system. We just we have a so-called reward function or loss function, depending on which field you’re in. You’re just like, here’s better, here’s worse.

It’s very simple. But for us as humans, it’s not like better or worse. Is that simple? Now, like, what is better versus worse on a word that you just produced? There’s a word that might be like, you look at my face. I frown when you say something. So like, that’s one source. But also you might be like, well, that didn’t quite sound right.

And you might be like, this a strategically problematic word in this context. And the listeners on my podcast will not like this one. So there’s all these things layered on top of one another in AI is clear. There’s just like one function that kind of says how good or bad there was for us. There’s a lot of things layered.

For example, we know there’s other neuro modulators, say serotonin or something, let’s say more related to like, hey, that mattered. You should probably pay attention and not forget what just has happened. So there are these extra dimensions on it. And I think in AI we are building them into systems.

Now where we are weighing where like this an important stimulus and this not a gradient descent. There’s a little bit of that, but we’re trying to like do that even better. So there’s these layers of complexity that I think are much better for biological systems than they are for AI.

Juan Benet

Yeah. Going back to the kind of building larger and larger networks. And so there’s the algorithms that we might need to run, like learning, maybe digging into learning. Most artificial networks effectively have one learning paradigm. And maybe you can look at the current transformer architectures as maybe patching some additional learning processes on top through RL and all these other things.

But for the most part, like the pre-training process is just one main single algorithm. What do we know about the brain and how does learning in the brain happen? Is there a bunch of different algorithms? Are these maybe localized separately? Are the neurotransmitters triggering different algorithms, or is just tuning a single algorithm in a way.

Like, what do we actually know?

Konrad Kording

Let’s maybe start with the theory. So there’s always a theory space in neuroscience, a theorizing space, a sensemaking space, and a data space. Let’s first talk about the theory space, because a I know a lot about that space. And B that’s a field one is a starting point that people have. So there exists a space where people say, well, maybe the brain is relatively simple and know maybe the brain is evolutionarily already set up to do almost everything, and then you just need to add like a little layer of learning to it.

Tony Zador is one of the people proposing that where you could say, if most of it is pretty wired, then maybe we just need a little of a little bit of extra learning. That’s this one set of theories. Let’s say it’s relatively simple. Why is this great? Credit assignment is not much of a problem if the whole brain is just like its evolution gives us the things we just need to save.

A goes with B like no, like that’s the what? Blue. Go with color blue. It’s a simple learning space and maybe like a simple learning algorithm can work. Now most people that I talk with don’t believe that it can be this simple. And why can’t it be that simple? Because language is really complicated and we live in a really complicated world.

And if simple learning could succeed in this world, probably the world would be very different. Like, I believe the world is truly, unbelievably complicated. And I believe that because the outside world is so complicated, our brain is complicated. So if you then go towards more complicated algorithms, a lot of the algorithms that we believe could actually work in a complicated world require credit assignment.

Now, like I if you make a mistake, there’s like an unlimited number of ways how something could be a mistake like. And because of that, you need to find out what’s the reason for that mistake. And that goes into both directions. That goes into the inside of the brain, like which neurons screwed this up and it goes into the outside world, which is kind of like, how is the world such that this wasn’t a good idea?

And so if we have these more complicated algorithms, we need to solve the credit assignment problem. Like what was the nature of the mistake that we just did? And if we need to solve it, we need to we need a way of figuring out what went wrong. And there is this in gradient descent is how we do it in AI. And there exists a large set of proposals how the brain could do it.

Now let’s see what they have in common, like they have all in common that they build of things that is purely local, like the synapse only knows what happens locally. The neuron only knows what happens locally. Everything that it knows about the wider world. It must have been told that by its neighbors, not like because it only talks with its neighbors.

There is no global like an extra like an extra human being. Like, tell us what it is. Or at least to our knowledge at the moment. I mean, like, sure. Like maybe happens to divine intervention into all onions. But as long as we believe in like mostly physical world, the problem like kind of that information must somehow become locally.

So we have a lot of algorithms that explain how purely local things, neurons talking with their neighbors can ultimately still figure out how we should do credit assignment. And let’s talk about the simplest version of that. The simplest version is a way how the biological brain could actually approximate gradient descent.

Really, really well. Imagine it goes to the brain. Let’s ignore the fact that it always goes back and forth, or fourth, it goes to the brain. In the end, you do something, then you find out if this was a good idea or not, and then if it was a bad idea, you have anyone that says you did this and that was a bad idea.

And then you have for every neuron in the brain, kind of like a twin, which kind of says, I think you did something wrong. And then the set of like neurons with their twins. Think about it like, what do you know? Like in Star Wars now you have like the Sith, where there’s always the master and the apprentice and kind of like you have a master for every union like twin that basically says this was good or this was bad.

And then they tell the other neurons, well, the neuron that I’m the master of like, did it wrong and you guys were part of that. And then they tell that to like their training twin of them. And it kind of goes all the way that way. That way you produce something that’s almost exactly like gradient descent. It just requires twice as many neurons, and it requires the twin onions, and it requires the student neurons.

Juan Benet

And so this would imply there’s a, you know, a set of circuits of running the actual computation and a set of then parallel circuits going back, propagating the learning information.

Konrad Kording

That’s right. And of course, if you want biologically that’s totally realistic, but it’s totally not what we are seeing. Like if we stick an electrode somewhere in the brain, let’s say in the visual system, we have neurons that are very active. If we show them a little black thing, like, like a black line on white background or something, we don’t see kind of neurons.

Well, you should have been more active for that black line. Now, parts of that reason could be that we wouldn’t see them now like, because most of the time your visual system probably gets it right by the time you’re an adult. So it would be hard to see them even if it was like that. But the other one is, you can say it should predict that this kind of this subnetwork that kind of runs something else than the first network.

We’re not really seeing something like that. You could say, in terms of connectome, six ways of quantifying how networks like really make their connections. We don’t see kind of like these like these parallel networks. That’s one theory. Not like there’s an extreme theory that very few people would hold because the data doesn’t go along with it.

But there’s versions of that where you can say, well, I don’t actually need a twin. It could, for example, be the same neuron at a different time where you could say that it’s a part where the neurons tell you what to do, and then there’s another time where they’re like, okay, let’s like kind of fix ourselves so that we wouldn’t have done that.

And there’s a lot of algorithms that have been proposed as theories of how brains learn that basically do that, where you use the dynamic switch if you want, like the neuron has a time where it makes the decision and the time where it learns in the first one, it’s like if you want feature like it, it represents something that’s in the outside world and at the second time, it’s more like it tells you what was good and what was bad.

Walter Senn has a version of that algorithm. Blake Richards was involved. Yoshua Bengio has a version of that algorithm. It’s even me doing my PhD had a version of that algorithm. But these algorithms all have in common is there’s no blatant biology violation, and there’s no evidence that the brain does it like that.

So we don’t know from the theory side. So what we do know is that the brain could be doing it realistically. Now there’s this thing. If I think about evolution, we know that algorithms that do something like backpropagation, like gradient descent, are really great and work on real world AI prompts, and that algorithms that don’t do anything like that aren’t very good.

So in that sense, I believe that evolution is amazing and we have a very long history. And you should have expected that kind of real brains that do the right thing, that do good. Algorithms that if it’s if it’s realistic, if it’s easy, as in those algorithms, they all like really pretty simple algorithms that people propose.

If it’s easy, then biology should have figured it out. So that’s why people in that field realistic gradient descent or however you want to call that field. I think we all agree the brain should be doing something like that. Therefore, because of evolutionary thinking, we should strongly believe that the brain is doing something like that.

But the data, the evidence for that is really weak. Now let’s talk about what the evidence is that we have. The first evidence is that learning in biological beings is pretty amazing. You make a mistake, you grab like a coffee cup, you make a mistake, you’ll be like this next time you don’t make that mistake.

Imagine the cup is heavier than you thought. And I’ve done some experiments, I. You know, I’ve done some movement science in my past. You do some mistakes with your coffee cup. Next time you’re, like, 30% better, like half a second later. And in general, if you look throughout the biological kingdom at learning prom, like we always get better at things.

So. So therefore we kind of know that learning must be good, because if it wasn’t good, you’d be. It’s not. It doesn’t look like you try some stuff and like, oh, this time I was even worse at picking up my coffee cup.

Juan Benet

There’s very few examples, you know, compared to like, the millions that we use for artificial neural networks today.

Konrad Kording

That’s right. You get better at pretty much everything. Now, that is not true for a lot of neural networks. If you’re not very careful with it, if you’re like, make them learn very fast, you’re even worse. And that’s despite the fact that we built in a lot of like these tricks. So it seems that like the algorithm used by the brain is very good.

We know that it’s biologically realistic and the relevant experiments haven’t been made yet. So let me talk you through what we know biologically. So we know that there are Hep like phenomena. You have a local synapse. If the presynaptic and the postsynaptic neurons are active at the same time, the sentence might get stronger or weaker.

Juan Benet

Neurons that the wire together and fire together will fire together more.

Konrad Kording

Yeah. Yes, exactly. Well, why are together more so? It’s a little bit like gradient descent. So imagine the neuron that you get input from isn’t active at all. Why would you change the weight? Because the weight from the presynaptic neurons would not make any difference at all. Imagine the neuron after the synapse isn’t active at all.

Why would you want to change the weight? Well, the neuron is not active. There’s no change that will make it active. So you can say that if the brain wants to do something like gradient descent locally, things should kind of look a lot like Hebbian. You know, like no activity pre. No learning. No activity post.

No learning because gradients at zero. If nothing is active, we can go go further and say if the presynaptic neuron is active before the postsynaptic neuron, we have a lot more plasticity. Well, if your presynaptic neuron was active afterwards, it cannot change you because well, it’s not active before you do things.

Therefore there’s no influence and therefore the gradient is zero and therefore you shouldn’t be learning there. So if you look at these local plasticity findings, they kind of look like a little bit like gradient descent. Now does the brain do a gradient descent. We don’t know because the experiments for that have not been done because the people who do synaptic plasticity, they kind of look microscopically what’s happening in the cell, they usually wouldn’t have a even know.

They often use slices. It’s all very local. They wouldn’t know what the gradients would be, so they can’t quite test the big idea there. And then there’s a cultural thing, which is the people who work on gradient descent and brains, they tend to be adjacent. They tend to not sit in active biology lab. So the theories that they come up with usually don’t get tested.

So we are at this moment where if you want, we believe that brains do things that are pretty much like AI, and we have microscopic data that’s kind of compatible with that. But there is no, like, clean, convincing paper that the brain actually does it. And that’s despite the fact that it’s pretty obvious how to do that experiment.

It’s just that the that the fossils never align to make that question be answered.

Juan Benet

Let’s talk about the limitations experimentally. I mean, we have the ability to grow neurons, you know, separately from brains, we have some amount of ability to look into specific brains and record small areas of the brain. We haven’t figured out how to record large scale full brains, but maybe we can record like C elegans and smaller organisms, maybe talk us through, like, what can we figure out experimental today?

Like what is like the current state of the art and like our ability to record and understand generate the data that we need to then figure out how to test theories, and then separately, like where do you see this going? Like how is the measurement and recording landscape evolving?

Konrad Kording

Yeah. Let me first briefly sketch the experiment. How would we find out if the brain does gradient descent. And then I give you like a broader overview of like what are the tools in our toolkit. So what is gradient descent? It’s a very simple idea. It says if it would make your decisions less bad, if the neuron was more active, if it would make your decisions better, then the neuron should become more active.

And if it would make your decisions better. If the neuron was less active in that given situation, it should become less active in that given situation. It’s a very intuitive rule here. Now you can combine two experiments that a lot of people in this world do that would directly ask that question. The first one is if you want to know how, if a neuron makes you better or worse at a task, you can absolutely do that.

You can go in and you stimulate the neuron in a given situation that changes the behavior. Therefore, it makes you better or worse at the task. And a lot of people have run these experiments by brain region in an animal show how it changes behavior. And of course the behavior changes how good the animals are.

That. So that’s one piece that if you want mathematically is forward differentiation. You go into a function, you change something and you see does the output go up or does the output go down. It’s it’s and you just what you want to do is how much does it go up and divide it by how many extra spikes you did, which is then the derivative of the output?

The reward if you want after the activity of the neuron, that’s the first thing. That’s what we call a gradient. A gradient just means how much better it gets. If you make it more active, then what you need to know is how that change now happens. You can do a different experiment that a lot of people do. Where you say, I take the brain and record from it.

Then I give a learning task to an animal and I ask, do the neurons change? Lots of people have done that. If you don’t want to test if the brain does gradient descent, you just need to correlate those two. It means that those neurons were making them more active makes you better. The task should be the neurons that become more active.

If you actually do give them that task. Like it’s a it’s a it’s a direct test. Now we can do that because people are doing both of these experiments. It’s just no one does both of these experiments at the same time. There’s nothing holding back people from doing that. Apart from the fact that a lot of people have, if you want this idea of like take AI as a way of thinking about brains isn’t yet an established way of thinking in neuroscience.

So I think that’s why those six pumps haven’t been done.

Juan Benet

Interesting.

Konrad Kording

And now let’s talk a little bit about like our toolkit at neuroscience. So one of the things that make me so excited about neuroscience is like this dramatic explosion of toolkits. So in the past, we could record from a small number of neurons. Recording from 100 neurons was heroic when I was a PhD student.

Now you can buy one device, stick it into a brain, it records from a thousand of them, and of course, you can stick a whole bunch of them into it. So we can now record from lots of neurons. My lab discovered this thing called Stevenson’s Law, which is it doubles every six years the number of neurons that you record simultaneously.

But there’s now these modern techniques coming that promise millions potentially more. At the same time, we can do much larger numbers by going optically into brains than going going electrically. So there’s this thing we can get far more data out of brains when it comes to the electrophysiology of it.

We now have molecular tools to do these things. So we know there’s many different cell types, and we can make it that we can only record from the neurons that are like type 17 into neuron with the following molecular properties. It’s like the rest of the brain becomes invisible. And you see only those nodes that are of the molecular type that you currently want to study.

Unbelievably cool optogenetics. Deisseroth and Boyden popularized that. It was a protein they found in some algae. It basically is this molecule. You shine light on it. It lets in current letting in can make the neurons more active or less active, depending on what kind of current you like it So with optogenetics you can basically go in.

You illuminate a neuron and it becomes active. Now you can illuminate one neuron or group of neurons. You can illuminate them again. You can like make them go up. You can make them go down. It’s kind of like you can play it like a piano. Unbelievably cool. Even like these latch and release techniques, you shine in one color of light, and every cell memorizes what it was doing at that point of time.

Like, I was very active, I wasn’t active. And then you shine another light on it. And the neurons that were active basically like replay being active, they all get activated. The ones that were inactive don’t replay getting active. So you can kind of like recreate a brain state if you want, like unbelievably cool, not like.

And then come all these anatomical techniques. We were always able to use em to reconstruct a neuron. And I get my PhD in a lab where Kevin Martin was doing these experiments. And it’s unbelievable. Not like they were basically drawing the neurons by hand. And it took forever to do that. And it was really unbelievably complicated.

And now we have these automated machines. They have like a hundred beans at the same time. And we have AI algorithms that like draw the neurons and they don’t do draw one, you know, and they draw like a million neurons. And it’s just that speed of development is unbelievable. We’ve done some analysis. Imaging is getting cheaper by a factor of two like every 18 months or something.

It’s unbelievable. It’s almost imaging is getting cheaper almost as fast as AI is getting cheaper. So we have like all these techniques that are coming online and by and large we are asking the same questions we asked 70 years ago. We just like do it with monuments and with more beautiful neurons and now we can see them.

But how to really like gain deeper insights with it? I think we’re still working on it.

Juan Benet

Part of the answer there might just be a scale thing, right? Like a, you know, in artificial neural networks. You just have to wait until the networks were big enough to be able to compute some very complicated things. So maybe now that we have the ability to look at millions of neurons, then we now have the scale of data required to be able to kind of build much more robust models, potentially.

Yeah, that’s the idea.

Konrad Kording

But let me push on where we’re like failing at even seemingly simple things. There is this worm that has the arguably the simplest nervous system that exists. It’s called C elegans. It has a huge behavioral repertoire. It likes forages for food, it finds partners, it lays eggs. It does everything you and me do, minus the talking, kind of like almost.

But it kind of has a really rich repertoire of things that it does. It only uses 300 neurons to do that, and we can’t simulate it now like that’s 300 neurons. We can see them all at the same time. Why can’t we simulate it? Well, why can’t we simulate it is because and we know all the wires in it. We know we have a name for each of the 300 neurons that it has.

We know a lot about the molecular properties of that. We know every single wire that exists there, and we don’t know it just for one one. We know it for a handful of them. We kind of like, in principle, have all that information. And yet our simulations are barely worth the papers that we print them on. My own lab tried to simulate C elegans based on that information.

We’re horrible at it. And so basically we have these like explosion of techniques. And yet we are not able to simulate relatively simple systems.

Juan Benet

So why not? Yeah, there’s been a range of projects trying to both record and establish the connectivity between the neuron, like the entire connectome of the of C elegans and various different attempts to model it. So. So what’s going wrong there? Like why don’t they work yet?

Konrad Kording

So the first thing is a lot of data that we have is not perturbative. So and let me highlight why is this problem for exponential explanation. Imagine a record from all three neurons at the same time. You’ll be like shouldn’t you have all the information you have like all the neurons? Isn’t that cool?

We don’t quite usually. No one like is quite recording for all of them, but we’re getting very, very close to it. The big problem is they are all correlated with one another. Now imagine you had two neurons that would always do the same thing. You wouldn’t know is it neuron A or is it neuron B that does it. The problem is there’s not two neurons that always do the same thing.

But if one there’s some dimensions that carry a lot of the variance. Very large principal components, singular values depending on the language you’d like to use. And we know an awful lot about those. But if we want to get at the function, we’re effectively solving an inverse problem. Like I give you the activity of the 300 neurons as a function of time.

And I ask you, can you tell me how much each of them influences everyone else. If we do even linear regression, which is the simple step. We need to calculate this covariance matrix of the neurons with everyone else and invert it. Now that inversion requires basically if you have small singular value, then it becomes a so-called ill conditioned prompt.

So you can’t solve that. Therefore, you cannot know because there’s dimensions in which you basically have no variance. You cannot know what’s happening there, and therefore you cannot find out based on neural recording how they interact with one another.

Juan Benet

But these are the type of problem that current machine learning techniques are very good at modeling. Like if you had enough data and enough recording of worms through a whole bunch of different behaviors, and you have activity traces of all the neurons across long enough time scale, shouldn’t you be able to fit an artificial neural network to that?

Even like relaxing a bunch of the constraints and having a pretty large artificial neural network against it?

Konrad Kording

Yes, you can.

Juan Benet

But.

Konrad Kording

Here’s the problem. So there’s a forward modeling problem that machine learning solves, which is give me a prediction of what these 300 neurons will do is sometime into the future. And it’s amazing at that. We’re really good at making those predictions. And then there’s the inverse problem, which is what happens in the neural network to make those changes happen.

The forward prompts are very easy. And the more low dimensional the worm is, the easier, because like it’s basically just changing along three dimensions, like making up some numbers here and because it’s low dimensional machine learning works great. The lower dimensional problem, the easier machine learning.

We have a really good understanding of that, but there exists an infinite number of nervous systems that would produce the same changes of 300 neurons. And in general, when we use machine learning, we are great making predictions. We are horrible at understanding how the world works and we don’t need to.

We just need to make predictions. And how does machine learning do it? It basically says, well, there’s a lot of dimensions. Let’s like distribute what we do along all those dimensions we don’t know. But like as long as we’re similar enough in this situation it’s okay. We’ll just like put a little bit on all dimensions and we make really good predictions with it.

It turns out that for making predictions, if the world is low dimensional, it makes the problem easier. And if you want to understand how the world works based on your data, you care about the inverse. And therefore, if it’s low dimensional, it’s impossible to understand the world. And machine learning runs exactly into this palm.

And we know in machine learning, let’s say if we build predictive models, a lot of people are new to machine learning. In fact, almost everyone who’s new to machine learning makes that same mistake where you’re like, oh, I fit a machine learning model. It’s good. And therefore the machine learning model would tell me what would happen if I would reach in and change something in the world.

No machine learning is horrible at that. Machine learning generally doesn’t work well if you go far out of domain and reaching into the world and changing something is bad. And let me give you an example of like how that comes about. Imagine you build predictive model of like death of people and you’re like, okay, let’s take like things that we have in the electronic health records.

Maybe people who work out die less often. People take vitamin D, die less often in a given period of time. Now, why could that be? Well, not like a machine learning system. What do what it should be doing, which is vitamin D, predicts that you’ll kind of like live longer. It might also just be that people are rich.

Take vitamin D because they like fall for like snake oil science and therefore it’s an illusion. It’s just like vitamin D is a great predictor of your living longer. It’s just like people are like rich. Like they also work out more and they go and like see a doctor earlier. And that’s kind of all kinds of causal chains.

So a machine learning system, it’s like full of these like predict things from the wrong things. If the measured world is low dimensional, which it always is. And every machine learning problem, in every learning problem for brain centers we cannot solve the inverse problem. There’s like a million things that could all have produced the same input output behavior.

And so the fact that we’re good at making predictions is confused universally by neuroscientists and machine learning people alike, as evidence that we understand how this system works. No we don’t. We’re just good at making predictions.

Juan Benet

So just pushing on it a little bit to get a crisp intuition. Can you create a structure where you train models that can be good at prediction, and then you either distill them or you figure out, like the simplest possible models that can will still exhibit the right behavior. Shouldn’t that give you a sense of what actual computation that system is running to be able to run a better and better abstraction of it, that should match events outside of the distribution that you’ve measured.

So this would be the equivalent of saying you build some large artificial neural network model of C elegans. You kind of potentially even constrain the training of the network to maybe mirror dendritic trees of the actual neurons. So then figuring out like segments of this network have to be learning closely, close enough to what those individual neurons may be doing.

And if you do this over enough data distribution, you should be able to then distill out some representation of what functions those neurons are actually doing, or what’s the or is there no hope? How do we get. Like clearly there’s got to be some answer. We have to be able to get the data that we have and the computational model of the system into something that actually represents from an information theory perspective, like what the system is actually doing.

How do we clamp this down? How do we get to the answer?

Konrad Kording

What I love so much is that you kind of sketch my own thought patterns over much of my career, and I started exactly at this corner. Like, let’s just like use like simplicity. Let’s use prior knowledge. To kind of constrain things and maybe it will converge there. Let me kind of try and squash that idea. So if what we have is low dimensional information because we don’t see all those dimensions, then you can say it’s like we measure a small number of parameters.

If you measure a small number of parameters, but like you measure 100 of them and you live in a system that has a billion free parameters, there is basically a billion -100 set of possible models that can produce exactly those hundred measured dimensions. And now you can say you bring it down from this space to maybe which neurons connect to each other, neurons.

And then you go from a very, very large space to still a very, very large space. So you’d need to be able to bring it into that lower dimensional space. Now, like let’s take the key elegant space like we have 300 neurons. So in principle the covariance matrix is 300 by 390,000 free parameters. Now you could say if I use.

Juan Benet

Although from what you were saying before, if we were trying to model these neurons with artificial neural networks, shouldn’t, shouldn’t be more like 10,000 parameters per neuron.

Konrad Kording

It could be worse.

Let’s not let’s let’s not. Yeah, let’s let’s let’s let’s let’s let’s Steelman that approach. You can say there’s a large number of parameters. Now let’s use the full connectome now like C elegans neurons 10,000 ish synapses. That means like we only have 10,000 synapses. But now imagine that this the 300 neurons really slosh around in a five dimensional space.

You just measured 25 parameters in a space where even if it was all just linear weights, would have 10,000 numbers. So you measure 25 out of 10,000, which means that there’s like 9975 different like dimensions orthogonal to that. And you can basically pack any possible model to it. And this before we even touch like this sort railways like.

Yeah. Now they’re also all non-linear. And they also might have like multiple violations of our modeling assumptions. And so in noise in the in the data that we have now there’s better approaches probably like Andrew Leifer has been doing these really cool approaches where he goes in and stimulates neurons.

The problem is they still need neurons. If you could stimulate them independently, maybe you could get the data out. But the problem is it’s like kind of like hard to do enough experiments of that. The good thing is, though, they separate it now. It’s no longer sloshing around in like a five dimensional system.

It’s kind of like you at least go in for every one of the three other neurons, and you push that button. And, so far, at least in our hands, we can’t take data like that to produce like, really good models either. That’s probably related to there’s still a lot of noise on it. It’s still not because now we are talking about very, very small number.

We have a lot of neurons, but we have a small number of experiments.

Juan Benet

What are you thinking now? Like where do we go from here? What are we missing? Is it. Yeah. Increasing the ability to perturb the system. To extract, like, higher quality data and signal out of that? Or is it like different modeling approaches or like what?

Konrad Kording

I mean, look like? I tried it for 20 years of my career, trying to solve 25 years to try and solve this inverse problem, basically. How can we look at the output of neurons to find out how they compute? I now understand the inverse problem we’re solving much better. I’m reasonably convinced that in that space, there’s no credible solution for maybe C. elegans, because we could at least stimulate it independently, but probably not for larger systems.

Therefore, I think we need a different approach and a different approach. From my perspective goes through a connector mix, molecularly annotated tissue perspective and the logic that is very different. Instead of saying I observe the output of the nervous system, and I try to figure out what the inputs must have been or what the interactions must have been, you can rather go and let’s say, can I see signs of the interactions?

Can I see an excitatory sign ups between two neurons? Then I don’t need to kind of solve an inverse problem like I see an excitatory synapse. What if we could say look at it and say, well, there’s a big excitatory synapse. And based on all the molecules that I see there, it’s also pretty strong and fast. So I feel that we need a different approach.

And I feel that connectome or like the broader field of like imaging can give us that, but it forces us to completely rethink the logic of neuroscience.

Juan Benet

Concretely, it means instead of just trying to look at a smaller sample of the neurons that are given moment in time and recording those, you’re able to look at all of the individual weights in between each of the synapses and trying to establish what those actually are.

Konrad Kording

That’s exactly right. Now, like if we if we if we stress the AI analogy, the deep learning analogy, if you want like instead of figuring out the weights based on the outputs, which is hard and probably impossible, let’s rather see the weights. And the idea is that the weights will somehow be represented.

And there exists some data where people, for example, found that if I know what these two cell types are, pre cell type and the post cell type, I will often be able to at least. And I and I also tell you how big the synapses I’ll be able to note, like a fair bit of how strong that is. And people are now pushing that now, like, what if I tell you the transcripts of the basically which molecules are being produced in pre and post cell?

It’s a big step forward. I think that there’s a whole new field emerging.

Juan Benet

So this goes into like the broader economics, another omics approach to figuring out starting with C elegans and then Drosophila and then larger and larger mouse and then eventually primate and human. It would seem to me that getting a, a computationally valid system, where that demonstrates that we can recapitulate the actual organism and the full behavior of the organism, like we’re still pretty far away from that, in that we currently are doing the recording, and we might be able to establish the wiring diagrams of these, and maybe it’s molecularly annotated.

So you don’t know, just the two neurons are connected. You also know in the ways in which they’re connected or you have some indication of the parameter strength. Some parameter space will be able to model the thing. But how far we suppose that we do that. So whether we get like a good molecularly annotated connectome of either C. elegans or what’s missing from that, to then put it into a simulator that actually recapitulates the entire organism.

Konrad Kording

So I think what’s currently missing is what I call compilers. So at the moment we have these images and we have the connections, we have a list of wires and we have methods that efficiently get them to us. But what we don’t have is what’s between those wires. Not like between those wires. Between two cells there is a synapse, and the synapse has a certain strength.

If we go with the AI analogy, in reality, biologically they have like timescales and plasticity rules and then lots of like a maybe a dozen parameters or something. But what we currently completely lack is mechanisms that take these images as input and tell you how neurons influence one another. All we have is the wiring diagram.

So the focus in the past was really like, let’s make wiring diagrams. Whereas I think we should ask like, what are the things we want out of wiring diagrams? I think what we want out of wiring diagrams is we want to look at the neurons, and the images need to tell us how they interact with one another. And for that we need this concept that I call compilers.

We can say what goes in is if you want like the portray of the sign ups, show me how big it is. Show me. Show me how many receptors molecules it has. Show me how many of various second messengers it has. You name it. And can you tell me how strong it is? Can you tell me how fast it is? I believe so, no, this like a standard machine learning problem of the kind that we know we are good at solving is, which is let’s get how strong and fast and so forth.

A million sign up system. And let’s get at how they look like. And then let’s find out how we can predict how strong and fast they are based on how they look like and the molecules that are their molecular annotations. All these things. But compilers are super cleanly defined. Prom, which is basically you measure how fast they are.

You measure how strong they are. I give you the image and then after training I’m like, here’s the image, how strong is it? And you’re like, this the Peco and current that will come through it. And I’m like, okay, you’ve been right the last thousand times I trusted you that you have this down. And then I’d be very happy to build systems based on this.

But this bridge we don’t have. The way I view it at the moment is like neuroscience has these two halves, has the physiologist stick electrode and so brain so optical recording, see what’s happening while they do things. And then there’s other pieces which is like make wires, make a wiring diagram. But as long as we can’t convert the second into the first it’s Its kind of two fields that I mean, like they touch one another at the level of storytelling.

So if you look at the fly field, they do look at the ways and they’re like, look like if I look at these wires where they come from, the visual cells, like maybe there’s motion and they’re better than chance, but kind of there’s still this translation is extremely qualitative and I think compilers could make it very quantitative.

Juan Benet

Describe a roadmap or a path to solving this problem. What would you do or what is like in that technology tree of like how would you break down the larger problem today? And like what’s what should we be doing over the next few years for compilers?

Konrad Kording

What you need is primarily a training set or calibration data. The way it works is you do relatively old school synaptic physiology. You go in, you measure what the current is that flows to a synapse, and then you freeze it. Or like you stop all reactions there and you store it and then you reconstruct it where you know how it looks like, where all the molecules are, and so forth in it.

And then you don’t do that for one synapse, as traditional papers do. You do it for a million because like, look, we have like robotic patch clamping and we can do it optically. There’s a lot of techniques that scale. Let’s not go through all of them. But like you use techniques that scale to produce a very large number of them.

And then you use a machine learning algorithm for that. And basically what we need to do as a field to build that bright. What I’m trying to say is that bridge is necessary for progress. And it’s also clear how to fill that bridge. Like, you just need to.

Both sides of the bridge at the same time. And it’s not rocket science. It’s perfectly well established standard neuroscience where you’re just like, show me molecular annotated synapses and tell me how strong or fast they are. And then you predict one from the other. And it’s a very concrete research program, and it will have trouble getting funded because it’s so different to all the other things we currently do.

Juan Benet

Problems like these on a path to a much larger swath of solutions can get funded, right? Meaning if we do this and we understand organic neurons dramatically better, and we then can figure out how the brain is doing learning, and that distills out insights that we can then use to improve machine learning.

That already is like a super valuable contribution. Then will greatly accelerate the field of AI and like let’s pay back itself like many times over. Or.

Konrad Kording

Yeah, but hold on, these things are pretty far downstream. So let’s be clear what a compiler does for you in an image. And it just comes knowledge about what currents there are. If I told you for a given neuron what current every synapse produces, it’s relatively easy to then even do the fitting or kind of like find out what the neuron does, at least in simulations.

But that’s not been proven out yet. If you know how a neuron converts its inputs to its output, it should be relatively clear how to get to a socket or even a whole brain. If you had a whole brain.

Juan Benet

It.

Konrad Kording

Would be reasonably obvious how you could improve AI systems. But if you want, the compiler is just the best defined first place. It’s also a major enabler for neuroscience. And like you can ask a million question if I tell you like look here, this synapse does the following thing. You can say, well how do molecules make that happen?

You can say, how is this thing different in the disease? You can say, no, it opens up a huge space. But from there to and now we improve AI. There’s like a multiple there’s multiple steps still that are necessary for that value. And that’s why I’m a little bit worried about like the funding future of the field, that basically we can break it down into these steps like compilers to synapses, synapses to cell cells to whole brains.

They all come with their own set of benchmarks. And it’s clear if I if I had a bunch of people working on it, how do I tell them if they’re on the right path? But to get from there to value if you want, like all three pieces need to be solved, but you can only start work on the second once you solve the first.

Juan Benet

So it’s very concretely like, you know, if current capital expense to train the current machine learning models, like we’re hitting scales of capital that are beyond anything we’ve seen in history. Well, there are these graphs that show that a fraction of GDP, this similar to different industrial booms like the railroads and oil and similar type of industrial output.

But in terms of like a pure science and engineering sort of discovery processes, if we can figure out how to dramatically increase the learning quality of machine learning models, that can be worth, you know, tens to hundreds of billions on its own right now, which then suggests that it should be easy to make the case to these groups that if we can figure out how the organic neural networks work and larger and larger circuits that then enable us to truly understand how learning works in brains.

Maybe not even a human brain. Like just in a mouse. Exactly. If we can get a mouse brain and understand how learning works in mice, and there are concrete insights out of that then massively reduce the expenditure for quality and artificial AI like that. That’s super valuable, right?

Konrad Kording

And it sounds like a good but like look like there are lots of risk factors in the way. For example, now let’s go through some risk factors. Let’s talk about risk factors for Comillas. It’s possible that every synapse is unbelievably complicated. Yet if you can’t somewhat compactly describe how synapses work, you’re dead in the water.

Now, with everything that we know about synapses, it seems very likely that’s the case. But that could go wrong. It could be that you need more molecules than you can realistically measure to be able to do that. We expect that it would be upper bounded by the number of proteins that exist in human cells, but it’s still a very, very large number.

But it could be that you need 20 of them. We don’t know how many you need at the moment. So there is a risk factor on how complicated synapses are. There is a risk factor on how complicated neurons are. Now what do I know if like quantum gravity and light is actually the source of consciousness, then no amount of compiling will solve that.

Now, with the things we currently believe we know about neurons, that shouldn’t be the case. But there’s real risks about how complicated neurons might be. There’s real risk about like how the interactions and brains could be complicated.

Juan Benet

But all of that sounds, you know, par for the course for the kind of risk involved in, in artificial, you know, in AI research today where, you know, people are trying to stretch the boundaries of how to get the current transformer architectures to perform better, like just even like the chain of reasoning model, which initially started as like this big Cluj around the entire thing.

Just that was an incredibly expensive thing to try to do, and it worked out. And that’s the one that we’ve seen work. There’s a lot of other things that people have tried that probably were worth tens to hundreds of millions of dollars in training costs alone that never went anywhere. So, like the AI labs have to have a portfolio approach to research where they have to try a bunch of different things to see which of these will catalyze improvement rate.

And, you know, maybe 2 to 3 years ago, the capital expenditure there was in the millions. Now it’s in the tens to hundreds. And so that’s just like a very different scale. It’s like.

Paying $100 million to like, see if we can figure out, like how my brain learns. Seems like, like an acceptable research question that Frontier Labs should engage in today. And if not today in a year.

Konrad Kording

I’m with, you know, like, I’m not saying that this isn’t the right path forward. What I’m just saying is I can give you a long list. In fact, the compilers paper has an appendix of 20 different ways, so this could be not working. Yeah. Including like no, there’s ethical risk. Like is it acceptable to if you would simulate what, you know, like a mausoleum and maybe it’s not ethical enough for like everything gets shut down because kind of simulating a mouse in a not so great way is maybe unacceptable.

I mean, like what? You know, like there’s there’s a very serious. No, it could be that the resolution that you need that it really matters not just which molecules are there in the synapse, but where are they exactly? They are like a one nanometer resolution. If that’s the case, it would drive up, like everything by a couple orders of magnitude and we couldn’t deliver on it.

So there are risks.

Most things that we do in neuroscience, I can write down those risks for you. Exactly. Like here’s a list of things and here’s how we would know these things, and here’s when we’d know those things. So it feels like that approach of like, let’s do bottom up understanding of brains. So at the moment we try top down like we record the things and then we try and see like what kind of models are compatible with it?

With what I’m arguing, there’s just this huge set of different models that all produce the same input output, whereas the buttons.

Juan Benet

Are like describing the risk profile to be very clear and upfront and like, what’s the problem? Space. But you very strongly believe in this approach in terms of like the outputs that they’ll generate.

Konrad Kording

That’s, that’s that’s right.

Juan Benet

I like to generate.

Konrad Kording

But there’s like glue pieces that also carry conceptual risk. For example I told you about cable equations. If a cable equation just is really bad at modeling how a neuron works, then it’s possible that you run into like fundamental problems of producing the models. And if you can’t model it, then you’re blocked on your value chain and you can’t ever do anything out of it.

So yes, but it’s an orthogonal approach with a risk structure that is entirely uncorrelated to the rest of neuroscience.

Juan Benet

Yes.

Konrad Kording

Yeah.

Juan Benet

We’ve been talking about neurons and circuits and brains and so on. And we’ve been talking about kind of modeling them and creating some simulations, but we haven’t really discussed the simulating part itself. So. One part of mapping the circuits is to just kind of understand. How they work. But what is a great value that can come from this?

Like what are the range of possible applications that we can get from both? Understanding how the brain works or understanding how neural systems work and simulating them, or even simulations without understanding. Right? There’s a lot of utility in creating these simulations. So what is the universal possibility there?

Konrad Kording

I think maybe it’s good to start first, like stretching a little bit ways of understanding. Because ultimately what we do in science is epistemic. So at some level there’s there’s a logic behind what we’re doing. And I think there’s two ways of understanding that we often mix with one another. One is what you call traditional science.

I tell you principle neurons are cells, and between them there’s a synapse. And you’ll be like, okay, Roger that. Like, now I have something that’s useful. And if you want these things, do two things. They describe things in the outside world, and they help us think better. And in that sense, if you want like a lot of what we know in neuroscience is of that nature, not like we know about synapses and Hodgkin, Huxley models and neurons and simulations.

There’s a second type of science that’s coming up, which I call machine science. So like take AlphaFold. It’s great that you put in a protein and out comes how it falls. It’s a huge enabler for things. But if I give you this, this system, if I’m like, okay, look, here’s AlphaFold, run like, tell me about chemistry.

You’re like, I don’t know. Like it works very well. And that’s very dense. So I think those two are different and they’re valuable in different ways. If you give me a simulation even of a human, let’s say, of a of myself, I give you on a hard disk, simulate Konrad’s and you run it and it works perfectly, it’s useful in some ways.

For example, you could say, here’s two kinds of coffee. Well, this simulated Conroy would prefer coffee A or coffee B. Great wax is useful because you can make the world better fast, and you can generalize it like we move into an AI world. Would people be happier in this kind of a world or that kind of a world for that?

You don’t need to have this traditional science understanding you. It simulates Konrad. You’ll be like, I don’t know, there’s like a trillion variables on this thing and like, it just works. Like Konrad, there’s this distinction of understanding of like human understanding, which we don’t get, and simulation quality, which we do.

In that sense, the machine science is useful because you could also say for AI alignment, if I can simulate a Konrad, the problem usually in the world is you can never have people compare A and be in the world ever, because it’s going to be the one. Now that gets like one coffee and the one tomorrow that gets the other coffee.

Well, if I could have a simulation of you at the same time, no problem. I, both of you, get the coffee right now and I can read out later. I can ask each of them and I want to attend scale. How do you rate your coffee? So it would be a major enabler if we could simulate animals or humans for that matter. There is the human possibility.

It’s possible that if I gave you this full simulation of Konrad, you’d go like and like, look, there’s this piece of Konrad that I now understand. There’s this neuron that every time does this thing. The question is, how real are those stories now? Like we might be? If you wonder if we look at the history of neuroscience, we often have like, oh, look at this.

This does X only that. Then a decade later we found out that it also does y. And now we find out that it does x, y and so many different things that Simeon and I think this history like shows us the limits of human science versus machine science now, like they exist neurons, they are like super specialized neurons in your retina that show you that like right there, there’s something bright.

It’s great. Wax generalizes. If you go all the way in the middle of my brain. One of those neurons might be about your face, but also about how I feel about you and maybe about relationships and so on and so forth. In that sense, there might be areas where we can only have machine science. Now, we might still be able to say, build better AI with it.

Maybe we can take this Konrad simulation and we can be like, can we fit a loss function of what Konrad optimizes for? And the answer is yes, good quality, quality of coffee. But we may not see that readily, and we might be able to build systems now that have reward functions, loss functions that mirror how humans think are the humans that we could simulate if we could simulate them.

So I think there’s these two the human understanding and the machine science piece that are different.

Juan Benet

Logically, having a more concrete set of descriptions like what is like sort of like the downstream utility or value that people can get out of it, right? So like the machine science, understanding that you’re describing that you can get if you kind of understand how certain neurons work, then naturally you can start figuring out the nature of disease or the nature of some problem and start like potentially fixing things.

Or, you know, a lot of like the neurotech applications look like. We understand that the brain does this particular thing. We see this issue in how the brain signals are connecting to a part of the body or a set of sensory input. And so therefore like let’s kind of like patch this by finding some circuit around the kind of like damaged nervous system or like the damaged nerve or and so that’s one path to like repairing or restoring function.

So yeah, it’s sort of like the universe of utility that you can you see out there in, you know, suppose that we can do the, the machine science and, or the simulation. So maybe we have one or the other. Like what? What is it like the downstream value that people get out of it?

Konrad Kording

Yeah, I think that’s even both a human science and a machine science disease application. You could say a human science application. Take Parkinson’s disease. There’s some neurons that die because they die. There’s a certain function everyone can’t do anymore. Once we understand what that function is that’s missing, we can now build drugs that help with it.

We can stimulate the brain stimulation. It’s one of the best things that exist there. The existence of brain stimulators is like one of these things that would be magic in the past. And then also like just embodies, like the wonderfulness of neural tech. To me at least, it’s just like seeing the patients is like believing in the in the research program in the case of Parkinson’s.

And that’s very like human science stuff. Like there’s neurons, you can see them, they’re different from other neurons. And they have a pretty simple path to the disease. Now you could say from a machine learning machine science perspective. Maybe the way my reward function in my brain is implemented is really complicated.

Maybe distributed over 17 brain areas and a trillion neurons. And I could still use machine learning approaches to, say, distill them into like a good compact descriptor of what Konrad cares about or maybe what people care about. So I think both of them, you know, and you can say a lot of diseases might be related to reward functions are learning algorithms.

And so for example, you can say, if I have a disease of the brain, we usually view it as there must be something wrong with the molecules. But it could be just as well that there’s something wrong with the learning algorithm. It moves us into a space that is not a healthy space. and in which case what really happens is upstream of the neurons and we might just be working on the symptoms instead of working on the causes.

And,

So, so I think I’m just trying to sketch that there’s like a really diverse set of, like, values that we’d be getting from it, but you can still group it into areas. There is the group of, of areas cure diseases. There’s a group of areas build better AI systems. There’s a group of areas align AI systems with what humans care about.

There’s a lot of different buckets. And I’m at this point of time unclear where the main value is. And there’s the other thing that no one ever talks about. Like, if we could, like, simulate human beings or if we could even simulate a mouse, then it would just be such an unbelievably cool thing, you know, like, it’s kind of like there we have a system with like billions of parameters, and yet we can kind of at least simulate it and potentially kind of come up with approximate descriptions of that.

That’s kind of like in the in the brag sheet of mankind. Not like eventually. Well, the sun blows up and like people will look at like, what have we done in the time in between? And it’s like, understand our own brains. But for me, be very high up there.

Juan Benet

Yeah. I mean, the last one like, to me, it’s one of the things that attracts me the most into neurotech neuroscience. and to me, it seems similar to AGI and ASI as a concept. For decades, scientists and authors were able to look ahead of where science and technology were going and kind of predict, in broad strokes, roughly how machine intelligence would be built and develop.

And we got that early in the 20th century, like mid 20th century. We got a lot of sci fi books about written about it, but the technology was just so far away that we then kind of watered down the possibility space to the point where, like now, we are finally building these AI systems to be close to or surpassing human intelligence on a range of domains, and suddenly, kind of like the mainstream opinion finds it hard to square the fact that we’re literally creating these incredible results, which are incredibly cool and powerful.

But yet, because it didn’t happen for so long, it just sort of feels weird or distant still. And to me, kind of building a digital version of a of a human or a simulated version of a human has that same kind of flavor, where it was also predicted in sci fi and by both scientists and science fiction authors way ahead of many decades, becoming possible to the point where right now, people are just not really considering the real possibility of what happens when we do, in fact, get there.

And like how incredibly cool that future landscape could be. And so I don’t know if you have thought about that landscape, like, how could that be a kind of optimistic vision of the future? What does that enable us to do? What is it kind of like a digital existence or not necessarily necessarily digital, but like a computational existence, open up for humanity the promise of the potential.

Konrad Kording

I mean, the question is what’s not the potential? So yes, if no, if it could if we could digitize an animal, we could transport it from A to B, no problem, because we delete it from one place, we instantiate it in the other place. It’s unclear how we should feel about that process. Like there’s ethical problems, there’s like logical problems, there are value problems.

But every change of technologies always forced us to rethink what everything is about. Like when when machines came, when the steam engine came around, people whose value came from, oh, I am so strong, must have felt miserable. And now that AI comes around, there will have been, there will be. People will be like, well, I knew everything about ex while in somehow on the internet.

X was there and then like now, like everyone knows about X and there will be proms and I think logic of if we’d be able to have simulations would have these similar trade offs that are weird. That being said, not like the first applications will not be oh, now we become like a digital only species, but the first users would be I can try drugs.

I can simulate what they do to a brain, and I can screen them before having to go to lots of patients. It feels like it would. Being able to simulate brains would allow us to speed up a lot of the things we do. That includes like drugs, electrical treatments of brains may even be possible to ask how? Sort of like how to optimize our environment for the brains that we have.

It might be possible for us to say which brain states are desirable, which for which brain states are not. It would open up a new set of capabilities, and people are often worried about these things, not with AI. There is this notion that just AI, because it’s so different, will be the end of human civilization.

A lot of people go around like AI, AI is coming, and then we all die. But I don’t think that’s the logic of the world. Like why? Like, hello? We are building it. We wouldn’t build it if it was just like. And then we all die. We’d build it because we think the world will be more magical, more wonderful with AI.

And we work towards simulations because the world will be more magical, more wonderful with the capability to simulate brains.

Juan Benet

Yeah. Let’s talk about AI for a while. Of course, the capabilities today are dramatically better than they were over the last five, ten, 15 years. You can now talk about building AGI like 15 years ago. We couldn’t. Right? Like, it was kind of in close rooms. People were talking about it, but like, you wouldn’t be taken seriously as a scientist if you like, even in the deep learning field, if you kind of said that you would build a machine close enough to human intelligence, and yet this happened like we are here now, like we now have models that can outperform humans at a broad range of cognitive tasks.

They still fail dramatically at a bunch of straightforward things. So maybe, you know, as you’ve observed the capability space growing here, what are some reflections that you have, like what has worked better than you thought it might this quickly or what has been way slower or like, you know, give us a sense of like how you look at it.

As a neuroscientist.

Konrad Kording

I think the main realization that I feel is missing in the debate is the extent to which artificial intelligence is different from biological intelligence. So humans kind of live on one or much of human intelligence lives on one axis now. Like, if you are very intelligent, you probably are good at reciting Shakespeare and solving integral equations and reasoning about rowing above a river in kind of distinguishing the superficial features of something from the deep features.

So there is this continuum. There’s this manifold of human intelligences. And when people think about AI, they often like try to project it on that manifold. And like the whole idea of AGI is like, there is this like axis of intelligence and intelligent humans that like far out are like, hi there. And like not so intelligent people like lo there.

And AI is somewhere on that axis where in reality AI is like just totally different. So AI’s are better than you at lots of things. They have been better at you and lots of things for a long period of time. And computers have been better at multiplying big numbers than you like since you were born. Webern. It’s not that there’s this one access and I’m like and at the same time, it’s then assumed, well, because they are great like summarizing a paper, you know, a complicated paper that therefore they could solve my short term strategic problems in life.

Yes. If you as a human being are like really great summarizing this paper, but like, sure, you could plan how I get to the airport from here or something, but AI doesn’t live on that manifold. And in that sense, that’s a that’s also one of the things that gives me a lot of hope. You know, like if AI’s were just like us, then yeah, that would be kind of set.

But they’re not they’re like really great some things and really not so great other things. And that’s why, at least for the time being, humans are great if you couple them with AI and like I like I use AI every day and I’m a much better me. Much higher quality is the things I produce are much higher quality because I use AI, and I think in the future this will just be much more the case.

And so, so like the main observation from the place where I come from is that this naive, like it seems intelligent and therefore it’s like us. That is a something I really have trouble with.

Juan Benet

We were talking about learning earlier the kind of neural network concept of learning, which is how do you adjust the parameter space of neurons to have the system learn some set of functions? But that word learning comes from a different concept, which is kind of how humans absorb conceptual information and draw relationships between these concepts and then extract some.

Like additional skill like that, you are able to generate an additional mimetic skill in your brain out of some prior context. And so humans are really good at that. That type of learning neural network version of learning, which seems like a much lower level algorithm. These two are maybe cast is the same thing by a lot of machine learning and artificial intelligence research, but to me they feel qualitatively very different.

Like, there seems to be like something is happening in the neurons, and something altogether different is happening in the conceptual space that the neurons are simulating or emerging. So I don’t know, how do you think about this? Like, like clearly we’re able to not just learn from a few examples, which is the characteristic case that people point to with AI relative to like the millions of examples that neural networks require to train or like artificial neural networks require.

But on top of that, we’re also able to reason about a complicated landscape of ideas and then come up with new ideas and or learn that some idea is wrong, and then shift our behavior and understanding mathematically. And that seems very different than a basic kind of like reward loop and weights adjustment.

That seems like some other computation that’s happening.

Konrad Kording

Yeah. Now, like it feels to me at least that in human behavior

We have a history of all our thoughts. And we not merely know how to solve the problem, we kind of know how to reason about paths that failed and paths that worked. And to have this like meta awareness about what a good thought patterns and what are not. And I think that is, you know, basically a neural network has no record of the thoughts that it had produced.

It might see its output tokens, but it doesn’t really it can’t really reason about the thought pattern it uses itself. And above all, I think we have. So we have these world models now, like I can close my eyes and like I can imagine I’m in a coffee shop and I sit there with my friend and I talk like, kind of like it.

We have these world models, and somehow we managed to get the information we have in our neural network things into this, like discrete world model where there are baristas and coffee machines and all that. And then once we play this thing, we can go back and kind of like put it into weights in our brain that we then use for like online decision making.

And I think artificial neural networks and like the whole deep learning field has nothing like that. And I think that’s why we have these complicated ways of managing context, because kind of we can’t hold it all in there, obviously in a large language model. So we have we try and build these elaborate memory systems.

I’m with you. There’s something fundamental that is missing. And again, like if we could like produce simulations of nervous systems, we might be able to really say something about it. And the brain has this structure now, like it’s not that it’s like layer layer layer, not like it has like stuff goes to thalamus and stuff then goes to cortex.

But when we sleep, nothing goes to thalamus. And there’s still kind of a world simulation happening and we can turn on and like we can kind of route information a much more freewheeling way in brains. So it feels there’s a big logical piece that’s missing there.

Juan Benet

Have people tried building large scale artificial neural network systems based on the regions of the brain and the functional utility that we think exists? Like taking a functionalist approach and kind of draw these kind of broad boundary boxes and connectivity between them, and then or is just completely fraught and like, unlikely to like, is it better to just kind of emerge everything from scratch because it is pretty crazy that the current Transformers are effectively recapitulating the entire process of evolution every time from scratch.

Right. Like, it’s just kind of like building a brain from zero. And, you know, it takes a few months to and like tens of billions of dollars worth of compute to then, like, output a new brain. But clearly, evolution learned a lot about how to build intelligence in that, you know, many 100 million year process that abstracted a bunch of pieces.

And these seems like those pieces are qualitatively very useful, but somehow we haven’t been able to, like, simulate these.

Konrad Kording

Yeah. Look, there’s a bottleneck. Every generation. The information we can give to our progeny is upper bounded by the amount of DNA that we have. And arguably most of the DNA isn’t even dealing with those situations. So the situation there is quite different. And if you think about it like a lifetime of a human being relative to what we feed to an LLM, it’s just this absolutely, incredibly tiny bit of information.

And yet on a lot of tasks, human experts are still much better than our labs. So there is this disconnect that clearly there’s a lot missing in it. Now you can say, is it consequential? Now you can say maybe over evolutionary timescales, we did get as much data as the Llms have, so that somehow we efficiently use the data from our long ago ancestry.

And it’s clearly there, like and in that sense that makes us also like much less fragile, which is we deep down through evolutionary paths, we remember the flood of something at some level, like we have like we have an implicit knowledge of all these cataclysmic events that were evolutionarily important.

I wanted to put a pin a little bit in evolution. So evolution I think, gives us deep insights about human minds. And I think also it’s helpful to understand brains. So the idea in evolution is like that. Systems are in a way you get good at the niche in which you’ve been growing. And there is an interesting way of thinking about AI, which is what’s the niche of an AI system, what makes an AI system succeed?

So when people talk about it, they often they think about it as if it was a species living out there. Some may live and others may die, and the ones that live will have babies. That is not how the evolutionary system for AI system works. The AI systems for most of the time before the Frontier Labs works like this.

They exist code that runs an artificial neural network. It’s being run on some computer by a PhD student, and if it has desirable traits, which is it wins at a benchmark, it’s going to be copied onto the laptops of a few other PhDs, will change a few lines of it and then submit to the same conference as next year.

So. So in the same thing as of course, even within the frontier lab that’s happening, there’s going to be some code bases that people like and others that they don’t. That is an evolutionary niche. Now what makes one LLM outcompete another one? Well, it must win competitions. What other competitions?

What are the benchmarks that we use? Or like that Frontier Lab series. While they use things like don’t lie to us more than needed, more than our customers actually like, which means that the evolutionary landscape for them is just like

Lying is bad for you, because then you’re much less likely that you’re going to be used in the next generation. And so this evolutionary process will give AIS the things that they’re selected for. And the same thing is true for brains. Not like you can say what’s the best waste that we have about brains?

The single best idea about thinking about human or animal brains is let’s just assume they’re pretty good at solving the problems in the niche. In neuroscience, we call that normative models, where we’re basically, let’s assume you want to be good at estimating how far away something is and how big something is, and if this one object versus another object, if we build that into into a models, and we used to build that using Bayesian systems.

Today we use deep learning systems. But the key is we built in just like here’s our understanding of the niche. Give me the best solution to that. And if you solve for the best solution, that’s like the single best, most useful, most reusable model of brain of brains that we ever had. So there’s thinking deeply about the evolutionary landscape is what gives rise to like, well, walking models and people don’t usually think about it.

Juan Benet

I really like the evolutionary frame on top of the current kind of generation of models. Certainly we’re also selecting for a lot of economic utility, right, where these labs are under intense economic pressures to deliver broad capability and utility, which today is shaped in terms of the knowledge work that people are doing day to day.

These systems are very bad at navigating the real world. So like robotics is so far behind. In fact, we should probably dig into that a little bit in a moment. But yeah, how do you how do you think this selection process is developing over time? Like, what are we. We’re clearly selecting for some cooperative structure and cooperative game.

Yeah. How do you expect this to flow over time?

Konrad Kording

Well, of course we should expect it to be become more and more aligned with the financial incentives that they’re there. But let’s be fair to those financial incentives. If I’m a company, I wanted to solve my problems. I don’t want to. I don’t need it to tell me that I’m like, so freaking awesome. It’s enough if it’s like, here’s the problem solved.

And by and large, CEOs do the right decisions so they will buy AI that solves the problems that occur. They have preciously little interest in buying AI that will take to that will try to take over the CEO job, but there will be really happy with AI that like if you call and complain that the piece wasn’t delivered.

Censor that piece. So at some level we should expect in that environment. And that means that the evolutionary environment is give the companies what the companies want and the companies as an evolutionary landscape. I would give the customer what the customer wants. So ultimately alignment is like built into that whole process.

And like the evolutionary landscape is improve alignment.

Juan Benet

That’s a very interesting way to put it. One counterpoint narrative to that would be sure, the kind of there’s like this outer alignment loop that is training and developing these systems, but there’s just a raw, increasing capabilities that they’re developing. And at some point they’ll cross some threshold of broad capability across a very wide range of domains.

So then, you know, far exceed what humans can do in a range of things. And that plus some misalignment in like internal goals where there might be like a, like a misalignment between the, the internal belief structure of a model, that’s kind of where the problem lies.

Konrad Kording

So let’s just see if we can answer it from an evolutionary perspective. There’s going to be some eyes that are more likely to try and break out of their sandbox than others. Which of these two systems is more likely to like win the long term, and which one in fact? Well, when? Relatively quickly? Well, like the scientists are looking for violations of their sandbox.

And let’s just say it’s not good for you if you’re an LM system or like if you’re one like it makes it makes it that way. Every generation we select for alarms that are less likely to break out of their sandbox, given that it’s.

Juan Benet

Isn’t it actually selecting for llms that not getting caught, which could include llms that are simply not doing the thing, or llms that get really good at doing the bad thing, but hiding it?

Konrad Kording

My hunch is not like they is something

About species and evolution that at some level you’re being fact checked by the rest of your species and. And if you want, I believe that there’s a general scaling law on intelligence, which is if you put more intelligence onto a task, the extra benefit you get tends to accrue in a highly sublinear way.

So if you’re one like the probability that one system, if you make a if you give more flops or something to a system that does something, whatever it is that doing something is the benefits. Well, usually at first you get a lot of benefits from basically not completely screwing up, and then there will be less and less.

So in that sense, if that take off like sublinear intelligence scaling is correct, then if you’re one like the other, llms will of course like be calling out that one LLM that tries to cheat and will be very successful at that because because well, the one even if that one would be more, would be better, more successful, it would gain a very small benefit from it.

And that’s kind of why I come out in that discussion as feeling rather bullish about the future of artificial intelligence, because basically, I believe on any one task there is sublinear gains to it, which means that kind of like the system is inherently stable. Now, you could you could hold a different view where it’s like it’s up to some intelligence, like you don’t get anything.

And then if you if you exceed some threshold, then you gain a very, very large amount. It’s just people are not doing a great job at giving me examples of that. Like cracking is maybe the best example that people have where like, you have a system that some point of time discovers it, but this growing process is, is more like if you’re one, you get like that one extra bit when you reach that and the reaching it on an individual not like it has this jumping off of capabilities.

But it’s not that it’s a clean thing. If you have like the right amount of compute, all of a sudden it jumps up. It more is like you discover a dimension that you didn’t have before. That dimensional benefit, though, is usually on an overall scale. It’s a relatively small benefit. It’s just like if you select for only those cases that this benefit was speaking to, then it looks like it’s a big jump.

But on the grand fitness scale, it feels like these are like generally small, small things.

Juan Benet

So far, we’ve had a lot of success where in the areas that we have an enormous amount of data because of the dynamics of learning and the fact that these systems require so many examples and learning vast data quantities, plus easy verification loops, predicting the next token. This where we’ve seen a ton of the gains.

And meanwhile we haven’t seen correspondingly good gains on things like being able to navigate the environment, which leads to the classic more of a paradox type argument, which is, you know, the dramatically easier to compose a symphony than walk down the hall and not break the door. Right? Or something like that.

And so clearly, robotics is way behind. And so we’re having kind of a an impact on knowledge work but not yet on physical work. Having done a ton of work on like just the motor areas and so on, like what’s going on there and like, what do you think? How do you think we’re going to get past this? Do we really need to do what the robotics labs are doing now, which is kind of, again, brute forcing the generation of just oodles and noodles and noodles of data?

Or could we glean some insights from neuroscience to be able to do this problem better or like yeah, how do you.

Konrad Kording

So there is.

Juan Benet

A.

Konrad Kording

Big cognitive effect which is that people are drawn to causal explanations. So if I ask you to explain something to me, you will immediately use causal language. Like that’s how humans talk with one another. in the motor domain, when I walk around the corridor and I’m trying to not break a door, it’s all about causal effects.

And like what? Well, me doing something to do or not do to the door in by design. Our systems are correlational, like it’s machine learning. It’s trying to predict the future from the past. That means that it knows what’s a good predictor. It doesn’t know what’s a good causal influence. And in motor control, everything is about causal inferences.

And in fact, if you look at babies start moving. A lot of this they’re trying to find out what happens if I move my hand, what happens if I do this? What happens if I chew my own hand? They kind of causally explore the world. And yet our machine learning system, causality, doesn’t really exist in them.

You can say in reinforcement learning it exists, but it exists in a pretty rudimentary way, which is we have probabilistic policies. If you run a probabilistic policy, you can say there’s a little bit of noise on your policy, and that’s kind of what you use to figure out how the world works. But it’s a very inefficient way.

Not like it’s not that. Like if a kid explores a room, it’s just like there’s random stuff in the room. I was like, oh my God, there’s a switch. Let me try what’s happening there? And you’re like, oh my God. Like stop lights turning the light on and off. But it kind of like it’s a directed exploration aimed at understanding causality in the world.

I think it’s super salient, at least in human cognition. Alison Gopnik has a great analysis of showing that, like, kids are, like, deeply attuned to that. And yet our AI systems don’t even know it. And there’s no doubt that human brains kind of know the difference between that. And in fact, you see it in like, super young babies that things like intentionality, you know, like, we are agents, we have goals, we do stuff to get out towards our goals.

AI systems just try and explain the regularities. There is no real sense of agency and no real sense of causality and like, like to the level that it helps predictions. Like, of course, like there are cases where causality does help prediction a lot. But then when it comes to like intuitive physics, AI systems aren’t quite up to snuff yet.

And when it comes to more complex like, like, like it look like causality is this magical thing. The way it works in our in our world, which is we have causally specialized object like a cup is for drinking coffee out of it. It does one thing, and there’s one thing that there’s one causal chain that is supported by cups, which is put coffee into our mouths.

And so we live in this world where causality runs on a very small number of sparse lines through time. And we are, like, so attuned to kind of discovering these causal lines and then being able to manipulate it. Whereas a lot of the world alarms are much better at describing the regularities than I am, but I kind of know how I can, like, get coffee in a good place in a way that probably AI systems that have great trouble with at least when bodies are involved.

Juan Benet

But so that implies that the brain is building these very sophisticated world models, these causal world models, and then are able to therefore to path, plan and navigate and so on. And I think that makes sense for humans, make sense for probably a lot of animals with fairly sophisticated brains. But if you go down the other direction to a much simpler and simple and simple animals like, eventually you get to like certain insects who can navigate very effectively and can have amazing swarm behavior, like bees, for example, will have tremendous swarm behavior and are able to do very complicated sets of things.

But they probably don’t have a very sophisticated world model or I mean, I guess we don’t really know.

Konrad Kording

Yeah, I don’t I don’t think they need to know. So you can say there’s always two possibilities. Either a can evolution can give you a reward function and you optimize for that Casio learning to get good at optimization. But you will get a policy that’s better in the end. Or alternatively give you the policy and then maybe just learn at the bottom of that.

I think insects are much more in the let’s let’s let evolution figure out the policy and then they just run policy there. You can say when it comes to navigation.

Juan Benet

That’s why it’s like so little to a very, you know, weird environment like it flies characteristic like are following light to some degree and like they can’t handle windows very well. Whereas like I think wasps do it better where like wasps I think have like some different policy where they can.

I think one of them handles light and one of them handles wind differently. And so there’s like these classic cases where you can they will respond very differently to potential predators. But it seems like very hard coded into the brain.

Konrad Kording

Yeah, I think I think it is. And, and look like the policies you need for, for 3D navigation if you’re a drone. I don’t believe that they’re crazy complicated and like the reason why our life and our niche is so important. It’s so, so complicated is because other beings have kind of like this. Other humans have like the same mental depth to them that we do.

And that produces kind of a fundamentally complicated world. And in one way we need models because we kind of come just assume that one policy is going to generally win. Like we can’t just say the same thing. Yeah.

Juan Benet

There’s all these theories about what led to the evolutionary run off process that gave humans this level of capability and intelligence. So, you know, there’s theories like the social behavior did it or like complicated hunting patterns or like, you know, all kinds of pressures. yeah. How do you think about this?

Or like, what do you think if you had like, guess or I don’t know how you explain this, this rise of intelligence.

Konrad Kording

I mean, let’s let’s talk about the evolution and the nature of compute. You can say if you have relatively short lives and you have relatively stable environments. Then in relatively simple environments, then evolution. The best path at some level is built everything in. Now you could say if on the other hand we have systems that are relatively intelligent way with intelligence.

What I always just mean is being able to choose the actions that are appropriate or that like reward generating in a given situation. Then the situation is different. You can say if, if instead if you want. For simple insights, evolution might want to mostly build in run this policy and like what I know the be like use this such strategy and you and they have the waggle dance and for communication.

But it’s hard coded. And alternatively if you have an abundance of intelligence arguably humans arguably AI systems. A lot of it is about gut reward design, and people are always like, yeah, like that sounds like that’s so simple. LeBron must be doing something much more complicated, unless you start to realize that it’s probably hundreds of reward functions that in your brain are like your brain knows to not have too much salt in the blood and not too little salt to, like, have enough water and not too little water to have not to have the not too hot, not too hot, not too cold.

It’s kind of like its Goldilocks, but not unlike a dimension or three. It’s like on like a hundred dimensions at the same time. We regulate our emotional content, we regulate who we hang out with. And a lot of it is undoubtedly. And we know that the hypothalamus has a good number of those, like we have this like rich set of, of, of rewards.

And then we have these rich sets of like learning things to it. So for example, obviously we don’t want to have to be vomiting because we don’t want to eat bad food. But here’s the problem now. Like if you eat bad food, that will be only there three hours later. And not only that, if it was just a reward, what do you do?

Like three hours later, you’re like everything. No. Like, are you going to remodel your visual system? No. This like a biological system isn’t like. Oh, there was a big negative reward. It is like there’s a big negative reward of type food thing, which should be attributed to my food preferences, which I now want to adapt, but I don’t want to adapt all the other things.

That is something that like

That is if you just have a numerical reward, like every milliseconds you get how good or bad that millisecond was. That is very hard to capture with that like machine learning setting. And yet it’s like very natural if you think about like a more biological credit assignment.

Juan Benet

It also seems to suggest that we may not get sophisticated artificial digital beings or like with sophisticated, complex behavior. Until that level of complexity and reward functions is sort of built in. Like the current reward function might be so straightforward and simple that it inhibits a sophisticated and creative landscape.

Like maybe that these models are really bad at extrapolating and generating new things, simply because their task is just so bounded to like predicting the next token model.

Konrad Kording

Yeah, it feels like there’s clearly something missing there. And I think it’s also about values. No, you can say AI cannot give you values by itself because where is it from? Like the values, the reward functions. It’s implicitly like imitate what humans do. And then it is now like if we look at how we train our lives now, there is a first stage like your value.

You’re the only thing valuable in your world is that you talk like humans and then like, yeah, and also like talk in a way that humans find desirable. That’s the second part. And then it’s like, yeah, and you can also talk with yourself. But like the long term goal is just to make that humans find you valuable.

So and then you can also use tools, but you can also only use tools with the one goal that humans find you valuable that I mean, it’s a good from a society perspective because we want these things to be tools that satisfy our desirable desires that make us happy. If you want, that’s good, makes it safe in a way, but it clearly kind of misses the richness of what we care about.

Our value is at least I hope my value isn’t just like do like everyone else, and it’s also not let’s do whatever makes my audience most happy. Not like I’m opinionated in a way that is much more opinionated than I should be for my audience is sake. And I think I think we all have that. And I think that’s just like it’s another example of not like we talked earlier about how intelligence is so different between humans and these llms.

And this just another of these dimensions which like, yeah, we’re not on the same axis. We’re very differently constructed, differently evolved systems.

Juan Benet

It also potentially is the kind of pressure that would yield the richness of conscious experience and qualia that we that we have where beyond kind of our, our own individual utility to the rest of society, there’s the kind of richness of our own individual perspective and experience. You know, it’s hard to reason about quality or consciousness eventually be of models trained in this very narrow landscape.

Konrad Kording

Yeah. And if we can even talk about it in those terms, I wanted to like, just talk about one thing that I think is important here. So evolution gives us reward functions, not many of them. Basically it’s because it can’t reach into our lives. It kind of tries to approximate fitness functions with reward functions.

We like coffee because it kind of it serves something else. And, and here’s the thing from evolution’s perspective, we clearly are breaking out of the sandbox. Not like we’re supposed to maximize fitness, but we’re not. We’re maximizing something else that kind of arguably breaks out of the sandbox that evolution had intended for us.

And in that sense, because we broke out of Evolution sandbox. So we might not want to allow AI to do the same.

Juan Benet

So now, thinking about AI today, we’re talking a bit about capabilities. You talked about how you use AI and how it sort of improves your work and so on. What are some of the ways, like you’re using AI for science or AI to accelerate either your own work or your lab’s work or like, how are you seeing sort of like the impact in the field?

Have you seen a lot of shifts or, you know, some fields are getting accelerated more than others. So here’s how you’re seeing this.

Konrad Kording

Yeah, maybe. Let me start with how I use AI. So a lot of my AI use is to produce friction. A lot of people use AI to remove friction, and I think it’s a huge mistake. Standard use. Write this paper for me. I’ve come up with an idea for me. Or do this science for me. For me, the use is almost always the other way around.

Here I wrote something. Tell me where the logical problems are with this. I cite papers in my paper. Can you check that they actually support what the sentence says? Where I cite them from. And it’s shocking. Like in the past, I could never have done that. But today, I check every single use of references for me.

And you’ll be shocked at how large the proportion is where I get it wrong. Like the sentence I’m like, yeah, this supportive of that. But the paper doesn’t quite say that it’s in the right direction, but not quite. And that is for things that are my field where I’ve been in for 20 years. So if I basically find myself often somewhat borderline reciting things, I’m sure that’s very common.

So, so this and this at so many layers, you know, like I do some coding, I might even do coding with an LMS. But for every command where I’m like, code this thing, for me there’s like ten commands. So I’m like, okay, now like look for bugs in this. And how could my key leak in this case and analyze if this an if this uses the same design patterns as the rest of my code base, and it’s kind of it kind of like I use it to add friction instead of removing friction.

And in fact, I build an app for students that called it’s called Planar Science

That adds friction to their way of planning science. So it asks, so what’s your science question? And then it pushes back on that. It’s like, well, these three words have no meaningful definitions. Or it might be like, well, there’s this Smith et al, 2015 paper that sounds like mightily similar to this, and it kind of gives you the pushback.

And then once you establish a question, it asks you about your hypothesis. And it’s like, hmm, would be nice if those hypotheses were mutually exclusive. And it’s not quite as snarky as I, but basically it pushes back, it gives you a fraction. And I believe fractions is so important. And so this one of the users that I really promote in my field as well, because I see it all the time now.

People test our hypothesis and there’s like ten other hypotheses in the field, but they don’t use them. They don’t know about them. They don’t take it seriously despite the fact that it would make their paper more interesting. No, it’s it’s not that we’re like always in a competition situation, but like, we’d want to know which of them it is and kind of people miss that.

So I think AI for friction is important and that’s important also from my hypothesis about people and life. If you’re one, you want to become stronger. You’ve got to start lifting. And if you want like if you’re intellectual, want to get stronger, you get to start intellectually lifting, which is you get stronger by people pushing on you.

This also maybe one of the reasons why I’m, like, happy to give my field a little push back at times. I believe the field gets stronger by me pushing back against the field. I don’t believe that the field will just crumble and die just because I disagree with some statements that we are making. And so this like add friction I think is a super important philosophy to me.

Now, there was a second question you asked. I’m trying to cut your your questions into pieces, which is how is the field using it at the moment and how is it positively using it. So there is this possibility now that you can replicate papers much more easily and test paper. So for example, I have a student I’m working with who is trying to you upload a paper with the method section and you upload the code base and it just asks the question, hey, can you check?

Does the method section do things that the code base doesn’t? And does the code base do things that the method section doesn’t?

Juan Benet

I’m so scared to hear the results.

Konrad Kording

And obviously there’s a fair bit of like this deviation happening. And I think there’s a general feeling of excitement that we can do these things. Now this another example of like adding friction. But it’s it’s again constructive friction.

Juan Benet

And that’s fantastic to throw at like the entire literature right. Like there’s there’s a lot of literature across all fields that probably has enough embedded in it, especially over the last 10 or 15 years that have a lot of code where we could get a view into the reproducibility of all of these papers everywhere.

Konrad Kording

Absolutely. And it would be great to do that. And yeah, I should start writing grants for like, medium level compute for that. But I mean like this.

Juan Benet

This DeepMind and OpenAI would be thrilled to like, do this.

Konrad Kording

I think they should do it. In fact, it’s one of the few reasons why I’m sometimes jealous of not working for those companies. But basically, yes. Not like you can do it for a dollar of compute per paper. You could basically do the whole literature in and make a complete list of, of which figures replicate, which figures don’t replicate which method section.

And there’s a lot of differences between what people say they do and what they do. I ran funded data sharing grant at some point of time, and good proportion of the literature can’t even produce their own figures. And that is these professors were sending it to me. It’s not that.

Juan Benet

Yeah. Even when there’s like all the best intention and you try to do those like good work. Yeah. You still like it’s.

Konrad Kording

It’s so easy. It’s gotten much better now. Like now you, like, put the whole Python repo online. It’s much more reliable. But back then, at the time where people were using Matlab, there was a lot of just difficulty. And as professor, you can’t replicate it from your students. It’s just so incoherent and just replicating it would be wonderful.

Juan Benet

It’s probably a large thread that we could spend hours talking about. We both care an enormous amount about the tree of knowledge. What do we actually know? How do we know it? What is the evidence? What is the science that we’ve done? What’s reproducible versus not? What is the likelihood of whatever claim exists somewhere in the literature actually being true.

Konrad Kording

And an alternative interpretation and not like a lot of papers are like A versus B, and you’re like, hold on. Like there’s also C as a possibility. And I bet you there’s lots of papers where C is much better than A or A or B.

Juan Benet

Yeah, yeah, yeah. I think a lot of us are very optimistic about applying AI to a ton of these methods, and accelerating science quite a bit. You also have been thinking about the economic impacts of AI, and you recently co-wrote a paper with your wife, Ioana Marinescu. Yeah. Which, by the way, this super awesome.

The two of you came together and built this amazing and combines a lot of your thinking, both in how intelligence works and how it gets applied to knowledge work and beyond, and then how economic systems work. What are some of the interesting insights or consequences the two of you think this will produce?

Konrad Kording

Yeah. So this paper comes kind of from a contrast of two very different ways of thinking. Like if you if you look at the AI worlds, you’re seeing that the price for equivalent tokens. Yep. Exponentially gets cheaper. It’s like a crazy you one year or sub one year time scale. So that is one way of looking at the world.

And then you can look at the world from a very different perspective, say the GDP of a country. And then you’re like, yeah, we’re like growing and like it takes us 15 years or something to double. These two frameworks are just so very different. And we started talking about what would happen when those two meet.

And let’s talk about how the two of them meet. So imagine you have a task a real task. You want to like take dot out of one place and bring it to another place. For that, you need intelligence and you need physical things. You need an excavator, and you need someone who sits there and controls the excavator.

And if intelligence is very bad, then the physical machine that you have, the excavator will not be used well, like you’ll be wasting lots of time digging or the excavator gets destroyed. Now the question is, what happens if we now saturate intelligence? We add intelligence and we add even more intelligence until we have unlimited amounts of intelligence.

What’s going to happen? At some level, this machine here is going to be upper bounded by its physical properties in how much Earth it could move. At some level it’s physically bounded. There’s a sudden there’s a combustion motor in it and like it cannot move more than a certain amount of Earth power, energy that it burns.

So in this task of like having to move Earth, you need a physical piece and you need an intelligence piece. Economics has a history of modeling such phenomena in that classical framework that people use for things where you need both is labor and capital. Not like we know from ancient econ literature that you can say we if we want to run a factory, we need a bunch of workers there, and we need a bunch of like, physical goods in the company traditionally capital.

And now how is the interplay between them going to be? There’s a certain amount of substitute ability. And like imagine people are very expensive and excavators are pretty cheap. Then let’s build one mega excavator with just one person controlling it. Instead of building 100 excavators that are small with like one person each.

At the same time, imagine that people get to be like super cheap. Let’s hire a lot of people with shovels. Shovels are very cheap. And this tradeoff between capital and labor. The economist model that with an idea of called constant elasticity of substitution CS which basically the idea is that there’s a certain friction to like moving from one side to the other.

So, so basically replacing the fast human with an excavator is very cheap, replacing the hundreds or A human with an excavator is very expensive. Representing all humans is impossible depending on the setting of the parameters of that substitution dynamic. But for a lot of parameters of it, it’s basically impossible, I guess more and more and more expensive as we go in that direction.

When we then talked about physical and intelligence capital, it feels like it has a certain similar property now. Like if I build AI into like that excavator, like there’s a certain limit to it. And so you can say what we what we then did is we said, well, let’s, let’s think of the world as four pieces. They exist physical, they exist capital, labor.

Think about excavators. They exist like human, a human physical, which is think about humans with shovels, but also think about like human doctors who have hands. Think of human nurses that have hands. Think of teachers that have a face that can frown. Like. Like anything that has the body. We call it physical and anything does it.

That doesn’t require a body or anything physical we call intelligence. Now the idea is that in production you can say there is a fixed like physical capital think excavators and there’s like a human physical. Think of human bodies broadly construed. And there’s a human intelligence and there’s an artificial intelligence.

And then what happens if, like we substitute in those areas? And now the question is how can we substitute like physical with intelligence? And the idea is very much not much because like, look, yeah, we can have robots, but the robot is the physical piece of it. And it turns out also and this something that I think the AI community often goes wrong.

Robot scaling isn’t intelligent scaling. It doesn’t have every eight months. Robot scaling is extremely slow and we built millions of industrial robots. And robots are just capital wise, very expensive. And the reason being, like there’s this long supply chain that we need to build to build robots.

Can we do it cheaper? Yes. Can we do anything in this universe cheaper? Yes. Does it take time and effort and all that also? Yes. On the intelligence side, the scaling is very different. And there’s even people who say, well, like one hour now like that, we all use Claude Code. And now that, like all the frontier labs use Claude Code, they will get better at building intelligence.

That maybe or maybe that’s not the case. I’m not sure you could make both arguments. You could say the easy things are being done fast, and so maybe we’re still at the face where we do the easy things that we can do with internet scale, compute, compute. But maybe that’s not. And we don’t know. And that’s why we want like a modeling that like brackets that out of the rest.

And now you can say how can we replace physical intelligence? And in most cases it’s very unclear how you could do that, like build a house,

Put more intelligence in it. How much faster is that how it’s going to build itself? Very much, not at all. Build a better straight. How much is intelligence going to help? Not at all. If you want to build bigger robot factories, you need to build streets and street scale like streets. They don’t scale like AI.

So you can say no. And people are like, yeah, but I could build like a $100,000 robot and put like super intelligence into it. Yes, if you want the robot that’s already there. But if intelligence was like went to infinity, all of a sudden that $100,000 robot would be a lot more valuable. In any case, the idea is that replacing physical fund agents is crazy complicated.

Juan Benet

The reason is kind of like until the scale up of the robotic supply chain can catch up to the demand, like, like what you’re kind of getting at is like, there’s going to be this process of sure, intelligence costs might go to zero, but that just increases the value of a whole bunch of places where you’re applying human work, which includes both physical and intelligence.

And then eventually, once the robotics production kind of catches up.

Konrad Kording

Yeah, but it’s not just the like. Like, robots are not unique in the way that AI people think. Robots are unique. Like a factory is a robot. If you think about it like it has a it has a ways of automating much of it. We’ve been automating factories like for 100 plus years. Yeah. And like they get better every year.

They’re more efficient. But it’s not that there’s super dramatic, unrealized possibility.

Juan Benet

There’s always like this very long tail of tasks that are just incredibly difficult to automate, that keep rising in price. Right. So like you keep automating and you keep finding the next constraint and.

Konrad Kording

That’s that’s right. And not only that, but also for a lot of places like robots aren’t even substitutable. You cannot replace a teacher with a robot because there’s something about them being human that makes them be attractive as a teacher. And let’s just say if you look at big tech CEOs, they don’t have like their kids, be robots.

Zoom taught. They have them be in very high intensity, human taught environments. And then you can say, what if we now only look at the intelligence production? do we find this phenomenon that basically humans get replaced massively with AI? That is entirely unclear within that sector, because you can say even within it now, how replaceable are there?

It turns out that least so far, it looks that AI doesn’t replace the human programmer. Maybe it replaces the entry level person, but the people who are good at working with AI, they’re just so much more productive now. So therefore, it’s unclear that the existence of AI kind of like moves all the humans out of it.

And there’s precious little evidence that the tech CEOs are actually being replaced by robots for the time being. And so but the key is because of the existence of this physical sector, if the intelligence thing would go to infinity, if the price of any intelligence would go to zero, the economy still wouldn’t go to super productive.

Sure. What do I know? GDP might double, but it wouldn’t like 100 fold. It wouldn’t tenfold. It would be like probably a medium size change. And it very much depends on your assumptions about how close to IQ tapped out we already are. But just give me the real world tasks. More intelligence would actually like produce mono like your car factories are already full of and robot factors.

They are already full of robots. It’s not that they don’t use robots, it’s just that yeah, they need the humans for the outlier tasks, both intellectually and physically.

Juan Benet

But it does suggest that there is some threshold after which you can get the recursive self-improvement of the robotics factory’s going well enough, and there is a very long tail there that I think everyone in the AI community is maybe underappreciated. But at some point, whether this might be, you know, certainly not in the next 5 or 10 years, but potentially after you can get production so well automated that you can then create like a much faster production of physically intelligent capital.

Konrad Kording

Yeah, but that is not a new phenomenon. You can say since the time of the Romans, we had exponential growth, and exponential growth has been in what, you know, like the 7% range for quite a bit. And now imagine it goes up to ten. Imagine it goes up to 15, which would be unbelievable. It would still take a very long time.

And like the Romans, if there would plug exponential curves, would be like, oh my God, in 200 years every peasant has unlimited food. and knowing that kind of ended up building a big civilization with all kinds of other things. And so I think, like exponential move, there’s no reason that it will stop. And I think like there’s the degrowth movement, which I think is an insanely bad idea.

But basically there’s going to be ongoing growth and we will find new things to spend that wonderful growth on, and it will make our lives much cooler.

Juan Benet

This will be like a super interesting area to see and measure over the next few years. I think it’ll be like a very big feature of the conversation.

Konrad Kording

Absolutely. And what in the context of the paper with my wife is really interesting, is if we see kind of the canary in the coal mine thing, there’s no signs at all for the substitution of physical for intelligence, because that’s very, very difficult. But even within the intelligence sector, like it’s not that we see a strong job market drift where this happening now, you can say a lot of it in a lot of our simulations there.

There will be an early period where if you want increased productivity, kind of drowns out any effect of substituting people. But even there, if you want like the upshot is we are kind of somewhat safe because there’s so many jobs that require bodies, and we could have a lot more of those jobs, like we could do a much better job at teaching kids, for example.

Like if we had like extra human beings that are jobless, then we probably don’t want classes of size 40 to 1 human teacher, which we could probably one on, one on one teaching. And so I personally, my take is that we’re rather moving into a very positive world where. So for example, take all that software engineering software is horrible by and large, just like software is just the worst thing ever.

And therefore you could say in the future is it likely that people have like expectations of quality for software? Yes. And that would be a good thing. And so it will take a lot of people to get that to happen.

Juan Benet

Very good. Optimistic view on the moment because it will just kind of we’ll be able to reap the rewards of all the growth along the way. The economic transition is not going to be as rocky as everybody currently fears it will be. And so yeah.

Konrad Kording

Right. Yeah. And there’s also this phenomenon though, people are like, yes, if we build robots and we invest all proceeds of building robots into building more robots, then wouldn’t it be like so fast? But that’s not how the real world works in the real world, if your robot factor is making wonderful profits, you’ll spend it on like a heli skiing trip in Alaska or whatever else you spend part of your money on.

And so the economy tends to, like, not reinvest all proceeds. It tends to reinvest 20% or something of it. And there’s no reason to believe that AI will all of a sudden change that. If people have money, they spend money on doing the things that they believe makes the world better from their perspective.

And that could be trips or yachts or like investing in neuroscience progress.

Juan Benet

If you were to kind of give a, you know, an optimistic vision for the future, you know, I don’t know, ten, 20, even 40 years out. Like, what do you find like very exciting, inspiring cool out there? Like what? what do you dream about?

Konrad Kording

Well, there’s lots of layers. I think I do dream about understanding brains to the level that we can simulate them. I think the idea that simulations are too far away. Well, that simulations have been too far away for 70 years. And we have, like, so much better techniques. And the idea that simulations must always be a far future dream should go away.

We should start simulating brains. I do dream of a world where AI is used, by and large, in a positive, constructive way that helps people live better lives. And I in that sense, I. I dream of a world where AI seen less as a competitor with humans, but because I think it’s not, and more just like something providing something of value in the same way that a machine, a car is providing value.

And in general, I dream of a world where we use technology to solve a lot of mankind’s problems. And I believe that as engineers, we’ve always been doing it, and I think we will continue doing it. And AI is just part of that.

Juan Benet

That’s wonderful. Well, thank you so much for coming to chat with us today. Thank you for great time.

Konrad Kording

Thanks for having.

Juan Benet

Me. Thank you. I hope you enjoyed this episode. This a new podcast, so we need your help to get the word out. Please like, rate and subscribe on your favorite platform and share it with people you think would find it interesting. Thank you. See you next time.

Displaying _FULL__Juan_Benet_Podcast_-_EP05_-_V04__edited__formatted_.txt.

15 million people live with motor impairment. The Stentrode lets people operate their phones and computers through thought, and could restore independence to people who’ve lost the ability to control their devices. Tom sees BCIs as a major technological leap that will help human flourishing, by enabling better communication, decoding and conveying emotions, and enabling us to leverage the great capabilities of our computing infrastructure.

This was a great, wide-ranging conversation. We discuss the origins of Synchron, the endovascular approach and its benefits, their next-gen system designed for high-channel-count recordings across distributed brain regions, the longer-term possibilities of helping people communicate better, how Tom developed as a founder and how he leads the company, how BCIs could unlock powerful mental states similar to psychedelics and meditation, how neurotech will transform humanity in the 2030s and 2040s, and why Tom thinks the US will lose the BCI race to China unless the US greatly accelerates. Hope you enjoy!

Links to this episode and references below.

Topics Covered

• 00:00:00 Introduction

• 00:03:18 Carl Jung, stroke surgery, and the road to BCI

• 00:06:58 The endovascular approach: reaching the brain without surgery

• 00:10:37 Getting Synchron off the ground

• 00:17:23 Reading the brain: channels, signal, and noise

• 00:36:41 The numbers: 15M patients, FDA, and Medicare

• 00:43:17 Cognitive AI: foundation models, the data economy, and the 2040s

• 00:52:55 Consciousness, psychedelics, and the extended self

• 01:02:51 Agency, addiction, and geopolitics

• 01:06:51 The optimistic vision: unlocking the subconscious

• 01:10:49 Building a company: 696 no’s

• 01:21:30 Losing the BCI lead to China

Other Links to the Podcast

• Juan Benet Podcast on YouTube

• Juan Benet Podcast on Spotify

• Juan Benet Podcast on Apple

Links From the Podcast Episode

Guest + Organizations

• Tom Oxley

• Synchron

• Tom Oxley on X

• Synchron on X

Research Papers + Technical References

• “Neuronal ensemble control of prosthetic devices by a human with tetraplegia” (2006)

• “Minimally invasive endovascular stent-electrode array for high-fidelity, chronic recordings of cortical neural activity” (2016)

• DARPA Revolutionizing Prosthetics program (2004–2008)

• Meta open-sourced sEMG (neural wristband) datasets at NeurIPS 2024

References Mentioned in Conversation

• NeurIPS conference

• Woman with cochlear implant hearing for the first time

Books + Media

• The Battle for Your Brain: Defending the Right to Think Freely in the Age of Neurotechnology by Nita Farahany

• Nexus by Ramez Naam — BCI sci-fi referenced by Juan for its depiction of shared emotional states between humans

• Do Androids Dream of Electric Sheep?

Links

• Protocol Labs

• PL Neuro

• Disclaimer⁠

Transcript

Juan Benet

So my guest today is Tom Oxley, MD PhD. He’s the founder of Synchron, a BCI company developing an endovascular implant, the Stentrode. He studied neural engineering and medicine at Melbourne and neurosurgery at Mount Sinai Hospital. He has performed over 1,600 endovascular neurosurgical procedures, which is amazing. He has published over 120 peer-reviewed articles and journals, filed over 100 patents. He’s also a professor of medicine at the University of Melbourne, and I’m sure other accolades as well. Honored to be here speaking with you and very excited about all of the tech that you’re building. Thanks for coming on.

Tom Oxley

Thanks, Juan. Happy to be here.

Juan Benet

So let’s just talk about neuroscience and neurotech in general. When you think about neuroscience or neurotech, like whichever angle you want to pull it apart from, like what do you find interesting about the broad field, like what are the kind of core questions that you find motivating in the field?

Tom Oxley

How the brain works, still talking about how the brain works. Last 20 years a lot of non-invasive imaging studies trying to piece together elements of functional localization in the brain, a lot of systems neuroscience work, a lot of computational work, and then still a lack of a cohesive model that brings everything together. That was probably the inspiration that got me originally interested in the field, and then how that generates the experience of being alive and having an ego. I think it’s still a question that’s being asked. What is consciousness? That’s still a very hot topic. I think the nature of these things are being answered somewhat by what we’re doing in basic neuroscience examinations. But now that we’re getting into clinical trial domains of BCIs, we’re starting to play with some of these systems in closed loop environments. And I think it’s really interesting to kind of reverse engineer the brain in real time and understand how to interface with it. That’s probably my motivation.

Juan Benet

Yeah. Were there any kind of key grand challenges that in the field that sort of got you very fired up? You just mentioned a number of really key questions, but many of those might be still kind of far away, but are there any kind of core grand challenges that either are short to midterm that you feel like we’re like now grappling with and potentially able to solve in the next 5, 10 years, 15?

Tom Oxley

Not in the next 5, 10, 15 years. I think motor BCI is seems to be the first wave. I mean, you’ve had a hearing cochlear device for hearing, I think, is a BCI. Now there’s a wave of motor decoding in order to enable movement intention control mechanisms to restore function. But the motor system has kind of been understood probably the longest, you know, since Penfield’s experiments in the 30s of, you know, Zaprain body does this, almost 100 years ago now. So the motor system is pretty well understood. It’s quite mechanical.

[3:21] Carl Jung, stroke surgery, and the road to BCI

I started in psychiatry actually and I think my early, a motivation, I was going to go down a psychiatry career. I was really interested in the biological nature of the mind and Carl Jung is probably my favorite author and I think the thing that still I am most interested in reading about is the nature of the subconscious, how that interplays with our day-to-day, what’s happening in the sleep life and then how that expresses itself through emotional content during the day. I think BCI, once we get into the frontal lobe and start interacting with some of those areas that connect back to the amygdala and have represented emotion, I think it’s a complete area of life and human interaction which is still in the dark ages. I think BCI is going to get there eventually, but I think no, it looks like motor BCI is sort of wave one.

Neuralink’s going for vision, which I think is interesting because it’s sort of going into stimulation. But from a sensing data sort of tokenization point of view, it looks like it’s going to be motor, speech, and then attention is kind of closely linked to hearing. I think there’s some really interesting things in intentional states, but then I think we get into emotional kind of emotional content decoding. And I think the the ability to I think that’s going to, yeah, I think that’s going to unlock a lot in how we communicate with one another, but also how we, I think, interact with technology.

Juan Benet

How early did you get into neuroscience?

Tom Oxley

So I did psychiatry for a couple of years. What I was experiencing clinically didn’t tie in with a clear philosophical framework that made sense. It was very social engineering, a lot of syndromes that are, you know, describe symptoms, but then whether or not you’re coping in your social life or your working life. I think now what’s happening in psychiatry with the psychedelic movement is extremely interesting and I think starting to actually get to some of those questions that I was really interested in. But I started, I tried to read Jung when I was 18. I couldn’t penetrate it. And then I think I was trying to read, you know, more core neuroscience. As soon as I went into medicine, I was trying to understand neuroscience, but it was all very scientific, core, neuronal level, structural. And so I was always more interested in the psychology piece as a sort of window into how the mind is pulled together.

Juan Benet

And when did you start thinking about the interplay with devices? Like what, at what point did that interest go from kind of understanding all of these emerging phenomena to hey, there’s a concrete set of needs here that can be addressed by, you know, a set of devices, and hey, actually the technology is not ready yet, so I’m going to have, you know, take this leap of, like, go build that in Australia?

Tom Oxley

You go straight from high school into undergrad medicine. There’s no college, there is now, but there didn’t used to be. And so I finished medical school, I went straight into working as a young physician, and I just started read, and I was then reading, getting up to speed with BCI. I first read about BCI, I think properly in like 2008, and I came across Leigh Hochberg’s paper, 2006, of the first Utah array. As it turns out, it’s been, you know, work in monkeys for decades and decades. Nicolelis, you know, a bunch of people putting in recording electrodes in brains and showing that you can decode real-time activity. So that I only started reading about that in 2008, and then my entry was opportunistic. My vantage point in 2008 was I was just doing my internal medicine training and the field of neurointervention was just starting to emerge.

[7:00] The endovascular approach: reaching the brain without surgery

Intervention is a term in medicine that means catheters, really using catheters. Interventional cardiology is putting in stents, hadn’t really made its way into the brain until kind of around 2000. You could treat aneurysms by going up and dropping a soft coil in the inside of aneurysm so you didn’t have to do open brain surgery. That was, I think, approved in 2000. And then from 2005 to 2015, physicians were starting to figure out a way to send these devices up into the brain, grab onto a blood clot, and pull out a blood clot. And that was the biggest breakthrough in stroke medicine ever. Before that, it was like using a medicine to dissolve a clot. So then it was like, “Oh no, you can put like a trap in, like an expanded, grab the clot, pull it out.” And if you do that within a few hours of the stroke, like a stroke is a blood clot that blocks oxygen to the brain. If you do that within a few hours, it’s the most incredible thing you’ll see in medicine. People coming in, can’t talk, tongue out of their mouth, unconscious, being wheeled in, go into the angio suite, 30 minutes, go up, pull out the clot, they sit up on the table like, “What happened? What are we doing here?”

Amazing. And if they hadn’t got it out, they’re either dying or they’re going to a nursing home. So, it’s this unbelievable thing. And so, that kind of, that looks like it’s working around 2008, but it hadn’t been proven. It actually, the actual hardcore randomized control trial proof didn’t come until 2015, but I was seeing that coming. And so, Rahul, my co-founder, is an interventional cardiologist. So, we were early in our training. He was obsess, intervention is very sexy in medicine. And it’s like fast-paced. It’s, it pays well. It’s very, the technology moves super quick. Because there’s a lot of innovation happening. Very rapid innovation cycles. And in heart it had exploded. Pacemakers, stents, now valve replacements, all without doing open heart surgery. And there’s a quip in medicine that, like, the brain’s always 30 years behind the heart.

So anyway, I was seeing this come through and Rahul and I would hang out a lot and we’re like, surely if those techniques come to the brain, because cardiology is kind of, it’s hitting the sort of top of the S-curve now. It’s kind of, you know, there’s probably more to come, but for the brain it’s just starting. And so the stroke thing happened, I, that’s what I came to the US to learn to do ‘cause I couldn’t train in that in Australia. Came to New York in 2015, the year that all got announced. And but I’d already been thinking, if this is true, then if what happened in cardiology happens in neuro, they’re going to solve this by not opening the skull, and this is going to be the solution.

And then during the middle of medical school, I’d done an honors degree in using transcranial magnetic stimulation in patients with schizophrenia because I was in psychiatry, but they were just doing magnetic pulses to the motor cortex, looking at patients with schizophrenia versus patients without. And we were finding these differences in the way the motor cortex activated. That was important because I got a motor cortical electrophysiology education.

And so I kind of, electrophysiology was super cool because it’s kind of the basis of consciousness, and there was a lot of electrophysiological breakthroughs over the sort of course of me going from early medicine to where we are now. So that kind of, keeping track of that, loved electrophysiology plumbing is interesting but it’s still plumbing. So I was thinking if the plumbing combined with electrophysiology and it looks like BCI is a real thing that’s clearly, that seems inevitable and world changing, what if all those things together was the ultimate solution in BCI? And that’s been the thesis for Synchron.

[10:43] Getting Synchron off the ground

Juan Benet

So you had this idea, at what point did you decide, hey, this isn’t just a good prediction about how the future is going to go, it’s rather, nobody’s doing this, we’re going to go and do it?

Tom Oxley

Well, what happened was I did my internal medicine boards. I had a year off. I went traveling and I was going to spend the last 3 months in the US and if you want to kick off, like, a new endeavor, you need funding. And I wasn’t really thinking about starting a company at that point. I was thinking like I would love, I had an ambition to be, and, you know, to be a professor, to be an academic. And so I was watching where the funding was going, and the US defense agency DARPA, this was dual use work, this is not like now, it’s kind of gone more weaponized work. So from like 2005 through 2015, hundreds of millions of dollars were deployed by DARPA and ONR global and army in a range of medical applications for neurotechnology.

Actually the history was interesting. It was like, Geoffrey Ling, by the way, was like if you, most people will point everything back, goes back to Geoffrey Ling. He was the guy, he had the idea in, he was a neurologist. He’s at, sorry, Johns Hopkins. He’s at, he was at Walter Reed. He had this vision of building a prosthetic, robotic prosthetic limb for repatriation for soldiers. Obviously if that works I go back to the field as well. That was a DARPA program called Revolutionizing Prosthetics, 2004 to 2008. Then he handed that off to Jack Judy and they said, “Well, we built a robotic arm, but now you can’t control it. So, we have to figure out how to get the control signal out of the brain, cuz otherwise what do you, how do you control the thing?” And the backdrop of that was desert warfare, Gulf warfare, IEDs, and the new emergence of post-traumatic disorder in combination with head injury. So there was this kind of defense problem of brain health which then converted into this neurotechnology infrastructure investment by the government, which then led to all these DARPA programs. I mean, a lot of the companies now have, were kicked off from that.

So anyway I targeted him and I said hey, how come no one’s, everyone’s opening up the skull? Like that doesn’t, like why are they doing that? Like if you can get in and deliver electrodes without opening the skull, that’s probably going to be the way that people are going to want to do it, if we can figure that out. And he was like, “Oh, that’s interesting. There’s, you know, yeah, if you go home and put a team together, we’ll give you the first million.” So, that was, that was kind of it, was opportunistic and product of the things that I’ve been exposed to in combination.

Juan Benet

So, you mentioned another co-founder, Nicholas. Maybe tell us a bit about them, like what sort of, like, their backgrounds. How do you compose as a team?

Tom Oxley

So, the reason this was possible in Australia, DARPA had not given a grant to an Australian group for over 10 years. It was for something in submarines 10 years ago. The reason DARPA were willing to support me in Australia was because Cochlear had originally emerged out of Australia. It was quite a couple of decades ago, but Cochlear had been very successful. There was then this kind of signal processing talent, electrical engineering signal processing talent in the University of Melbourne. And then Kevin Rudd was the Australian prime minister. I think it was 200, I’m guessing 2006, 2007. There was like a $50 million grant to create a visual prosthesis. I don’t think it worked out that well ultimately, but it generated a huge amount of academic infrastructure and there were multiple groups, lots of talent. So there was this sort of continued, this reverberation of signal processing, academic talent in the University of Melbourne and in Sydney under Anthony Burkett’s leadership.

So I got home, I started my neurology residency and then my Terry O’Brien, who was my, basically my boss, my professor of neurology, he walked me over to meet Anthony Burkett who was the guy, head of the visual program, and he introduced me to Nick, who just finished his PhD doing some, a component of hardware in the prosthetic eye. And so that plus Rahul, Rahul was my, we were very close friends, and he helped with the original concept and this financing for the original patents. And so the three of us got together and we thought let’s give this a nudge. But, you know, so the first million was a university grant, so we’re actually in our university capacity supporting that work, but I formed a company, it was sort of a shelf company, but we needed that for patents. And then I had that goal and that ambition, but really it was not for another, that was 2012, and then for the next 3 years we, you know, raised, turned 1 to 5 million to 10 million to 15 million, all in government grants, Australian government, US government, ONR global, more DARPA. We’re very lucky. And it all culminated in a Nature Biotechnology paper in 2016, and so that was, that was enough to then get the series A financing. But had a lot of trouble raising capital in Australia, moved to New York, and then started my new interventional fellowship at Mount Sinai, and then, but within 6 months we had the first financing done.

Juan Benet

That initial government funding that funded the university work, was that kind of just proving the beginnings of the technology, or like what, what the first generation of the device?

Tom Oxley

It’s a long time ago now, isn’t it? It’s still like, we’re still obviously not commercially FDA approved. You know, to this, there’s us and Neuralink who are implanting permanent devices in bodies for BCIs, and that’s after a long, long time, a lot amount of capital. That’s how hard it is to get to the point where you’ve got a device that is safe and effective for what it’s meant to do, and we’re still not there. So, what were we doing back then? We were trying to prove, like, we were trying to solve a very mechanical, electromechanical problem of how do you attach and push electrodes that have insulation, that are like electrically, you know, isolated. Cuz you know, the lovely thing about going through the bone is that you can put your package right there and you have your feed throughs right there. And so all of your electronics can be done where it’s hermetically encapsulated at the point. And so you don’t have to solve this challenge of connections. That’s been the huge challenge for Synchron. That’s been the huge limiter on, on our, the ceiling and limitation of the technology through our catheters has primarily been how do you handle the noise of having a very long cable coming down? How do you, and then how do you achieve good insulation, and how do you get more channels in, and in the end the limiter was around how many individual cables you can put into one device that can be left inside a blood vessel.

Juan Benet

Let’s go into exploring the device. Just maybe redescribe the device, giving people maybe a view into Stentrode in general.

[17:24] Reading the brain: channels, signal, and noise

Tom Oxley

I guess the idea with the BCI is, you know, if on this podcast you know probably, but just briefly, you have to detect electrical activity close to the source of signal generation that’s associated with some signal that’s valuable, and in this case it’s the motor signal, the intention or volition of motor signal. It’s because it’s internally generated. You’re in control of it. You can therefore control an external signal if you can capture, decode, and bring that signal out. So, how do you deliver a sensor with enough fidelity close enough to the motor cortex that can be used to drive an external device? So our technology is what blood vessel is closest to the motor cortex, and what’s the first blood vessel that you would target and is big enough and is safe enough and has that, people have put in things before, and where’s the best place to put it. And so we used a traditional stent architecture, cuz a stent is a metal scaffold that expands. It pushes against the wall and then we figured out how to put multiple sensors on that stent. So we have 12 sensors on the device that’s currently in front of the FDA. And knowing, and so you know we made a bunch of design decisions that we knew would be manufacturable at scale, cheap, cuz it’s very expensive, these systems. And then the question is and remains, like what is the minimum use case that is needed to drive control of a system that’s actually useful for people, because if you overshoot with redundancy, you’re really paying for it. You’ve got, if you’re not using all of your channels, or if you’re designing for an abundance of channels.

Cuz Neuralink came in with a very ambitious, you know, target, and we kind of came in from the bottom end saying, well, you know, what would be the minimum needed for a first-gen system to let us move more quickly, because we’ve got some future generation systems that we are very excited about, but we kind of wanted to first explore what does redundancy look like and what’s the minimum use case. And so maybe the short answer to that is now, like, the relationship with Apple has been critical, because we’ve kind of identified a discrete low number of, like, I’d call them gestures or kind of control, you know, embodiment thoughts, that would be enough to allow someone to navigate their way around a useful system such as Mac OS or iOS. And so that’s, yeah, so that’s where we’ve started.

Juan Benet

Yeah. And that can, like, massively broaden people’s interaction with the whole world, because if you can actuate a computer, you can then start controlling all kinds of other systems and you can communicate with a lot of people. What are the lives of patients that need this, like before and after this kind of device?

Tom Oxley

There’s been an interesting dialogue in the field about who could benefit from this, and as we’re moving through uncharted territory, the definition for the FDA and for Medicare, who are going to pay for the device, really matters, and the use case really matters. And so we’ve kind of, it’s been a bit surprising hearing from people that I might not have thought would have had benefit, but I kind of categorize it as, if you lose control of your ability to control your device, you normally use your hand or your mouth. There are other accessibility features that use eyes and maybe head control. Some people have a sip and puff, but there’s actually a range of conditions where people lose the ability to normally control their system. And then there’s a range of different accessibility products out there that are not standardized. And working with the Apple team, particularly the accessibility team, has been pretty incredible. Like the different use cases, the power, like the different mechanisms that they have are very broad. But I think that what’s exciting about BCI is that you try and standardize everything towards a single, because you go back to the, so it’s not like oh I have a little bit of finger, I’ve got a little bit of movement in my cheek, or I’ve got my eyes are okay but they do this sometimes, my voice is getting soft but sometimes it works. So everything else is very bespoke, but if you go back to the source of the motor signals and the brain is still okay. And so for a lot of people the motor cortex is viable, is like alive, but you’ve had a subcortical stroke, you’ve had a brain stem stroke, you’ve got deination MS, you’ve got cerebral palsy, you’ve got movement disorders, you’ve got Parkinson’s disease, you’ve got ALS, neurodegeneration, you’ve got a spinal cord injury, you’ve got a nerve injury, you’ve got a muscle injury, you’ve got joint problems, and then you’ve got muscle disease. So there’s actually, it’s been incredible to hear from the community with the level of enthusiasm around the idea that there would be a implant that you can get that gives you back control of your Apple device. We take it for granted, and there’s been a lot of somewhat tech negativity recently, and there’s certainly BCI’s struck with the dystopic stick.

Yeah, but when you speak to people, and I can only imagine what it’s like to have lost your autonomy, then that’s what this technology is about: restoration of independence and autonomy using devices.

Juan Benet

Yeah. And in the core, personal computing devices are the launching point to many other kinds of things. I can imagine people might even be able to like run programs that then enable them to move around or like control a chair or maybe have some like basic locomotion or things like that, all mediated by, you know, their personal computing devices. But what sort of stage is the device in now? Like you’ve had a range of trials already. Give us a sense of what you’ve proved so far. What’s next?

Tom Oxley

I’ve been watching with interest the Meta’s neural band, and that’s been a long journey to get to that. A lot of similarities to actually ice. So they’re doing gesture classification with the neural band using an EMG band and they’ve got like a handful of gestures that can be used, and they’re trying to achieve zero shot classification which is very low burdensome, low training required, a lot of training. So what had to happen there was a huge data set, good experiments, and then it was the development of ML pipelines that could be used to achieve this goal of zero shot gesture classification, very similar to what we’re trying to do in the brain. So to get to that point, you have to have hardware design freeze. Then you have to have data collection. And so the big challenge with BCI is that you can’t train until you have someone with the implant doing training. So it’s been a huge impediment to progress, fast progress in the field. But I think what that means is that when scale starts to happen, there’s going to be a takeoff that’s going to be super interesting in this field. So it’s going to be going slow, slow, and then it’s going to go wow, and then with scale is going to come better procedures, better ML pipelines, better features.

So, we’re now at the point where we’ve done two clinical trials. We’ve learned a lot. We think we’ve figured out what is the base level of gestures that can interface with Apple’s accessibility platform. We’ve got a new Bluetooth HID protocol, which enables back and forth computer to brain, which is awesome. And we’ve hit a design freeze now, and we’re now going to scale up for more implants over the course of the next 12 to 18 months, with the goal of getting to a pivotal trial and then commercial approval. So the way we’re thinking about this is we’ve got a target in the brain. We’re trying to reduce the dimensionality as much as possible. Target in the brain, area of the motor cortex, we know, same blood vessel, same size, gestures we want to achieve, the use case we want to achieve, show that it helps people be independent. I feel like it’s been a long time. We’re now getting very focused on what we think success will look like in the clinical trial. So there’ll be this period, FDA said they want 30 to 50 patients in a pivotal study. Then you’d submit the pivotal study to the FDA and you’d get commercial approval, and then you can start selling, which by the way you need Medicare to cover the payments for, because most 90% of people who have severe motor impairment or paralysis are on Medicare. So that’s another element which I think the field is maybe undercooking a little bit. That’s all ahead of us in the next sort of 3 to 5 years, and then all going well, then I think there’s going to be a big ramp up towards large numbers of users. And so I have a vision where there’ll be a software system, there’ll be a training apparatus, there’ll be a number of use cases, and then there’ll be a constant release of new features or gestures, and we’ll be building out an app where there’ll be a certain interaction methods on offer to engage with, but ideally what our goal is to enable our users to jump out of the app and use the natural ecosystem of Apple’s iOS, and then we’ll move on to other platforms. But we’re enjoying working with Apple’s platform a lot. With my open source hat on, would be great to develop very good portable interfaces there so that any kind of device can be plugged in. And yeah, so Meta has released all of their neural band data and there have been other groups, like Apple has been, you know, working to try to do better zero shot classification. So I think that’s, I think in BCI as well.

Juan Benet

I imagine it does raise, it raises a bunch of questions though about, you know, who owns the data, how do you release it, you know, what consent did you get to do that, who benefits?

Tom Oxley

I guess with Meta, they’ve handled their privacy issues with the neural band really well, but they’ve found that line of like doing on-device inference but also releasing data for aggregation for improved, so you know, that’s actually a good model I think for the field. The current device is read only.

Juan Benet

It’s read only?

Tom Oxley

Yeah. Yeah.

Juan Benet

And will you be able to do write in the future?

Tom Oxley

We have a plan to do stim. The challenge with stim is that it takes up about twice the amount of space on your PCB and a lot of energy. So, as everyone’s trying to get smaller and smaller and further and further, I think the question that we’ve been grappling with is cuz Neuralink’s gone in for stimulation and vision, which is an interesting move, but I think there’s, I think it’s all a trade-off, and especially for us, like, with our next-gen system, space is at a premium because we’re trying to push chips through catheters now, which no one’s done before on the scale factor that we need. So I would say that I think stim, I think write is somewhat important. I think if you’re going for vision it’s extremely important. I think for most other areas in the brain it’s less important. For motor, haptic feedback really matters, but you don’t need a huge amount of discrimination for haptic feedback. And you can send bits in, for speech it can like tell you about speech interruption. For hearing it can be interesting because you can, you know, provide prompts or just feedback. But I don’t know, like the brain has got very good to write, like we’ve got very good ears and very good eyes generally. I think the bigger challenge has been the output. So I still think there’s going to be more value generated from sensing and reading at scale, and because once you sense and you read then you create, you identify features, and then you tokenize, and then you’re building layers that can interact.

Juan Benet

With the current generation of device, how much can you read out? Is it sort of a number of different neurons around the blood vessel that you’re able to tap? What’s the scale there? Like what’s the kind of equivalent to where other BCIs might be talking about, I don’t know, number of channels and things like that. How do you think about the IO?

Tom Oxley

We’re actually in a class of our own from that perspective. I think everyone, it seems, jumped straight into the super high channel count world. I think because Neuralink did that, EEG scalp based EEG at times can do things. So, I thought there was a pretty interesting medium between doing a lower channel count starting point. As long as you can serve a use case, it’s going to be a lot cheaper and more scalable, and I think probably safer than others. So, we’re in a kind of category of our own with this first-gen system. It’s going to have 12 channels. It samples at 2,000 Hz, and so we’re, you know, we’re recording up to like 500 Hz activity. So, it’s like intracranial ECoG. The sensors are laced within the blood vessel around the precentral gyrus in the left and the right so it can detect, and we’re also we’re seeing a little bit of stuff from preoter supplementary motor, but primarily primary motor cortex, and we believe that’s enough to generate a handful of gestures that enable a level of device control. And then for our next-gen system, we’ve got some extremely exciting breakthroughs that we think are going to enable super high channel counts in distributed regions. And so one thing I’d say, I think for the field to keep growing I do think there’s going to be a Moore’s-law-esque thing, but to do that you have to find a way to scale up the number of sensors, but it’s not necessarily about the density. I have a belief that spatial coverage really matters. Because the brain has its own redundancy built in. And so what I think is that there is a sweet spot for coverage, because you need unique information. You need unique information from different brain regions, and you don’t want to over sample. So then if that’s the case, if you hold that thesis, then the question is what is going to be the best mechanism to get in there, navigate around, deliver packages, and then get out without destroying local anatomy. And I think that’s going to be one of the challenges that skull-based approaches are going to face as they want to get to more and more regions of brain. What do you do there? How do you, so you know, but there might be ways, there might be ways to achieve that. So I think the vision here is that we’ve got this first-gen system that we’re kind of solidifying a use case. We’re going to build these ML pipelines. We’re going to build a system that controls Apple products. And then our next-gen system, we are have to solve the issue of tiny chips across a distributed range where you can navigate around and not be limited by space. And we think we’ve got a pretty exciting solution for that.

Juan Benet

With that jump in regions and channels, what are some of the applications that you’re thinking about being enabled?

Tom Oxley

Yeah. So, so motor, a motor system is about predicting the intentional activity from the motor system. If you go to speech, it’s the same thing. Actually, the speech decoding right now is kind of also just doing motor decoding of the creation of speech. There probably is then speech, semantic speech decoding. So then you may be getting into concepts or ideas as you’re moving into the frontal lobe. But then you if you’re into the frontal lobe, then you’re getting to emotional content. All the non-verbal things that you communicate are kind of taken for granted a little bit. You know, it’s the little gestures, the little the eye, the facial expression, the tension, the triggers, the annoyance, the frustration. If that can be decoded, I think that could be really transformative. And I and people might say, “Well, [ __ ] I don’t want to be I don’t want to have my emotions.” And I get that presuming that you can find a way to solve privacy. It feels to me like it would be a mirror because you go to therapy to basically get told that, well, you reacted in this situation. You know, you should think about maybe it was because of this. But if you had a system that in real time could feed you back like what is going on with you, that could be a hugely beneficial thing.

Juan Benet

Presumably, you could also detect all kinds of early signals of potential problems or emotional disorders and things like that.

Tom Oxley

Yes, there’s probably some long-term health benefit, but I think the value of BCI is primarily going to be in the live kind of contextual utility. Imagine that you had a system that started to build knowledge of your typical behaviors of react. Now, it has to have context to the environment, which is is kind of an obvious statement, but the BCI will eventually train on making predictions on how you would likely react in a certain environment. That would be, I think, an emotional. Well, actually, that would extend to like things like navigation as well. I think there’s a big portion of your brain that’s there for navigation. So, I’m sure there’s going to be some navigation input that means you don’t have to check your phone for stuff. But I really believe the emotional sort of subconscious piece. I guess it’s like there’s the mirror piece. There’s the for people like I have a lot of people with families with people with autism who write to me saying if you could help like transmit the internal state of my child or my brother we could much better understand what’s going on because we have a, because the use of language with people with autism is very challenging. So we get a lot, we get a lot of people. So like that ability to oh like actually his emotional state is this right now, I thought he was angry, or, you know, so I think there’ll be conditions where there’s kind of a decoding element to it, but then I also think, like, you know, I keep thinking like as robots become, physical AI becomes ubiquitous in our world, you kind of want it to know how you’re feeling. In the future, I think this technology will, if you choose to let it, you’ll have mechanisms to express yourself in a way which the world around you will be able to react to. I’m also a little bit scared of that myself. Like from an ethics perspective, I think there’s huge challenges in getting that right. But if we did that right, I think it’s going to help us overcome one of the biggest challenges of being human, which is that we all piss off each other so much. So bad at like communicating emotionally with one another and with ourselves and like being aware of yourself.

And one other thing I was going to say, it seems to be the case that we’re learning that doom scrolling, like TikTok, is having an impact on young children, but it’s also having an impact on everyone because it’s very addictive. And so there are these patterns of mental behavior that are very bad for you that you will do. I think the technology could actually be helpful in identifying when and how you’re using your brain in a bad way and actually just be a bit honest about it and help you manage it better.

Juan Benet

Yeah, in that good, once you have some amount of read write figured out you can also help implement better self-regulation pathways and all that kind of thing. There’s a range of people that are very excited about even things like being able to regulate sleep and, like, just kind of induce sleep through a BCI and things like that. Would you be able to do speech with the next-gen device, or like, further on?

Tom Oxley

Yeah, our next-gen system is going to be able to reach all corners of the brain and deliver high channel count systems. So yes, we’ve, the question is where do we, and what do we do with it?

Juan Benet

Very exciting. Super cool.

Tom Oxley

We’re very excited. We’re going to, we’ll, we’ll probably be making some announcements about it in later 2026. It’s been like 12 years of thinking on this problem, but there’s, I’d say that there are some techniques that have been used in cardiac which are not obvious which will let us get to regions that were not thought possible before.

[36:41] The numbers: 15M patients, FDA, and Medicare

So, in all the conditions that you mentioned already, just even with the current generation device, there’s likely millions of people globally. I mentioned a bunch of conditions from a motor perspective. We think there’s 15 million people global that have either moderate or severe motor impairment. From the range of conditions I said before, on the very severe end of the spectrum in the US, we think that’s about 3 million, and then we think there’s about 700,000 people on the severe end of the spectrum in the US. This has been an active conversation in the field right now. We were talking about it at your retreat, and we’re hearing that Neuralink wants to implant people that do not have injury or disease. My view is that that’s all a little bit rushed and it’s going to take a bit of time to show that it’s safe and effective, and I think we’re going to be in the domain of motor systems primarily, and I would include speech in that, like motor speech, as the first wave for the next, say, 10 years, 10 to 15 years. I think that’s vision with prima from science with vision as well. Yeah. Yeah. I’m kind of putting vision in a different category cuz we’ve talked about this cognitive AI thesis.

I think to build a framework like that can predict the next kind of cognitive token, that’s more of a sensing thing. The feed, the write, is interesting, but I don’t think it’s as interesting as the architecture of a sensing system that can help predict the next cognitive action. So then, you know, we talked about motor and speech and emotion and maybe attention. I think attention is a super interesting thing because if you can start to cover, so there’s been a lot of discussion of whole-brain BCI. So you know, what is whole brain? Well there’s a lot, the brain does a lot. So if you can cover multiple different cognitive domains including those ones, you can build models that can kind of create predicted features of the user given a certain context, and you feed that into the model, it starts to begin to potentially supplement your cognition in any moment. So then it’s like, okay, who is going to want that? And I don’t think it’s just going to be a jump to like normal people. I think this is going to take, you know, 10, 20, 30, 40, 50 years. And what I’ve been thinking about is the next wave after, say, motor impairment. I’ll include speech in that because speech impairment is a subset of a smaller subset of motor impairment. I was thinking there are people with age related cognitive decline and maybe mild cognitive impairment, and then you’re getting into the hundreds of millions of people. Some people have more cognitive decline than others, but there is a very common phenomenon of people getting into their 40s, 50s, 60s starting to report cognitive slowdown, some moving into mild cognitive impairment. If these systems can supplement your cognitive processes, help you engage in elements of communication that includes emotion, or with technology, and it can kind of fill the gaps that start to happen, it can sustain you at a cognitive level. I think that is the sort of thing that a large population might start to adopt because it’s, you’re not, it’s not a like a, it’s not a utopic BCI superhuman vision. It’s more like how can we preserve a level of cognition over a longer period of time. Now, I’m not, I’m not saying that it presumably goes to a level that is superhuman in some way, and maybe there’s a becomes a market for that. But there’s also a lot of push back on that. There’s a lot of questions over how, what happens to society when you start to do that, and who are the people doing that.

Juan Benet

And so, but couldn’t this, just to push back a bit, Steve Jobs used to say the personal computer is the bicycle for the mind, and I’ve been thinking of the BCI as the car for the mind, where it’s like a much harder to make device, much more complex, much more difficult, more broadly impact society and reshape society, but as soon as you sort of like get it working at scale, the benefits globally just become so dramatic that that kind of becomes like a key thing for lots of people to have. And so what’s the scale of utility that a device like this can give such that yeah like people broadly say hey like this would be a way of interacting with each other with the world with AI with all these systems that’s just kind of a, you know, broad capability expansion, and kind of like how far away that is. But I would imagine as soon as there’s thousands of people out there already, maybe tens of thousands, maybe even more that would want to have that kind of mind expansion sort of human experience broadener.

Tom Oxley

I’ve thought about that. I’ve the car analogy I think is interesting. It’s a little bit dangerous. It helps you get there faster. It’s like it, when the car first appeared it was, people were terrified. It was like a horseless carriage.

Juan Benet

Yes.

Tom Oxley

And there was, and there were moral questions about it. There are no more moral questions about cars. So, I’ve reflected on that. Although, I think the ethical questions because it’s your mind and there’s now skepticism around how technology is used from a privacy perspective. I think there’s potential much more push back on this technology than, and maybe adoption, and even like discrimination in how it’s going to be rolled out, time scale. I don’t know. It takes 10 to 15 years to go from inception to commercial approval in the current state. Maybe China is going to do it quicker than the US, which is a very high likelihood. And so we’re still in the domain of like motor systems right now. And you know, Neuralink’s not covering much territory unless they’re going to go and start putting holes everywhere and putting multiple in. That seems really difficult. Tons of holes over the skull. Like it seems like a not super scalable. Yeah, I think Elon wants to remove the whole skull and just put a helmet, but I don’t think that’s going to scale into the population. And I think that, you know, so I think we’ll probably have digital humans uploaded before we probably have large volumes of people replacing their entire skull. Yeah. The problem is to you need context. I think to train these systems, you can’t just do like a go into an imaging, go into an MRI and just get a scan done. It needs to track your brain. You need to train it against how it’s reacting in real time, kind of like a car. So I still think it’s going to take time. And so to build these systems that are going to have such great digital twin capabilities, you have to have coverage in many different regions over a period of time where it’s being trained over a period of time. It’s a major design challenge because you don’t actually like once you’ve trained it, you don’t need all that compute. But the problem is you have to build it all in in order to achieve the training, but then you can run it on a much lower order of power and compute. But you have to get there.

[43:18] Cognitive AI: foundation models, the data economy, and the 2040s

Juan Benet

So, yeah, let’s go into the cognitive AI thing. So you already touched on this a bit but kind of the broad idea is hey, across a range of patients you can collect a range of neural data and start building foundation models that enable you to have, like, the zero shot learning for additional applications and so on that require very little individual training to be able to use the device. So, so you can kind of onboard faster, but there’s probably also, you know, a range of applications that become unlocked also in terms of integrating data from multiple devices potentially and being able to get better models of how the brain works and how various signals work. Yeah. How have you thought about this and kind of what is sort of like a timeline? How much data do you think we’ll need? What are some of the early fruits that we might see?

Tom Oxley

Yeah. And then and then the third part to that list was then potentially these models being useful when taken out of the loop and then using to, you know, so if these models get so good that we’re learning things about how the brain works that we hadn’t otherwise learned, then they can potentially inform out of the loop models to control physical AI. I think it’s going to take time. So Nvidia is, so Jensen’s very interested in this, and Nvidia are kind of watching the space wondering, this is a new data economy emerging, and they wanted to sort of step in when compute will accelerate the progress within BCI, but you come back to that challenge of you need the data and you need a lot of data from many people, and so this is the, this is the arm wrestle, it’s going to take, that’s why I think it’s going to take a bit of a bit of time. You need distributed systems, you need them live, you need contextual features, and then you need some interaction with the system that enables some model to train and learn, and it needs to be delivering useful features all the time to the users. I think it’s going to take time. I think, you know, we’re in the first innings. We’re seeing first generation, I’d say our systems, first generation system. You’re seeing second generation systems starting to emerge. I think third generation systems, I’d consider our next thing coming, a third generation system is getting into the whole brain domain, and then maybe fourth generation systems are beginning to, you know, be safe enough, scalable enough, easy enough to use and useful enough to start generating that level of scale of data. In terms of neurotech right now, the most voluminous neurotech implants in the world are the cochlear implant, and they’re doing about 40,000 units per year. They’re pretty low numbers, like cardiac are doing a million stents, a million pacemakers. So, neuro still got some, this, this, you know, everyone’s approaching this slightly differently, but we’ve still got a way to go before we can scale. Even the infrastructure, the way the hospital systems are set up, the number of neurosurgeons, the neuro-interventionists, the places you can get things done, people don’t want to have to travel a huge amount to, so there’s still a lot of maturing to happen. That’s why I think like we’re looking at like 2040s, 2050s for some kind of an event horizon where, you know, there’s some, I don’t know, I don’t want to call it singularity, but that there’s, you know, a moment where the system starts to surpass what the human body is capable of doing.

Juan Benet

And I think you were kind of getting at it from a network effect, but I do think there’s going to be a moment where there’s enough people that can start to interact with one another in a way which is not capable using the normal human body. And again, back to emotions, the most obvious example of that is I can somehow broadcast or throw my emotional state so I don’t have to explain to you what’s going on. And so when you get to that level of, and that’s very human by the way, that’s a very intimate human engagement. So I think there could be the potential for these systems to enable a level of human interaction that is maybe incredible, like feel amazing and like very intimate. The Nexus sci-fi book has a whole, it’s a BCI oriented story, but it has a great description of the just like the human experience of being able to sense each other’s various emotional states and being able to tap into like this kind of like shared reality that just isn’t possible otherwise once you have kind of very high bandwidth connectivity between human brains.

When Android’s Dream of Electric Sheep, the, that was a book that Blade Runner was based on. I think in the first chapter they dial up each other’s, that’s a bit more dystopic cuz that’s not like the husband and wife are fighting and they dial up each other’s mood for the day.

Tom Oxley

Yeah. There’s this kind of bad effect where in order to kind of sell a story well, you just need elements of high drama. And usually when some technology isn’t clearly serving a purpose to the story, it tends to get cut. So you have this selection pressure for stories to become overly dramatic around technology. So, so you just select for sci-fi to have dystopic lenses over the possible future. And so you end up with this distribution where most of the sci-fi that you read feels dystopic, and so we have like a dearth of extremely positive visions of the future spelled out in sci-fi.

Juan Benet

That’s interesting.

Tom Oxley

But there is like some archetypal dystopic narrative that is like I think ingrained genetically in the human brain. We see a handful of quite idiosyncratic like predictable psychotic delusions that occur with people who have psychosis. And one of them is that you have put something in my brain or you’re taking things out of my brain.

Juan Benet

Fascinating.

Tom Oxley

So I’ll get an email once a week about someone who thinks I’ve put something in their brain. What I think everyone, all the companies are getting this. So that’s why, that’s why there’s no signage on the building.

Juan Benet

Wow.

Tom Oxley

Death threats, like, a lot, so there is this kind of cultural embedded fear around like stealing stuff from my brain or putting things in my brain that I think fuels it as well, but there’s cultural obsession with BCI.

Juan Benet

That’s the first time I heard of that. I can imagine it’s similar to like 5G chips conspiracy theories and so on.

Tom Oxley

Probably, and we’ve got Bill Gates as an investor, so that was a double, that was a double whammy.

Juan Benet

Yeah, maybe going into AI and neurotech, how it might co-develop.

Tom Oxley

Certainly, a lot of AI has already helped BCI tremendously by being able to decode signals much better and solve a lot of the hard challenges that we were, we thought we were going to have around how do you even extract signal from the brain and so on.

Juan Benet

But what are some of the other ways that you see kind of AI accelerating BCI development now?

Tom Oxley

I was at an Apple event and there was lots of conversation around like the word, the term foundation model was, is being used a lot now, and so these foundation models that enable this zero shot classification I think the idea that you’ve got a mechanism to train models that are increasing in size that can improve your ability to separate signal from noise in a way that you probably couldn’t have done with a supervised method, seems to be a huge driver of space. So the ability to use large compute with unlabeled data with self-supervised techniques I think is going to drive, because you know we see signal and it looks like there’s noise but it’s not actually noise. It’s all different signals coming from the brain and you don’t know how to parse it all out. So I think the ability to use high amounts of compute to solve that is going to be a really big deal again, which means more data is going to have better features. That’s on the ML side of the decoding. But I think the other interesting side is if the BCI enables interactions with your device, like OpenAI is building a new device now, right? So is is typing in on a keyboard into a prompt to activate the next thing, is that going to be the best way, especially as the AI, so I think the interaction between the BCI and prompts is what’s going to be super interesting, especially when if the BCI is predicting your next cognitive move and that can feed in in real time to a prompt and you can have real-time interactions, then you might have very fluid and really quick interactions with the AI in a way that didn’t require you to sort of sit down and use your hands on a typical interface with a keyboard or a mouse.

Juan Benet

At what point do the blending of these systems start feeling like an extension of you, as opposed to like a separate device?

Tom Oxley

That’s already happening. That’s already happening. I’ve seen some examples. If you want to speak really quickly in a conversation, then you might skip the keyboard altogether and just choose what is being generated as rapid prompts in real time, even if they’re a bit goofy or not quite right, the user will much prefer to do that because they want to be live in the conversation. And so even if you can because like we’re talking, I don’t know, 120 words per minute or something. So even if you’re typing on a keyboard at like 60 words per minute, even 30 words, 60 words per minute, you’re still not actually keeping up in the conversation. So there’s a huge driver of people who have disability or impairment to be live and active in a conversation and they would prefer to say well I’m not, I’m not speaking for everyone here. I’m speaking for some people that I’ve spoken to. They’ve said I would much prefer to say something even if it’s a little bit goofy and to impart an effect on the conversation and have people react to that and then redirect and keep going. Then you’re like well was that really what you wanted to say? Like was that exactly, I don’t know. For me, I’m always struggling to find the exact words in my brain anyway, but I’m not diminishing that. There is definitely an agency. Nita Farahany feels very strongly about this. I don’t know if you’re talking to her, but she wrote the battle for the brain and she talks a lot about cognitive liberty. She’s very worried about what it might mean to have the AI, the combination of the BCI predicting your next move and the AI predicting, and then your agency having a couple of steps removed, and then what that will do to impact you and yourself.

[52:56] Consciousness, psychedelics, and the extended self

Juan Benet

Yeah. And there’s hard questions here around, you know, how many layers of brain are involved in that anyway, and at what point does the tool and machine that you’re building just become an additional layer to the brain as opposed to a separate system. You know certainly like the human neocortex is a totally different system than like, you know, the much earlier, you know, the cerebellum or like, you know, the brain stem and so on. And somehow our brain is able to integrate these systems and experiences to form a coherent whole, probably as a result of the degree of connectivity bandwidth like this. I’m reminded of the corpus callosum experiments where they can find different personalities in the different lobes of the brain. And so that’s a sign like when you don’t have that kind of degree of connectivity, then your shared consciousness is potentially split into like separate components.

Tom Oxley

Mhm. But that suggests that if you have a high degree of connectivity between your brain and some other system, you may be able to cohere that into a single system or a single kind of conscious experience. I don’t think we know enough about consciousness yet to be able to tell whether or not that will happen, but we’re probably going to experimentally reach this sometime in the next decades.

Juan Benet

Have you seen these explanations of how desynchronization, especially the default mode network, which is kind of thought to be this sort of resting state network that enables your consciousness to exist, the high-dose psychedelic drugs, mushrooms, DMT, acid, cause a desynchronization, the resting mode goes away, and they, I’ve seen some descriptions on MRI studies of a desynchronization event where you lose that resting state and then you actually go offline and go to some spiritual place. So there’s, there’s lots of us, there’s still a lot to learn.

Tom Oxley

That reminds me of studies around certain types of meditation that reach like probably very different behaviors but it’s the meditation effect is also a desynchronization event, the same as the psychedelic. It’s the same, seems to be the same process of desynchronization across the brain and high connectivity, but it somehow that results in a loss of ego and then going to some other realm.

Juan Benet

Yeah. Yeah. Yeah. Once you start kind of interfacing with all of these systems with AI or other people or other humans in high bandwidth, like that that just kind of starts opening up a whole range of questions about, I don’t know, the future of humanity and like the species and how we develop, like how much do you either think about that, or are excited about it, or?

Tom Oxley

I think about it a lot. I’d say I’m cautiously optimistic but I’m a little bit worried. I’m worried in particular for the things that we don’t handle very well as humans, which is the addictive component of the technology. So the idea that a BCI would enable interactions with a system that is puts you into that kind of passive scrolling mode and reduces your agency, I think that’s a major problem. I think it’s probably going to be about the user and the product, but we’re learning with social media that there is this very addictive element to the way that we use these systems. I think we’re already seeing that now though with chat basically. People are kind of offloading cognitive work. You’re now receiving work from people where you think it’s their work and you can tell it’s like, but it’s okay, like you’re like, okay, I know that you’ve used chit to generate this, but I can still, you still generate the themes and it’s a different slant, but you have to sort of think about it slightly differently. So we’re definitely going in that direction. I’ve definitely already encountered cases of that. There was a whole conversation with multiple people, all of which was ChatGPT talking to itself for all parties.

So, individualism is still a bedrock of Western society that I think we have to uphold, and I think BCIs could go in either direction. I think they could be profoundly individualized or they could result in, you know, more group think or group connectedness.

Juan Benet

Yeah.

Tom Oxley

There’s geopolitical differences emerging. China’s got a very different approach to this, probably, to Western societies, and they’re probably going to be deployed in very different ways. So I worry about privacy. At the same time, I think that there are elements of, like, using the system to preserve your, like, not privacy but maybe something like sovereignty, you know, like your agency. Yeah. Your agency. Like ironically I think you can use these systems to determine whether you are actually behaving with agency or not. So that’s kind of ironic because I think it could go in the other direction, but at the same time I think they could be used in a really good way to have an honest conversation with, you know, how human are you being right now.

I mentioned discrimination, how society’s going to handle pluralism. I saw Steve Bannon comment about a very vicious anti-Elon rant that he had about the potential of BCI like he was talking about Neuralink resulting in major cultural divisions because, you know, there’s going to be like, you know, anti-BCI people or like, so, or they’ll, you know, discrimination, either you couldn’t get one cuz you couldn’t afford it, so like equity problem, but also discrimination in that people begin to like detest or, you know, feel hatred towards people who might have these technologies. But the other side of it, the positive side of it is, like, self-determination, like more agency, more autonomy, more decision making, more protection of yourself, like more extended human behavior, that you could, like you were describing the car, I think I do think about that analogy. More capabilities, more capabilities, more forms of expression, more forms of experience. Yeah, I do think that it could open up parallel streams of cognition, which we kind of do a little bit. You like you multitask and you think about different things at once. But actually, but usually when you’re doing one task, there’s one kind of dominating thing. But if the device is running, if the brain is running in parallel, you always got a thought in the back of your head or something you’re worrying about or you’re distracted, you know, are you listening to me right now? I do think the BCI could potentially enable multiple, like if agents come online, physical AI, it’s now using your cognitive tokens to predict next actions in some AI integration. It’s start, and you’ve given it permission. It’s starting to act on your behalf. Oh, you’re worried about where the car is. I went and made a phone call. They’ve topped up your parking meter. I spoke to the police already. You know, it’s doing five or six things at once and taking care of stuff for you. I think that could be amazing. But then it’s like, well, were we built for that? How’s the how are we going to handle that? Like,

Juan Benet

Yeah. And if we can figure out the, how do we input that information back? So, it’s not another thing doing it on your behalf, but it’s rather it becomes kind of like an extension of you, and you have some degree of awareness of what’s going on and control over it, and so on.

Tom Oxley

Nita Farahany talks about, what does she call it, when the fiduciary, the same way that you treat your physician or your lawyer, you need that level of trust in it.

Juan Benet

Yeah, totally agree that if it’s a separate thing that’s acting on your behalf, you definitely need that level of trust. But I’m also thinking of a case where it just doesn’t feel like a different thing anymore. It just becomes part of you. If you can increase the connectivity between that process and yourself, then kind of, you can kind of, we can potentially integrate it. To your point around, are we ready for that, or are we built for that?

Tom Oxley

I don’t know. We’re not built for, we’re in the middle of a concrete city with like skyscrapers and so on. We definitely didn’t evolve for this. Like we were, you know, our genetic evolutionary path can’t adapt fast enough to the technology that we’ve been building. And so early hominids were never meant to like, I don’t know, build steam ships and go around the planet.

Juan Benet

Yeah.

Tom Oxley

I think as with any tool that is very powerful, it can go in different directions. The thing I worry mostly about, those elements of humanity that already, we’re already seeing causing suffering in this. But there’s also a dialogue in BCI that it’s going to be some solution for AI alignment and safety with AI. I don’t think that’s the case. Like it doesn’t make sense to me that like it’s going to definitely expand human potential, but along the trajectory of ASI, it’s not, it’s you’re still interfacing with the biological.

Juan Benet

Yeah, maybe back that, and for listeners, so try to steel man the argument that you’re, you’re the steel man would be, evolution of humanity is slowed down on the scale relative to what technology we’ve created, the potential, and so evolution was powerful and we like keep improving, but the pace is so slow we need to find a way to evolve humanity so that we can keep up with AI. And if we do that, we might be able to interact in a more safe way and continue to steer AI in a safe direction so that we never lose control of the AI and we don’t become its slave.

Tom Oxley

Yeah. Right now, we’re on a path to potentially create a separate species, that will be first generally intelligent, and kind of superhuman in some jagged sense, but then later superhuman in maybe all counts, and then eventually like ASI. And if we build it as a separate thing, then we have this hard question around which species is in control, which is like a very uncomfortable question. And so the part of the steel man argument is maybe there’s a path by which if we can figure out connectivity between the human brain and the AI systems, then as they develop, we don’t build them as a separate thing, but we build them as part of humanity.

[1:02:54] Agency, addiction, and geopolitics

Maybe just a quick divergent going back to the things that I worry about is that the end destination with everyone getting a BCI is that we do all become connected in some way and then there’s some loss of individual agency from in that process. There might be cases where you have such high connectivity that you do, or there might be cases where you might still be able to preserve individuality in that environment. I really hope so cuz I think that’s like a key thing about being human. That was Jung’s kind of apex, and Nietzsche actually like finding the apex of individualization, was actually how the entire species moves forward because we’re all strive towards our own very unique experience and those highly differentiated unique experiences are what, where we learn things a little bit different, and that’s how, like, the edge case that helps us learn. So, I worry about losing that.

But with the ASI thing, like I like it. I like the idea that okay, it’s like we’re somehow more connected to it, but that doesn’t mean that there’s not a separate ASI being that we’ve created. Maybe we’re a little bit touching it a bit more. But the idea that the ASI’s trajectory relative to the human brain trajectory, which is interacting still with a static kind of biological lump, doesn’t, it doesn’t make sense to me. Even if we manage to interweave AI systems and the human brain and interface them to some degree, the interface boundary still like quite different, and so these might not fully integrate, and even if we do, it’s likely that separate from all of that, you might still be generating an ASI in a data center that is not connected to humans, and so that whole process might be just fully unavoidable.

Juan Benet

Yeah. Or you’ve actually made it worse by connecting to it because we’re a species that’s less powerful and now connected.

Tom Oxley

Yeah.

Juan Benet

And then maybe exposing some degree of like attack surface in a sense.

Tom Oxley

Exactly. Which you’d go straight for the dopamine system, like if you’re, you know, but that’s not to say, like so detaching it from the AI safety alignment problem. It just puts such a big burden on the field and it’s kind of, it feels to me it detracts from the focus on what BCI could do to be really awesome for humanity if you just think about it from the lens of, well, if we, you know, it’s there purely to engage in the fight with ASI, that’s, it’s a different way to think about the problem, and I think there’s a lot of, yeah, it doesn’t do justice to the degree of amazing impact that just the BCIs on their own can, for both like people with a range of conditions today, or people who are they going to be able to expand their experience as we were talking about before.

Juan Benet

Yeah, like flourishing.

Tom Oxley

I do think I do think this is a mechanism by which we improve human potential and human flourishing, and that can be beautiful and good and positive and is kind of in a separate thing as to what do we do about ASI.

Juan Benet

Yeah, it’s going to be a very interesting next few decades, I think.

Tom Oxley

Yeah, I think the 40s is going to get interesting.

Juan Benet

Yeah. Yeah. Yeah.

Tom Oxley

I think stuff’s going to happen in the 40s.

Juan Benet

Yes.

Tom Oxley

Yeah. It’s pretty funny that this was well predicted by, you know, a number of people including Kurzweil and others. Like they just, yeah, they described like the 40s is when, the 40s. He just predicted that the singularity was going to happen in the 40s. He later, I think one of his predictions was in the late 20s or early 30s, and then orig,

Juan Benet

Yeah.

Tom Oxley

And then other predictions were in the 40s. Yeah. Yeah. Yeah. And I think he like, I don’t know the exact descriptions but he’s much more right than I think everybody else gave him credit for at the time. Like I think when he was first publishing all this stuff, everyone was like that’s not going to happen. And now we’re like sailing into mid-20s and we now have LLMs that can reason quite well.

Juan Benet

Yeah. No, I’ve always been really inspired by his writing.

[1:06:55] The optimistic vision: unlocking the subconscious

Juan Benet

What’s your kind of optimistic vision for the future? Like what can we unlock as a species?

Tom Oxley

I think there’s a lot of suffering in the world from a lack of self understanding of the way that your brain has architected itself on account of all the traumas that you’ve collected. I’ve got a four-year-old daughter and she’s probably starting to form memories around now. She’s been a person for several years now. It’s so weird that like you walk around for 4 years and you can’t remember what happened and yet things that happen, you form these like fundamental basic philosophies around the world based upon what happens early in your life. That’s just the beginning of what becomes like the way that you are built. And then you know people get therapy for a lifetime trying to unlock, and we all have these very predictable same triggers that you then spend your life trying to manage in your relationships around you and probably going to war over and all sorts of things. So I think that this technology, yes I think it can help people who have injury or disease overcome that. I think it can help people communicate, but I think the things that it could do that nothing else will do is unlock those subconscious domains within us that there is no other way to do that. I think that could have a huge profound positive benefit on society. Kind of like I keep thinking about a mirror because like I when the mirror was first invented, but like if you hold a mirror up, you can just if you could hold up a mirror to your own actions, they’re very hard to manage. They’re very hard to reflect on and they’re very hard to change. I think that’s one of the most powerful things it could do.

And then I do think you talked about creativity. I think the ability to share nonverbal creative outlets, like almost chaotic, that’s the other thing about the subconscious, there’s this like you got every night we go to sleep, there is this, it’s not fully chaotic like there are these sort of archetypes bubbling deep underneath, some which are more pronounced in us than others, but there’s very predictable archetypes told in story form in fairy tales, and the ability to get down to that level of like chaotic storytelling that’s happening deep within us as this kind of engine underlying ourselves. Like it sounds a little bit weird and crazy, but there are these like drives underneath the way that our brain can communicate now that are just not at all not seeing the light of day at all. So I think there are going to be mechanisms of storytelling, of creative outlet, of communication, of emotional integration, of connection, of productivity, of all human things. These are all human things. You know, when people say, “Yeah, but that’s like very unnatural.” It’s like, well, actually, you know, even the development of speech, when do we, when did we start talking? Like a million years ago.

Juan Benet

A million years ago. Yeah.

Tom Oxley

Probably depends on how you count it, but like more advanced speech.

Juan Benet

Yeah.

Tom Oxley

Half a million years ago. So, that wasn’t like the brain, like it evolved to that state. But like we’re starting, we started to use it in a very different way. And you’ve now got like if you look at the brain like you’ve got this huge amount of processing power and then it comes through this tiny band of fibers down your brain stem called the pons. It’s like the output of the motor cortex. If that gets damaged like you can’t, you’re locked in and your whole brain is working fine. That’s like, you know, the patients that need BCIs the most. But it kind of tells you that there is this whole world going on in there that we have this kind of not very good mechanism of expressing that I think the BCIs could bypass. And I think at that point we get to levels of self-knowledge, interaction, communication and hopefully positivity that we couldn’t do before and all the bad side of humans as well. But humans are humans, imperfect. But I think this becomes a tool of self-expression fundamentally.

[1:10:54] Building a company: 696 no’s

Juan Benet

Just in terms of building a company, like you’ve been able to span the whole R&D pipeline, like going from, you know, early science to building a device and figuring out the possible applications and developing it, and like building a team and creating a company, and like, you know, fundraising for it, and like steering the ship, and there’s enormous amount of work of spanning the whole pipeline to be able to do something like this. How did you learn all of this landscape? Like do you kind of just learn by doing, learn by studying others? Like what sort of pointers you would give out to other people following in your footsteps of, you know, how to get there?

Tom Oxley

So maybe going back, research, jump from research to entrepreneur was, research is asking questions and being very curious and hoping that the question that you’ve asked and the discovery that you find will be, I think, significant enough to feel like you’ve really made a difference. The difference with pursuing development or entrepreneurialism was you then have to choose one thing that you think is worthy to then dedicate yourself to. In the case of me, it’s been like 12 years, 15 years or something. And that’s tough because you stop asking the same question and then you start like banging your head against a brick wall to try and get things right over and over and over. Like pitching, I’ve done, I’ve done five, six, 700 pitches now. And I remember early on when I started doing, I’m like, “Oh, why? I can’t say the same thing again. This is absolutely exhausting and I hate the sound of my own voice.” And but then you’re like, “Well, you know, that time I used a slightly different word and I got this reaction.” And so then you have to sort of enjoy the storytelling and like the feedback, the interaction really matters. So you have to enjoy the performative nature of it and learning. So that that was a big jump. I think the resilience, I think you’re just a resilient person or you’re not. I don’t know how you learn that. So, that just probably comes from childhood. But being told no over and over and over and then failing over and over and over and being kind of stubborn and driving forward despite that, that’s a really important, a lot of people stop. A lot of people give up. A lot of people believe cuz you speak to like that’s, we’ve done, actually we’ve just announced the series D. So that’s four rounds of financing, so that’s four yeses and like 696 no’s, and they’re really smart people, and if you start believing the really smart people, like of course I’m going to fail, why would I do this? So there’s kind of an element of stubbornness, but like, yeah, persistent stubbornness,

Juan Benet

contrarian,

Tom Oxley

contrarian, contrarian, I think contrarian and right perspective.

Juan Benet

Yeah, yeah,

Tom Oxley

and then, but that kind of clashes with building a team, because then like building a team is like a different challenge, you have to let people fail and you have to hand things off and you have to like trust people, and that’s also a very different skill set. The three things that I’ve reflected on are the three most important features are resilience, common sense, and tenacity. And then I think Peter Thiel said this, but you have to really look into yourself and ask the hardest questions of yourself. You have to be willing to really go into the ugliest worst side of yourself and take on the worst element and be willing to like work at it. A lot of people find that hard. Everyone finds that hard. But you have to take on the ugliest part and like and figure out how to turn it around because it will express itself as you go, and as things build, like that’s going to express itself, and you bring everything with you.

Juan Benet

Yeah, kind of like the companies tend to reflect the strengths and weaknesses of the founders, and so if like whatever your worst weakness is, the company’s going to have in spades, and so then you have to like figure out what that is and like work against it.

Tom Oxley

And then quitting, you got to quit things, and so, you know, people drop out of college, I had to quit medicine. I spent 20 years learning how to do those procedures. I came to New York and then I after 20 years getting like, I did a lot, I did medical school, I did internal medicine, I did neurology residency, I did a PhD, I did fellowship, and then I quit to go all, who’s done it has sacrificed something.

Juan Benet

Have you studied any specific founders or companies to like learn from their...?

Tom Oxley

Well, I mean I’m inspired by Steve Jobs, inspired by Elon. I mean Elon’s created brutal cultures in his company, but he’s like he puts the mission above everything. And that’s a pretty powerful way, that’s it’s grueling though, and it’s like having a family is an impediment to your absolute full blasting at work. So I think finding that balance has been a real challenge, and then Jensen and Nvidia is just an incredible leader. He’s a very different style.

Juan Benet

Do you learn from them by both, like I don’t know, watching their public discussions, or reading biographies?

Tom Oxley

I think reading biographies you learn more about what’s really going on.

Juan Benet

Yeah, and you also learn to pick apart like what seems to be like, you know, a good feature, and what are bugs, right?

Tom Oxley

You can, in learning about other great leaders, you can figure out like which of the pieces you want to take on. Well, Andy Rasdal has just joined this company. So Andy, he sold AB to Medtronic, became president of Medtronic. He went and started Dexcom. He’s had an incredible career. So I’m learning from him. He’s coming in in a chief of staff role here. He’s made me realize what I’ve been thinking about recently is the ability of a leader to stay, Brian Chesky from Airbnb, he’s great, the ability of a leader to stay high level but then choose the thing to zoom in on. You can’t zoom in on everything, but you choose the most important thing and then you go right down and you get to the technical bottom of it and then you probe until you get the questions that can’t be answered and then you just focus on that and fix that and then you zoom back out and then you go around, look again. That’s Elon’s secret. Like that’s what he’s, it’s unbelievable. He walks in, figures out the problem, goes straight down to the IC engineer, talk to them, and then zooms back out and then leaves, or fires a few people, promotes a few people, then leaves, comes back in a month. I can’t do that. But like that concept of being able to zoom out, zoom in. That’s really hard.

Juan Benet

Yeah. Marc Andreessen has a good description of this where like Elon will intensely prioritize solving whatever the company’s biggest problem is right now and hyperfocus on that and do that, solve that problem, and then move to the next one to the next one to the next one. And part of how, why that works for Elon is that he also has very strong management teams also, right? Like at SpaceX, he has Gwynne and a whole team that are able to do the parts of the function that he doesn’t want to do.

Tom Oxley

You know a lot of large organizations get the management structure becomes so ossified or strong that then it gets overprocessed, and then this Elon style ability to cut through all of that and get to the core problem and solve it is like becomes extremely extremely useful to be able to really refocus a team on the things that matter, and maintain, like they’ve got high turnover, but the ability to maintain the intense focus, the mission driven, but then handle a culture where there’s very high turnover, and it’s kind of you worry about failing, otherwise you’re going, you get it like that. That’s that’s that’s that’s hard to do.

Juan Benet

The kind of speed of development that Elon is able to achieve is extraordinary. Steve also drove that. The iPhone was kind of, they decided to do the iPhone and announced it like 2 years later. It was like a 2-year extreme sprint. The iPod, when Tony Fadell joined Apple, the iPod was built in like 9 months or so, like going from, you know, Tony Fadell joins to the iPod, is, I think shipping in 9 months, or at least announced, but I think it’s shipping. It’s extremely fast development cycles, very difficult to drive a team with to that kind of velocity. A lot of people question whether or not the velocity matters that much. And in reality, what ends up happening is if you don’t run a team at high velocity, all of the timelines will start expanding by very large factors and then whole ranges of things just become impossible. So the really extraordinary R&D companies achieve this extremely fast clock cycle on the whole thing, and they’re able to learn and develop so fast relative to others that they’re able to like really accomplish great feats because they’re just able to accomplish so much per unit time, and it’s just so hard to drive that kind of fast R&D in general, and to keep kind of teams very motivated to these kinds of goals.

Tom Oxley

And with medical devices it’s over a sustained long period of time where you’ve got impediments to rapid cycles because you if you change too much you back to the drawing board and you have to go back to a very long cycle of benchtop, animal studies, human testing.

Juan Benet

Yeah. So you have kind of these longer horizons and you’re trying to keep things moving in really fast pace but gated in a way which is it’s it’s feels grueling. When you think back on the whole process of like just the core invention to developing the first prototypes to like now at the stage that you’re at, or even looking ahead, what stage do you feel is the hardest or highest risk, or what would you change about the nature of how we organize ourselves as a world? You certainly mentioned the slowness of the regulatory structures. What other things come to mind?

Tom Oxley

With regulatory innovation comes risk. So there’s one question of like as a society, how tolerant are you of risk of something happen? How much ethical responsibility do you give to the physician doing the consent for the procedure versus having a more paternalistic regulatory body? In Australia, the regulatory body isn’t paternalistic like the FDA. They actually hand the responsibility at the level of the hospitals with the physicians, and they scrutinize the consent form, and so they make a determination there, whereas the FDA has a very high bar at a federal level. Trump’s spoken a lot about deregulation. What that would mean is taking on a little bit more risk. So have we got the risk level right in BCI? I’m not sure. I haven’t seen anything bad happen yet. So that that’s one thing that’s been on my mind. We’d set the goal of delivering electronics without cutting the skull. So we had a lot of early failures. That that was, I think, in the very early days though it was very, it was because it was research, it was a little bit more, there’s not, it felt like we didn’t have much to lose, and so we were throwing caution to the wind a little bit more with how we were doing designs, how we were trying new things. As we’ve gotten on now, we’ve now got a lot more to lose. And any little one little mistake or one little error on the manufacturing line, one little problem can slow down everything. So now it’s a very different skill set of getting the requirements right, getting the manufacturing reliable, getting the testing right, not over testing, not having too many requirements. I honestly don’t know how we do this quicker. I will say, I was in China recently and I was blown away.

[1:21:35] Losing the BCI lead to China

Like there is a top-down strategy in the government, BCI, robots, AI, energy, solar energy, the CCP released the strategy, it therefore immediately become social capital to be working on you know that topic, and I was in the hospital, and the chairman of the hospital has regular weekly meetings with the CCP, and they’re like, so you know what what are you working on that’s aligns with the policy of BCI. Yeah, I’m working on BCI. What are you doing? So, the galvanization of the society towards these strategic imperatives, and then the, I think it’s called the NMPA, the Chinese FDA, they have the mandate. The hospital ethics committee have the mandate. Now, I don’t, you know, the ethics of it is questionable there. Like is that going to impact safety? But the top-down alignment on how to make this move fast was unbelievable.

Juan Benet

Do you have a sense of what the goal landscape looks like and what approaches they’re pursuing?

Tom Oxley

There’s a broad vision that BCI intersects with AI and results in technological superiority of the Chinese people and therefore we have to win. And I don’t think you’ll hear the US government saying that. I think you’ll hear venture capitalists saying that. I think like there’s an industry saying that, but the speed with which that industry is moving in China is going to outpace the US, you know, in the next five years. I think they’re still behind, but they’re moving at a rapid pace and there’s lots of capital flowing now. Now, they don’t have the reimbursement system. They don’t have the healthcare system. They don’t have, you know, there’s lots of other things that are maturing in China. But maybe like in an ideal world in the US, there would be like, Israel does this well. Israel has like a department of strategic innovation. They actually decide, here’s what we’re going to work on. I guess US is doing that a little bit, but in the 70s, and in the 60s and 70s, that’s kind of how it really worked in the US, but no, I feel like in the 60s and 70s there was like, I was even reading Benjamin Franklin, like even Benjamin Franklin had like a, became, sorry, it was Jefferson, was Jefferson the third president I think he was, it was Jefferson, and he came in and he like even said, we’re going to just spend on a scientific strategic initiative is like it galvanizes people, it makes them real. You know, it’s, we need, we need a bit more of that in the US, and obviously lately there’s the NIH has taken a big hit. But BCI, like, we’ve got a, US is way ahead in BCI, but we are, we, but we could lo the lead, we could lo the lead, we are going to lose the lead, we will lose the lead to China unless we have a significant acceleration. Yeah, and AI is great, AI’s accelerating massively, but I feel like I don’t think, you know, BCI’s been, I don’t think the potential for BCI has been recognized at all.

Juan Benet

Do you think there’ll be like a Sputnik moment where China unveils some interesting significant capability and the US just immediately reacts?

Tom Oxley

I haven’t thought of that. That could definitely happen. There’s a lot of skepticism when the reports come out, like was it really, was it really doing that? But I think that I could definitely see that happening in the next 5 to 10 years.

Juan Benet

It might also be, you know, too late in the game when, if you wait until that Sputnik moment happens.

Tom Oxley

But oh, the US can galvanize pretty quickly if it wants to.

Juan Benet

Yeah, definitely. And it’s being politicized a little bit now because of Neuralink and Elon.

Tom Oxley

So it’s, you know, I’m noticing some kind of, you know, political now impacts on the space, particularly emerging around privacy and the treatment of neural data, which seem to be politicized. That’s unfortunate. I think something that will help there is a number of the other companies coming out with their devices over the next two to three years and fleshing out the space, because today Neuralink is by far the most visible and the loudest, and so it’s easy to think that that’s most of the space when in reality it’s just the one company that is very public about everything that they’re doing, whereas there are many other companies that are developing a range of technologies. It’s just kind of more the traditional medical device pathway, or biotech oriented direction, where you kind of keep everything much more private until you have a moment to, you’re getting close to releasing it commercially, and then then you have like this big public splash. Something I think would help a lot here is much more public video about patients using BCI, like just so that people have a much more clear view into the benefits and the impact on a daily life of kind of what happens. There’s just this incredible set of videos. One that was very moving for me was watching a woman getting a cochlear implant turned on for the very first time, and she’s hearing for the very first time, and it’s just this incredibly moving moment of the change in perspective and experience that is made possible by that. And so I think collecting those kind of stories and being able to help people understand the scale of impact here, that would be very helpful.

Juan Benet

Any last words or advice for other folks building in the space or dreams for the future?

Tom Oxley

Well, I think the differentiation in BCI companies right now is playing out in the electronics, in the materials, in the delivery, the medical delivery, in the physicians. So I think occupational therapists are going to come into the sunlight with, I think the occupational therapists are going to be the therapists of BCI. I think rehabilitation physicians are going to have a new tool that’s very exciting. I think neurologists are going to have an ability. So I think I can see a lot of different subsets of medicine that are going to get very excited about and galvanize when all this comes together. And I’d encourage anyone in any of those domains to kind of begin to read and get involved, because I’m still surprised at the low number of people in the clinical domain that are kind of putting their hand up right now, and the ones that do are going to have an ability to become experts. I think, like I said before, I think it’s going to go slow, so slow, and then I think in the next, probably not in the next 5 years, but maybe within 5 to 10 years, it’s going to suddenly hit a hockey stick, and it’s going to suddenly be around, and there’s going to be a lot of it for entrepreneurs. I’m seeing a lot of companies try to do software plays primarily. I think that’s interesting. I think every system generates very different data. The neuroi field is super interesting. I think it seems to be emerging primarily at NeurIPS, the conference in December that’s coming up, that seems to be a hotbed of, you know, convergence of ML, AI, neuroscience, computational neuroscience, which is super cool, foundation model development, that seems to be emerging as a really super interesting academic heart of BCI. Yeah, there’s lots of different entry points to look at, and I just encourage everyone to get involved. It’s going to be a lot of fun. It’s going to be a very fun next decade or two.

Juan Benet

Yeah. Well, thank you very much.

Tom Oxley

No worries. Thanks for chatting about all of this.

Juan Benet

Thank you. I hope you enjoyed this episode. This is a new podcast, so we need your help to get the word out. Please like, rate, and subscribe on your favorite platform and share it with people you think would find it interesting. Thank you. See you next time.

One program is built on the idea that brain disorders are circuit problems, and funds tools to target those circuits with molecular precision across the whole brain. The other aims to deliver high-performance neurotech to the brain non-invasively or at most in a 30-minute outpatient procedure.

We dig into the engineering and biology behind both programs, potential scaling unlocks for the field, how ARIA programs drive breakthroughs, Jacques background, the role of media in shaping the future, and much more. I hope you enjoy the conversation!

Other links to this episode and references below.

Topics covered

• 00:00:00 Introduction

• 00:01:22 Why 20 years of neurotech breakthroughs haven’t reached patients

• 00:04:08 The two variables that determine whether any medical technology gets adopted

• 00:09:17 Brain disorders cost the UK £100B/year and we’re barely treating them

• 00:16:15 Using stem cells and gene therapy to build better brain interfaces

• 00:21:40 Self-regulating gene therapy that helps the brain quiet its own seizures

• 00:24:03 The non-technical reasons transformative neurotech fail to reach patients

• 00:31:34 Watching a 30-second brain ablation stop severe tremors

• 00:38:11 The case for delivering brain implants and therapies without opening the skull

• 00:50:56 Why high technical uncertainty makes distributed teams better than vertical integration

• 01:02:55 Why the UK keeps producing world-class neuroscience but not world-class neurotech companies

• 01:11:04 What AI-driven hypothesis generation means for breakthroughs per pound

• 01:20:40 From quantum computing to improv comedy to running £119M government brain programs

Other Links to the Podcast

• Juan Benet Podcast on YouTube

• Juan Benet Podcast on Spotify

• Juan Benet Podcast on Apple

• Juan Benet Podcast on Amazon

Links from the Podcast

Jacques Carolan

• Website: https://jacquescarolan.github.io/

• On X: https://x.com/jacquescarolan

Jacques’ Programmes at ARIA

• ARIA UK

• ARIA UK on Substack

• Opportunity Space

• Program 1: Precision Neurotechnologies

• Program 2: Massively Scalable Neurotechnologies

Research Papers + Technical References

• Physical principles for scalable neural recording (2013)

• How advances in neural recording affect data analysis (2011)

• Personalized brain circuit scores identify clinically distinct biotypes in depression and anxiety (2024)

• Predictive validity in drug discovery (2022)

• The “Sewing Machine” for minimally invasive neural recording (2019)

• Professor Gabriele Lignani on closed-loop gene therapy for epilepsy

Videos + Demonstrations

• DBS Tremor Surgery Demonstration

• John Nelson on DBS for Depression at ARIA Summit

References Mentioned in Conversation

• Harvard Designated Driver Campaign partnered with more than 160 TV shows, including Cheers and Dallas, helping popularize designated driving before Friends premiered in 1994

• Mahnaz Avarneh — Neurotechnology Is Inequity

Books + Media

• The Hitchhiker’s Guide to the Galaxy — Douglas Adams

• The Expanse — James S.A. Corey

• The Idea Factory — Jon Gertner

• Imagined Worlds — Freeman Dyson

• We Are Legion (We Are Bob) — Dennis E. Taylor

• Pantheon

• Neu World — Ryota Kanai / Araya

• Analogue Quantum Simulation (Jacques’ book)

Links

• Protocol Labs

• PL Neuro

• Disclaimer⁠

Transcript

[Cold open]

Jacques Carolan

The cool thing is that is actually a Purkinje cell. I spent my neuroscience postdoc working on these neurons. Yeah. So it’s this beautiful like massive dendritic tree. They’re in the cerebellum and they like integrate inputs across that entire like dendritic structure. It’s so freaking cool. [laughter]

Juan Benet

My guest today is Jacques. Jacques is the founding program director at ARIA, the UK’s version of ARPA. Prior to joining ARIA as a founding program director, he was a discovery fellow at UCL and a Marie Skłodowska-Curie fellow at MIT. Jacques’s work involves applying the principles of physics and engineering to create next generation scalable tools that aim to radically change our understanding of the brain and ultimately to be used to repair it.

Jacques Carolan

Awesome. Thanks for having me, Juan.

Juan Benet

Welcome. Very excited to chat. Awesome. So, let’s dive in. Part of what you do enables you to have a broad view of where the entire field is going. you see recent breakthroughs, you see the original breakthroughs of the field, you see the bottlenecks. So from your perspective, how the neurotech field developed over time and what were the major breakthroughs that proved that this was a transformational field and what were the bottlenecks that either still exist or we managed to break through recently. So just give us a perspective from your mind.

Jacques Carolan

I think the you know my job as a program director is to kind of come into a space like you said and see where that white space is and I think there have been incredible things in the neurotech space like what

[1:27] Why 20 years of neurotech breakthroughs haven’t reached patients

Jacques Carolan

are the things that you know have been impactful we’ve seen incredible advances in brain computer interfaces you know from the early days in the early 2000s and now to contemporary time where we’re seeing individuals with severe motor impairments severe speech impairments being able to communicate for the first time in absolutely incredible breakthroughs.

We’re also seeing those similar things in neurom modulation, right? Technologies such as deep brain stimulation, DBS, originally used for movement disorders like Parkinson’s. Potentially, [snorts] we can use those same technologies for severe psychiatric conditions, things like treatment resistant depression. These are also breakthroughs in just the past few years.

And for me, I really think we should be quite open with our definition of neurotechnologies. I think there’s incredible work in the cell and gene therapy space as well. You know, we’ve seen incredible work from uniQure with their cure for Huntington’s disease or potential functional cure. So, all of these for me just kind of, you know, in the last 20 years, but really accelerating in the last couple of years just point to the huge potential of interfacing with the brain in a targeted way for reducing human suffering.

So, so given all of that, like what can we do? Like where is the white space? Where’s the interesting things that we can leverage? I think for me we need to take an honest look and see can we get these technologies to the people that need them the most.

So let’s think about deep brain stimulation for Parkinson’s disease. If you look if you do a make a plot of a number of people with Parkinson’s in the US and number of DBS procedures the number of people is probably about.1% that actually receive this DBS. and given kind of you know some kind of eligibility criteria significantly underpenetrated and that is a technology that’s been FDA approved 25 years it’s well-reimbursed we have an idea of who’s going to respond to it and it’s still massively under penetrated and not even increasing at the rate that we want. So my kind of starting point is that if that’s the case for one of our most well- tested neurotechnologies, emerging new technologies for new indications, they’re going to fare even worse.

So the thing that kind of occupies me and really what we’re trying to do in this space is like how do we get these things out into the world? Like what are the levers that we can pull? What is holding people back from adoption if they have been shown to work and shown to increase the quality of life for people? What are the adoption bottlenecks?

I think of it as two axes. On the first axis, you have efficacy. Deep brain stimulation for Parkinson’s. It works well, but it doesn’t cure Parkinson’s. It’s a neuronal loss. There are many other side effects. We have examples of therapeutic interventions

[4:13] The two variables that determine whether any medical technology gets adopted

Jacques Carolan

that you know are serious surgical procedures but get out into the world because they’re slam dunks. I think about like hip and knee surgeries. In the US, I think you do something like a million of these, over a million of these a year, and they’re a pretty involved procedure. But because they’re so good, because they’re so effective, they get out into the world. I think you can drive on the efficacy axis, and on the other axis, there’s something I call procedural burden. How easy are these technologies to deploy?

So, you know, we have examples of technologies such as things like SSRIs with mixed efficacy that still get out into the world. Tens of millions of prescriptions are written every year in the US for these. So I see those as the kind of two complimentary axes and they actually map reasonably well to our two programs that we’re leading.

Juan Benet

As you think about technological breakthroughs, are you thinking of trying to reach similar levels of efficacy or greater by trying new paradigms that maybe make the tech smaller, easier to like make it easier procedures? Or is it taking existing technology stacks and instead of trying to invent something new just see how you can cut those down to get to the kind of easy low burden of procedure or you kind of trying to explore both areas simultaneously or

Jacques Carolan

yeah I think we’re trying to push both sides of the face space so maybe just to begin with talking about this kind of efficacy side so many brain disorders are really disorders of circuits right so what do I mean by that I mean that there are different brain regions involved there are some like really awesome studies where they take individuals with a kind of diagnosis of depression. Many of them put them in fMRI scanners and you can actually see that there are certain circuits that are either hyperconnected or hyperconnected and potentially that could be a new therapeutic target. So different brain regions is critical multiple brain targets and cell type selectivity.

It’s actually kind of awesome having these pictures behind us because there are many different cell types within the brain, whether they’re neuronal cell types, excitatory cell types, inhibitory cell types, non-neuronal cell types. I love these kind of structural images, these EM images for connectomics where you just like do these electron microscopy slices and the brain is just jammed full of neurons and I think people really appreciate that. The neat part is that many of these cell types drive radically different downstream processes.

Something that I think a lot about is there’s a cool example in the striatum. It’s kind of a deep brain structure and there are two types of medium spiny neuron and they look exactly the same. Like morphologically you couldn’t tell the difference and they’ve just got one subtype of dopaminergic neuron receptor that’s different. Okay. If you turn one of them on D1, they’ve got these cool experiments. The mouse turns clockwise and if you turn the other one on D2, they turn anticlockwise. So there’s a radically different downstream pathway. And and these neurons are really really important. So cell type selectivity.

Juan Benet

You know, in recent years, a lot of the world has become much more familiar with artificial neural networks, which are slightly like initially inspired by organic neural networks, but are fundamentally very different in that the mathematical model is fairly simple. you almost apply the exact same structure across the entire network, the same sort of rules. What you’re getting at is like with all these different cell types, you get different behaviors that might be coming from different ways in which the neurons themselves wire together or different ways in which they learn or is not clear at all what the parameter space is actually doing. So curious if you can expand on that a bit.

Jacques Carolan

Yeah, absolutely. I mean these different subtypes have very specific functionalities. Maybe some of them are doing some kind of long range projection. Maybe some like inter neurons are just kind of locally controlling inhibition. Like so much of the brain is this kind of complex dance between excitation and inhibition. I even give you an example like when I was working in neuroscience, I was doing these optogenetics experiments and you like turn on the brain, you know, turn on like a few neurons and you see those things excite and then you see inhibition just dampen it down. So all of these circuit motifs are just critical and we understand lots of them but we’re also discovering new cell types which also kind of blows my mind as well because our technologies are advancing and when it goes wrong that can lead to pathology.

So epilepsy is a great example here. If there is this kind of buildup of excitation or if you shut down inhibition it can lead to seizures. So controlling this excitation inhibition is really critical.

Juan Benet

Maybe to have a bigger context on how significant these disorders are for humanity globally. What what’s the what’s the scale of the impact here? Like how many people are affected by this globally or even you know just in the UK or

Jacques Carolan

Yeah, absolutely. So the numbers I have best are in the UK even just from I’ll start with economic and I’ll change it to people like economic it’s estimated to cost the UK probably over a hundred billion pounds a year. That’s both direct costs. So hospital costs, this is neurological neuropsychiatric conditions and losses of earnings, things like this. Okay, but actually if you just look at the most common conditions, prevalent conditions, you have things like mood and anxiety disorders. You have things like

[9:21] Brain disorders cost the UK £100B/year and we’re barely treating them

Jacques Carolan

addiction, you have things like psychosis, you have things like epilepsy, Alzheimer’s. I think 1% of people have epilepsy, onethird of that is treatment resistant. The numbers are much much larger for psychiatric disorders. So just the disease burden, the impact, we all know people who have experience with these things or we experience them ourselves. So I think that’s the thing that I keep coming back to and the thing that just like man, we need to move quicker.

Juan Benet

Yeah. Yeah. And the cost of the quality of life of so many people around the world is gigantic. curious how much this also lines up in your interest in not just to sort of repair conditions and kind of get back to a baseline or kind of fractional return to a baseline but to potentially just improve quality of life even beyond the baseline like just getting figuring out kind of what leads to not just happiness in one particular domain but you know there’s a wide range of what the human experience is and neurotechnologies can greatly enable people to explore that entire spectrum. I’m curious to what extent like that is a role in how you’re thinking about things.

Jacques Carolan

Yeah, I would say like within ARIA it’s, you know, it is very much focused on this therapeutic endpoint. I think for that longerterm vision, it really does need to go through unhealthy individuals first. That’s how we’re going to generate those large scale data sets that are really going to allow us to be like, well, how do I perturb mood, for example, or maybe drive towards empathy? I don’t know. But at least within ARIA, I think the near term is going to be the therapeutic use case.

Juan Benet

What are some of the maybe recent advances that maybe sparked the possibility space in your mind? Meaning the things that either happened recently that you think, oh, these things need to get translated or need to get enhanced or need to get put together, you know, across different groups or what are some of the bottlenecks that you’re seeing in the in the near term of that you’re specifically trying to trying to target?

Jacques Carolan

Yeah. So I spoke about the kind of importance of circuits and that was really the driver for our first program precision neurotechnologies really the thesis there is we need to increase the precision not just like spatial temporal but the molecular precision the scale across the brain so what made me think this is possible I think there’s huge opportunities by leveraging advances in biology okay we’re funding a number of projects in the biohybrid space can you grow stem cell derived neurons like if my goal is to be able to really increase my you know make sure I can get this part of the circuit I can get the inhibitory part the excitatory part [snorts] cells naturally do that. So what if we could grow stem cell derived neurons and use that as an interface and say hey I want to control dopamine in this brain region or I want to control GABA or I want to control some other kind of neurotransmitter. I’ve always been inspired by that field. There’s been some kind of early demonstrations in it. So we’re actually funding a number of people to look at this.

Can you regrow damaged pathways in the brain? In Parkinson’s, you get the loss of the nigral pathway. What would happen if you could regrow it? Maybe you wouldn’t even need a kind of active batterypowered interface. Maybe it would just go back to a physiological state. We don’t know yet. I think that was like a big motivator.

And maybe the meta point there is just like I really think biology is our new unlock for this. We’re seeing a lot in the gene therapy space. You can now design AAVs which can potentially cross the blood-brain barrier, target specific cell types. Lots of problems that are kind of protein design problems are now much more tractable than they were just a few years ago. I think that’s really promising. Maybe even using things like ultrasound to actually target the different circuits and we’re funding a number of projects around that idea as well.

Juan Benet

How do you structure a program like this? So currently you’re leading this opportunity space at ARIA sub programs within that. what is the structure of a program because probably a lot of people out there are not familiar like what are the how do you think about running a program over time that you’re deploying some amount of capital in it you have technical areas kind of walk us through that structure

Jacques Carolan

yeah so a program this is kind of you know very much inspired by the ARPA model DARPA model there’s been a huge you know rich history of these organizations driving advances in neurotech the kind of model is that technical scientists technologists like myself come into the agency and we get the freedom to build a focused R&D effort of our own vision. Okay. So you want to set some north star some northstar where you’re like if this is true if people can do this we believe the world can change. If I knew how to do it I would just go and start a company and do it. So there needs to be some kind of technical technological uncertainty about the exact solution to it. You want to have some idea that it might be possible. You don’t want to be totally impossible but like right on the edge of possible as the kind of sweet spot. We often think about like how many miracles do you need to have to happen to make this thing work.

What we then do is I say okay this is the north star. and then I put a solicitation out and we picked projects towards realizing that. That can be across industry, it can be across academia, it can be across nonprofits, hospitals, the whole spectrum of things. I think for me the key part is that in a program so each program to give a sense of scale is about50 million pounds to $80 million to what’s like $70 million to $100 million something like that [snorts] some magic should be able to happen in a program that isn’t possible by funding these folks in isolation like there should be some kind of nonlinear things that can happen and basically like my job is to orchestrate that is to make sure we can drive towards that.

So within the program we then set different strands that we think are important and we can go into detail about that within our programs but you know that might be saying someone needs to solve the delivery problem and then someone needs to solve a performance problem. You can break things up into kind of what might be tractable subtypes. Yeah. And then we fund these things and coordinate them. I meet with the teams. If we have if there are things that are really exciting we can double down. If there are things that are not so promising we can pivot. we can terminate projects. That kind of interaction is just critical.

Juan Benet

So let’s dive into program one. you already started to give a sense of it or like some of the things that you’re funding but just for you know how would you describe it as a whole?

Jacques Carolan

Yeah. So really the vision there the insight there is that like I said many brain disorders are really disorders of circuits. Can we build better circuit level technologies? The way we’ve ended up structuring it is mostly around two main technical areas. So the first technical area says, can you just give me these breakthrough technologies? And one of the neat things is when you’re like building a program, there’s always a moment when you’re like, I think this is important. I feel this is important. And you put a solicitation out and you’re like, I don’t know if we’re going to get anything and then you see what you comes in and then you have the opportunity to really shape like where

[16:17] Using stem cells and gene therapy to build better brain interfaces

Jacques Carolan

you see a bit of momentum. So just to give you a sense of that, yeah, I spoke a bit about some of the work in biohybrids. We’ve got three teams working on biohybrid neural interfaces and I think it was kind of cool. They actually on this point of how do we get these nonline interactions the so one of the teams is based in the US and two are based in the UK. The US team all of their team members came over for a kind of biohybrid summit and it I was like sitting there I was like man I think this might be the biggest collection of people working on biohybrids in the world. I was like hell this is so cool. And then you see like you know like one student is like oh we solve things this way and they’re like oh great here like and they were showing each other doing giving each other lab tours it was so cool. So we’ve got efforts on the biohybrid approach. I think that’s like one interesting way to get you know but probably a little bit kind of like later stage to get that interface really circuit specificity.

Another one we’re looking at is like what if we don’t have to touch the brain at all. Can we think about fully non-invasive ways to interface with the brain? So one of the teams that we’re funding is a team called Sonalis and they’re using something called the acoustoelectric effect. The idea there is that you know ultrasound can be focused through the skull. You get some loss but you can focus it through the skull and [snorts] potentially through endogenous ion channels you could also modulate the brain potentially you could also read out through blood Doppler shift. These are kind of indirect proxies for neural activity.

Okay. What if you could actually read and write electrical signals anywhere in the brain? So this is their project and it’s based on this idea called the acoustoelectric effect which says that if you put an acoustic field into the brain, you focus down to the brain at 1 MHz and you then apply an electric field at 1 MHz + 10 Hz. You’ll actually get a beat frequency in the middle, an electrical drive frequency of 10 Hz. the difference. This is just something that happens in ionic media. Kind of wild. And they’re looking if we can scale this up for whole brain. Really, really exciting.

Juan Benet

Yeah. What are the limitations there? Like it’s a super interesting piece of technology. Like fully non-invasive going across the skull. You said scaling it to the whole brain like how much readout can you can you get?

Jacques Carolan

Yeah. The the challenge with any kind of like ultrasound thing is that the skull scatters. So you do get energy through it and you know and the power you get essentially as you increase your frequency the absorption gets larger. So you know and as you increase your frequency your wavelength gets smaller. That’s to say having really precise access to the brain is hard like kind of the physics is hard.

What’s kind of cool with this team is one of the professors so it’s a spin out from Imperial College. One of the professors has a background in geophysics. And in geophysics, they go around with boats and they like drag these large ultrasound transducers around the back of the boats to try and find gas pockets. And you have to solve a bunch of inverse equations to figure out where these gas pockets are. So he was like, can we apply this to the brain? Can we like figure out what the scattering matrix of the brain is, then apply the inverse of that with some kind of helmet system? Yeah, there are a few people looking at this, but I actually think it’s a really tractable problem. Then the question is like how do you do it cheaply? Like do you need a CT scan? Are there more tractable ways? But like I think it’s possible.

Juan Benet

And would this be for read only or could it could you do read write?

Jacques Carolan

You can do right as well. That’s the neat part. So you get the reverse effect which is that if you then apply a kind of like offbeat electric field, you should be able to record an electric signal at the difference. It’s hard and the electronics is hard and the ultrasound is hard but it’s not impossible. That’s kind of what RS is set up to do.

Juan Benet

That would be gigantic, right? Especially if you can target any region of the brain, you could do the deep brain stimulation type.

Jacques Carolan

Yeah. And and actually like I’ve seen more and more now really interesting work in the DBS space that requires mapping first. Pain is a big one. There’s a big pain study maybe a year ago where they initially map using these sEEG leads that are implanted these individuals who have chronic pain. They look where these pain centers are and then can target the modulation based on that.

If there was a way that you didn’t have to go through this implantation process and be in some you know center for two weeks similar for epilepsy mapping if you could do that just through this non-invasive approach potentially could be really huge similar experience to you know getting an MRI type thing and then you end up figuring out exactly how to target the pain center and restore function.

Juan Benet

Yeah. And you can even like in situ like target that thing be like oh yeah that’s the one. And how much of that is read study and then perturb or is it kind of like trying to build integrated systems of read the state and try to kind of sample try perturbing and until you kind of narrow down.

Jacques Carolan

Yeah this is a great question. So we funded this goes back to this kind of technical area structure. The TA1 was really around like let’s build these breakthrough technologies. We have biohybrid approach non-invasive approaches. we have people developing gene therapies or like micro implants things like this and then I thought to myself I was like okay if we have these like so what right and actually I think there’s an interesting point which is like can we demonstrate something that’s not possible in terms of you know reading out from the brain and controlling brain states we then funded in this second technical area a number of teams which have breakthrough technologies but also bring in computational modeling. Really, the goal

[21:41] Self-regulating gene therapy that helps the brain quiet its own seizures

Jacques Carolan

here is to say, let’s take some brain state and let’s see if we can perturb it to some desired potentially physiological away from a pathological state. I give you an example of one of those. There’s a couple of cool ones actually, but one team we’re we’re funding is led by a professor at UCL called Gabriele Lignani. They have a really novel approach which is a closed loop gene therapy. So, you know, I spoke about epilepsy earlier. It’s this buildup of excitation which leads to this seizure. They came up with this technology which said deliver a gene therapy and then the neurons can sense if there is an increase in excitation due to a seizure and if there is they can then upregulate potassium channels which kind of puts the brakes on the neurons.

So it’s like only the neurons that are involved in the seizure and once the seizure stops then they kind of transitely down express and it’s like awesome man this is so wild and it turns out there are lots of conditions that are somehow disorders of excitability schizophrenia dementia a spectrum of things.

So we’re funding them to say, can you do large scale modeling and figure out, you know, to figure out where to deliver these gene therapies to and then use ultrasound to actually target those gene therapies. And if you can do that, is it possible to actually fully restore you back to a pathological state? I don’t know if it’s possible, but I think it’s really really fascinating.

Juan Benet

And you also have a third technical area in program one around patient and stakeholder engagement. Can you talk through that?

Jacques Carolan

So maybe I can just jump into one of the TA3 projects that we’re funding. so we’re funding a someone at the University of Sheffield to look at inequities in neurotechnologies. So if you just look at the kind of like look at DBS implant Parkinson’s it’s almost all exclusively white.

So this incredible researcher Mahnaz Avarneh at the University of Sheffield she’s working with communities that typically don’t get these types of implants and trying to understand like what do they know about neurotechnologies like how like what would be the limits of them want to do something really towards just kind of addressing the disparities and inequities in neurotechnology like we’re trying to build things that everyone can access and really understanding this is

[24:03] The non-technical reasons transformative neurotech fail to reach patients

Jacques Carolan

going to be critical. She actually put together a really cool video. So if you like search I think you can find it through maybe the ARIA website as well but if you search Mahnaz Avarneh University of Sheffield neurotechnologies inequity like you’ll find that great and we’ll put the link on the description.

Juan Benet

Awesome. I think it’s a really key thing in the adoption of any technology is truly understanding how people perceive it, how they think about the cost benefit, what are the blockers. sometimes it might be slight tweaks that maybe the technologists never thought about that could dramatically increase the adoption rate.

one of the things that has been on my mind recently is the invisibility of devices. So when a device is visible, it tends to cause some amount of social stigma because you look different and you seem like you’re carrying something. And you can think of a range of technologies once they got to the invisible part of the spectrum, the adoption went through the roof.

A good example of this might be braces, right? So initially think about the old braces, not just like the on the teeth ones, but the ones that had like the wire going around and just like a like a very difficult experience that you’re putting somebody through to carry around this type of device and compare that to something like Invisalign where suddenly it’s entirely invisible or close to invisible and suddenly the people’s experience of it changes dramatically.

So yeah, curious what if you’re seeing similar things in these kinds of studies or what other vectors are you thinking about exploring?

Jacques Carolan

Yeah. Well, I mean like I think that’s a great example, but you take that and then you convolve it with a stigma that is already associated with neuropsychiatric conditions, right? Like and in certain communities that there is a huge amount of stigma having a psychiatric condition and then you have something that can like demonstrate that.

So yeah, we’re like we’re definitely seeing you know there’s some of our teams are looking at non-invasive approaches and understanding adoption there actually. Yeah, like shaving your hair is a really big deal for a lot of people and it also breaks down over different like you know if you’re a young girl you definitely don’t want to shave your hair.

The TLDDR like is from what you know I’m really not an expert in this but the thing that comes up all the time is that it’s just complicated.

Juan Benet

Yeah. And different people and different communities will have different things they care about also intersecting with different levels of severity for that indication as well. So it’s like yeah it seems like a super valuable data set for the entire field. Are you guys sort of collecting that and going to publish the whole thing or

Jacques Carolan

I was you know I was actually so surprised these things don’t exist like you know of course the field of like lived experience engagement and all of these things it’s a field and there are professionals that do this but especially when I was building this new program I was like what for something like Parkinson’s people been looking at cell therapies people looking at gene therapies people looking at DBS I couldn’t find any data sets which did large scale patient engagement across these variety technologies which is kind of wild.

So I think within our everything we produce along these lines in our TA3 is absolutely being shared publicly

Juan Benet

and is that looking primarily at the UK or you looking broader

Jacques Carolan

or most of these studies are in the UK

Juan Benet

and perhaps you can give a blueprint to of how exactly you conducted it. So then you know people elsewhere can maybe broaden it to other countries.

Jacques Carolan

Yeah.

Juan Benet

So I would imagine that there might be significant differences in you know across the world.

Jacques Carolan

Yeah. So we actually we did it in kind of a unique way which was like my instinct was that like this is an area that’s important and there are other people who are better experts than me at driving this forward. So we put a really broad call out for this just said like we’re interested in these areas like give us the best things that you have and got some really incredible efforts come through that.

So I don’t know what the mission I think that’s a unique thing that a government funer can do in this space is kind of set that and there are like you know there are organizations like the James Lind Alliance that does priority setting for different types of technologies and different types of like medical conditions. So one of the people funding is looking at priority setting for brain computer interfaces for people with motor neuron disease.

So I don’t necessarily have a good answer to your question, but I think there are people there that are experts in this and it requires some kind of forcing function to drive them in some right direction.

Juan Benet

And I also saw you include media as part of the technical area. What are you thinking there? Is that kind of is that kind of like a production of media that helps explain things to patients or what’s sort of the range there?

Jacques Carolan

Yeah, so this is and this is for our second program. So this one isn’t out yet. on the media side, it was really the observation that you know there are so many dystopian stories about neurotechnologies. and we’re in a unique moment, I think, where there are individuals who have neurotechnologies of one form of another. It just feels that there is a need to kind of get their stories out into the world.

In the UK, there’s a bunch of awesome podcasts around Parkinson’s. There is there’s like loads of people making podcasts. That’s really cool. Can we actually, you know, fund these folks or communities that work with them to create YouTube videos which kind of just show the realities of their conditions and also the technologies themselves. Like there’s probably great parts and there’s probably annoying parts and just get those stories out into the world.

So it might be videos, it might be podcast, we want to access like the influencers. I just think there’s this kind of space that I think would just be awesome.

Juan Benet

How much of it might be also normalizing it in mass media? There’s this story, I don’t know if I haven’t verified it whether it’s true, but there’s a story that I think the Friends show included a reference to drunk driving and having like a designated driver instead. And like they just threw a small comment of just making it a point that whenever they were, you know, going out drinking somewhere or whatever to just always like reference the DD. And that just that light introduction, the story goes that became a huge vector for mass adoption.

Now, I don’t know if this is exactly true or not, but I would imagine that portrayals in mass media can be super impactful to how people perceive this kind of thing broadly. I’m curious if you know to what extent it’s like reaching out to these kinds of video production or these groups to kind of include help educate those groups on like the importance of highlighting these kinds of things.

Jacques Carolan

Yeah, I think that’s really fascinating. I think there’s actually there’s a couple of examples and I’m sure there’s examples in the US but examples I know of in the UK of like famous newscasters that got Parkinson’s and got a DBS and now like and now become you know inadvertently an advocate for that.

So I think like it’s really interesting to think about how do yeah how do we leverage that to kind of normalize these types my kind of takeaway is like it’s almost like a phase transition. It’s not like one lever that’s like okay hell yeah this is going to drive it. It’s going to be, you know, both understanding the needs and the kind of boundaries of that. It’s going to be sharing stories. It’s going to be the whole spectrum of things. And then slowly we see that and I think that’s the unique thing that we can kind of do at ARIA.

Juan Benet

Yeah. And let’s maybe jump into the second program is for massively scalable neuro technologies. What is the northstar for this second program and you know kind of what was it kind of open space that you weren’t uncovering in program one that triggered you to kind of shape this new one?

Jacques Carolan

you know when I spoke about that kind of that axis that you know this like this one axis we have like efficacy and then the other a procedural burden so what do I mean by that implanting something chronically into someone’s head is a really serious thing it sounds trit to say it maybe as a footnote here

I went to go and see a couple of DBS procedures be done at the end of last year the first procedure I saw be done was called RF ablation so it’s kind of an old technology so it was a person who had very severe tremor like an essential tremor.

What you do is you go into the you put a probe into the thalamus and then you kind of do a very small

[31:39] Watching a 30-second brain ablation stop severe tremors

Jacques Carolan

ablation to kind of disrupt that circuitry. And it’s one of these remarkable things because they do it while the person’s awake. You ask the person in the surgical theater to draw a spiral or write their name and they can’t do it just because their tremor is so significant. They kind of they go in, you know, they plan this trajectory. They plan they want to go. they go into the thalamus and they like press a button for like 30 seconds and then they say, “Okay, can you draw a spiral?” And it’s just a perfect spiral.

Juan Benet

Wow.

Jacques Carolan

And you know it’s and this person is like absolutely stoked. It’s like it’s remarkable. It’s like the closest to magic that I’ve actually seen. So I feel very honored that I got to experience that.

Juan Benet

That’s amazing.

Jacques Carolan

Yeah, it’s wild. And so that was one procedure I saw be done. And I saw another procedure which was a bilateral DBS implant. So you do it in both hemispheres. And of course the person’s under anesthesia and it was just so much clearer that was a really serious surgical procedure from like a sterility point of view from like everything. [snorts] Implanting something in someone is like is really serious. Unpacking that what do I mean? I mean that there is singledigit percentage of you know likelihood of a bleed of infection. You know these things are real. What does that mean? That means that drives patient hesitancy. that potentially drive clinician hesitancy for prescribing these types of things. And then I think about the barriers of just like where these things can be done.

Yeah, in the UK we have you can’t get the number. It’s just kind of hard, but we probably have a few dozen functional neurosurgeons at consultant level that can lead these procedures. You know, changing that is like a generational time constant. Like it’s a bit better in the US, but it’s not a ton better. And just where you can do these things as well. you know basically it always requires some like large scale urban medical center where you know it’s expensive to get to and tend to live nearby and these things so I think it might be similar if it’s a brain computer interface I think it might be similar if it’s a cell and gene therapy like basically have to do the same things

that was really the starting point so I became fascinated in like what are transition points where you know we can like radically increase the scale of technologies and there’s some really cool stuff looking at like early pacemakers if you had like a heart block bas have to do like a massive kind of thoracotomy put leads on your on your on your heart and the transvenous lead essentially revolutionized that you know so you kind of go in through the veins and you radically changed the surgical procedure so you went from a kind of you know maybe [clears throat] half a dozen of these done a year to now millions of those things

Juan Benet

yeah and that’s with the stent right like you bring in a stent and

Jacques Carolan

Exactly exactly and it’s actually kind of a cool case study as well because initially these things kind of sucked like leads broke, you know, they were using like nuclear batteries, like all of these things, but it created this like innovation ecosystem just because it was so radically more scalable and that’s ultimately what drove it.

So I was looking at this and the kind of proto pacemakers were around the kind of like 1950s and in 1947 was the invention of the transistor. kind of cool like on my eyes like what is the new technologies and breakthroughs that will could drive mass adoption of technologies like what is the transistor of our modern age and I’m going to go back to something I said earlier like I really do think it’s an engineered biology like we can engineer you know vectors whether that’s cells whether that’s AAVs we can combine that with like hardware like all of these things can access the body and they can be programmed like I think there’s so much there so That was kind of a long answer. That was really the nugget by which I was like, I think there’s something here to radically increase the scale of neurotechnologies.

Juan Benet

How are you defining or thinking about what makes a neurotechnology scalable? So, you referenced the operating room and the difficulties there. You referenced potentially making it minimally invasive. So, potentially going through other access media like what is the full spectrum here that you’re looking at?

Jacques Carolan

This question of scalability is super interesting and you ask 10 people what scalable means and you’ll get like 11 different answers. So my kind of thought process is like what are the like technological leverage points which we can you know have a plausible argument that will potentially lead to scalability. Where I landed on this was this idea of you know can we avoid the craniotomy altogether and for many of these things you require you know to remove part of skull implants from it chronically can we access the brain and deploy high performance neurotechnologies without ever requiring that so [snorts] can we leverage blood vessels the vasculature can we think about leveraging the CSF can we think about like other access points we haven’t even considered

Juan Benet

what’s the CSF

Jacques Carolan

oh the cerebral fluid so you can get that through a lumbar puncture It actually goes all the way to the ventricles in the center of the brain. It actually goes all the way onto the surface of the brain. It’s like kind of an interesting space to access things. Can we think about the nasal passage? Olfactory neurons actually have processes that sit in the nasal passage. So, if you can get things to there, potentially they can kind of pass into the brain. There are many challenges with it. It’s really really hard. Our nose tries to keep things out generally. but does that kind of white space maybe be like I think there’s something here.

Juan Benet

Yeah. Yeah. And and what is the spectrum of devices that you imagine here? Is it you know primarily devices that would become implantation or it’s intersection between biology and engineering? how much of it do you expect to be just purely either chemicals or gene therapies or things like that?

Jacques Carolan

Yeah. So, and maybe I should also just like add in the north star to this program is really to say can we get high performance neuron technology to the brain in a 30-minute procedure without transcranial surgery really with the goal of like ultimately being able to do this in an outpatient procedure and people were like like why 30 minutes actually 30 minutes is kind of it’s pretty consistent with both interventional procedures and cell infusion procedures and like the amount of people who are like well okay like no craniotomy can I get like a small hole and I was like well actually like it’s important that we somehow constrain it because I do believe that like constraining things drives radical innovation.

Juan Benet

Yes.

Jacques Carolan

So yeah, your question was like do you imagine this is like devices or biologics or I imagine it’s the whole spectrum of things. We’ve only just recently closed for concept papers and the full proposal is going to open in a couple of weeks depending on when this airs. What I can say so far is like at least what inspired me that this might be true like we are seeing some interesting work on the devices. Of course companies like Synchron are doing incredible work there. How can we think about accessing deep brain targets? You

[38:12] The case for delivering brain implants and therapies without opening the skull

Jacques Carolan

know, many of the things that we care about in neuropsychiatry and neurology are in these deep brain areas. How do we access that? We think about yeah, new devices that might span different pathways to do that. On the biologic side, one area that I’m really fascinated is in immune cell engineering. So, immune cells naturally cross the blood-brain barrier during regions of inflammation. So, we’re seeing advances in synthetic biology. Could you imagine applying synbio to these things such that they can track to the brain, monitor pathology, secrete biomarkers out into the blood like potentially or even dropping off cargo like these types of things. there are early signals that it might be possible and I think it’s kind of a moonshot that’s maybe worth taking

Juan Benet

and do you imagine that being yeah the full the full gamut of possible sets of technologies of trying to kind of reach the same level of efficacy as the as the current approaches or are you also enabling I don’t know the exploration of like new types of targets or what

Jacques Carolan

I the way I set this up was we want the teams to target brain regions or brain targets where there either is like validated evidence You know, so there’s a bunch of those things in DBS, whether it’s subthalamic nucleus, whether it’s subcallosal cingulate for depression. Either it’s those sets of targets or targets where there is like a plausible kind of clinical path. And and why is that important? It’s important because like you’re solving a really hard engineering problem. If you’re solving this, you want to like at least know it’s going to be useful somewhere. That’s not to say like I think it’s really really important to find new therapeutic targets. these are just the ones that we’ve accessed and the ones that we’ve kind of validated some degree. Like there’s a whole space of finding new ones and that’s actually a bit more along the first program idea. But what we said is like let’s just hit targets that we know about.

Juan Benet

And in the in the three technical areas of this program you have like technical area one delivery and performance, tech area 2 prototyping and translation and technical area 3 adoption. Can you maybe walk us through each of these?

Jacques Carolan

Yeah. So technical area one and actually like going back to the state that’s that’s the core technology development part and my kind of thesis is that we really want to be delivery first like there have been a bunch of cool technologies for mice like and like it works really well in mice and then like it never sees the light of day like we’re serious about getting things out into the world. Okay. So I really believe that like delivery in this program is the gating factor for scale. So in the first technical area, we’ve split this up into two areas, two phases. The first phase is, can you show me you can solve this delivery problem, right? You can get to the right brain region. You could do it in a way that has the potential to fit in this 30-minute potentially outpatient procedure and it can be done safely or at least we have a handle on what the safety profile is. Even if it’s not exactly there yet, if you can’t do that, I’m not interested. like we need to get technology into the world and then based on that we can then advance the performance. So the way the way we’ve structured it is that we have this initial three-year drive where we fund a broader number of projects you know it might be across like really radically different technology streams biologics hardware nanoparticles who knows and then based on that we can then down select the things that we’re looking most promising and really think about that translation path. So that’s technical area one.

And maybe just another thing I’ll say on that as well is like you know if you speak to anyone in some kind of like ARPA agency you know they love metrics like everyone like loves a metric and like they’re valuable and sometimes you can overdo it and so what we’ve said for this program is that actually we’re indication agnostic like I want teams to select the indication that they care most about that they think they can drive the biggest you know change whether it’s in neuropsychiatry whether it’s in epilepsy they then set the metrics like what is the readout bandwidth you need to monitor Alzheimer’s. It’s probably different from the readout bandwidth you need for a motorbi, right? Like teams then set that themselves. Okay, so that’s technical area one.

Technical area two is kind of it’s a bit of an odd thing, but it’s basically looking for a series of teams to support our TA1 teams. Okay, we’ve actually seen a pretty interesting use case of that in the first program. So through a program we’ve got called activation partners I can talk a bit more about. We’re working with an organization called Emodo Design. They’re essentially an engineering design house over in Sheffield in the north of England and they can and they make like incredible hardware. They make incredible ASICs. They’re like super smart people. They work with our teams and they say you know this like widget that a grad student spends like a year and a half building like what if we could do that? what if we could give you that in 3 weeks and it’s going to work 10x better like and that mindset shift is just a really fascinating like kind of opener for people. So what we’re going to do is once we see the teams that we end up funding in TA1 we’re going to put a separate call out for these supporting teams. So it might be you know hardware engineering it might be high throughput protein design maybe in phase two where we’re thinking about regulation can we think about regulatory support can we think about organizations that can do lived experience engagement to understand what like like product profiles need to look like just that whole like kind of infrastructure around there

and then finally in TA3 it’s really thinking about you know how do we tell the stories of individuals who are living with these conditions who are maybe even using these technologies there is just a gap in my for not even like positive stories of neurotechnologies but just like realistic stories of neurotechnologies. So I really hope we can somewhat

Juan Benet

how do you think about robotics? So one way to address the surgery question which you know neural link has been pushing one of the first things they focused on was trying to decrease the cost structure looking at maybe like the eye surgeries like once they became robotic in nature then they could become like the 30-minute outpatient procedure whereas before you know it’d be incredibly difficult to try and produce that kind of change. How do you see robotics as a as an angle here? Like do you think that could be a vector to get to like the 30 minutes or do you think no specifically you’re trying to create a different innovation space where you think groups like Neuralink are already going to cover that and you instead want to try and find other things?

Jacques Carolan

This is an awesome question and I spent so long thinking about it. Yeah. And as a footnote, you know, the kind of like proto sewing machine came out of like a DARPA which is kind of cool. Anyway, actually that’s the story there.

Juan Benet

That’s fascinating.

Jacques Carolan

Yeah. So I don’t know the exact story behind it but it came out of a postdoc’s name who I’m going to forget now. It might be Tim Hansen. So had this idea implanting electrodes automatically. Flip’s lab which I think was at UCSF was like a neuroscience lab. It’s like an electrical lab had no reason like no business doing like fab or doing like deposition like totally bonkers and you know like Flip’s incredible person. He’s like yeah sure let’s try it. They got some money from DARPA and they kind of were doing these things in electrophysiology lab. That’s the story that I’ve heard. I just think that’s insanely cool.

Juan Benet

Good taste from a PI goes a long way.

Jacques Carolan

Yeah. And also been able to like support an incredible postdoc who’s got an idea kind of out there just like man go for it.

Yeah. I thought a lot about this robotic side like absolutely it’d be great if it can be automated. Where I landed on this is I actually think there’s a lot of efforts in this already and there’s a lot of capital deployed towards like some of these like medical robotics companies have like billions of dollars and are just about starting to turn profit. Like my sense was that one probably the amount we could put in there would just be a rounding error. Like I couldn’t really see the differentiated thing. Like we’re interested in things that are like catalytic and nonlinear. I couldn’t quite see that in the robotic space. like it’s cool and if it’s going to happen anyway, it can just plug into this stuff. So it’s like a catheter procedure like you know there’s a field of people developing machines for mechanical thrombectomies for stroke to remove these clots like if that’s already happening that can probably apply to devices. So given all of that I just didn’t quite see it.

Juan Benet

So you know as you do all of this and you just mentioned the process of coordinating all these groups and the way that you’ve designed and shaped the programs includes thinking about who are the labs that are bringing the kind of innovative new technology or new science and who are the groups that are going to be able to play important roles in building components of that. You know, maybe it’s like fabbing a specific device or maybe it’s solving some concrete problem or doing some kind of mapping or bringing in algorithms and AI expertise to be able to you’re trying to coordinate incredibly talented people from across many different organizations very much in the you know age-old tradition of government agencies like ARPA and others which you know at the time in you know in the ‘ 50s60s which was a very different time they managed to produce a very fast pace of engineering or very fast pace of R&D. Today we’re seeing tremendous successes in vertically integrating systems. Groups like Apple and then now Elon and others have shown that like if you try and vertically integrate the entire process, you can go usually way faster than the rest of the market. There’s something there’s some breakdown in the coordination structure of the broad contracting structure where time delays sneak in through all of the negotiations leading to, you know, this super high cost measured in time.

And so you know Elon has gone really far in the last 10 20 years by saying just no to the you know conventional wisdom of decoupling pieces and letting the market sort of solve the problem and instead just solving it yourself. and that is a huge counterpoint to how you know ARPA in the and even the precursors to ARPA you during the war efforts built so much of the of the deep R&D required to build incredible programs including all of the machinery like all the tanks and so on. then later all of those insights were set up for the Apollo program. They also influenced how the internet itself was built. There was so much R&D that was driven this way and it was actually quite fast by modern standards maybe especially by modern standards given how long everything takes today.

But and so you’re trying you’re trying this older approach and coordinating a lot of talent in a time period when other groups are having massive success by taking the counterpoint. where are the parts where like it works really well or have you seen these delays or no actually you disagree with that point of view and instead you actually have seen measurably faster approaches this way. Yeah. I just want to get a sense of how you think about this entire

Jacques Carolan

I think it’s like the first thing is it depends at what level of like technical uncertainty you’re at like when you’re ready for takeoff and to like vertically integrate you know you’ve kind of reduced the error bars on your technical directions. So I generally think like where we’re sitting with these programs is we’re kind of earlier in that pipeline where there is still technical uncertainty. Now a win for us is if we can show like one of these technologies is like a goer is massively valuable then the kind of infrastructure whether it’s through venture whether it’s through un like non-dilutive funding whatever is there to support that and allow it to take off.

My sense is that like in neuroscience at least things that we’re thinking about it doesn’t quite make sense integration precisely because we have this high levels of technical uncertainty then the question is like you know I spoke about this kind of nonlinear magic like we still do want to get to that so how can we drive that within our program towards accelerating the technologies we’re developing and I think that happens in a few ways the first is actually just kind of by like serendipity so you want to increase the surface area of interactions between the various teams such that they’re like, “Oh, you’ve got a problem to this. You’ve got a solution to this problem.” We had a workshop a few weeks ago, which is the annual meeting for our precision neurotechnologies program. Someone was like working on some kind of human tissue samples and like someone else was like, “We really need access to human tissue.” But like, I can help you out. They were actually already collaborating on a separate project, but didn’t know they had these capabilities. So that’s like serendipity and we can kind of you know I think about it as like a thermodynamically we can like increase the temperature and do that

and then there’s the other part which is like are there core technologies that are being developed across these various programs. So for example, the computational modeling side, there are a number of really incredible computational neuroscience teams that are all trying to do a similar thing. Trying to take some highdimensional data sets, trying to do some kind of dimensionality, correlate it with something you care about, and then perturb the brain towards a new state. We want to get those folks to work together to be able to like help accelerate each other. So we can kind of drive that as a fun.

And then there’s the final part which is like resources like hardware. So I’ll go back to you know this team that we’re funding a modo design they can almost be you know

[51:02] Why high technical uncertainty makes distributed teams better than vertical integration

Jacques Carolan

that like play a little bit of a role of a vertical integrator they can work across many things they have in-house mechanical design firmware design as design like they can do those things and do it quickly and do it well

Juan Benet

earlier you were mentioning bringing together a set of people and then finding you know the serendipitous insight that you know one group could help another go faster and but this was only found kind of at a workshop isn’t that also an argument just for putting all of those teams within one organization and trying to like go as fast as possible kind of like in the Elon style or not just Elon this maybe he’s the most visible version of this recently but you know many other organizations apply the same thinking of like just get all the super talented folks put them under one roof don’t do it remotely actually force everybody to move into the same facility you know this maybe even has Bell Labs hallway errors right like if you get and concentrate all of these groups into one place you can increase the collisions and get to the innovation rate faster Maybe because of the technical risk in that like when you’re contracting outside you maybe want to have things more definite working across many organizations.

Yeah. Yeah, I don’t know like how do you think about that versus maybe the reality is also could be around hey these talented people are already in a range of organizations and they’re kind of stuck in that environment today and the prospect of pulling them out of those environments you know into a separate single organization is just so energy and capital intensive that unless you have billions of dollars you can’t quite do that and so therefore like you’re you know stuck having to coordinate them in this kind of more remote environment. Yeah, curious where you land.

Jacques Carolan

You are completely right. Serendipity is highly inefficient. It’s the kind of like bane of my life. And and I actually think I think you’re totally spot on. I think what it comes down to is incentive and incentive alignment. Like a startup has venture investment. They’ve got things they need to do and they’re aligned towards doing that. Like academics need to produce papers and need to graduate and need to get tenure. Like hospitals have things hospital things they need to do and like how do you coordinate across that?

Like I think the thing I would push back on is like I still think we’re not at a point yet where we can like there is the need like because there is still so much technical uncertainty about like what is the pathway going to be that we’re ready to like stick these folks in a single place and be like hell yeah let’s do it. I love Bell Labs but I think we have a lot of nostalgia for Bell Labs as well and I think the kind of modern version of that will probably look a bit different.

Juan Benet

Yeah. It strikes me often that a lot of what makes certain companies work so well and drive so such fast R&D it’s more softer type changes things about the culture things about what meetings you actually have you know how much you protect individual contributor time to not have meetings and instead work however you provide crucial fast highly iterative cycles where they have to present results and so there’s like a lot of pressure that forces the whole team to advance quite quickly you know week to week to week you know getting up in front of the whole company every week and you have to have some results. So, I don’t know. I’m curious if you think about that sort of softer culture oriented organizational structures.

Jacques Carolan

Yeah, I think about it a lot. It’s and there’s kind of a few different examples and I would say this is like probably less towards getting you know just to be clear like we fund a bunch of what we call projects typically those might be a few different research groups or different companies working together. What I’m going to describe is less about finding connectivity between those folks but more about like driving innovation. I think first I think mission is just so important and just like driving home the reason we’re all doing this.

At my recent meeting we were working with a guy called John Nelson. John’s kind of like lived experience like advocate like a real activist incredible storyteller. He actually came to the summit. So if you search ARIA Summit John Nelson, there’s a beautiful talk that he gave and he’s someone who’s lived with treatment resistant depression for many years and he went through this kind of experimental procedure with the clinician Dr. Helen Mayberg to implant a DBS into his brain and the way he talks about it is that it saved his life.

When I see that like it’s I find it so so powerful and it’s like we need to move even quicker. And so we actually got him to come to our annual meeting and run a session around lived experience and brought in various folks. You know, someone with epilepsy, someone who was a carer for someone with Alzheimer’s, someone with Parkinson’s just for them to talk about their experience and like you know, you go to these workshops and everyone’s like on the phone or like laptop or doing stuff like just you could silence. You’d hear a pin drop and like there you know, there were lots of tears.

Why is this important? one, it does light a fire underneath you and it lights a fire under technical teams who actually very rarely interact with these folks. It helps you build better things. We’ve spoken about this before like if you’re trying to like build something for someone like you want it to be good and you want it to be usable and you want it to like people to buy it. So you chat with people that are going to buy it like you do that if you’re building any product, right? So I just think it’s I think that like in terms of driving it that’s just so important.

So what I really try and do within my programs is like create a sense of a team. You know, yeah, we’re distributed around the UK. We’re distributed around the world, but we are a team that’s like trying to solve this problem. And I genuinely believe we might have the most ambitious set of neurotechnologists in one place. In that there is some magic and just really try and impart that across them.

And then you’ve got the kind of maybe like less glamorous part of this which is teams have to report every quarter. So we get technical reports every quarter. We meet with them. They have really concrete milestones they need to hit. If they’re not hitting with them, then we work with them to try and debug it. So we are really really deep in the technical details and they’re presenting to us and every year they’re presenting to the team. So there is that pressure and that’s kind of critical to the model. So it’s a real balancing act to say, you know, it’s ambitious. We’re asking you to do really hard things in short periods of time, but it’s also supportive. So that’s the needle that we try and thread.

Juan Benet

What do you think is really special and unique about this kind of coordination approach that just can’t be done in more traditional single organization, single company setting kind of you already spoke a bit about finding a set of groups and orienting them towards this you know finding deep expertise having a wider kind of field of experimentation where maybe you couldn’t run that many programs in a single company because you have to the focus will drive narrowing of perspective but here you maybe try a lot of different bets. What are some other organizational cultural insights that you’re finding interesting or valuable about this approach that you know as a counterpoint to other structures?

Jacques Carolan

I mean we’re not driven by profit which is also interesting like I think does play an important role. It’s like the most important thing is that we have breakthroughs that we plausibly believe can change society. And it’s actually interesting as a footnote when we fund teams like fun like startups or even established kind of industries. They’ve often got the thing they need to work on because they got venture investment, they’ve got a board, they’ve got things like this and then they’ve got the thing they really want to work on. And actually just like a few million pounds or dollars can actually unlock that bigger which is kind of cool.

So we are looking for the tail of the distribution. You know if everything that I fund is wildly successful then I probably haven’t been ambitious enough which is kind of an interesting point to be in which might you’d have to have a very special type of company that can leverage things like that and can really drive with things.

So I think that combined with being really deep in the technical details and ultimately we’re not committee based like you know I work with experts to get insight into different areas. if I think that something is promising and everyone you know there’s a big group of people who are like tried it in the 80s and it never worked and I’m like I think there’s a reason it might be able to be solved now like we can take those contrarian point of views that is not a given for a government organization and I feel very fortunate that ARIA has been set up in such a way that we really do have that embedded into our DNA we really do have those levels of protections that level of ambition like it’s really in there and I at least hope that it kind of filters out to the teams.

The reality is like we’ve got some awesome programs across many different areas. Like we need to see breakthroughs on a decade horizon. Like that’s what we need to do.

Juan Benet

Can you speak a bit about the innovation ecosystem in the UK? Like one of the most significant breakthroughs on the entire planet in the in recent decades has been DeepMind and the UK had the talent insight maybe contrary and perspective at the time where Demis Hassabis and others were able to build an extraordinarily talented team there and grew DeepMind there and then you know that eventually joined up with Alphabet and so on the whole there’s a whole other story there but that to me shows an example of being way ahead of where Silicon Valley was on AI at a particular moment in time and you know historically there’s a tremendously talented scientists and engineers and so on all over the UK and you know in the complex of universities around London and so on but perhaps on the startup technology company side the UK has not been as strong counterpoint of DeepMind has not been as strong in recent you know decade or decade or two how are you seeing the ecosystem now how are you trying to maybe change that or kind of boost the technology translation to yield a better set of outcomes or kind of stimulate the growth of more neuro techch companies locally or how are you thinking about all this?

Jacques Carolan

Yeah, like you’re absolutely right and in kind of in you know of course the DeepMind but in like the neuros there’s been a bunch of like half a dozen Nobel prizes for neuroscience kind of UK or UK associated things right like neurosurgery was kind of incepted in Queen Square in London there’s incredible neuroscience generally there’s great engineering we have these like core components and when I started ARIA you know I like went all around the UK the thing that struck me the most is actually like is this like entrepreneurial appetite like there are so many young folks who like don’t want to go into academia they don’t want to go into like established industry they want to like build something wild and I actually wasn’t expecting that and just that like level of like appetite and enthusiasm

I’ll just like give you an example Imperial College London has a student society student run student organized everything and they’ve got a WhatsApp group and there’s like a thousand people on this WhatsApp group like How cool is that? And it’s kind of like a small example, but I really do believe there is this kind of like undercurrent of people wanting to do things.

Now, I see ARIA’s role as kind of like activating that talent base. There’s a kind of like slightly handwavy version of this and there’s a kind of a concrete version of this. like the handwavy version is that like I do believe that ARIA has given people to permission to dream big you know to really like think about doing a moonshot in a kind of like a non-traditional environment that’s kind of handwavy but like I get a sense of that the second part is like what actually can we do to drive that broader ecosystem and very early on we realized that was going to be an important thing in ARIA how do we actually get you know our technology into the world like can Can we build new talent pipelines? Can we have new capital in flows? Can we build new communities, new capabilities?

So, we essentially launched a very broad call for what we called activation partners. These are like organizations that we just had a feeling were like going to be important in kind of increasing momentum in our spaces. There is a bunch of different folks that are kind of towards this like activating that community.

to give a couple examples, you know, so we’re working with Convergent Research, you know, like for coming they’re developing FRO’s. they launched their first ever residency program in the UK. We’re actually

[1:02:57] Why the UK keeps producing world-class neuroscience but not world-class neurotech companies

Jacques Carolan

funding one like one FRO around in vivo connectomics. Really really cool idea and like wouldn’t have happened otherwise. Like this is just awesome. In terms of like other talent pieces, we’re working with a number of deep tech VCs based in the US. Pillar VC, 50 Years to actually run their programs in the UK taking found taking scientists turn them into founders taking top AI talent put them in dropping them in labs. The amount of people that apply to this are as much if not more than in the US which is like kind of cool. So I think there is just this like strong strong appetite like I give a couple of examples specifically in the space which is we’re also partnering with University of Cambridge and they’re setting up a neurotechnology accelerator now. So you can come on it’s a year-long residency the whole thing is paid for. You can run you know build prototypes you can chat with clinicians lived experience the whole thing towards at the end trying to raise some money. It’s still an experiment. It’s still to be seen if this is going to be the thing that drives it.

Yeah. Like I go back to this image again like it kind of is like a phase transition. We push on our technologies and of course we can reach all of our technical milestones and they don’t get out into the world. We can push on talent. We push on community. Another thing I didn’t mention is on the regulation side. One of the challenges in the UK is that neurotechnologies are still pretty new and just the kind of guidelines aren’t there. like you chat with folks and they’re kind of confused about how they submit things like this. But parts of the UK regulatory system is really innovative. Like we had the first ever CRISPR drug approved anywhere in the world. Kind of wild. So what we’re actually doing is partnering with the University of Newcastle. We’re actually seconding some people into the MHRA which are our regulators like the FDA in the UK to develop that guidelines, you know, and it’s like a small lever but potentially could be really impactful.

So yeah, it’s just I think that could be gigantic. So from my perspective, the kind of the regulatory blockers to a lot of this is where so many ideas end up dying or where the prospect looks so bleak that groups don’t even try. and so I think if you can get a much better landscape for neurotechnology in the UK and make it a just a dramatically better place than other places to advanced technology companies, that could be a huge edge in being able to reach patients dramatically faster, save way more lives, get these breakthroughs into the world.

Yeah, I give an example as well. You know, one of the things we’re funding through the program is a collaboration between Forest Neurotech, you know, nonprofit FRO developing ultrasound systems for read and write a the University of Plymouth and a hospital in London called Queens Hospital. The ultimate vision of this project is can you use ultrasound to both read out brain states associated with mood and perturb the brain to more physiological states. I think just last week they did a clinical study with about six participants in London using this like in my eyes like state-of-the-art brain computer interface and as a footnote it’s kind of neat is because like explaining the challenge of like the skull scattering ultrasound this surgeon works with people who have had their skull removed if you have traumatic brain injury you can do these large craniectomies to relieve the pressure and it means you basically walk around without bits of skull for like months at a time. So he recruits them into this clinical study and you know they’ve recruited a really large amount of people. So that’s exactly to your point like I actually think there are some like levers which can be really impactful both in regulation and clinical trial.

Give you another example. So we’re funding a collaboration between a neurotech manufacturer neurotech company in the UK called Mint. they have like incredible ASICs incredible boards and Motif Jacob Robinson’s company in the US. As part of that, a motif set up a UK office. They hired someone, but actually realized the talent is so good and to be totally honest is cheap and you can get like incredible people. They’re actually starting to hire more people in the UK just based on that. It’s like that’s great.

Yeah, it’s kind of I think there’s a load of things happening around there.

Juan Benet

Yeah. I wanted to dig into the AI for science question. you mentioned it a little bit earlier, but this is a field that I’m also extremely interested in. I see a lot of people in the in the AI field themselves are super interested in this ultimately why many of the people in the AI field are pushing on AI because they know that the breakthroughs that are that will become possible once we get to you know full AGI levels and ASI and so on are so transcendental that we’ll be able to greatly advance how so many fields go everything from like the early advancement that we’re seeing now where you’re getting LLMs capable of being great research partners in helping you do a lot of the literature reviews and thinking about ideas in the interstitial spaces to you know I’ve been fascinated by the stuff that Terry Tao and others are doing with math where like they’re actually pushing the frontier and also not just LLMs but you know Lean and other automated ways of producing results like we might see versions of something kind of like that if we build better models of the cell or better models of connectomics and so on.

So I’m curious how you think about this entire area. What are some of the interesting projects that you’re seeing? What are like your hopes and dreams for the next few years? Because I think the frontier is so ripe. the capabilities have advanced so much in recent years like even just like the last 6 months. And the capabilities will continue expanding over the next 6 12 18 24 months. And so what you say now could inspire somebody to go and like do that for you.

Jacques Carolan

I think it’s like so fascinating. I think AI for science or AI science can mean so many different things. And of course like having labs that can like be automatable that can be closed loop that can be hypothesis generation like I think that will happen and I think certain problems are very amenable to that protein design is very amenable to that like high throughput like biological screening material science there are definitely like things where you’re like I think that’s just going to happen applied to the brain I actually think it’s almost lost all meaning but like the idea of a foundation model for the brain and but like what do what do I mean with that and what do I see as like the interesting cuz I’m like starting where like I want to see like scientific signal that something is possible and for me then my spidey senses go and I’m like okay we need to drive on this and there’s some work I found really inspiring from Leanne Williams at Stanford.

Leanne Williams is a kind of like a psychiatrist like a precision psychiatrist and they take individuals with a diagnosis of depression and the challenge with firstline psychiatric treatments it’s just like basically a dice roll so you take individuals with depression put them in fMRI and then show them images and do some kind of simple tasks. And then with all of this data based on functional connectivity alone, you can actually subtype people just based on functional connectivity. And those subtypes better predict whether someone’s going to respond to a first-line treatment or not.

Juan Benet

Wow.

Jacques Carolan

And it’s like early and it’s kind of like it’s like marginal, but like it’s there. And like I think it’s just really fascinating to be like if that’s true, what does that mean? And actually like I really do think this huge leverage point is going to be in human neural data at different spatial temporal scales at different like like I just think that is going to be a critical flywheel. It’s going to help us identify new therapeutic targets.

I speaking to my friend Milan Cvitkovic and he’s telling about this idea of predictive validity. So I’m stealing it from him, but Jack Scannell has this theory where they show that like the biggest like predictor of whether a drug is going to be clinically successful is whether it’s pre-clinical models recapitulate what happens in humans. Almost sounds like trivially true, but actually the key point is that even just changing that by a few% can radically change the outcomes of whether it’s successful or not. And the challenge is that like the best model for the human brain is the human brain. So I really do believe that we need to collect large scale human neural data sets, you know, in a safe and scalable way to identify new

[1:11:07] What AI-driven hypothesis generation means for breakthroughs per pound

Jacques Carolan

therapeutic targets to have much more personalized interventions. Like I think it’s the only way. So it’s not necessarily a kind of like AI for science, but like that is just a problem that is so ripe that we need to leverage all of our tools towards.

Juan Benet

Do you have a sense of which types of recording will produce like the most interesting or most valuable data sets like what avenues like what devices or do you think it’s like new ones that might be coming through?

Jacques Carolan

I think a variety will be valuable. We want it to be able to be like accessible. So it’s not ideal if you have to like the only place you can get this recording is like in an fMRI at Stanford. you know can you do it in a lowcost way that can actually be distributed around places where there isn’t like a tertiary center. I think that is really important.

I think figuring out ways to go between different modalities like how can single neuron spike data be useful for like fMRI or focus ultron data or be you know be like upsampled for EEG like that spectrum I think is I don’t have a concrete like answer but I feel that’s going to be important.

Juan Benet

One of the questions that’s been on a lot of our minds in the last year is could we just record ample amount of neural data across a wide variety of behaviors so then be able to collect enough data points so then figure out like what matters for what types of conditions or do you think it’s like it’s much more about robust data sets on the very specific because I think most of the neural recording that is happening today is very specific and targeted to the to the a particular pathology or something like that and yeah I’m curious where you how you think about this.

Jacques Carolan

Yeah, I think you need I think you need broad access across different brain like regions. I think just I know I’m particularly interested in like therapeutics and pathology and I think you need to be able to access deep brain structures.

I think it’s a really interesting question like how much can you get from just like like behavior and physiology alone like in mice you know if you record a bunch of cortical data it turns out that pupil dilation explains 50% of the variance of neural like spiking right because basically relates to attention and a bunch of other things and like so that’s not to say that like neural data is not important but like you can actually get a huge amount from physiology and I don’t think we’ve really understood the limits of that yet.

Yeah, it’s a little bit of a copout, but I think the whole spectrum of things are going to be really important.

Juan Benet

Well, it seems to me that having very scalable recording across the whole brain with sufficient precision is this massive bottleneck in the field, not just code economics, but all the neuro AI questions. And I’m curious how much of that is either part of the current programs or potentially a future program or something like that.

Jacques Carolan

Yeah, I would say it’s like almost the it’s almost the intersection of both the programs in a way. You’re absolutely right. like if we had that it would unlock so much. I don’t know we have a path to it yet.

Juan Benet

Yeah. Yeah. You you go into it and you mentioned recording but a lot of times it’s so linked to the pathologies or so linked to a specific target. Both of your program writeups have like this very cool section around ways to plot like different metrics and how the different technologies show up when you measure them a particular way. like you see the parallel frontier of like where the technologies are now and like kind of the types of devices that you’re trying to inspire and it seems to me like we’re missing something like that for scalable and precise measurement like where you can scale like really scale up the invivo reading while retaining the precise measurement.

Jacques Carolan

Yeah, I agree like yeah, there’s this like plot I put together which is essentially like precision some metric of it’s kind of like spatial precision but it’s a little bit different and how I was a bit cautious about using the word but how invasive the procedure is and you know things that are implantable have much better precision things that are non-implantable like kind of makes sense just like the physics makes sense because you’re further away from your sources like so there is this kind of white space of like how can you have something that has like high levels of precision that can be deployed in like a scalable way I don’t think it’s impossible possible. I don’t think it’s impossible, but it’s just a really hard problem. And that’s ultimately what we’re trying to push in the program.

And there’s another axis to that as well. Like in some way, you put together a graph and you have to like make a bunch of assumptions. And [clears throat] on the y-axis of that graph, you know, I have like spatial precision, but I don’t care about spatial precision necessarily. Like I think there are other things are important, but what I really care is about functionality. Like how useful is something? Another like axis is also like what are smart ways you can process data or like churn through data like by building large scale models maybe you need to collect less maybe you need to only collect lower resolution data so I think there’s like another axis that is kind of I would love someone to like put this together about how like invas like a threedimensional graph of like invasiveness of like functionality and like data something like I don’t know there’s something there.

Juan Benet

Yeah. Is there a robust current model of where all those technologies are across all these vectors? It seems like a useful resource to maintain to like you know have like a website artifact that you can go to and see you know the plot today and you know maybe you can see the plot over time and how it changes. It reminds me of like the cost structure graphs in clean energy like when you know you were tracking the cost structure of solar coming down or like various different sources of energy and then you can start making reliable predictions about where certain technologies would be or like why invest very deeply in batteries today as opposed to you know a decade from now because like they’re actually close enough to like this frontier where suddenly they become like actually very practical to scale and so on.

Jacques Carolan

Yeah, it’s a it’s a great point that like that doesn’t exist to the best of my knowledge and I think that would be massively valuable. Like the closest that comes to it is I think it’s Konrad Kording and Stevenson where they have this graph of like number of neurons recorded as a function of time and you know you start in the 60s and people like patching cells and it’s like one or two neurons and then like like electrode arrays happen and then like optical imaging happens and this thing kind of increases exponentially and you can be like okay mouse brain is here like human brain is here how far away are we like I think it’s kind of cool but actually I think there’s a much more I think there’ be a much more useful thing that you talk about would actually like where like what is the prito frontier of our current technologies in terms of I don’t even know what the metric might be like what is it going to be like cost per like unit recording in human or something like that yeah someone needs to do that

Juan Benet

let’s just shift into your story I’d love to hear about how you got where you are now like studied versus a physicist that you’ve done a range of things maybe take us back to how you started I don’t know learning what inspired you what you end up setting across sort of over time what brought you to different fields so

Jacques Carolan

I think starting as a kid I was like always very very curious I think like I was a kind of you’re like that kind of like intro to the question is like a nice way of saying like easily gets distracted [laughter] like and I and I think that’s like I do see my kind of history as a kind of like jack of all trades and master of none like so I think I so I grew up with just my mom and I think I very much saw like the importance of education. My mom going to university like changed our circumstances in like a very very f like a very real way. So I think like the importance of education and learning not in an explicit way but like implicitly was like always kind of installed onto me

and then like when I went to university I was like you know like what were the things I was doing I was like interested in the kind of foundations of physics. I was like man you can learn physics it tells you something about the nature of reality like that’s so cool. I remember to going to going to university and I was like studying like physics and philosophy together. I got to a point where I was like I kind of want to do something useful with these tools that I’ve built and that’s when I kind of found out about the field of quantum computing.

So the first part of my like career was in applied physics and quantum computing and I was just like fascinated by this idea which I think still sticks with me today that physical systems can compute.

Juan Benet

Mhm.

Jacques Carolan

It’s kind of a bizarre thing, right? Like whether it’s capacitors in your phone full of electrons, whether it’s like a water wave computer, whether it’s quantum bits, like you know, I think it’s just really I find that just deeply fascinating. I think it tells us something very fundamental or whether it’s biological neurons.

So I did a PhD in physics and I was trying to build quantum computers using light. funnily enough I, you know, when I finished my PhD, my adviser actually moved over to Silicon Valley with a bunch of the teams there and set up this company some people might know called PsiQuantum. They’re like the biggest quantum computing company now.

And that was kind of like it’s interesting. I think after a while, after about a decade of working in that field, it became clear that like to have impact and to like solve problems, you just need a huge amount of capital. It’s semiconductors. You need billions of dollars to do that.

And I actually went on the faculty job market. I was like, I’m going to be a physics professor. I’m going to be a quantum physics professor. That’s what I want to do with my life. I like went on the job market and it happened around the time the pandemic hit and I came back and I was like, I just don’t know if this is thing I want to do. Kind of decided it when I was like 21. I like love the people. I love the problems. It’s like is this the highest leverage, most impactful thing I can do with the rest of my life?

I’d always been like interested in biology. Like I don’t have high school biology. I don’t know I didn’t know anything about biology and I actually remember during my PhD I read this paper by Adleman of like RSA and Adleman had this like cool experiments in the ‘9s around DNA computing. So you can like like the story goes he was like reading about you know a DNA polymerase in his bed and he’s like kind of like a Turing machine and you can do these cool experiments where you like somehow get DNA and you like shake it up and you can like solve a traveling salesman problem like

[1:20:44] From quantum computing to improv comedy to running £119M government brain programs

Jacques Carolan

freaking cool and I read it and was just like oh this is wild. So that was always in the back of my mind and then I around the time the pandemic hit and I was like I just don’t know if I want to be like a physicist always been fascinated in the brain and the idea that like biological sweat wetty things like can compute that’s wild.

So I basically spent some time trying to be like is it a crazy idea to become a neuroscientist? So I ended up just like reading a bunch. I read a bunch of papers, spoke to people. It was during the pandemic, so like people had time on their hands. This like random dude in a field they didn’t know would like get a message and be like, “Hey, can we have like a virtual coffee or something?” All the physicists I spoke to were like, “This is a terrible idea. You got a good career.” [laughter] And then all the neuroscientists were like, “This is an awesome idea.”

And actually one like I remember one of the conversations I was like hope I don’t like embarrass him but I read a paper by Adam Marblestone scalable principles of physical recording like a legendary paper very yeah legendary and you like the author list is wild as well right like Dario’s on it, Shapiro, George, anyone who’s on it and for me that really resonated because it was like using principles of physics to understand like what is the limits of what we can do and I actually spoke with Adam and he was like dude you just do it so I ended up doing another postdoc in systems neuroscience playing with Purkinje cells and actually yeah I learned so much about the brain it was it was awesome yeah

Juan Benet

and there’s like a similar to the what’s the expression about math is like the unreasonable effect effectiveness of mathematics in explaining the world there’s also an unreasonable effectiveness of physicists in advancing many other fields so I think like you actually it was the right thing to study physics and then go into neuroscience I think Dario has a similar point about how in early anthropic they were primarily selected physicists because they ended up being very able to like then go on and learn everything else.

There’s something about how physics itself is taught. I don’t know if this is about like the people that it draws or the or the some of the lessons of the field or some of the complex intersection between like deep math and practical real world where like you don’t get this behavior from just mathematicians alone. There’s something about how physicists develop that ends up making the whole class of you like unreasonably effective at like working in a bunch of different fields.

Great that you made the leap like you translated from physics into neuroscience and are now helping advance it. So great.

Jacques Carolan

It was it was a journey and actually on that point as well is like a lot of the physicists I really admire like they’re so good at doing quick back of the envelope calculations you know because they can keep various equations in their head. They can keep various physical constants in their head and can play with those to orders of magnitude and arrive very quickly whether something is plausible or implausible.

I would say like when I went into neuroscience it was definitely like a wakeup call like I was the only physicist there. I’ve got this wonderful photograph which is the first time I ever pipeted. And my colleagues were like cracking up. I was like, I don’t know what this thing does. And you know, I’d done a bit of fab. I And then I went into like doing mouse surgeries and like learned how to do this stuff.

My main takeaway was like a lot about being an experimental scientist is like knowing what your error bars are. you’re trying to jiggle some complex system to optimize some cost function and like at some point you need to stop jiggling and I think when I moved into biology in physics if I built something it didn’t work it’s cuz I built it wrong in biology things don’t work because of biology so the error bars were so much bigger that took me a while to like really grow and it’s actually useful doing this job now because I know like it’s just a it’s a different it’s a different beast

Juan Benet

so you got a postdoc in science and then what

Jacques Carolan

yeah so I was working there for about 3 years in London and like learned a bunch. Everyone around me was a biologist and then it was around that time that ARIA was being spun up. So I kind of had heard about it and actually I’d met some DARPA PMs when I was on the faculty job market and doing interviews and there was this kind of like you know kind of reverence associated with them and so I was like reading about it and I was like this is such a cool model and then they just

Juan Benet

is that when you first found out about that model or had you already been kind of exposed to the

Jacques Carolan

that was when I first yeah so probably only really when I so I did a postdoc in the US so I was aware of it there but like only really the kind PM le model I only found out about when I chat to these people. I still have the I found the job ad on the way back machine which is kind of I forget exactly what it is now but it was something like you know do you want to like like direct £50 million and change the world? It was like the most like audacious thing I’ve ever seen. like, “Wow, this sounds cool.”

And then went through the application process and like it was basically what like I can imagine was just like a six-month personality test like various like and I just met so many incredible people. You know, the people that had set it up, Ilan and Pippy, they just really believed that we could do something different. They were really trying to create a culture that could enable that. And I think that was the most like striking thing just like how intentional they were about the culture they were creating internally.

I remember going to this like finalist day. So all the finalists come together. This is a unique thing about the model is that they recruit in cohorts. It’s actually like awesome because you just have these group of people that are just like on this journey with you. I went to this finalist day and like 20 of us rocked up to the British Library and there were some really impressive people and I was like I don’t know what I’m doing here. And then yeah went through the day and like none of it was really technical like bits of it technical but it was more like how do you work together like what are you driven by? Yeah. And got offered the job and I think like it was that culture they created where I was like you know what like let’s just do this.

And the interesting thing was is that like all of us had kind of taken a career risk being there. were like, you know, they were like successful professors, like successful like founders, founding team, all of this had kind of like had really good careers and like rocked up at this thing that could just be a car crash, [laughter] but that creates something really special when everyone is just all in. You must see it with founders all the time. Like it’s a really special thing. There’s some m some magic happen. So yeah, I feel really fortunate for that.

Juan Benet

Yeah, it’s really magical to create a the right culture of an environment like that. And also like what an exciting journey to be a part of in shaping a new agency like that from scratch like like like you’re saying it could be you know many of these types of organizations and agencies end up ultimately not working for some reason or whatever and being able to work in the beginning of it to try and give it the best shot possible at having like the large scale impact. Like that’s awesome.

Jacques Carolan

Yeah. And you know we were like sitting around a table being like how the heck do we launch a program like what are the most important things and just like working on that together having ownership over that together like the biggest impact I might have will be like does ARIA exist in a decade and is it still exciting and do people around the [snorts] UK around the world be like man I want to be a program director like that’s real

Juan Benet

what are some of the cultural things that Ilan and Pippy or others in ARIA set up that you know in retrospect you think were super valuable.

Jacques Carolan

Yeah, it’s a great question. I think there’s a bunch of things. I think always pushing us to be as ambitious as possible. Just being like, you know, is this the singular most ambitious there? So, I think all of us that came into it had our ambitious levels raised. We had an idea of like a thing we might want to do, but like what is it a few levels on from that? And just like really instilling that into the agency and doing what you need to do to get that information to let you make that decision.

Of course, that’s like ambition in science and programs, but also in things that might not seem so sexy like procurement. Are we doing procurement in like the most ambitious way? Are we like doing contracting in a way that like can happen really quickly, really seamlessly? I think that was like really instilled into the entire agency.

The cohort part was a big part of it as well. It’s like the best job in the world, but it’s also really hard and you’re wearing a ton of hats and actually just having a bunch of people that like have your back and you can just go and say, “Man, I had a really rough day.” And like they’ve got you and you’re there for them. Like there’s so much power in that. So I think that culturally was a was a really big piece.

And then their question is like that then hopefully filters down to the teams that we’re funding. Like we’re trying to say that we’re doing things differently, that we’re there to support and there to work with them. So I hope that comes across but like the external projection of ARIA like it really is the case internally. I feel very lucky for that.

Juan Benet

That’s awesome. There is something also unreasonably effective about cohorts in shaping organizations and cultures. You see it in universities in militaries in you know in YC in things like that. There’s something about having a group of people going through a similar experience together and learning roughly the same set of things plus having the different classes of that cohort such that you can build more senior cohorts that then can help advise the younger ones or whatever. There’s something about that structure overall that I think outperforms a similar type program that just changes that one variable where it just hires people whenever it’s it doesn’t kind of build this cohorting setup and just kind of maybe has an open call and hires people immediately in their own timeline per position. Something culturally really changes when you have a group of people temporally going through the same kinds of challenges.

Jacques Carolan

Yeah. Yeah. I 100% agree and I could not sure I can articulate what it is, but it’s think about like learning rate is like you can multiplex across failure, right? Like someone’s like, “Man, I tried this. This didn’t work.” Okay, cool. We all update. And like having that component, I think, is really important and it builds relationships in a way that like I think the other approach doesn’t.

Absolutely. Like I think all of those program directors like, you know, I’d walk through Firefall and I know it’d be the same. And you’re trying to move really quickly. You’re trying to put like documents out like you’re trying to build this new thing. Being able to like iterate quickly with people that you trust, I think trust is a big part as well is so critical.

And yeah, and now we have like new program directors coming in and like we can work with them and they’ve got like awesome ideas. I actually think the you know there’s also a finite tenure at ARIA. I think that’s also really important because it brings new ideas in like you know I don’t want to be the old program director who’s like better in my day you know like so like having that iteration I think is so so critical and I’m really glad they built it in like that.

Juan Benet

Yeah, I see also that you wrote a book, the analog quantum simulation. talk to us about that.

Jacques Carolan

Oh, wow. Yeah. So, this was in my kind of philosophy days. I was always interested in the kind of philosophy of physics and then when I was doing my PhD, I kind of like left philosophy physics behind. And honestly, it’s like hard like being a philosopher is really hard. You have to like write arguments and I was like never that great at it. And I got approached by an incredible philosopher Karim who was really interested in analog simulations.

So to give an example of that like a fun one is special relativity. So it turns out that water flowing down a plug hole has similar equations like analogous equations to special relativity. Maybe general relativity. I forget exact I think it’s general. There are water analoges of like Hawking radiation. Right? So then the philosophical question is like what can water tell us about spacetime like and so that’s the kind of like line of inquiry and I was working on a field of analog quantum simulation like taking some kind of physical system maybe it’s like atoms or photons manipulating them and trying to learn something about chemical structure or learn something about a material like what can we actually learn from that

so this book was a kind of collaboration between myself Karim and another theoretical physicist just to like take some of these state-of-the-art examples and just like really unpack them through this philosophical lens and it was like it was a blast. It was actually really fun. I think we probably have I don’t know if we’ve sold a single book but it was a really fun thing to go through.

Juan Benet

How do you find like the right trade-off between applying some analogous structure too far or just right? Meaning some of like the early things that maybe held back physics in the classical to quantum leap was being too married to the analogies of when wave physics mattered and trying to think about quantum systems as either particle or wave instead of thinking of them as a totally different class of objects that had some of the properties like cur. Yeah, curious how you think about this.

Jacques Carolan

Like one of the interesting parts of kind of analoges is also just like a kind of mental playground, right? Like I think about things in this way. I don’t know. Let’s take like the classic like analog between electrical circuits and water flow, right? Like each of these things have analoges like voltage is flow rate like all those things. Actually, that’s pretty useful for a lot of situations just to think about it as like water. That’s helpful.

And there are things where it kind of falls over and you know the kind of like Kuhnian idea is that like at some point that enough [clears throat] evidence comes up and you’re like this is no longer a good model. Then you have to build new models. So like I actually think a big value is just like and we always do we all do it. We all have certain like analoges in our head of how things work. I do it in biology. I have oversimplified models. Of course we need to do that. Turns out this is maybe why I wasn’t a great philosopher. I was like I’m just a pragmatist. Like is it useful? Does it allow me to like build things or like do an experiment? And if not, like I don’t care so much.

Juan Benet

Yeah. One of the ones that I have found most interesting is or most useful is the analog leap that Dawkins made between genetic computation and then memetic computation of extending hey maybe all of the different rules and mathematical structures that emerge out of gene replication out of DNA strands and protein translation and sexual selection and all of this kind of stuff or the different types of selection over genetic strands can tell you something about how ideas propagate and replicate.

and form larger memetic complexes and you have like you know the analog of both a gene affecting and yielding a phenotype in a population and you have you know a memetic complex expressing itself in a large population of humans as it you know replicates and rips through some population. So I have found that like a super fascinating type of analog thing. Both interesting and valuable to find what properties about the two systems are similar enough where like some idea carries over or what are the practical realities that are different about both systems that then make it behave very differently. Things that are laws in one system don’t apply to the other or apply differently or like the constants are different or you know frictions are different.

Jacques Carolan

Yeah. Does it matter if it’s exactly the reality of what’s going on? Probably not. But is it useful? Is it explanatory? Like that is there’s so much power in that. I love that example. That’s a great example.

Juan Benet

You you also spent some time as an improviser.

Jacques Carolan

Oh yeah. [laughter] Some of that. Wow. Yeah. So this is once again just like jack of all trades and master of none. during my PhD I got into standup, nerds doing standup. It was like scientists doing standup and it was like very fun. And then I moved to the US for my postdoc and I was like I don’t know if I want to do standup here. There’s like cultural differences, like humor is different, but also like, you know, you spend a lot of time hanging out in some like pub by yourself or hanging out with other stand-ups. I don’t know how many stand-ups you have watching this show, but yeah. And it’s not always super fun.

So, I was like, I want to do something that’s community building and also comedy. Improv was so big in the US. It was like big in Boston where I was studying. So, I went and joined this like improv team and it was kind of like a Ponzi scheme, right? You have to like pay to do the classes and then you pay and then eventually you can like perform but like and then so it’s funny I haven’t hadn’t described it that way but it’s like yoga.

Juan Benet

Yeah. Yeah. 100% 100%. You get to the top and then like Yeah.

Jacques Carolan

You It’s just so good. You have to [laughter] keep doing it.

Juan Benet

Oh, exactly.

Jacques Carolan

And it’s kind of got a cult feel about it as well. And so I did it and like I absolutely loved it. And you know when you’re doing science it’s just like you just something to totally switch off. It’s just adults playing makebelieve. It’s like so dumb and it’s so fun and then when I moved back to the UK, London actually has a pretty big improv scene. It’s like a new thing and it’s like so I joined some of the theaters there and now I have like teams and it’s funny and it’s like also funny how it interacts with like my work. So like if I’m like this meeting’s dry, all right, everyone stand up. Let’s go. And I even do it with like my teams as well sometimes.

One lesson I take from improv, which I actually love, is that like your job as an improviser, it’s a team sport fundamentally. Your job as an improvis improviser is to make everyone else look incredible, right? Like if you come off stage and you’re like, “Man, I freaking kicked ass. You weren’t funny. You’re an asshole.” Okay. So, like if we can make other people and if you take that throughout life or take that throughout a workshop or a scientific endeavor, there’s actually so much power in that. Yeah. I don’t know what your thoughts.

Juan Benet

Yeah, I so I also did a bit of improv. I don’t keep up with it as much as I really should, but I found it to be not just the most fun but most practically useful activities that I ever undertook. I first got introduced to it through a class at Stanford where there was also this kind of like unreasonable effectiveness for improvisers where many of the people that took that class ended up being just dramatically more effective at a ton of other things. because they were able to jump into any group of people and just immediately adapt and handle problems and be able to approach all kinds of situations with this yes and mentality of that ends up being so effective and they tend to be able to teach other people. So improvisers because of the nature of the game you’re trying to help each other to construct a good scene and so on. You end up in this very helpful and friendly environment where once any team that has one improviser in it will just naturally kind of function better. So and did you end up doing it afterwards as well?

Jacques Carolan

Randomly here and there like play with friends but not I wish I wish I did.

Juan Benet

Yeah. It’s and also like there’s something about if you’ve like died on stage like you can do a lot of things like you know and that is actually you develop thick skin and it’s kind of wonderful.

Jacques Carolan

Yeah. Yeah. Yeah. [laughter] Yeah. Highly recommend it to really everyone for everything. It’s just one of those kind of unusually good things to do that tends to be applicable in almost any walk of life.

Juan Benet

Where can people watch you perform?

Jacques Carolan

Oh, okay. Wow. We actually do I normally don’t say them publicly. I’m actually not gonna say it publicly on

Juan Benet

So, so now the really committed people like seek you out, you find it, we’ll get the drink.

Jacques Carolan

I think there are ways. I think there are ways.

Juan Benet

Have you ever performed at the Fringe?

Jacques Carolan

I have never performed at the Fringe. I’m like so bummed. Yeah, we’ve got a monthly show for maybe like more tidbits for people, but I’ve never done it. I really want to, so it’s on my list.

Juan Benet

Yeah. Yeah. Well, well, we should do like a neurotech-oriented art show at some point.

Jacques Carolan

Yeah. We’ll have like, you know, wearable BCIs like responding. I think that would be so cool.

Juan Benet

Are you a reader of sci-fi?

Jacques Carolan

I’m actually like a kind of reasonably new reader of sci-fi actually. I still would argue that the great once again things need to be funny. I would argue like one of my favorite sci-fi which I read when I was a kid and I can pick up anytime is Hitchhiker’s Guide to the Galaxy.

Juan Benet

You know where this is headed.

Jacques Carolan

I just love it so much. There’s very few books I can just go back to. I’m like, “Oh, it’s extraordinary.”

Juan Benet

Yeah. And you keep finding layers. You read it again. You’re like, “Wait, hold on.”

Jacques Carolan

Yeah. I totally missed it the first time. Oh, it’s so good.

So, like I am really enjoying that and actually what I’m reading at the moment is I’m reading the Expanse books and I’m like half I’m like like book three or four and it’s like a blast and I watched the TV series but I just wanted something to get lost in and I’m really enjoying that.

Juan Benet

Yeah. Yeah. Any others?

Jacques Carolan

There’s a great book on Bell Labs. Is it like this is called the imagination factory?

Juan Benet

The idea factory.

Jacques Carolan

Yes. The idea factory. Like that.

Juan Benet

It is required reading at PL.

Jacques Carolan

Oh, really? I love it. And it’s also like one hell of a story as well. Like that’s I really really like that. And I go back to that.

Another one is like is it I imagined Worlds?

Juan Benet

Yeah, Imagined Worlds. Yeah.

Jacques Carolan

Another one is Imagined Worlds by Dyson. And it’s just like an incredible exploration into like future technologies and like what they can be. And I think he like wrote it in the 80s or something and you like look at it now and it talks about like molecular machines and like so much cool stuff. It’s like super readable but just so visionary and I think things which like this is part of the reason why I’m really enjoying sci-fi like things that expand the mind of like what is possible I just find so fascinating and it could be science books it could be like sci-fi books just that whole spectrum I think is incredible also has this other books I don’t know if it’s Imagined Worlds might be a different one it’s just a collection of essays that like take some like interesting question and take it really far that I found like super fascinating.

Juan Benet

There was like just this super contrarian perspective where Freeman Dyson was able to just like reflect back on some dogma of the scientific or academic establishment and just take a very strong counterposition and just point out how something was like broken or wrong or whatever and then just you know pursue that idea like pretty far.

Jacques Carolan

Oh yeah.

My other like recommended reading just like a call back to our early discussion is like if you are a physicist or like mathemat whatever like interested in maybe getting into neuroscience read physical principles of scalable recording. I just like that I just people should read.

Juan Benet

Yeah. Yeah. And yeah, again legendary author list. In fact, one of your PM David is also one of the authors from sci-fi. You might enjoy We Are Legion, We Are Bob, which is a story about a person who gets uploaded into a van probe and he’s trying to save the world and hilarity ensues.

It’s quite It’s like comedy as well, like Yeah. Yeah. It’s a pretty funny book. It was a little bit slowed for it probably would be slow for most people’s tastes in sci-fi because I think there’s so many advanced ideas that I don’t know if this is the case but I think the author was explaining it too much and maybe but then he realized after releasing the first couple books that the audience was totally right there with him and knew all about it and so like the later books in the series just get very fast and pretty fun. So has like a slower start in like the world building, but it is very good.

Jacques Carolan

And yeah, I feel like all of your guests will probably like suggest this, but like obviously Pantheon is incredible.

Juan Benet

Yeah, Pantheon is amazing. They did a phenomenal phenomenal job. I kind of feel this kind of sense of loss of kind of Japan had cultivated this media side to it where like when Japan was leading technologically in a bunch of areas, it cultivated this area of media through both manga and anime that explored really deep questions about reality and seeded so much of like what you know are modern stories that end up being very impactful both in sci-fi and fantasy and so on.

And as like the bubble burst in Japan and it sort of started falling behind in the technology, I feel like it’s sort of missing. Like a lot of those stories became less frontier oriented, less thinking about where the future was going. I mean, they ended up inspiring other things like Pantheon and beyond. but you know, I always wonder like what would have happened if like the bubble in Japan hadn’t happened, like if it hadn’t gone so crazy and bursted so hard, if you had just kind of kept being this intense leader and how society would be different today.

Jacques Carolan

Yeah. Well, it’s actually in like you Ryota Kanai who like is funding stuff at Araya in Japan actually funded a number of manga artists to write and illustrate pieces about BCI.

Juan Benet

Oh wow.

Jacques Carolan

And like imagining what the future it is. I can’t remember the name of it. Like I’ll just send it to you. I’ll share it with you. It’s like it’s really cool and I was like oh I just love this idea of like what we can do as a funding agent. It’s like a kind of kind of cool.

Juan Benet

Yeah. I would love to have a ton of amazing artists envision very positive futures where neuro technology is like an ever everyday sort of thing. Like sci-fi has this problem where because when you’re imagining the future and you’re thinking through the lens of some technology the it has such a big power in the story cuz as as a writer you tend to extract elements and remove elements when they’re not doing something important for the story. Therefore, for something to remain in the story, it ends up having to play a very important role.

And so, you end up in a situation where because of the level of power that some new technology gives you. And I think because I think humans are just tend to be more worried about like downsides than trying to paint positive pictures. You end up with a selective selection pressure where all media about the future tends to skew very negative because the if the technology isn’t in the story, it has to be there for a reason and that reason must be to create some dystopia you have to fight. And so we end up with this very bad skewed perspective where all of our media about the future is just like all these dystopic narratives and it really sucks. It doesn’t give us like good things to aim for.

And so we kind of need to like forcibly counterbalance that by getting a ton of artists to just paint stories of the future where there are these elements but they’re not they actually think about all the positive cases like where are all the transformative amazing stories about neurotech being able to like regrow a spinal cord or something like that or like a love story that happens to have neuro tech in it or like

Jacques Carolan

Yeah. Yeah. I’ I’ve agreed. I’d love to see that. That would be great.

Juan Benet

any advice that you would have for either people in the field today what should they think about or look at or maybe people going through university now or postdocs considering shifting fields like

Jacques Carolan

I don’t know I will say like yeah maybe I’ll give my advice to folks that are working in a different field like I think it’s valuable to like find those people that inspire you and maybe have taken a particular career path like there’s so many people who have taken really nonlinear career paths that I you know that really inspire me some of my friends some are not my friends and just kind of see if you can get time them. See if you can understand like what has driven towards that. You have a you’ll have a pretty good gut sense of where you want to go and people will tell you advice and listen to the things that resonate. Ignore the things that don’t. That’s totally fine. And it’s totally possible to make that jump.

I think be voracious, read, be excited. There is so much work to do in this field. There is so much left to do. I think that’s easy to forget given the excitement that we see in certain corners of the neurotech field. There’s so much and we need incredible talent in there. So yeah.

Juan Benet

And and if you were to you know think out to 2030, 2035, 2040 like I one of those do you have some like crisp view of how the world would be very different then and you know tell us a bit about it.

Jacques Carolan

I think if I do find it hard to project really far into the future with like neurotech and everything that’s happening but even like a kind of 5 to 10 year I do believe that we’ll I hope there will be a big c this is like a footnote a cultural push whereby neuropsychiatric conditions and neurological conditions don’t have the stigma that’s associated with them. I do believe we can be in a situation where we have significantly better treatments through large scale phenotyping through new precise tools that can be much more personalized, much much more effective. I believe that is possible like the science en like says it’s possible and it’s just a coordination problem and the signals I’m seeing is that like I think we’ll get there.

Juan Benet

Yeah. Awesome. Well, thank you very much for joining us.

Jacques Carolan

Thank you so much. It was a blast.

Juan Benet

I hope you enjoyed this episode. This is a new podcast, so we need your help to get the word out. Please like, rate, and subscribe on your favorite platform, and share it with people you think would find it interesting. Thank you. See you next time.

Precision is building a minimally invasive brain-computer interface (BCI) that reads from thousands of points on the cortex without penetrating it. The Layer 7 device is implanted through a one-millimeter slit in the skull rather than the larger borehole other approaches require. It is also fully removable.

Precision seeks to help the 5 million people living with severe paralysis in the US (including 800,000 new stroke cases per year). In March 2025, Precision received FDA clearance for a temporary wired version of the system. Over 85 patients have been implanted with and used the device in clinical studies. Wireless implants are planned for 2027.

We go deep on the history of Neurotech from the 1980s to the ML inflection points that triggered Neuralink’s founding, why surface ECoG was a contrarian bet that’s now paying off, the path to treating paralysis and stroke at scale, and why Ben believes neural data is at the same inflection point genomic data was in 2000 — a whole class of biological problems about to become tractable as computer science problems.

Sections

• 00:00:00 Introduction

• 00:04:39 Paralysis as a lens to understand the brain

• 00:05:36 The 1980s breakthrough: population encoding and the birth of BCI

• 00:14:36 Google Translate, ML, and the founding of Neuralink

• 00:23:08 What is the long-term vision of Precision Neuroscience

• 00:31:56 Layer 7 and why transformative technology always looks impossible at first

• 00:50:21 The surgery: a slit in the skull, not a borehole

• 00:55:19 The clinical program: who are the patients

• 01:04:16 FDA clearance and the path to wireless implants in 2027

• 01:08:32 The patient population: paralysis and stroke at scale

• 01:16:26 Neural data as the new genomics

• 01:30:06 BCIs, AI, and the future of the human-machine interface

• 01:31:22 From medical necessity to lifestyle technology

• 01:40:36 Precision as a platform — and an optimistic vision

Links from the Podcast

Precision Neuroscience: https://www.precisionneuro.io

Layer 7 BCI: https://www.precisionneuro.io/our-technology

Icahn School of Medicine at Mount Sinai: https://icahn.mssm.edu

Podcast Episode Links

Precision Neuroscience

Juan Benet on X

Juan Benet Podcast

Protocol Labs

PL Neuro

Episode Links

• YouTube

• Spotify

• Apple Podcasts

• Amazon

• PocketCasts

• Player FM

• The Podcast Index

• X

Disclaimer⁠: https://bit.ly/PodcastDisclaimer

Transcript

Juan Benet

My guest today is Ben Rappaport. He’s the founder of Precision Neuroscience. He previously co-founded Neuralink and Symbiotic which was acquired by Apple. He’s assistant professor of neurosurgery at the Icahn School of Medicine at Mount Sinai. He leads research in brain computer interfaces and serves as scientific director of Mount Sinai Biodesign. He studied physics, mathematics.

He got a PhD in EE and computer science, went to med school and then became a neurosurgeon and treats lots of individual patients. He’s authored over 50 papers and 40 patents in neurotech, and medical innovation. So he’s just a legend in the field. I’m very honored to be chatting with you today.

Thanks for taking the time.

Ben Rapoport

Thank you so much for having me. It’s a pleasure. I’m flattered by the introduction.

Juan Benet

Thank you so much for all of the work that you’re doing for people globally. It takes a lot of individual drive and effort of people working for decades, both in the core science and the core innovation to actually get tech translated and available.

So thank you from all of the people that are benefiting from it. So let’s dive in. Let’s talk about the device, which I think is a beautiful work of art. Maybe walk us through the design of layer seven and kind of what were the core insights that you designed around, like why this particular form factor?

How does it sense neurons? Why is it in this particular way? It’s very different than than many others.

Ben Rapoport

It is different. Nothing is totally emerges out of a vacuum. So it’s similar in certain ways and, different in certain ways, but you’re totally right in the brain computer interface world when we started precision, it was heretical in a sense because the whole field really had developed around micro penetrating intracortical, micro electrodes that were developed to record kind of one neuron at a time. In the ‘80s, ‘90s, 2000s, in the academic field of brain computer interfaces, there was a real emphasis on trying to record from individual neurons. The neuron being thought to be the atomic unit of information processing in the brain and the atomic bit of information being the action potential, so-called colloquially referred to as a spike. So a one millisecond duration signal, like the ones and zeros of the brain. That was kind of how the brain computer interface world, academically was thinking about, interfacing with the brain.

In the medical world, there has been for many decades, routine use of electrocorticography. So surface contacting electrodes that sit on the cortex without penetrating in, and are used to map neural activity, with very precise timescales and, in the clinical domain, usually much less fine spatial scales but known to be safe and quite flexibly useful for many decades.

And one of the insights that I think we take for granted on the medical side is sort of that most of the conscious experience of human beings takes place at the cortical surface. So the brain is of course, this three dimensional structure, but really it’s optimized for surface area. If you’ll take a look at the neocortex, it’s kind of like a flat sheet that’s crumbled up into a ball and it has these hills and valleys, as a way of optimizing for surface area.

And the reason is that, at the surface, that’s where all of the conscious thinking happens. A lot of what’s on the inside is essentially white matter or wiring that connects what happens at the surface of the brain to the outside world through the brainstem and spinal cord. The intuition at the beginning of all this was that we really want to be sensing what’s happening at the cortex in a way that maps in the greatest spatial extent, relevant to what’s happening at the cortical surface, but not damage the brain by penetrating into it if we can avoid it.

What if there’s a way of doing that? Well, electrocorticography already existed. No one had built electrocorticography systems that were at comparable scales to the intracortical penetrating electrodes that had been used in the world of brain computer interfaces. So we said let’s just do that. If we do that and we get the electrodes down to the scale of individual neurons, we should be able to sense high quality neural information.

We also knew there are some kind of like in every field, there’s black magic, just like in AI. Like in early machine learning, there’s all the black magic, like when do you stop training?

How the actual stuff gets really done? There’s the stuff that you write papers about and then there’s the stuff that you have to learn by doing which nobody talks about. So one of the things that nobody talked about with intracortical electrodes was that very often you would stop recording from individual neurons, or you’d start recording from one, and then it shifts to another because the electrode moves a little bit or the brain responds a little bit. The impedance of the electrode changes and so you can’t record from individual neurons and you’re really recording from a bulk average of tissue. All of these things were happening and anybody who was in the field knew that was the case.

The reality was that recording from a local average field and recording from individual neurons was basically equally informative from the standpoint of can you decode what the brain wants to do?

Juan Benet

Mm-hmm.

Ben Rapoport

It’s not equally informative if you’re a neuroscientist trying to study the behavior of individual neurons. But from the standpoint of how the brain decides how to make the body move through the world, it’s equivalently informative.

Juan Benet

Is that primarily for motor cortex cases? Or will there turn out to be other areas, of thinking or behavior that end up with these populations of neurons?

Ben Rapoport

I’d say that it is definitely true of the motor cortex.

Juan Benet

Yeah.

Ben Rapoport

It also seems to be true of many other areas of the brain. Are there processes that require more and less fine detail? Probably. A hundred percent. But is there a huge application space that resides in this scale of hundreds of microns of spatial and few milliseconds of temporal resolution?

Absolutely. This shouldn’t surprise the machine learning world either, right? Because like, what was the TPU and the GPU all about? It was about reducing precision in order to increase the scale basically in order, and so the fact that we can lose a few bits of precision in order to increase spatial scale and scalability and addressability of different areas of the brain. and, and just to make a

Juan Benet

safer device

Ben Rapoport

isn’t easier. Yeah,

Juan Benet

Yeah.

Ben Rapoport

Without compromising efficacy at least in these class of problems.

Juan Benet

It seems to me like a classic case of a great contrarian secret when building a technology company. It’s like the rest of the field thinks and behaves a particular way. You have this core insight that, hey, actually the problem gets a lot easier and you can tap a whole range of use cases by applying this other case. You might not get the full holy grail of hundreds of thousands of individual neurons, but a whole wide range of applications are achieved.

Ben Rapoport

That’s not my Holy grail.

Juan Benet

Yeah, yeah,

Ben Rapoport

Like you asked me at the beginning, what’s my holy grail? My holy grail is making products that work for people. My holy grail is not studying every individual neuron in the brain simultaneously. For some people, that’s the holy grail. It’s not mine.

Juan Benet

How many people out there might qualify for this kind of treatment on the road?

Ben Rapoport

For the most severe forms of paralysis. Basically, spinal cord injury from the neck down, severe impairment of both hands. We’re talking tens of thousands of people in the United States per year newly affected. If you looked at a cross-sectional study of the United States today, about 5 million people existing in need. So of 5 million existing and a few tens of thousands per year newly affected. That’s for the most severe forms. So people who basically are paralyzed almost completely from the neck down and don’t have adequate use of their hands to engage in routine desk job work.

My hope is that, as long as the technology is safe and effective, other forms of paralysis will be amenable to treatment with this type of technology. I think that the Precision technology is well poised for this. I’m particularly thinking about stroke, because stroke is the most common form of paralyzing injury.

It’s different from a spinal cord injury in that it is more often incomplete. It does not usually affect both sides of the body in quite the same way. It has, in some ways, less focal impact because of the way disrupting blood flow to the brain works. But nevertheless, I know there is a huge need in restoration of function after stroke. There’s a big segment of the medical community that really believes that one of the next stages in the impact of brain-computer interfaces is for people with stroke. Because this is a little bit different from the AI futurist application world of how the user interface change and we can get there too.

If you think about it in terms of the medical need, in the last 10 years the treatment of stroke has been totally revolutionized. When I was growing up, when you were growing up, stroke was a medical disease.

There was nothing procedural to offer. That changed slowly over time. And in the last 10 years, the ability to intervene acutely and remove a blood clot from a blood vessel that is causing the paralysis, that is causing the stoppage of blood flow and oxygen to the brain, changed. That and a few other things made stroke, as a whole, a more procedurally amenable condition.

And I think there is a very influential stream of thought in medicine today that feels that, even though that has saved life and limb for many millions of people, many of those people whose lives are saved and improved live with a form of paralysis that we do not yet have a way of treating. There is a sense that what is waiting for those patients, and in the United States that is almost a million people per year who are affected by stroke.

Juan Benet

Wow.

Ben Rapoport

This is and was a contrarian position. At the beginning, we were told, “You can’t do that. You can’t get enough information out of the brain just from the surface.” And then, of course, like anything, you have to prove it. So that took time.

Now people ask other questions, and they don’t say, “Can you do it?” They ask, “How well can you do it?” That’s a natural evolution.

Juan Benet

Classic case of you can’t possibly do this. Well, here it is proof. Right. Okay. But you can’t possibly do this other thing. Right?

Ben Rapoport

But to your question, and to where you started, that was the insight.

So the notion was, what do we need to do if we want to get to those scales? We need to make electrocorticography arrays that are on that scale. Well, how do you build something at that scale? Actually, like we said, the real problems are not so easy to solve sometimes.

Because when you have to put a lot of electrodes in a plane array, how do you connect to them? How do you make such a thing? How do you make it small enough? How do you manufacture it in a reliable fashion? How do you manufacture it in a safe fashion? How do you route all the electrodes out? How do you power that device, and so on?

So the answer to how do you make a small patterned device was easy. You use photolithography. That’s how you make all kinds of small electronics.

Easy to say, but not so easy to do because, although well known and universally used in electronics, photolithography is really not used in medical devices. The number of medical devices that use photolithography in a sensor context, you can count on one hand. Precision is now one of them.

But also, this thing has to be flexible because it has to conform to the surface of the brain. By the way, the brain is delicate. It has a Jell-O-like consistency. So the interface itself has to be correspondingly delicate.

And all these electrodes have to be uniform in size, shape, and impedance. Then we had to solve questions like, how fine do you make them? What material do you make them out of? And how big should they be in order to record from the electrodes? These were all questions that had a theory behind them and that we then had to go and solve empirically.

But the fundamentals are that the modular device is a polymer thin film that contains tiny platinum electrodes, most of which are about the size of an individual neuron. The module that we use most frequently contains 1,024 electrodes, but we can place many of them side by side to tile the brain.

The traces are very thin platinum wires, essentially, that are embedded in this polymer thin film. Those traces then connect into a hermetically sealed can that contains the electronics that amplify, digitize, and condition the signal, and ultimately transmit it wirelessly outside of the body.

Juan Benet

It is beautiful and fascinating. You had this idea for how to produce this kind of device. How do you start? How do you make it?

Ben Rapoport

How do you make it at scale when you’re building something that demands incredible quality and reproducibility?

The truth is, there was no supply chain for this. There were research-grade houses that could make small batches of things, but not usually in a controllable fashion, in a reliable way, on time, and under quality controls.

When you’re building something that is going to be implanted in the human brain, it needs to adhere to medical-grade quality control. It needs to be highly well documented, predictable, testable, biocompatible, and meet all of these specifications that we appropriately require of a device that is going to be touching and implanted in the human brain.

We had to bring that capability in-house. We now own and operate a microfabrication facility just outside of Dallas. It is amazing.

Juan Benet

Do you build it from scratch or do you acquire it?

Ben Rapoport

We acquired it.

Juan Benet

Yeah.

Ben Rapoport

It’s very hard to build one of these from scratch.

Juan Benet

Yeah. You wanna make a BCI device, you must first get a fab.

Ben Rapoport

That may be the case.

Juan Benet

At least for the short to medium term.

Ben Rapoport

I think that’s true. As of five years ago, there was no supply chain. During COVID, many of us became acutely aware of manufacturing supply chains and how they limited the development of technology and the progress of industry.

So we started all this early in COVID, and we were thinking through supply chain with incredible sensitivity at that time. And there was, and remains, no piece of the medical technology supply chain that involves microfabrication.

Juan Benet

Especially if you want to tweak the process or do something that is a bit out of the ordinary, you then have to do it yourself.

Ben Rapoport

Yes. So if you have a locked design, then maybe that is one thing. But all of us who are seriously in the field know that it is evolving. So the ability to conduct R&D while we are building product is essential. You hit the nail on the head there.

Juan Benet

Yeah. If you’re serious about your hardware, you have to be serious about your fabs.

Ben Rapoport

Definitely.

Juan Benet

Yeah.

Ben Rapoport

Yes. You have to be serious about the R&D. You have to be building for today and also planning for tomorrow.

Juan Benet

So you design the device, and in order to make it, you try working with external groups. It does not work as well, so you need to bring it in-house and get your own fab. And so now you have this awesome device. We now have some insight into the approach in sensing and sensing populations of neurons. Potentially, some of these can get small enough to sense one or a few neurons at a time. But it also has this very super-flat. What is the thinking there? You also developed this new surgical procedure that is very different from what a lot of the other BCI groups are doing. What is the thinking there?

Ben Rapoport

Okay. So let’s unpack that.

But before we continue, as you were asking the question, I wanted to pause for a second. Because you frame a lot of the questions with “you,” and I want to clarify that I represent a plural “you.”

This is not just me. “You” represents a whole team of people, certainly at Precision and in the field as a whole. So I want to acknowledge that.

We were talking about the microfabrication facility. We did not just acquire a facility. We acquired an incredibly experienced team that had actually been working at that facility, for some of them, for more than a decade.

That team knew how to operate the facility and knew how to work together as a team. So that is not to be taken for granted. Of course, all of this science and technology is developed collectively by expert people. We are not quite at a millennium of experience yet, collectively, but it is a lot. It is a lot of people working together as a team.

So I am here having this conversation with you, but representing a huge team.

Juan Benet

And that know-how is part of what makes fabs so hard to reproduce and hard to scale, rather than just the depth of expertise of a team that is able to run them very well.

Ben Rapoport

No question. And it applies to basically every piece of what we do.

There is the clinical side, which we touched on. But the team that does the clinical interface, working with expert physicians and surgeons and patients and families, that itself is an expertise. The whole medical system and academic medical system in this country and abroad that enables early-stage technology to enter the medical system at an inchoate level, that itself is a whole other conversation.

The electrical engineering that happens on the R&D side, the machine learning, all of these are groups of experts, and the development of the technology and its rollout into the world depends on that.

So, you asked a series of questions, which include the electro array being planar, what the thinking is there, how it is delivered onto the brain and how that is different from how other groups have thought about it, and what signals we are actually recording from the brain and how we think about that.

We talked earlier about how the brain, conscious experience, kind of unfolds in the two-dimensional surface of the cortex. That was the reason the electrode array is flat, so to speak. We’re really focused on neocortical real estate for the most part, although we’ve also interacted with the brainstem and spinal cord. But again, the computing happens mostly at the surface, and the wiring happens under the surface.

So with that insight, we can build interfaces that focus on the computing elements of the brain and on sensing and stimulating it at those interfaces. That’s obviously a simplification, but it’s a really useful simplification.

The planar nature of the electrode array also allows us to deliver the array in some innovative ways. First of all, it makes the array placement in the brain, or on the brain, incredibly safe. Because basically, the array just caresses the brain surface. It sits delicately on the brain surface. It basically adheres through capillary action, and it does not penetrate the brain surface. And yet, it records with great spatial and temporal fidelity.

That means we can place the electrode array, and it can be moved along the surface. So a little bit of trial-and-error adjustment is possible in ways that are not possible when you have a needle-like electrode that needs to be withdrawn and replaced in order to optimize its position.

So that carries certain advantages. It is also possible to be extremely modular. We can place a lot of these electrode arrays simultaneously without having to worry much about whether there is a blood vessel or what have you.

We can also get into the nooks and crannies and folds of the brain, so we can get around onto surfaces of the brainstem that are inaccessible to conventional electrodes. We can place electrodes within the sulci, the folds of the brain, which are really not accessible to penetrating types of electrodes.

So there are certain things that this planar, film-like nature of the electrode allows us to do that are not possible with traditional electrodes.

One of those things is that we can reach parts of the brain that we do not need to expose. What does that mean? Traditionally, in order to place electrodes or sensors in or on the brain, you need to expose the part of the brain that you want to address by making a hole in the skull or cutting a piece of the skull out, and then penetrating the brain’s surface underneath where the skull has been removed.

The surface array that we have can be placed through a little slit in the skull. So we designed it to be placed through a slit in the skull, like a little incision in the skull and the dura, and it can be slid across the brain surface. It can be placed directly adjacent to this slit, or it can be placed many centimeters away.

And we have done that, or rather, our collaborating surgeons have done that. They have made small apertures in the skull and slid the electrode many centimeters away. That can be done with one electrode or with multiple electrodes.

So what the technology allows us to do is really decouple the invasiveness of the surgery from the total amount of information that you can exchange with the brain.

Juan Benet

Yeah.

Ben Rapoport

That was also kind of an “aha” for brain-computer interfaces: the minimally invasive nature, not just in terms of the way the electrode contacts the brain, but also in terms of the procedure itself, and hopefully the risk and risk-reward profile for patients considering undergoing the procedure.

Juan Benet

Maybe to help listeners understand the difference, describe how some of the other surgeries happen, how the devices are implanted, and then the approach you’re taking.

Ben Rapoport

In the research world where this all started, it was required to cut a piece of the skull away. That is a traditional technique of brain surgery. So when patients need brain access to remove a tumor, seal an aneurysm, or something like that, there are tried-and-true ways of accessing the brain that are totally appropriate when there is a need to get in.

But of course, the less you open the brain, from many perspectives, the better.

All of the prior-generation technologies were based on needle-like electrodes. The only way to get a needle, even a very fancy needle, into the brain is to expose the place where you want to put the needle. You need to expose the area directly underneath where you are removing the skull, because that is where you are going to be placing the needle.

So if you are trying to get to the surface of the brain, you need to remove skull over that whole area you want to address. And so that becomes a pretty invasive procedure.

That applies to Neuralink, for example. Even if it is a very small removal of skull, to scale the process, you do need to remove more skull in order to get more electrodes in.

Juan Benet

It seems like most of the BCI companies right now are following this kind of procedure, where they’re doing a borehole right on top of the area they’re trying to sense, especially if they’re trying to scale it out to multiple different sites. That would mean many holes in the skull.

Ben Rapoport

Exactly. So it would be nice to decouple the number of holes, or the size of the hole that you want to make in the skull, from the total number of electrodes, or from the functionality of the device.

And that’s what we did with the Precision system. We can make one small aperture, which in our case is either a small burr hole or a slit. Through that one-millimeter slit or very small burr hole, we can slide many electrode arrays, so many thousands of electrodes can be placed.

That decoupling of the invasiveness of the procedure from information exchange with the brain was a Precision innovation.

Juan Benet

That also seems like another great contrarian secret, where the rest of the industry is going in one particular direction. When you scale that up, or take it to its conclusion with many of the later applications, that seems like a long-term, maybe brittle path. And if you take this minimally invasive approach instead, and you focus on how you achieve scalability through that, that seems like a very good long-term strategy.

Ben Rapoport

That was deliberate on our part. And I think it comes a little bit from the starting point. It comes from combining an engineer’s orientation to the field with a physician’s sense of what is known to work and what is acceptable to people with a problem.

Juan Benet

So you have the device. You have this innovative surgery for delivering it. What is the range of applications that you’re using it for now? You’ve implanted around 50 people. What were those cases, and what were you looking at so far? And if I understand correctly, currently you implant it for a period of time, use it, and then remove it. It does not stay in there for now. What are the kinds of things you’re learning at the moment, and where is this headed?

Ben Rapoport

One of the things that we tried to do when we started Precision was to avoid getting into a situation where brain-computer interface is already high-complexity and high-stakes field. We’re talking about developing a fundamentally new technology for implantation in the human brain, and in the brains of people whose brains may already have been compromised in some way. So the risk tolerance is pretty low. How do we ensure that what we’re developing is going to work?

One way is to spend a lot of time building something in the laboratory, then prove that it’s safe through some incremental process, and then start clinical trials with the end-state device. But we know that something is not going to be quite right, and you would have wanted to know that at the beginning and fix it before you got all the way down the path.

There’s almost no field of technology in which that is how things are developed. You don’t build a multibillion-dollar rocket with many millions of parts without testing subsystems in the real world. You don’t build a high-performance automobile without iterative test laps and a team that can pick the thing up on a lift, get underneath it, and tweak stuff after five minutes of testing. Same thing with software, with consumer electronics, and with almost every other complex system.

So we knew that we wanted to be in a situation where we could test, validate, and verify pieces of the complex system in parallel. And the problem with medical technology, including brain-computer interfaces, is that the end user is not really part of the engineering team.

The end user is a patient, a surgeon, a neurologist, or some combination thereof. So how do we bring that part of the user experience into the design loop?

One of the good things about the Precision technology is that the electrode array we were just discussing is removable and reversible, because it does not penetrate into the brain. It can be moved and removed without traumatizing the brain. And so the risk profile of the device is fundamentally different from that of penetrating electrodes, which, however safe they may be, still do some amount of cutting into the brain.

So the ability to do that in a temporary fashion, or in a research-type fashion, is fundamentally different. Because of the nature of our electrodes, we were able to design and supervise studies with clinical partners at major academic sites, in which patients undergoing standard-of-care neurosurgery, sometimes asleep, sometimes awake, could volunteer to participate in an early-stage study involving the use of the Precision technology.

In that way, we learned a tremendous amount about how the devices interacted with brains in the real world, in real clinical settings, and what the signals looked like. Importantly, even from a very early stage, we proved that in people who were awake, and most of whom were able-bodied, they could execute and imagine movements, and speak, and that we could, using the machine-learning side of this technology platform, which we have only barely touched on, decode the neural correlates of intended, attempted, and executed movements.

That’s the fundamental insight and motivation of our clinical program. Number one, it is to validate the technology, make sure that we are getting the signal quality we need, that the signals we are getting are decodable, and that we have a signal-processing algorithmic pipeline that really works. So we could de-risk that, and not only de-risk it, but really push it forward in a nuanced and accelerated fashion, even as other parts of the system were coming online.

We are well beyond that at this point, five years into the company. But that was the insight early on. Let’s make sure that we are decoupling risks and testing pieces of the complex system together.

So, to your original question, who are the patients who are involved, and what is the nature of the surgery? With different clinical sites, we have different types of procedures in which the device has been used. I’ll give you a few examples.

In some brain surgeries where there is a tumor near language-related areas or movement-related areas, part of the standard of care is to electrically map where the so-called eloquent areas are, where the language-related or movement-related areas are, so that those areas are not removed along with the tumor.

Juan Benet

This is localizing what areas to make sure to really avoid.

Ben Rapoport

Exactly. Which areas to protect, and which areas represent abnormal tissue that can be removed. There are traditional techniques that are used to do that. So it has been of interest to understand, can higher-resolution electrode technologies help do that better?

And also, while that is happening in parallel, once you have identified those eloquent areas and are operating on the others to remove the tumor, can we place the Precision electrode on what is known to be a movement-related area or a language-related area, and while the patient is awake, have them engage in a behavior, in speech, in movement, hand gestures, and so on, that allows us to test and validate that we can identify distinct signatures of those neural behaviors, those intentions, those behaviors, and really make sure that when the technology rolls out to people who can intend the movements, who can intend the speech, but cannot execute them because they are paralyzed, the device and the system will be able to perform.

So a lot of the early work that we were doing involved situations like that. Some of the patients had brain tumors and were being operated on partially awake. Some of the surgeries were for Parkinson’s disease, in which an electrode was being placed in a different area of the brain, but nevertheless that allowed access to the movement-related areas of the brain in people who, for a portion of that procedure, are awake.

And I just want to, as a sidebar, note that in this type of clinical study, the people who are undergoing the surgery are volunteering to do something pretty extraordinary.

Juan Benet

Yeah.

Ben Rapoport

They’re really partnering with us and with their surgeons to say, “Look, we understand that there are people a little bit like us, or who we may know, who down the line are going to benefit from the data, knowledge, and insight that we’re contributing. And that is going to help other people who are not us.”

So we’re willing to engage in this, to have our surgeon, neurologist, or treatment team spend a little bit of extra time, in parallel with our treatment, to help push the boundary of what is known and help somebody else down the line.

And I think, to me, that is part of how medical science advances. It’s amazing.

Juan Benet

It is extraordinary pioneering work that just benefits so many other people downstream.

Ben Rapoport

Totally. I think that has to be acknowledged. These people and their families are amazing. It is a very special person who does this, and I think that needs to be said.

Juan Benet

Yeah, absolutely. So many people in the future will be deeply thankful to those people who took action now to help push the frontier, understand what’s going on, and shape therapies.

Ben Rapoport

They are usually energetic, incredibly fascinating, and curious people you want to be friends with and partners with. They feel as invested as we do in learning, pushing the boundaries forward, and contributing something to people beyond themselves. But nevertheless, those are the people who are really on the spot at that time, and they usually have their own things that they’re worried about.

So to zoom out a little bit and think about somebody who may come along later, who you may never know or meet, nobody will know your name and you may not know theirs, but somehow you’re contributing something, it’s special. It’s altruistic, and I think not everybody realizes that that’s how it works.

Those are the kinds of interactions that we’re having. Earlier this year, we did receive FDA clearance for a version of our device, which is a percutaneous, wired version. Of course, the permanent implant is a wireless version of the system that records the neural signals and wirelessly transmits them outside of the brain so that we can control computer systems.

But some of the R&D versions of the system that we rely on a lot in our R&D are wired, including the ones that we use during the temporary implantations. So we now, as of March, have FDA clearance for a temporary version of the system that can be implanted for up to 30 days. That type of system is the type that can be used for this kind of R&D work.

I have to be careful about exactly the language that I use with respect to what is actually cleared by the FDA. So just understand that the R&D work and our collaborations with all of the clinical sites are all investigational studies, which are not part of the clinical use of the technology, but they give you a sense of what the capabilities of the system are.

Juan Benet

Yeah. You’re doing this range of R&D work now, this set of studies. At what point do you make the wireless device for long-term implantation and actually start treating patients?

Ben Rapoport

Yes, as intended. Treating people with paralysis.

The current plan is that we have a wireless version of the system that is currently undergoing the validation and verification required for an implantable medical device. We expect that, in 2027, the first of those devices will be implanted in human patients in what’s called an early feasibility study.

That’s our current plan. That first group of patients, assuming things go smoothly, will proceed to a slightly larger pivotal clinical study. Between now and then, basically in 2026, while we’re doing the validation and verification required to advance into that feasibility study in human patients with the wireless device, we’re also spending a lot of time, in parallel, in clinical studies at partner sites to robustly and validate the algorithms and the software, so that we can decode intended movement and do all kinds of things that represent the functionality we would want to provide.

We want to be as certain as possible that we will be able to provide that functionality to the patients who enroll in the study starting in 2027.

Juan Benet

How long is that study? If I’m thinking about people who may have the forms of paralysis that you might be able to treat down the road, what does the timeline look like for them? So you’ll go into that 2027 study. At what point does this become commercially available for them?

Ben Rapoport

Let me say that, approximately, the 2027 feasibility study will be on the order of 10 patients, maybe fewer. And the pivotal study will be, we don’t know exactly, but hopefully a modest number of patients.

So I think we’re talking about a single-digit number of years before. Both of those studies need follow-up. You need to enroll the patients, and they need a certain amount of follow-up. Some of that is still to be negotiated and agreed upon with the FDA, so I don’t want to say too much in a premature fashion.

But I think it’s reasonable to say that we’re looking at a single-digit number of years, if all goes well, to devices becoming part of the standard of care and starting to treat people with paralysis, and being available for people with paralysis.

Juan Benet

Amazing.

Ben Rapoport

So it’s like 800,000-plus people a year in the United States have a stroke, some of whom will fully recover, some of whom are extremely severely debilitated, and about a third of whom go on to live a full life expectancy but with a significant paralyzing deficit. I think many of us know people like that.

So the question is, what do we have to offer them today? And there isn’t much, actually. So there’s a stream of thought, and I subscribe to this, that brain-computer interfaces will have something to offer that group of people. And that’s a lot of people.

There are other areas, but to me, if you think just about paralysis alone, without thinking about mood disorders, the visual system, the sensory system, executive functioning, memory, and all the different things that the brain does that can go right or wrong, or can be optimized or augmented, to me that’s an area where I see the future of medicine changing through brain-computer interfaces.

Juan Benet

Do those numbers look similar in the rest of the world as well?

Ben Rapoport

Yes. I would say, as a matter of prevalence, yes. It’s stroke. There are some segments of the world where it is a little more or a little less common, but stroke is a common condition the world over.

Juan Benet

So the outlook is a set of trials now through 2027, maybe 2028, and beyond, with a single-digit number of years before hopefully being clinically available for a range of potential problems, both in the US and abroad. That’s a phenomenal outlook. Once you get through the studies, how does that scale from there?

Ben Rapoport

I’ll say two things. One is that I also think one of the things we’re learning, that we did not set out to learn, is that by engaging with the world of experts, with medical and surgical experts and with patients, and by getting early versions of the technology out into the world, we’re discovering that there are other smart, knowledgeable people with experience and ideas about what to do with the Precision system as a platform technology.

Juan Benet

Yeah.

Ben Rapoport

We don’t pretend to have all of the ideas or insight. But what we are finding is that people are coming to us and saying, “We’re seeing these signals. We’re seeing this technology that you have. Can we use it? Can we partner with you?” And that is a kind of signal that you seem to be on the right track.

So we have to stay extremely focused on building one particular product, but we are also spending some time and capital on building out future use cases and supporting a community of users in doing additional exploratory work. Some of that we have done a little bit of ourselves.

I think we do have a sense that we do not have all the answers, and that what we are building has a platform nature to it. Of course, in order for there to be a platform, you have to have one application. I do not want the cart to go before the horse. But we do hope that what we build will be a tool that enables discovery, and that enables clinically oriented discovery into areas that may not be our first use case, and that we may not be thinking about now, but certainly we are able to access and gather data from all kinds of areas of the brain, brainstem, and spinal cord and we want to enable others to do that work, either independently from us or in partnership with us, to build on the platform, especially in a software-enabled fashion.

So that is part of the vision and I do hope that some of the other applications of the technology will be thought of and moved forward by others too. We want to set the stage for that.

In practical terms, we do have quite a few partners. Some of them are clinical research partners, and some of them are other types of partners. We are trying to balance how to prioritize the main thrust of what we have to do with building for the future.

Juan Benet

It seems to me that there is this huge bottleneck and overhang in the capability set, because as soon as you are able to have devices that can read and write to the brain at some reasonable level of bandwidth, that opens up a wide range of use cases and possibilities that just have not been done before.

So there seems to be a lot of not exactly low-hanging fruit, because it is very difficult to build these kinds of devices, make sure that they are safe, and make sure they work really well for patients and so on. But there is mid-hanging fruit, and a lot of it, that could open up over the next 5 to 15 years once you are able to cross this hard part.

Ben Rapoport

Yeah. Let me give two examples of how this is helpful.

One is MRI, magnetic resonance imaging. If you had told somebody in the early days of MRI, “You’re going to go into a superconducting magnet that costs many millions of dollars to build, and it needs to be maintained with liquid helium at four degrees, and you’re going to get a high-resolution picture of your brain because of it,” they would have said, “What are you talking about? That’s crazy.”

Maybe they had heard of nuclear magnetic resonance being used to study individual molecules for chemical structure. But if you told them, “I’m going to put my brain into that, and it’s going to give me an image of it,” they would have said, “You’re crazy,” because you can’t scale that up. How are you going to get the resolution? Too expensive. Crazy. Okay.

But over time, at scale, MRI is now an everyday reality that you can pay for. It is cost-effective. It has totally changed the way we diagnose diseases of the brain, and other parts of the body.

Juan Benet

Saved millions of lives if not tens of millions.

Ben Rapoport

Absolutely. And by the way, within probably a mile of us, there are probably a dozen MRI scanners. So it’s accessible to everybody. Insurance pays for it. It’s cost-effective. People know how to run them. It’s push-button, in a sense.

So at the beginning, in 1970, you would have said, “That’s insane. No way.” And who needs it, because there are other technologies that allow me to diagnose all of this stuff? So that’s one example. Maybe a little mundane, but nevertheless the technology behind MRI is nontrivial. It’s a lot of physics, electrical engineering, and understanding of biology and chemistry that goes into making that a real product.

Now, is it a consumer product? Not quite. Is it a medical standard of care that’s universally available, not even just in the developed world? A hundred percent. So it takes time. But that reality can come.

Another example that is a little more directly relevant is genomics, gene sequencing. This is something that, in your life and in mine, totally transformed medicine in ways that, at the beginning, in 2000 or 2001, would have seemed impossible. It cost many millions of dollars to sequence the first human genome. Many people were involved. It involved chemistry and computer science and biology, and a massive infrastructure and logistics to get all of this done.

That’s interesting because it’s a biological code. Once the code was read, the tools of computer science could be applied to do something in biology that was never possible before. And I think in neuroscience we’re seeing something similar. Not exactly analogous, but similar. The availability of data in a format that is portable, relatively uniform, and that packages neural data in a way that makes it amenable to the tools we’ve developed for other things, including image and video processing and modern machine learning in a highly parallel compute world.

Unlike the linear and fixed-in-time structure of the genetic code, the neural code is basically a two-plus-dimensional code. It unfolds on a planar surface at a timescale of a thousand hertz, on physical scales of tens of microns, hundreds of microns. But we now have a technology that interfaces on those spatial and temporal scales. We have the ability to store data on that scale. And the Precision technology, we’ve learned, is easily applied in many contexts to harvest data, save that data, and compute on it.

So I think we’re starting to see the edge of an exponential. It’s hard to say that 50 patients represents exponential growth. But if you think that, as of 2024, the number of patients who had ever been implanted with a brain-computer interface was about 75, so 50 patients is almost like a doubling of that, in one company in two years.

I don’t want to predict too much, but I think there will be a lot more patients in a year’s time. In one sense, that is a huge amount of data compared to the petabytes of data you might get in other fields, it is a thimbleful, but it definitely starts to add up over time. It shows you that there’s significant growth here.

Juan Benet

Let’s get into the neuro data and AI segment. As you collect data and use the device to map and understand what different signals correspond to, you’ve talked about being able to aggregate this data across a number of patients over time to build more robust models of what’s going on and then enable more patients.

We’ve already had some good results where, in a range of applications, it might take a lot of time for one person individually to train to use a particular BCI. But if you’re able to aggregate the information across many such users and build better ML models, you can cut that down significantly, to the point where you’re able to extract much better signal from brains in general, and supporting each additional patient is much easier.

Plus, you might also be able to decode other kinds of signals that previously were super difficult to decode, or that people did not think you could decode.

So where does that vision take you, and what are you thinking in terms of scale here? How many people might this take to start yielding really strong results, and what path are you thinking about taking over time?

Ben Rapoport

Absolutely. To both directions that you articulated, I would say yes.

Direction one is can we achieve accurate neural decoding more quickly in patient n+1 on the basis of patients 1 through n and what we have learned from them. The answer to that is absolutely yes. We are already seeing evidence of that.

By the way, that is incredibly powerful and not necessarily a given. It is easy to see in 20:20 hindsight, but it was not. In the early 2010s, when every patient was an n of one and a lot of time was spent calibrating and recalibrating these systems, for those who know the field closely, it was not a given that you would be able to build generalizable models that would be able to decode neural signals in a patient’s brain when they had never been seen before, sort of off-the-shelf models.

An analogy that makes it seem obvious is speech recognition, where Siri or Nuance PowerScribe, or whatever your chosen system is, works out of the box and works better the more you use it. So it is true of brain-computer interfaces, and we have proven that now.

To your question of how many it took, we got an inkling of it in the first handful of patients. But by patient 10, I think there was a sense that we were seeing some things that generalized, and we were able to show that to our own satisfaction. Of course, more is better, in a logarithmic fashion, like you would expect from image recognition and other areas.

So definitely, and that is really important from a product standpoint. It is not just true that you get things that generalize. It is also true that more data improves performance. Accuracy in particular, and in some cases speed, but definitely accuracy, benefits from having more data. That is not a surprise, but it is really important. Then, to your point of whether we learn that we can decode things that we did not think we could decode before, I will say yes to that, but I do not want to go into detail.

Juan Benet

Interesting.

Ben Rapoport

For the future, but it definitely is an extremely exciting area for us, and more on that will emerge over time.

I will say that genomics is a good example here. This is something that we’re really trying to do somewhat within Precision and definitely external to Precision with our partners. If you think about what we can do with the data and what we can do to move the field forward in a federated fashion, we cannot do everything ourselves. And we want to position the community for success because everybody is going to benefit from this.

This really means thinking about what it takes to map the electrical activity of the neural code across people’s brains, across time, and across function. This is not as discrete and finite a problem in neuroscience as it was in genomics, but there are many analogies.

In genomics, you kind of have the elements of your genetic code, your C, T, G, and A’s, and all the annotations there and then you have phenotype. Phenotype could be traits or diseases. We have all kinds of things. In a sense, the phenotypes are infinite, or certainly a large finite number. Those are your annotations, and you’re trying to map sequence elements to phenotype. But that sequence is linear and fixed in time.

In the neural code example, we think of it as a two-dimensional dynamic code. It is two-dimensional and unfolds in time. What are the annotations? From an ML standpoint, you want to make sure that the data is high quality, relatively uniform in some sense, and that you can compute on it using primitives and techniques that are relatively familiar. But it also has to be annotated. You really have to learn against something.

For most of this conversation, the implied annotations were related to movement or language, moving my hand in a particular way, making a particular gesture, moving my arm. We can talk about making a gesture, but how do we define that? Do we define movement in three-space? Do we define it in some other way? But you get the general idea that there are some elements of movement that are the dynamic, in-time annotations of the neural data.

But there are also other annotations that we can provide. Those may be disease-specific annotations, like whether somebody is having a seizure. There are certain ones that we care a lot about. Speech and language are some. Memory, vision, and sensation are others. And there are many more.

Juan Benet

You can probably find precursors to certain problems.

Ben Rapoport

Yeah, exactly right. So degenerative disease: is somebody going to develop Alzheimer’s disease? Is somebody going to have a memory problem? You can think of any number of things that you might want to annotate.

Those may be transient states, like movement is a transient state, or they may be less transient states, like a disease state. Some disease states are very permanent and are there all the time, and some of them, like an epileptic seizure, are also transient. But maybe there are features of abnormal activity that are present.

So part of what we are trying to do by mapping this space is to map the electrical activity patterns, label them on the anatomy of where the electrodes are receiving data from the brain, and on top of that layer in what these features are, speech, movement, and so on.

And to do that in a way that is curated, well structured, and accessible to people. I think we are in the early days of that. We have to build some of that pipeline for ourselves, and we are building a very limited version of that for our core products. But we are starting to build that in a more general-purpose fashion with some of our collaborators.

I think that is a tremendously exciting resource for the community. And I think it will be something, if we do it right, that really will be as impactful in this space as the genome was in cancer, infectious disease, the understanding of human origins, and many other things.

Juan Benet

Yeah. I would imagine also the data will be able to be aggregated across many different device types over time, and it’ll just kind of contribute to higher and higher quality models.

Ben Rapoport

So I hope that’s the case, and I have a little less insight into that. I know that you guys at Protocol Labs have been interested in trying to feel a way forward for the field to collaborate in a way that allows multiple platforms to interplay. And I think that’s a good aspiration.

I’ve spent a lot of time thinking about how we can do that for data coming from the Precision system. I’ve spent less time thinking about how it could be done for a general-purpose device. But I think there are ways. There are some open platforms now that people use for multiple device types.

Juan Benet

Yeah. And this might follow the path of other ML modalities, where certainly the data from the same device will be very useful for being able to use that device. But there will be a set of applications that become possible or useful once you start aggregating of these different modalities.

Ben Rapoport

I definitely see it from that perspective. If you have data on a problem from multiple sensor types or multiple device types, you should be able to learn about that specific domain from the different sensors. It does not matter what imagery you use if you are taking pictures of the relevant problem.

Being focused on my problem, it is not always clear that data gathered with somebody else’s sensors in my area of the brain. The truth we’ve learned from a lot of what other people have done.

Juan Benet

I’m sure it may be the sort of thing that only turns out to be very useful down the road. So first we have to go through this large collection process and start using it for the shorter-term applications, and then maybe later it becomes really useful.

Ben Rapoport

Yeah. There are quite a few areas, medical imaging and genomics and well beyond medicine, where federated approaches to sharing data, harmonizing data sets, and making sure that annotations are uniformly applied in a high-quality way have been a huge boost to the field. All of modern ML basically has groups like that to thank.

Juan Benet

Yeah. And this might also be a case where the Precision technology can be used to collect tons of data that may not directly lead to short-term applications, where the data that you have is maybe not useful enough or sufficient. But then down the road, it turns out that you’re able to decode various kinds of information that people just didn’t think were at all going to be possible today with this type of technology.

Ben Rapoport

I would not be surprised. We’ve already surprised ourselves in that sense. Sometimes you do something by accident that you didn’t realize.

Juan Benet

Yeah. It seems to me there are areas of technology where, once you start opening a bottleneck and you get the ability to start instrumenting and get a lot more insight into the problem, all kinds of things turn out to be way easier than people expected. Certainly still super hard, but certain things just become possible that people did not think at all were possible.

Ben Rapoport

That’s definitely true. And I think one of the things, in a general sense, in biology, in cancer biology, infectious disease, and other related areas, is that we did not necessarily know it at the time, but it turns out those areas were really data-impoverished.

The data were very difficult to compute on. And now we have discrete, universally agreed-upon data standards. They are digitized and amenable to compute.

Neuroscience is now coming into that domain where what’s happening at a fine-grained scale in the brain is accessible. It’s accessible in a way that is digitized and can be computed on.

Up until five years ago, if you wanted to get your hands on a dataset of what’s happening in the motor cortex or in the sensory cortex when somebody is doing something pretty sophisticated or thinking, there were only a few datasets like that in the world. And now we have almost 50 of them.

So I think you’re right, and that will scale. I think part of what we’re doing here is, because we’re at the beginning of it, it may be hard to recognize, but part of what we’re doing here is digitizing neural data in a way that has never been possible before. And in doing so, making it amenable to compute, and making some of the problems that seem like biological problems, or even impossible-to-articulate or impossible-to-crack problems, into problems that can be framed as a computer science problem.

That’s what happened in genomics. That’s what’s happening now. And some things that are very hard in the biological domain, like developing a vaccine, in the computer science domain are tractable.

So I think that’s what we’re going to start seeing now. But we have to be deliberate about how we handle the data, curate the data, and make them available. I think you’re right that some of the things we do now, if we’re careful about how we do them, will open doors.

Juan Benet

Yeah, it’s super exciting. The next 5 to 10 years of development and application finding may turn into being able to interface with computers, AI, the internet, and each other better.

It seems like the Precision technology will be able to do more than just restore basic mobility and the ability to navigate basic computers and digital life. It is also arriving at a time when the nature of interfaces is going to radically change as AI becomes much more capable.

So this opens the door for a whole range of other applications and uses of BCIs, where it becomes less about, or in addition to, all of the repair use cases, and more about being able to use a different range of interfaces to interact with AI, or with each other, and so on.

There are a whole bunch of possible use cases that have maybe been explored in sci-fi and in various vision-paper-type settings. It really feels like we’re on the cusp of that range of use cases becoming possible.

How much do you think about this in the long term, and how do you think this might play out over the next 10 to 15 years?

Ben Rapoport

Yeah. I find that a little hard to predict, and maybe I’ll be criticized for not being imaginative enough. But the way I see what has happened in the past being relevant to what is happening now is this: if you think about the way electronics have interfaced with the human body over the last several decades, things that were avant-garde a generation ago are now just part of everyday life.

Examples of that are implantable cardiac devices like pacemakers and defibrillators. A generation and a half ago, that was science fiction. Then it became an ordeal, but something that was possible. And now you may have a relative, or maybe several relatives, who have such devices, and it is basically an office procedure. They have a computer that controls their heart, and nobody even knows about it. They forget to tell the doctor that they have one. And by the way, their cardiac data is being uploaded and downloaded by their cardiologist, and maybe even by them on a smartphone.

So that became pervasive over time, and more highly functional, and almost kind of a lifestyle change for many people. It is basically a lifestyle choice at this point. It used to be a major surgery.

Cochlear implants are another example, maybe one that is a little closer to home for brain-computer interfaces. They were initially for people with severe deafness or congenital deafness, really a very medical type of application. They improved dramatically, and a big inflection point that was very influential in my own life was seeing the deaf community accept that parents of a deaf child could make a choice to have a cochlear implant implanted in an infant just a few months old.

To me, that signified that the technology was so robust, reliable, and effective that even a community that knows it has its own culture and language and grammar and literature could nevertheless decide that parents could reasonably choose to have an elective major surgery in an infant. And now, by the way, cochlear implants are used as effectively as a hearing aid. You have people who decide to have a surgery to have a cochlear implant implanted to essentially augment their hearing.

That process, going from cochlear implants coming into the world in the 1980s to becoming acceptable for people with a condition in the early 2000s, was 20 years later. Now we are another 20 years later, and you have people of our parents’ generation, or even our own generation, deciding to have a surgical implantation of a hearing aid device.

So that is the path this technology has taken, from explicitly and only medical technology, maybe an avant-garde medical technology, to kind of lifestyle modification. And there are other examples of that. Glucose monitors are an example. I could point to quite a few.

For all of these, in a way, this is the body interacting with the digital world. They are not quite what we are talking about in terms of interacting with modern artificial intelligence and the internet and so on. But they are advanced implantable electronic technologies interfacing with the world and shifting from something that is a medical necessity to something that is essentially a lifestyle choice.

Could I see something analogous happening in brain-computer interfaces? One hundred percent. We are at the very beginning of this process, where no matter what we do, opening the door to that possibility requires showing safety and efficacy in a medical context. So the next five years are devoted to that. We need to do that safely, effectively, and responsibly in order to open the door to possibilities for what may come next.

So it is fun to have these conversations about what may be here in 5, 10, 15, or 20 years. But I think what we need to do today to make that possible is clear.

Juan Benet

Yeah.

Ben Rapoport

Right. So the speculation is important. It’s fun, it’s interesting, and I think some of it is what we probably want to be doing now, because there is ethics, data curation, and some fundamentals that I think we probably want to put in place. No matter how the field evolves, I think we will be glad we did it right at the beginning.

But yeah, I could envision that. And by the way, people’s concept of what’s acceptable also changes. I’ve said this many times, but I have two kids, and they know a little bit about what I do. To them, it’s normal. So the next generation, the people who are five years old today and who are going to be the young engineers, physicians, and computer scientists coming of age and deciding what to do, when there is already a platform for development in this space in 5, 10, or 15 years, their concept of what’s possible, what’s normal, what’s acceptable, and what’s taboo is going to be very different from ours.

Juan Benet

Yeah, absolutely. But it sounds like you still imagine that taking many decades to develop, whereas a range of other folks in the field, and I personally, think it might happen a lot faster because the different piece now is that we have a much more online world, where people are interfacing with these systems much more actively, like the diffusion of something like ChatGPT, going from zero to “hey, this is possible now,” to percolating through society, happened in a few months to a year and a half or so, depending on how you look at it. But now it is widely used.

Ben Rapoport

Totally. This is an interesting conversation to have, because the diffusion of different technologies is limited by different things. Something that is purely an internet-based software application can diffuse globally in weeks or faster.

Juan Benet

Yeah. And obviously, the whole process of making sure that a device is safe to implant, which takes a long study process and so on, makes sense. What I’m getting at more is the level of shift that humans might go through over the next 10 to 20 years as AI becomes more capable and starts seeping into our societies and economies. There are a range of people now who are starting to have very significant and meaningful relationships with AI companion or counterpart. This is becoming a phenomenon now. So we’re on some shift curve that might end up being faster. I’m just kind of pushing on that.

Ben Rapoport

Yeah. I think that’s definitely true. There is no question that there are predominantly software-based, and wearable- and portable-device-based applications of artificial intelligence that we are all experiencing. You and I both experience the impact of changing AI tools many times a day.

So in the wearables and software world, that is a definite reality. I think the question is how implantable brain-computer interfaces play into that. I think there is maybe a slightly different timescale involved.

But certainly, once there is a cohort of people with implanted technologies, the ability to improve what is possible using the technology is an economy of scale. This hearkens back to something we were talking about earlier in the conversation, which is that the more people we are learning from how to decode, the more everybody benefits from our ability to decode effectively.

So there are a number of these kinds of economies of scale that I think we will see. And certainly, when you have a hardware platform that people are using, whether it is wearable, implantable, or something else, then the software that lives on that, or lives partly on that, whether that is AI-based software or anything else, will impact their lives at a rapid scale, at a scale that we have not seen before in other areas of medicine. It will be more familiar to pushing out software updates.

Juan Benet

That’s a great insight. There’s this hard bottleneck now of establishing the first set of interfaces that can actually be worn at length, are safe, and work really well, and where the entire product experience of having this device implanted, used, and removed, all of that lifecycle, is really figured out.

And once you’re able to do that, then a lot of this turns into primarily a software problem or at least a range of applications become possible is just purely software problem.

Ben Rapoport

I think that’s accurate. Exactly what that looks like and how much time it takes is not totally clear. But I think we can learn some things from other implantable medical devices. The nature of what a brain-computer interface does is obviously different from what a cardiac device does. It’s different from what a cochlear implant does. It’s different from what a glucose monitoring device does. But the interaction between software and an implanted medical device can give you some insight into that. Once you’re on the platform, the ability to adapt, program, and change is dramatically different.

Juan Benet

What first got you into neuroscience? What attracted you about the field?

Ben Rapoport

Well, I’m kind of a child of the field. My dad is a neurologist who also has both a PhD and an MD, and he’s an expert in clinical electrophysiology. His career kind of came of age with the beginning of clinical electrophysiology, at the very beginning of when electrodiagnostic tools started to be used in patients. Of course, the link between electrical engineering and neuroscience goes way back, almost to some of the pictures that are on our walls here. It really goes back to the beginning of the 20th century. So the electrophysiologic nature of the brain and nervous system has always been a part of neuroscience. But until the 1960s and 1970s, it was more a part of research neuroscience than clinical neuroscience.

My father started his career as an electrical engineer and became a neurologist, and his specialty became, and still is today, clinical electrophysiology, at the beginning of that era in the 1970s. So I grew up with that as a backdrop to my life. And what was emerging when I was in my late teens and early twenties was next-generation electrophysiology, which today brain-computer interfaces are a part of.

Juan Benet

When you think about the field of neuroscience, or neuroscience and neurotechnology together, what are some of the grand visions that the field has been trying to chase? Which of those have come true and been realized, and which ones are still forthcoming?

Ben Rapoport

Well, I think broadly speaking, the interest of neuroscience has been to understand the human brain. That is kind of a grand vision, and the specifics of what that means can be boiled down to different targeted questions. But in a broad sense, that is the grand vision.

In clinical neuroscience and medicine, the question is often: how can you understand what is going on in the brain in a way that heals a brain or a nervous system when something is not quite right? That means you have to understand the brain in the normal state, and you have to understand the disease process or the disorder well enough to address it and treat it in some way.

My focus growing up has always been that part, really understanding the brain in and of itself is, of course, super interesting. It is endlessly interesting. I do not think it will ever stop being interesting. To me, it is the most interesting thing in the world. That has driven my entire life and passion.

But more specifically, trying to understand when something is not right in the brain or nervous system, how do we fix it? That has been my lens.

And so the grand vision changes a little bit in a generational context, as sometimes problems become more important to a generation for one reason or another, for historical reasons.

In the early 2000s, various forms of paralysis became very important in the national and international conscience, partly because people were surviving injuries that were not survivable in other generations. And partly it was a technology push. So sometimes it is a clinical pull and a technology push.

There are various forms of paralysis that exist today that stem from stroke or spinal cord injury, or diseases like amyotrophic lateral sclerosis, so-called Lou Gehrig’s disease, which is a neurodegenerative disease. So paralysis ends up being an incredibly interesting lens through which to try to understand the brain.

A lot of the way we interact with the world is through movement, even in ways that we do not totally understand, because speech is movement too. That became a road into the modern world of brain-computer interfaces. So one of the grand challenges is: how do we treat paralysis?

Juan Benet

And how did you first become interested in building technology to treat these conditions? Were you encountering a lot of these cases as a neurosurgeon and then thinking about the problem space and realizing, “Hey, actually we might be able to start building some devices that can help patients here,”

Ben Rapoport

For me, it actually started in reverse, and I think this is true of a lot of people in the field. In the 1980s and 1990s, it was both a pure science push and a technology push, and some of what was being discovered then opened up possibilities for treating disease.

That is true of the beginning of the field of brain-computer interfaces. In the 1980s, it really was not understood how movement gets computed in the brain, how it gets planned, how it gets executed, and how the brain coordinates fine movement of the hands, the limbs, and so on.

One of the major discoveries in the field at that time was that fine motor control is coordinated by populations of neurons, not by one neuron at a time, and not by one bulk aggregate of brain tissue, but by populations of neurons firing in coordinated patterns to basically tell the muscles what to do in a time-synchronized fashion.

That set of discoveries was made in the 1980s, and those mechanisms were elucidated at that time, partly with the help of being able to bring multiple tiny electrodes into the brain, simultaneously record from them, and begin to parse out what the time-varying signals from multiple tiny electrodes meant.

So that population encoding became the basis for a whole movement into brain-computer interfaces. Really, it was basic science at that time, the basic science of how the brain coordinates movement. And early on in that process, people started to envision that if we can understand this, maybe this is a way to treat paralysis.

Juan Benet

In that case, these sets of neurons that are firing together, are they firing in parallel through different downstream sets of connections? Or are they ending up being joined by other neurons?

Ben Rapoport

One of the nice things about movement is that it is kind of an isolated system in the brain. Much of what happens in the brain is feedback among multiple circuits that interact with one another in ways that lead you to ask what is the chicken and what is the egg. It is very hard to untie these feedback loops and really understand causality.

Of course, there are some nuances to all of this, but in the coordination of movement in the brain. People who are knowledgeable in this area are going to listen and criticize, but let’s simplify. There is an area called the primary motor cortex, or M1, and that is located in one of the hills of the brain, right around here. Neurons in that area of the brain basically project long axons that go down the spinal cord and make one synapse to another neuron that actually connects to a muscle.

Juan Benet

Then to a single muscle fiber?

Ben Rapoport

Basically, yes. If we’re getting really technical, they’re called layer five pyramidal neurons. They’re large neurons. They’re called pyramidal because they have a pyramidal shape, and they project long axons, a meter or so long, that go all the way from the surface of the brain down through the brainstem and spinal cord. In the spinal cord, they connect with another neuron, an interneuron, and that interneuron then connects directly to a muscle endplate. So that one-two connection is what gives rise to movement.

Of course, one neuron firing is not enough to yield movement of a muscle. It takes the coordinated activity of many neurons firing together, synchronized in time and space, to generate a muscle endplate potential that contracts a muscle. But basically, that’s how it works.

On the sensory side, the body understands when a muscle has moved, and on the planning side, it understands how to move muscles in time to yield the desired movement. All of that is pretty nuanced. But the basic aspects of how movement happens work like that.

Then you get into asking, how do you actually move in a coordinated fashion to move a hand or an arm in a particular direction? Do you do it in order to achieve a result, or do you do it in a coordinate frame? Are you really moving in terms of kinematics, or are you moving in terms of concept? There are all kinds of questions that you can ask.

But the nice thing about movement is that a lot of the action happens in this one clearly labeled part of the brain that we can access and study. We learned in the 1980s that it was groups of neurons that gave rise to these movements, and that by electrically listening in on what they were doing, you could get a sense of what movement was being planned and executed.

The motor cortex, because of what I just described, kind of does it more or less in real time. The planning happens hundreds of milliseconds before, and the sensation and feedback happen 10 to hundreds of milliseconds after. But the real-time aspect of what the brain is doing to move is happening in the primary motor cortex. So you can listen to groups of neurons to understand what the brain is actually executing in real time.

That was a breakthrough in science. The corresponding breakthrough in engineering, more or less around the same time, was that whereas up until then neuroscientists had been listening to neurons, groups of neurons, muscles, anything electrical in nature in the nervous system, one electrode at a time, one amplifier at a time, there emerged the possibility to listen with many electrodes at a time.

There were many technologies used to do this, and part of it was that amplifiers became more readily available. Neuroscientists have always kind of been physicists and electrical engineers all rolled into one. Whatever was the data science of the era was also being used. Whatever was the computer science of the era was also being used.

When my father was starting out, everything was a physical trace of a pen on a rotating wheel paper script. Digital storage did not become commonplace until the 1980s and 1990s. So the ability to really apply what we think of as sophisticated compute did not come around until the 2010s.

Juan Benet

Wow.

Ben Rapoport

Fast-forwarding to where we are today, if you’re thinking about how brain-computer interfaces got to where we are today, it’s really because of the availability of multi-electrode arrays, of which there are different technologies that we can talk about, and the ability to store and compute in real time on all of that data, which didn’t come around until the GPU was basically a desktop reality and easily programmed.

Of course, it began with some understanding of the physiology. That was the first piece to emerge. The enabling technologies came next. And then, as you mentioned at the beginning of all of this, understanding that there were tools here that could be used not just to study the brain, but to understand how to try to heal the brain or the nervous system when there has been an injury or an insult of some kind.

That in itself is not just one enabling technology. That’s a team effort of many scientists and technologists. And even beyond pure science and technology, there’s policy and other realities in the world to actually try to get something from feasible to real and part of the standard of care.

And for me, going back to your grand challenge question, the goal is to have this technology available as part of the standard of care in a way that really touches people’s lives and helps to heal. And I think that is a palpable potential reality now. I think we’re very close to it.

We’re engaged now in clinical studies and with real patients who are either being helped or helping to help others who will soon come after them. And to me, that’s an incredibly exciting reality of today.

Juan Benet

Fantastic. Through studying the brain and learning about all of the advances in technology that could enable a range of things, it sounds like you then had the perspective and insight, and surely many other people in the neurotech field also thought this way, that you could finally start producing devices that could really help people heal. So from there to actually building one of these devices, inventing it, building a company, and so on, there were lots of steps. You co-founded Neuralink. How did you then go from there to Precision?

Ben Rapoport

Always thinking back through it, you try to make sense of it in a linear fashion.

But as I mentioned a few minutes ago, if you think back to what was happening in the mid-2010s, the first applications of machine learning were really starting to hit the public consciousness. One of the first, arguably, was Google Translate. People do not talk about that so much now. They talk about AlexNet, images, and computer vision. But one of the first overnight changes that happened was Google Translate.

Google Translate used to be terrible, almost a joke. And then almost overnight it was completely revamped to the point where you actually had high-quality translation across multiple languages. Very quickly, that moved from text-based translation to almost real time. There were apps where you could translate spoken words. This happened in the mid-2010s.

The reason I use that as an example is because, basically, in brain-computer interfaces, we are really solving a translation problem. One of the most important applications of machine learning in brain-computer interfaces is solving a translation problem between the electrical, electrophysiologic data streams that percolate in your brain and mine and everybody else’s, which are a similar grammar, but with different words and accents and so on. Nobody’s neural code is exactly the same.

So if you want to have a device that translates that neural code and makes it able to control other digital systems, you need to translate the individual neural code of somebody’s brain into the control output that you want to use.

Juan Benet

That whole generation of machine learning, including the transformer is critical in enabling this tech.

Ben Rapoport

Transformers didn’t even exist back then. At least not as, we think of it now.

Juan Benet

Certainly not the scale of transformation that we have now.

Ben Rapoport

But what was happening was that the last piece of the technological puzzle to fall into place was high-quality compute with an understanding of basic machine learning models that were obviously applicable to the translation problem that is neural decoding.

And with that, it kind of dawned on a number of people simultaneously, including Elon, that there was an opportunity to take what was a solution in the laboratory and make it into a technology that could actually do good in the world, and also perhaps have something to do with the way humans interacted with artificial intelligence in the future.

I think once people started to see a clear exponential growth in the capabilities of artificial intelligence and the speed of development of artificial intelligence, a number of questions emerged. One of them was, how is the human brain going to interact with artificial intelligence?

And that is still, as you know, a question that a lot of people ask with nuance and thought. What does that look like today, tomorrow, in five years, in ten years? And how does the technology we use to interface with artificial intelligence, the human brain interfacing with artificial intelligence, change over time? Is it fundamentally different from the way we currently use human-computer interfaces? Is it conceptually different? And I think there is this intuition that it is going to be very different.

So a number of groups at that time, back in 2016, Elon, Mark Zuckerberg, Brian Johnson at Kernel, and a bunch of others, decided to invest in trying to bring those emerging technologies out of the laboratory and into the real world, and to put significant investment dollars into doing that.

And I think there was also a consensus in the scientific community, which was small at that time, that the technology was as ready as it could be in an academic sense. A lot of the problems had “been solved”. And I say this with a laugh because it is never really quite the case.

Juan Benet

I describe it as R&D pipeline myopia, where no matter what part of the pipeline you’re working on, whether it’s the basic science, the invention, productizing, or even selling it into the market, the people working in that area encounter such large challenges that it feels like that’s the hard part. And that everything else is implementation detail.

But all of these parts are actually really hard. Doing the basic science is super hard. Doing the core invention is super hard. Doing the engineering is super hard. Building really high-quality products is super hard. Figuring out the strategy to deploy into the world is super hard. It’s all hard. One of the core things that I think ends up causing these super important technologies to be built and scaled is people who are able to span multiple segments in the pipeline and do the actual translation work.

Where this kind of breaks down is when you have such a division of labor that people do not span those areas well enough, and then things get stuck. There’s a lot of important science that is stuck in paper form, where the “hard problems of science” have been figured out, but there is no invention or downstream technology being built.

So by having that span over the R&D pipeline, you’re able to pull things forward, and you have the depth of knowledge of both the science and the technology and engineering to actually figure out what the core problems are to solve. So in that time, in that era, there were a range of people figuring out whether we could take this from the core science, and from the stage where things were, to now starting to build initial versions of products.

Ben Rapoport

“All the hard problems have been solved,” like you said, and all that was needed was capital to take it to the next step. That was not so far from the truth. A lot of really important problems had been solved.

Around that time, quite a few people who had been engaged in the science of brain-computer interfaces, which was a very small community then, much bigger now because of everything that has happened over the last ten years, pretty much everybody knew everybody, if not by one degree of separation then by two.

So if you were one of those people trying to get something started and bring it out of academia into the real world, it was not so hard to locate people who could help get things going.

Those of us who were a little younger at the time, and not established professors or established professionals, and at that time I was also single, the people who were most mobile, highest energy, and newly minted PhDs or professionals or engineers were the likeliest people to help get something going. That sort of describes me at that time in my life.

Of course, I was incredibly passionate about it. I definitely believed, as I do now, that there was something incredible to be done in brain-computer interfaces. And maybe I had, like you said, a slightly different skill set to bring to bear.

It obviously takes a huge team with a variety of different areas of expertise. There is the core engineering and the neuroscience, then there is a clinical understanding of how the systems work, and there is how the experimental neuroscience works, how you design studies around it, and how you scale things up.

There are so many aspects to how a team needs to be built to bring something that is a proof of concept into a scaled reality.

So yeah, that’s how it started. I was headhunted, and I was lucky to be headhunted and to join an incredibly talented team that grew very quickly.

I think it also became clear that the space was so huge, and the potential to do really high-impact things in the area of brain-computer interfaces was more than any one group could do with any one enabling technology.

So pretty early on, I also had a sense that to really scale the technology, to access multiple areas of the brain, to study those areas, and to deliver the technology to patients who already had some compromise to the brain, which itself is pretty delicate, the trade-offs you wanted to be able to offer had to emphasize, in my view, at least for some portion of the market, a technology that didn’t penetrate the brain but was also incredibly high in spatial and temporal resolution.

And so that’s how Precision was born. I left Neuralink in 2018, and basically in 2020, Michael Mager and I, who had known each other quite well before, got back together and started Precision with that vision of really bringing neural interface technology to the medical community to treat neurologic diseases that currently don’t have a treatment, with an emphasis at the beginning on paralysis, because that’s the clearest, most well-defined need, but with a view toward doing much more.

Juan Benet

What’s the long term vision of precision? What is the range of things that you wanna be able to do in the long term?

Ben Rapoport

I would say I want to be very concrete. We want to have, in the next five years and beyond, an effective technology for, I cannot exactly say treating paralysis, but something to offer people with various forms of paralysis to restore lost function in a really meaningful way that brings people who are not currently able to engage in the world in the ways that you and I take for granted back into that world.

For some people, that is about communicating with other people. For some people, it is about financial independence. For some people, it is about different things. But what we know we can do is deliver seamless integration with the digital world that, for many of us, is a major part of the reality that we take for granted. And for people with forms of paralysis, it is not.

So that is our goal over the next five-plus years. Very severe forms of paralysis, like spinal cord injury, will be the first that we address. But there are many forms of paralysis, and that includes paralysis caused by stroke, which is not always whole-body paralysis. It can be semibody paralysis, or it can even affect parts of the body. And there are other diseases of the nervous system that cause paralysis.

We have developed, over the last few years, a much more nuanced understanding of how forms of paralysis affect people in their lives, families, and communities, and the people around an individual who suffers from a paralyzing condition. So that is our five-plus-year goal.

At the same time, we have discovered that the Precision technology represents a platform for understanding and studying the brain, and for interfacing with not just the motor areas of the brain that we started this conversation on, but really, essentially, the entire brain.

We have now implanted the Precision technology in more than 50 people, and the electrode technology has touched parts of the brain that almost no other single-electrode technology has touched, everything from motor and sensory areas to prefrontal cortex for decision-making, to inside the valleys, inside the sulci of the cortex, which is the thinking part of the brain, the brainstem, the spinal cord, and areas of the brainstem that are very difficult to address.

So what we have discovered is that we have basically built a platform for discovery, diagnosis, and intervention. I think it is definitely the case that we cannot even envision today. We have a pipeline of use cases that we are working on, but I think that the possibilities are very significant. I won’t say endless, but like very significant. And we want to enable others to build on that platform.

Juan Benet

And maybe as you look ahead into the future, what is your optimistic vision of the future? The stories that tend to be told most often have high drama, which is why popular fiction ends up with highly dystopian future worlds, because that sells much better. So it is kind of up to us, and a lot of other people, to paint really positive visions of the future so we can aim for them.

What are the kinds of things you think about when you think 30, 50, or 100 years out? What is possible that you are working toward, that you see other people working toward, or that you want to inspire people out there to work toward?

Ben Rapoport

Let’s try to project a generation out, even though that is hard for me. If I think about what I am able to work on relative to what my father was able to work on, and what I think my kids will see, I think we are going to change the nature of what it means to have certain kinds of disability in society and in the world. I think that is going to be incredibly powerful.

Exactly how we change that, and exactly what we make possible, is hard to say. But just as cochlear implants have profoundly changed what it means to be hearing impaired, I think brain-computer interfaces are going to have a profound impact on what it means to be physically disabled in society. That is a massive change.

I think we are going to see that in the next decade. Are we going to cure all forms of paralysis? No, not in 10 years. But are we going to change that significantly? Yes, I think in the next 10 years.

Also, in 10 to 20 years, we are going to have brought forth a very powerful technology for understanding how the brain works. I think some problems that seem hard now are not going to be problems people care about anymore in 10 or 20 years, because they will be taught in grade school. That has been the case with genomics, and I think it is going to be the case with certain problems in neuroscience.

Some of the things that we think of as profound, like questions of consciousness and questions of neural computation, are just going to be things people understand routinely. Even my kids know how to query some of the AI software now.

So I think it is going to be an amazing step forward that some of this technology unlocks. I think there is going to be a big impact on disease. There is going to be an impact on understanding medicine and neuroscience. And whatever application layer is built on top of that, I think part of that is for us to enable, and part of that is for another generation of scientists and engineers to contribute to.

But I have a very optimistic view of the future, and I really do not worry too much about dystopian future states of neurotechnology. There are a lot of thoughtful people trying to build this in an appropriate manner, and I think we are going to see tremendous strides forward in a number of areas.

Juan Benet

Thank you very much. This was a phenomenal conversation. Thank you for spending the time and good luck with all of the work you’re doing.

Ben Rapoport

Thanks for being a partner in this and thank you for having me on the show.

Juan Benet

I hope you enjoyed this episode. This is a new podcast, so we need your help to get the word out. Like rate and subscribe on your favorite platform and please share it with people you think would find it interesting. Thank you and see you next time.

In this conversation, we go deep on how both technologies work, how PRIMA restores vision, how the biohybrid BCI connects to the brain, what the next milestones are for neural interfaces, and what it would imply to add a new functional brain area to a human.

We also dig deep into how Max built and leads Science: his founder story, how the team drives Fast R&D, and how the team is able to speed through high-uncertainty, high-impact projects.

Hope you enjoy!

Watch on YouTube.

Timestamps

00:52 What counts as neurotech?

01:45 History of brain-computer interfaces and the iPhone dividend

07:25 PRIMA - How Science is restoring vision in blind patients

10:10 Why stimulating bipolar cells works when the optic nerve doesn’t

30:30 Are we bottlenecked by biology or engineering?

32:40 Expanding the brain’s bandwidth beyond 10 bits per second

37:00 Can we add new areas to the brain?

37:46 Biohybrid BCIs: neurons growing on a chip

39:20 What could neural augmentation look like?

01:13:20 How Science drives Fast R&D

01:44:00 How founders learn and level up

Referenced Links

• Woman hearing for the first time

• PRIMA Visual Prosthesis

• PRIMA patient filling out crossword puzzle

• PRIMA a global mission to restore vision

• Biohybrid BCI

• Patrick Collison FAST post

• iPhone launch keynote in 2007

• Pantheon on Netflix

Links

Juan Benet on X

PL Neuro: plneuro.xyz

Protocol Labs: protocol.ai

Transcript

Juan Benet

Today I’m interviewing Max Hodak, an extraordinary founder, scientist, engineer and investor at the frontier of neurotech. Max is the founder and CEO of Science Corp, where he and his team are developing breakthrough neural technologies, devices and platforms. Today, Science is focused on the PRIMA visual prosthesis, which literally restores sight in people with some forms of blindness, and they are also pioneering a biohybrid brain-computer interface, which I’m personally most excited about.

On top of that, Science Corp is building software and hardware platforms to drop the cost of neurotech R&D, enabling other companies and a new generation of startups. Before Science Corp, Max co-founded Neuralink and Transcriptic, now Strateos. His work cuts across neuroscience, engineering and philosophy, all united by the questions about how to expand human capability.

This interview is part of a series of neurotech, where we explore and discuss near and future breakthroughs, why they matter, and how they will benefit humanity. Let’s dive in. What exactly is neurotech?

Max Hodak

I mean, in some sense, neurotech is this incredibly broad idea because everything about our behavior is rooted in the brain. And so, I mean, there’s some interpretation where social media and the iPhone is kind of neurotech, but I think when we when we talk about it specifically in the context of brain computer interfaces, you’re talking about mostly implantable devices, some wearables, but instead of it’s — if you’ve lost the ability to use your hands or use your eyes or use your ears, maybe we can we restore those sensory or motor capabilities, can we bypass a broken, broken nerve or a lost function?

And there’s some groups that are starting to look ahead and we can understand really, how does the brain do these things and how does the brain form percepts and how does it reason and how does it remember. Then you can go from restoring to potentially extending.

Juan Benet

Can you maybe start with a quick history of the neuratech field, recently, like in the last few decades? From your perspective.

Max Hodak

A lot of what I consider brain-computer interface started in the ‘70s, the ‘60s and ‘70s. It was pretty quickly figured out that if you put an electrode in the brain of a monkey, if you got it to move a joystick, you could figure out how in the brain that was functioning. I mean, people have been doing this in humans really since the ‘90s.

So in the mid-’90s, the Utah Array was invented, which is a very classic neural probe. It’s like these silicon-like needles that can be placed into the brain to record neural activity. And in the late-’90s, early 2000, people were decoding cursor control and basic keyboard control in humans. Now that system is not really ready for prime time then, that probe in particular gets an immune reaction and it’s a relatively small number of electrodes. It doesn’t really match the material properties of the brain in some of the ways that you want. But kind of more to the point, it had a connector that came out through the scalp. So there was a pedestal that was anchored to the skull and then a connector coming out through the skin.

And the skin is a very important immune barrier. You really want the skin to be closed because if it’s not, there’s a risk that a bacteria could crawl down your connector and potentially into the brain, and then you’re gonna have a really bad time. And so a lot of the advancement over the last 10 years has just been miniaturizing the electronics and making them low power enough that they can be fully implanted, and the skin can be closed.

I think the fact that this is possible now is something that’s been built on the back of what we call the smartphone dividend. The fact that Apple and Samsung and others have poured tens of billions of dollars into building the modern electronics required to do this, and then we get to take advantage of. But the neuroscience is really not that different from what’s been understood. The basic neuroscience of motor decoding has been understood for decades. It’s the advances in the electronics and the materials that have led to the modern wave of new BCI.

BCI originally, I’d say, came out of neuroscience. And the field of neuroscience is concerned with things like what our brains, and how do they work. Like there’s a bunch of circuits in the brain, some that deal with affect or attention or motivation or memory or reasoning or motor planning or whatever. And if you want to understand what’s going on in there, then at some point you need to be able to see into the brain. You need to be able to record and detect all this activity, and the endeavor of neuroscience has been placing wires into the brain to record neural activity, basically since the discovery of neurons. And so BCI originally came out of neuroscience, and the tools that were being used in BCI were essentially neuroscience tools.

And I think only now we’re starting to see some new types of BCI probe approaches that could be used to understand things about the brain, but really begin to diverge a little bit. So, for example, the biohybrid work that we’re doing is, I think, a great example of this. With the biohybrid probes, we don’t place like a wire or an imaging system or a fiber into the brain. What we do is we load a device with living cells, with biological neurons, and then we make it so the cells don’t go anywhere. They’re stuck in the device, and then we can graft this into the brain. So the living cells project new biological connections. They form new chemical synapses with the host brain.

But they do this in a perfectly biocompatible, nondestructive way because the brain is a bunch of neurons. And so if you add some more neurons, they grow in and wire up. And this can be an incredibly powerful BCI technology, but it doesn’t necessarily help you understand exactly what those cells deep in the brain are doing, because you don’t necessarily recover the original representations.

You can in some ways, but you don’t know, like you’re kind of getting a bunch of information mixed together back at your device, or the way that you stimulate it to activate the brain is a little bit different than you would with, say, like a wire. And so I think we’re going to see a little bit of a divergence between neural engineering and neuroscience, even though they’re obviously highly synergistic.

Juan Benet

In some of the early neuroscience and neuro engineering successes, maybe led to things like — would you say a cochlear implant is like an early BCI?

Max Hodak

Totally. Like, absolutely. We include cochlear implants and retinal implants as brain-computer interfaces. The cochlear nerve is part of the brain. And I think a thing that isn’t widely appreciated is there’s already like a million, over a million people that are walking around with brain implants out there.

Yeah. here’s a lot of cochlear implants, and there’s also a lot of deep brain stimulators. So one of the biggest effects in medicine is a Parkinson’s patient who you turn on their deep brain stimulator for the first time, and they go from unable to hold a pot-like a thing of water to just steady.

Juan Benet

Wow.

Max Hodak

And it’s not — the disease will still degenerate. It’s not curative, but it is a huge quality of life improvement. It’s a huge effect size for a meaningful period of time for many patients. And yeah, I mean, we joke that it’s like the kind of the best patient testimonials or may your product demos be like a newborn having their cochlear implant turned on or DBS stimulator turned on for the first time.

These are effect sizes that you just don’t see in medicine, by and large. Like you do in a couple of cases. But this is one of the reasons I really like working with the brain — because when you have something that really works, the effect size can be huge. We’re not talking about extending a poor quality of life by two months in a way that is statistically significant, but not really meaningful to the patient.

Juan Benet

Yeah, there’s an amazing video. I’ll link it in the description of this where a woman has her cochlear implant enabled for, I think, for the first time, and she suddenly starts hearing people for the first time. And it’s just this amazing experience of expanding her sensory perception.

And it’s just like this wonderful testament to the good that all of these technologies can do. So with that, let’s get to the PRIMA device. So first, what is the problem that you’re trying to solve?

Max Hodak

There have been cochlear implants for decades. And even though things like those are amazing and are widespread, there’s still limits. By the way, I think better cochlear implants will be coming out — like they’re good enough to hear voices. They’re not necessarily good enough to take to a concert. And so there’s still ways to really expand applicability. But when you look at the retina, there’s nothing like that. There is no cochlear implant of the retina. And that is a huge unmet medical need. So sensorineural hearing loss is more common than some of these forms of blindness because the immune system learned a long time ago not to overreact in the eye because that is like really bad for the organism, but similar to Alzheimer’s, age-related macular degeneration is one of the things that many, many people are most worried about in getting older. And for these, there’s what we call outer retinal diseases, where the rods and cones, the the photoreceptors, the light sensitive cells in the back of the eye have died, but the optic nerve is still intact and the brain knows how to see because they they saw most of their life.

For these photoreceptor degeneration diseases, there’s millions of patients and there’s nothing available. So as a company, we’ve looked at a range of technologies for restoring vision in the retina. And we have a state of the art optogenetic gene therapy. And then we also have an electrical stimulator. The gene therapy is a little bit further away that needs to start clinical trials. And will take some time to really develop, but might be capable of really powerful things. And today we have a product called PRIMA, which I mean not to undersell it — the results are amazing. They finished a major clinical trial last summer. It’s the first time, to our knowledge, that some of these patients — the trial was in age-related macular degeneration — that some of these patients have been able to read again.

And I saw a video recently of a patient recognizing faces, which was a thing that some of the investigators weren’t even sure would work with this generation of device. We have new versions of the PRIMA implant already in development. We actually just got some of the first devices of the next generation back a month ago, and the next step there is to shrink the electrode size so that you can target fewer cells at a time.

But for the current version of the device, I mean, there’s clips online of patients who cannot see, definitely cannot read natively, filling in crossword puzzles. And that had never really been done before.

Juan Benet

What does a person with macular degeneration see? What is their experience? What is the form of blindness do, and then how does the device fix that?

Max Hodak

There’s three layers of cells in the retina that kind of matter. There’s the rods and cones that are light sensitive. There’s 150 million of those. Those connect to about 100 million bipolar cells, bipolar because they’ve got two add-ins. And those then can connect to a much smaller number, about 1.5 million retinal ganglion cells.

Those are the optic nerve. So the optic nerve cells reach all the way out from the eye, deep into the brain. And previously, many groups had had really primarily focused on stimulating the optic nerve cells. But the problem is that because the image is already so compressed, you can’t just excite them within, like a camera image that that doesn’t produce what we call like a form vision, like a form percept in the kind of in the mind’s eye that just produces these flashes of light that are called, that we call phosphenes.

Juan Benet

So that means the documented description is like somebody’s visual field when you stimulate too deep into the optical nerve, they have these flashes.

Max Hodak

You get these flashes of light that like, if you see an artistic representation of it, if you see a picture of a field of phosphene, you might see like, oh, there’s like a woman’s face or I can see letters from the flashes. But the way that the patients actually perceive this is really different than you seeing the artistic representation of it through your photoreceptors.

The brain does not piece phosphenes together into a really coherent whole. And so like ten years ago, there was a company that got a retinal stimulator to market. They sold 400 of them before it was eventually withdrawn. And they were able to stimulate these fields of phosphene onto the optic nerve cells.

The best patients could kind of read in the sense that they’d be like, oh, here’s a line, here’s a line. It’s connected — that’s an A. Here’s a shape. It’s bridged this other shape — that’s an N. But it wasn’t like they were reading off like sentences from a book. It was an empirical finding of this clinical trial that if you stimulate the bipolar cells. So the first way that our device is different is we stimulate that 100 million bipolar cell layer, not the 1.5 million optic nerve cell layer. And this requires us to place our device under the back of the retina instead of on the interior surface. If you go in through the eye, ironically, the layer that is most forward is not the rods and cones, it is that optic nerve cell layer, which is part of why you have a blind spot backward. So the blind spot is the point where the optic nerve dives and exits the eye. So our device sits under the retina, stimulates that first 100 million cell layer. And it’s an empirical finding of this trial. This had not really been done before that produces a clear form image in the brain.

Now it is black and white. We don’t get color with this. Color rolls up through the bipolar cells. We do think that in the next one or two device generations, we’ll probably get at least some red and green. But it acts as essentially an electronic photoreceptor. The implant is a light sensitive chip that has all these tiny little pixels on it, and the patient wears glasses that have a camera looking out at the world where you could get the video feed from anywhere, and then an infrared laser that projects into the back of the eye to strike the implant. And because your eyes are not normally light sensitive in the infrared, you can’t see the laser projection. But when it wherever the light energy falls on the implant, it just absorbs the light, converts that into an electric field that stimulates the bipolar cells and thereby kind of bypassing the dead rods and cones and getting the visual stimulus back into the visual processing pathway at the first possible opportunity beyond the dead photoreceptors.

Juan Benet

Basically, you’re creating a different technical pathway to bypass a layer that’s not working and then activate the layers below that are working and then generate the image. And so here, I think we didn’t describe the experience of a patient with the disease. You mentioned blind spots. There’s kind of like partial blindness, right? How does the disease work?

Max Hodak

Specifically, the trial was done in age-related macular degeneration. These patients are not in total darkness. They have some residual peripheral vision. They can use that to walk around and not run into walls. But only the central 6 or 7 degrees of your visual field is actually high acuity color vision.

This is the fovea, and your eyes are constantly darting around to figure out where in the scene should I look next in order to minimize my ongoing uncertainty. And what you see is not the image that actually falls on the retina. What you actually experience is this world model constructed by the brain, which is actually only relatively weakly updated by the senses.

This is one of the reasons that we’re so susceptible to things like illusions. And so, as the eye is darting around and filling in this world model, that is the thing that you end up perceiving. And so even though these patients, they have some residual peripheral vision that the area of their high acuity color vision is degraded — they think that they see.

Patients don’t perceive a dark spot because the brain fills it in. Now, if you’re really late stage blind, like in dark RP, that’s a little different. But the brain really wants to fill it in because it’s looking for information, even if it’s not real. One of the things that we can do with these patients is you can show a solid green bar that goes all the way through the visual field. And they’ll say that they see a contiguous bar that’s green, and then it turns white, and then it turns green again. But they see it as continuous because the brain fills in that area around the stoma.

Juan Benet

In a lot of these cases, what are the impacts on people’s lives. You mentioned they can maybe potentially walk around and so on, but certainly it sounds like they can’t read.

Max Hodak

Yeah. They can’t recognize faces. They can’t order off a menu. They can’t.

Juan Benet

Use that computer…

Max Hodak

They can’t definitely can’t use a computer, except to the degree that they can have it in their assistive device or talk to it or something.

Juan Benet

And so with a device you replace — you bypass the layer that’s not working. You use a laser mounted onto the glasses to then stimulate those cells. And now what do they see?

Max Hodak

This is one of the reasons that the crossword puzzle test is interesting. Because they’re — I mean, we don’t — but in theory, there are ways that you could cheat on other types of object recognition or these reading tasks. But here, this is a visually guided fine motor task.

Most motor control is not visually guided. It’s proprioceptive-guided, which is this feedback of your body position. You don’t need to look at your hands to catch your baseball or to type. But here, this is a visually guided fine-motor task. And by the way, the fact that this works actually says deeper things about how cool this technology is.

In prior devices where the electrode array moved with the eye as it moved, or if you put it, say, in cortex or in the thalamus, where as the eye moves around, the electrode mapping is the same, so then the eye movements are no longer meaningful, this breaks a very important connection to motor control, whereas with PRIMA, this is intact because the eye movements remain meaningful because they move relative to the laser.

And so the fact that you can do this visually guided fine motor task is very cool. But these are shapes and they have to see the lines of the grid squares. They have to read the letters, they have to fill in the letters between them. And so I think this is really very strong evidence that it — what does it look like? It looks like vision, like this is vision, structured form vision.

Juan Benet

Filling crosswords and like reading books.

Max Hodak

Yeah.

Juan Benet

Wow.

Max Hodak

They’re playing card games.

Juan Benet

Amazing.

Max Hodak

Yeah.

Juan Benet

How do the patients feel?

Max Hodak

The patients are pretty stoked. I mean, I have to be clear, this is a clinical trial. You cannot say that this would work for everybody. We’re going through the regulatory review process now. It’s not commercially available yet. This is an investigational device, but, I mean, the results from this trial are very, very — they’re great.

Juan Benet

Yes. Amazing. What are some of the people saying? I don’t know if you can talk about that, but, when I get a sense of the experience that people have.

Max Hodak

I’ve only spent a limited amount of time directly with the patients, but it’s pretty cool. Like when I was in France a couple of months ago, I went to a rehab session, and it’s pretty cool to see them holding a newspaper and reading it. And these patients are all pretty elderly, the average age of the trial was, I think, 81, and that it is a lot to ask an 81 year old to do anything, especially a blind 81 year old, to come into a clinic and engage with this clinical trial. And many of them, they really want to do it because this is a very special experience.

Juan Benet

So you mentioned France. And as far as I know, the clinical trials started in France and now you’re doing them in the US. Can you talk through what stage the tech is in and worldwide?

Max Hodak

The main clinical trial was done in five countries across Europe. It was 17 sites across France, the Netherlands, Germany, Italy and the UK, and we have currently submitted for the CE mark, which is the marketing approval system in Europe and a bunch of other countries. And we are going through the audits now. That seems to be going okay. In addition to the main pivotal trial for that approval, there is a small feasibility study in Europe and a small feasibility study in the US. So there’s a handful of patients that have been implanted in the US. We are currently engaged with the FDA to talk about the path to market in the US.

We also, like I said, have a next generation of the chip already being worked on, that we’re starting to think about human studies for that. It’s very similar to the current version, but better, and we’ll be doing a clinical trial for that, both in the US as well as Australia and some other countries.

Juan Benet

That’s amazing. And do you think that this device can then be used for a range of other diseases, or it is very specific to this one family of problems.

Max Hodak

Well, the family of problems are photoreceptor loss diseases.

Juan Benet

That sounds pretty general.

Max Hodak

There’s a bunch of different reasons people go blind. One of the most common are things like refractive errors, like cataracts that are easy to fix with surgery, done all the time — it’s one of the world’s most common surgeries. Then there’s glaucoma, which is loss of the optic high pressure, which can lead to loss of the optic nerve, or potentially other reasons why you would lose the optic nerve, say, trauma. If you’ve lost the connection of the brain, this does not work for that. But for a relatively broad range of diseases. I mean, one of the things that’s nice about our approach, both the gene therapy as well as PRIMA, is that it doesn’t really matter why the photoreceptors died, as long as you can stimulate the cells past them.

I think this is when we talk about neurotech, I think some people would say, well, this is ophthalmology. We’ve been doing ophthalmology for decades. Obviously people have been trying to restore vision for a long time. There’s all kinds of drugs, there’s gene therapies, there’s a bunch of devices — like what’s different. Neurotech is really — there is a different lens on the world. You can think about many of these things more in engineering terms than you could have before.

And it’s different from the other gene therapies and drugs that are out there because, like, we don’t necessarily care why the photoreceptors died. We just care that we can think in terms of the information and the representation that we need to inject into the brain in order for you to experience the visual information and get the visual experience.

And so beyond macular degeneration, we’re also thinking about retinitis pigmentosa Stargardt disease, which is something like macular degeneration affects young people. Diabetic retinopathy. There’s a bunch of different potential indications that will be needing to do separate studies on.

Juan Benet

Yeah. How many people does that represent globally?

Max Hodak

Macular degeneration is one of the largest of those. Glaucoma is big. But it’s in the millions of people.

Juan Benet

So if you if you succeed, if you get this device market, if it works out, if it’s successful, you could literally help millions of people around the world restore their vision.

Max Hodak

Yeah, that’s the idea.

Juan Benet

That’s great. That’s super awesome. I guess you have to go through the trials and that takes some amount of time. Can you walk through what that looks like? Because, of course, every day or year that passes without this kind of thing, there’s kind of significant loss there globally of like a day or a year that people don’t have these.

Mad Hodak

Medicine can move very slowly. And clinical research is like a very deliberate process. In some of the delays, to be totally honest, the risk benefits for the patients is if it takes another two or three years, they’re just not going to be there. They’re not going to get it. So clinical medical device R&D and translation is a very slow and deliberate process. The PRIMA technology was originally invented at Stanford by a professor there named Daniel Palanker. It’s been in development for, I think, a decade now. That’s the timescale that it takes to bring something like this to market.

Now, like I mentioned, we’re doing a trial in Australia. So the first version of almost anything is always pretty limited. The first iPhone was pretty cool. I mean, that was a huge advance. But the first iPod, they put together in six months. And that in itself, like if you look back now, like to that versus what we have now — the really cool stuff is in versions four, five, six and beyond. And so I think we now in PRIMA, we have this existence proof that you can create a form image in the mind and that this basic approach works.

But there’s a bunch of ways to make it better. We’re going to want every two years or so, to be coming out with a new version, to begin to realize the potential of the technology through these refinements and being able to do that quickly and easily to get feedback from human subjects and feed this back into technology development is really important.

We only work in sophisticated, modern developed health systems. We’re not going off the reservation to go do this somewhere sketchy, but we’re always interested in moving as fast as we can.

Juan Benet

There are probably some countries and jurisdictions that are starting to think about leaning forward and going faster on this. I’ve spoken to people in the UAE, for example, that I want to create pathways for neurotech to advance there. What sort of advice would you have for government folks globally to think about improving their systems?

Max Hodak

I don’t know that I would answer that specific to neurotech, necessarily. There’s on multiple levels opportunity for clinical trial reform. Healthcare, to a great degree, really represents a market failure. It’s something that everybody has to deal with. The willingness to pay is basically infinite. It requires big investments and very specialized labor and a lot of very specialized infrastructure. So on many levels it leads to these big economic distortions that need to be corrected somehow. I think one of the most valuable thing that any regulatory body can do is just engage actively, because if it takes two months to figure out if something is possible versus eight months, even if the conclusion is the same — whatever the rules are, the rules are fine — but being able to get feedback in two weeks versus two months versus six months, this has a really profound impact on the types of entities that can play in this. If it takes you a year to get even feedback on a clinical trial protocol, then the only people that can do this are super highly funded giant companies that think in decades, and it’s just too hard to venture fund a lot of this.

Even if the return could be there, bridging through that is very challenging. And so anything that we can do to compress these timelines is going to really encourage innovation. There are still biological time constants that can’t be avoided. Like at some point you need to just follow some number of patients for two years or three years and get the safety follow-ups before you go to more patients.

But just the paperwork. Like the paperwork — like if you send a letter saying, hey, we’ve interpreted this rule this way, and then there’s a three month delay that can make it hard to innovate.

Juan Benet

Exactly. The time that the biology needs to take and for safety. You have to allow the trial to pass for a certain amount of time. But if there are delays that are in paperwork or things like that, that seems like a huge problem. You mentioned upgrading the device over time. Would people then when the new upgrade rolls out then get another surgery and just replace the device? What does that path look like?

Max Hodak

Potentially. One of the things that’s nice about PRIMA is it might be upgradable. We’ve never done this in a human, but there’s animal evidence that this is possible. It’s a tiny, little fully wireless chip. PRIMA, being powered by the light that is used to activate it, is this, like, that’s a really cool trick.

Juan Benet

Yeah, that’s a great innovation.

Max Hodak

Yeah. There’s no cable. There’s no battery. Prior implants had a cable coming out of the eyeball. It’s like a titanium box. I can get rid of all that. You just have the tiny little silicon fleck. And. Yeah, we think that it’ll be possible to upgrade it now in practice. Probably in the first generation or two.

That won’t happen that often. And then as you start going to younger patients who are less disabled for a more powerful implant than you’ll, I think, start to see that more. Certainly, if you’re talking about like a 25 or 30 year old Stargardt patient, they’re going to expect to get a handful of upgrades over their life until we can get towards native acuity, color vision.

Juan Benet

Right now, the idea of having one surgery sounds intense. Multiple sounds even potentially scarier. How do we kind of enable this to be super easy and routine?

Max Hodak

It’s a super simple one-hour outpatient surgery. So this can be done under local anesthesia. It is often done under general anesthesia, but that’s as much for just nerves and for you can’t change your mind halfway through the surgery, but it can be done. You just make some injections next to the eye. It goes dark and numb for a few hours. They come in through the front of the eye. They leave the chip in the back of it. I remember when I got into this field almost 20 years ago — now it takes a little while to acclimate to photos of surgery and I think everybody has this experience going from like, I don’t really want to watch the videos of surgery to being, oh, check out that vasculature, that’s super defined and you can like see this part of the brain. And then when I moved to start working on the retina, it was a whole different adjustment to start looking at videos of eye surgeries. But you acclimate to it. And then you’re like, oh yeah, you can see that marker and it’s like, oh, you can see where the implant was placed relative to this blood vessel. It’s really not a big deal. The patients tolerate it well.

Juan Benet

And certainly the ability to recover a sense is probably a lot more of a difficult surgery and recovery process.

Max Hodak

Yeah, and this really is not a difficult surgery. I mean, like a brain surgery where you’re drilling a hole in the skull and tunneling into the brain, that is way more intense than this outpatient, leave a little wireless chip in the back of the eye. You don’t have to go through any bone to get to the eye. The eye is soft tissue. And so it’s surgically easy to access. You can look at it by looking in through the eye because the eye is clear. So on many levels, this is a much easier surgery than some of the other things out there.

Juan Benet

That’s a great segue. So let’s talk about the biohybrid. You mentioned it a little bit earlier, but let’s maybe start general and then dive in. So how do brain-computer interfaces work? I mean we’ve definitely been talking about maybe special purpose or special case BCIs, but what was the general vision of the BCI field, meaning being able to do read and write into the brain? What’s the possibility space here?

Max Hodak

I don’t know that the BCI field has a general theory of the vision of BCI. I think this is part of one of the issues right now where there’s some very classic BCI applications, like extracting motor representations to control a cursor, or a robotic arm or keyboard, or more recently, speech decoding, extracting speech from the brain from ALS patients. Or I actually saw some data recently where some non-verbal autistic people might have speech representations that could be decodable with with a speech prosthesis that hasn’t actually been done, but there’s evidence that that could be possible.

And that’s the type of neuroscience that isn’t new, but the devices are getting better. Now, does a better motor decoder — is that the big prize at the end of where a lot of people are imagining right now? I don’t know. There are vague gestures to maybe you could improve memory or maybe you could improve reasoning, or maybe you could have some other capability. I think that is kind of harder to reason about right now. I think that’s not really clear what that would mean.

But then at the other end, when I think about some biohybrid where you’re thinking, really, can you add whole new brain areas or can you add almost a new hemisphere to the brain? This takes you into even kind of more out there sounding territory. Where can you in some very profound sense, redraw the border around your brain? Can you add new things to that system? So then you go from thinking about just exchanging information with the brain, extracting motor information or writing sensory information, to kind of adding a whole new part of the brain it didn’t have before.

Juan Benet

Yeah. I mean, right now already, so many people are deeply integrated with their phone and their personal devices. These are kind of part of them in a way, but they just haven’t crossed the barrier. You still have to go through the eyes and your thumbs to type. But if you open the bottleneck and you then enable you to think with your computer — what does that look like? What could be possible here?

Max Hodak

So there’s a lot that I’d say is debated in the field right now. There’s some potential limits. Human language is about 40 bits per second. There’s different interpretations of that. But just think about the amount of number of bits of information that are kind of reduced, you’ve got a base of phonemes, like the words that can get put together. You choose one from that and some number of times per second. Some human languages are spoken more quickly and convey less information per token. Some are spoken more slowly and have more information per token, but in all cases it’s about 40 bits per second. Now there’s also some evidence that there’s this deeper cognitive bottleneck in the brain at about 10 bits per second for reasoning and experience.

So if you put someone with a photographic memory on a helicopter flight over Manhattan, then for an hour and then ask them to draw what they saw, and then you look at the details that they got averaged over the hour. It’s about ten bits per second. Or there’s many different ways to triangulate this number. And so that leads people to think that there’s this deep bottleneck which is co-evolved with all the different brain areas that kind of serializes thought at about 10 bits per second in terms of what you’re able to remember and act on. In that sense, just opening up some bottleneck may not necessarily help you.

You feel like you could speak faster, you feel like you have all these ideas. But the reality is that until you can until you serialize these in this way, they aren’t actually formed and you can’t actually use them now. On the other hand, there are even at 10 bits per second or at 40 bits per second. There are faculties that you don’t have or skills that you don’t have. This is kind of like the “I know kung fu” example. So that doesn’t necessarily cheat that limit. Those bottlenecks could still exist, but at those rates, the brain could have the ability to transform information in a different way than it would have otherwise had.

Juan Benet

So the “I know kung fu” moment from the Matrix, where Neo downloads this capability and is now suddenly able to control his body. That probably requires training the motor.

Max Hodak

Yeah, we don’t know how to do that right now.

Juan Benet

Yeah, that seems like surprisingly harder than, like, let me remember Wikipedia or, I don’t know, let me look at my message inbox with my brain. What are some of these other things that seem more likely?

Max Hodak

You can almost think of it as like a UI UX problem. What is the UI UX of a BCI? That is what you want for accessing something like Wikipedia, because like the iPhone, the iPhone is a crazy technology in multiple levels. If you want to talk about addiction, many people, if they’re physically away from their phone for more than a couple minutes, they get noticeably anxious. Like, that is a crazy technology for what that does to our brains and how it rewires it.

Juan Benet

Less a kind of dopamine addiction. More kind of a sense of self, evolving into it. Like, I’m addicted to walking with my legs, right?

Max Hodak

Kind of. Yeah. But when people rely on phones more, their memory degrades. Similarly, people that use GPSs to drive have worse spatial capabilities than those who don’t. And so the brain is definitely offloading. It’s something that’s not being exercised that it’s realized it can offload, that it can draw upon.

And so that is definitely already happening. But if you wanted to integrate this directly into the brain somehow, how should that be exposed? Like, should this be like an internal monologue? Not everybody has an internal monologue. How do you cue this thing? Do you want to just be able to have a thought, be like, oh, I wonder what is the GDP of like Micronesia or whatever? And then how should that be delivered to you?

Juan Benet

What do you think of I don’t know, maybe not a super long term future, but maybe a short near-term. I suppose that we get something like the biohybrid and we’re able to expand the connectivity. What UX capabilities would suddenly be unlocked?

Max Hodak

Well, to some degree this gets at like what is your unannounced research?

Juan Benet

Yeah, but what are some really cool stuff you want to tease.

Max Hodak

I think this idea of adding new brain areas entirely is underexplored. You have clinical needs for this in cases like stroke, where a patient might have lost some part of their brain that underlies a speech or some part of reasoning or some part of memory, and if you can restore that capability to them, then that is a really valuable medical need and also hints at potentially other types of expansion of capability.

Juan Benet

Concretely, this would be like adding a chip into your brain that would then carry along some amount of those capabilities. Would that connect to the phone or other devices?

Max Hodak

You could do it potentially either way. So with biohybrid, there would be a device that has these cells that the cells grow in — they form new bidirectional connections into the brain. And then those cells, we think of them as they kind of join the local cortical representations. For example, we know that if you have a bio implant in a mouse, and then you have a mouse and a treadmill, we can recover how fast the mouse is running through these cells.

We know that the dendrites of the biohybrid graft cells have grown into the brain and formed connections that join these, like Paw kinematic representations. And then also, if you have an animal in an environment, it’s making a decision. You can give it information that it can use to make that decision through a biohybrid environment.

So we know we’re getting these axonal synapses. And so when these things go into the brain, they form connections with the neighboring tissue. Biohybrids are pretty cool. They grow pretty deeply into the brain. You see connections all the way down to subcortical structures like the thalamus.

And so it’s totally possible to imagine that you could add a new cortical thalamic loop, that instead of going through a native part of cortex, kind of ends up in your device that then you could process either locally or you could send over a radio to some other larger model running elsewhere.

Juan Benet

And that would start hooking in entire subunits of computation into your brain. It seems like the brain could start calling out to these kinds of subunits to do complicated tasks.

Max Hodak

Exactly. Can you add new pretrained representations to cortex that you could draw upon is a really interesting idea. There’s a huge amount of basic neuroscience to do here.

Juan Benet

From a computer science perspective, this looks like opening up the whole operating system and programing application structure and then replacing interfaces in-between with calling out to APIs in the world or in your device, and filling in larger amounts of information processing or computation to then be able to access it as part of your intuition or your thinking.

Max Hodak

You kind of have to think of the brain as an information processing organ, and it’s really about these informational flows, and you can have this mental model of the brain as full of these abstract objects called representations. So there’s a representation of my hand that this state corresponds to whatever the motor control is and whatever the current proprioceptive state is. And you want to be able to access these so that you can decode them or recover them or drive them and create them. But this field is really young and moving pretty fast, and there’s a lot of basic research that we’ll need to do to understand how to use it in all of the implications.

And we do a lot of that internally, but we also try to collaborate with outside academics and get tools into their hands. So that, I mean, we’re a small part of a global community, and there’s a very deep well of neuroscience and biology to do with these types of devices over the coming ten plus years.

Juan Benet

Let’s go through maybe an explanation of biohybrid again, like walk through the description.

Max Hodak

The central idea is that instead of placing wires or optical fibers or anything like that into the brain, we can make a device that we can load in a dish before you put this into a patient and you have it in a dish, and you can seed this with living cells that will turn into neurons in a very predictable way.

There’s a bunch of different ways that you could do this. We have a bunch of different probe designs. But the key is that you have these cells growing really directly on the electrodes. Whether these are like actual electrodes or things like LEDs or imaging imaging pixels —

Juan Benet

How do you talk to those cells?

Max Hodak

So neurons store information in a voltage across their membrane. You can think of a cell as a fat bubble full of saline and a bunch of other molecules. Neurons are a type of so-called excitable cell, which have these proteins that go across the cell membrane, which allow them to pump charged particles around, basically, so they can control their membrane potential to store information and do computation.

This is detectable with an electrode. You can have what we call a capacitive electrode because of how the physics of the device works. And so if you’ve got a neuron next to your electrode, if the neuron pumps ions around, you can detect this with your electrodes. We can record the cell state with a conventional electrode.

Juan Benet

Like basically reading the state of a single neuron.

Max Hodak

Of a single neuron, yeah. You can also inject a charge through an electrode in order to depolarize a cell. So what you’re doing there is you’re kind of increasing the concentration of a charged particle next to the cell, which has the effect of moving its — the base of voltage is a differential measurement. If you’ve got, say, a 120 millivolts potential across the membrane of the cell that you can change by either pumping ions across the cell membrane or just changing the external concentration, which will have the effect of implicitly changing the membrane potential. So you can do that. That’s the most common approach. We call it micro-stimulation. We mostly don’t do that because you can’t stimulate and record at the same time. And there’s some other biological effects of that type of stimulation that aren’t ideal. So what we do is we express a special type of protein in these cells that make them light sensitive in a particular way.

So it’s a light-gated ion channel. It’s like one of those holes and the cell membrane that allow them to move ions around. Except this one is open or closed, depending on if there’s light of a specific color that is being shined on it. And so in addition to our electrodes, we also have micro LEDs in the device. So that if we want to fire a cell, you turn on the micro LED next to that cell, which will cause it to open the channel, and then that cell to fire.

Juan Benet

So, we have a tiny LED right next to the cell. It’s shooting photons into the cell that is activating.

Juan Benet

That opens the channel, causes it to fire. Yeah. Exactly.

Juan Benet

Yeah. Wow. And so that’s like the optogenetic opsin, right. Like maybe walk through why that was a breakthrough innovation.

Max Hodak

Those proteins are called opsins, and this is a field called optogenetics. They were originally discovered, I want to say almost like about 20 years ago, originally in some algae. And so optogenetics is a very widely used thing in neuroscience now, because it allows you to kind of causally influence these neural networks. So it’s one thing just to record neural activity. But if you want to say, well, what happens if I drive this cell or this population of cells in a particular way? Optogenetics is very powerful for that. And so we use optogenetics in a range of different things. We use it all over the place as a research tool.

But then in our products we use it in two places. One, we express our opsins in our biohybrid cells that make them light sensitive. And these opsins are a major area of research for us. We have an internal protein engineering group that has developed the world’s most sensitive opsins, and are continuing to improve their properties in a bunch of different ways.

Having very sensitive meaning, it takes a small amount of light to open or close them is important because if you want to have more channels and you have more cells, then if it takes less light to activate a cell, then you can have more LEDs in the same amount of power consumption and as importantly, in the same amount of, of heat generation, because one of the things that really limits what you can do with an implant is when you use it, how hot does it get, which is a function of how power efficient you are.

And so by having options that are very, very sensitive, that take much less light, you can have way more LEDs because each light, each LED is much dimmer. And so you don’t run in these thermal limits. We also use our opsins in the retina for, for our next generation retinal therapy. and it’s I think we have this intuition that nature is a great engineer.

And like, when nature has a problem to solve, it makes a protein. We should be inspired by this. This is something that’s really been enabled recently by the amino acid language models. I mean, this is really, I’d say AI driven innovation. We don’t talk that much about it like we don’t trust ourselves. It’s like, oh, we’re an AI company. And like, it’s AI powered and you should invest because of AI. But that’s a concrete example of how we use modern AI technologies like transformers for practical applications that can be really valuable.

Juan Benet

Yeah. It’s awesome. Without going super deep into the AI rabbit hole, how much do you feel like the latest models have accelerated your R&D?

Max Hodak

So in a couple of places they’ve been really impactful. For example, protein engineering is probably the single biggest impact. The other big place that we’ve used a lot of AI techniques is actually in regulatory documentation, and we have to generate thousands and thousands of pages of regulatory documentation.

And then we need to track that all of what we’ve done is compliant with hundreds of standards that all have super detailed criteria. And so that’s the type of thing that just as an automation tool can be useful. Everything we do is always like signed by a human. Everything is read by a human. But as a productivity lever it has been pretty useful.

Juan Benet

Yeah. Generating like thousands of drafts of things and so on before you actually settle on something. And it’s really happening on the other side of all of your applications and so on are probably being read by LLMs and distilled and summarized and so on before whatever response gets back to you.

Max Hodak

I hope so. I mean, they should be. That would definitely. I mean, anything that allows us to get responses faster would be great. And I mean, clearly these need to be reviewed by humans and signed by humans. But as a productivity tool, AI has a lot, a lot of potential.

Juan Benet

Okay. So then you have these cells that you can read from, electrically, and you can write to optogenetically. You can drive a single cell. The cell is living in a well in this whole sheet of silicon. You fab a chip with the LED to be able to drive it. And I guess you have, like a whole bunch of these cells in a grid and — how big do these grids get? Like how many?

Max Hodak

So we do have designs, which I think you’ve seen where there are wells that individual cells are in. We do some of that. We also have other designs. The microenvironment of a surface that these cells — it’s a pretty complicated co-culture of a couple different types of cells that are happy to survive, kind of at this silicon interface that is super detailed. A lot goes into making this microenvironment that these cells want to survive and thrive in. And we’ve learned a lot over the last few years. Probably the deepest area of IP on the biohybrid technology for us is actually the cells themselves, which we’ve heavily edited to make them compatible with patients’ immune systems and have things like these opsin proteins, but also have a bunch of other safety mechanisms built in.

But then second to that is a lot of what we’ve learned about building microenvironment is that these cells will survive including through the implantation procedure, which is a pretty harsh hypoxic, glycemic, other types of shock happening, these typically if you just take a bunch of cells and inject them into the brain, 99% of them die. And so there’s fairly deep technology around making them survive and thrive there.

Juan Benet

How big did the grids get? Like roughly how many like cells?

Max Hodak

We routinely engraft a million cells. Their standard probe right now is 4mm by 4mm. Cells are really small. There’s plenty of room at the bottom. The limit on the size of that end of the probe is just the brain is curved, and you want something that has more small little islands rather than a big chip that you’re going to try and squish on the surface of the brain. But within a 4x4mm probe active area, you can easily fit a million cells.

Juan Benet

Wow. That’s awesome.

Max Hodak

To be clear, that device right now does not have a million electrodes. Okay. It’s a way smaller number of electrodes than that. But it should be possible to get, and again, this is really limited primarily by power thermals and also our willingness to pay very large amounts of money on nice chip processes.

And so you can prototype this on some older chip nodes for hundreds of thousands to a million or so dollars and then start shrinking it. And then you quickly find yourself spending 10-plus million dollars at a time on a chip tape out. And I think we’ll get there within the next few years.

And that allows a very straightforward scaling. I think it is possible to get to a place where you’re talking about millions of channels with billions of cells. But one step at a time. We definitely get dendritic connections. We definitely get axonal connections. We’re translating this up to larger animals. We’re working methodically.

Juan Benet

Yeah. And the dendritic and external connections means when you implant a device, you are connecting into upstream of loops and downstream of loops? Like you’re making —

Max Hodak

Yeah, it just means that your cells are getting inputs and outputs, but they’re exchanging information bidirectional with the brain.

Juan Benet

So you take this grid of cells, you implant it on the brain. The cells grow their axons into the brain. And you were mentioning they go super deep.

Max Hodak

If you think about a motor decoder — so this is usually implanted in the pre central gyrus, so there’s this— Primary motor cortex is the strip at the very top of the brain. But from there to a muscle is like two synapses. The axons from the cells in primary motor cortex go all the way down the spine to the vertebrae. That’s at the level of wherever that connection is.

And there’s a synapse, and then it goes out to the muscle. And so neurons can be quite spatially extended. And I don’t know if this is

Juan Benet

Super amazing.

Max Benet

Yeah. Yeah. So neurons are used to being super long. They do this all the time.

Juan Benet

What sort of process drives this wiring. Do they just automatically do this?

Max Hodak

This is the developmental program. They do this — I don’t want to say totally on their own because again, the stem cell biology here is pretty cool.

Juan Benet

Neurons want to wire and learn?

Max Hodak

But yeah I mean this is what neurons do. When we graft these things, you see them wire up very broadly. And then they get pruned. And then there’s more things, active things that happen. But the neurons want to learn.

Juan Benet

Yeah. And so as soon as you connect it and so on, what have you detected from the types of systems that they wire into? What kind of experiments have you done?

Max Hodak

The things that we’ve published that we kind of have announced so far are very focused, like enabling studies on you want to show that the inputs to these cells are growing into the brain and forming functional connections. And you can recover information from the brain.

Juan Benet

Maybe even before the more sci-fi use cases of being able to kind of interact with a computer and integrate and so on, it sounds like this could even help with a bunch of the motor problems, like being able to wire parts of your brain to, I don’t know, across a separate spinal cord or something like that.

Max Hodak

Yeah, potentially. One of the labs that was into — I mean, they didn’t call it biohybrid at the time, but it was really the same idea — they called it living electrodes. There’s a guy at Penn, Kacy Cullen, who’s been into this idea for a long time. One of his ideas was, can you get axons to grow along these channels to build things like jumper cables in the spinal cord to bypass the break?

And so the spinal cord jumper cables idea is an old one that people have made progress on. And there’s some animal models that have really cool results. So that’s not something that we’re doing like in that embodiment specifically. But others are.

Juan Benet

If they’re wiring this way, couldn’t you start creating additional sensors like this that goes through your new brain area sort of thing, but could you add a sense?

Max Hodak

That’s certainly more speculative, but that’s exactly the type of thing that I’m excited about. You need a theory about, like — our senses like, where do they come from? Why do you see and hear and feel and not other stuff, which I have thoughts on, but getting to elaborate that and show some of these things.

I think that it’s been bottlenecked on probes and electronics. Like we just haven’t been able to get these bidirectional connections into the brain at a scale that allows you to really do this. And to seriously ask these questions about like, what would a novel sense be? I think that we’re going to be getting into the stuff within the next decade.

Juan Benet

Yeah. And what is the path? What is the science and engineering and then eventually the product-building look like there? Sounds like there’s a whole bunch of fundamental science to do right now with these new devices to figure out, what do you integrate into and so on. You’ve already done an enormous amount of engineering of figuring out the device and how to make it work and how to increase the channel counts and so on.

Seems like there’s a clear there’s probably a whole bunch of safety and health oriented testing to do. What does that look like? What does that track? How many years is that?

Max Hodak

You can have a debate about whether biohybrid is more or less invasive than other options. In some sense it’s less invasive. It does no damage. It’s perfectly biocompatible. The brain is happy to accept these things. On the other hand, you’ve had a biological cell therapy which is engrafted, which is forming these like very deep — you can’t really easily remove that unless you chemically cause these cells to die.

It has a bunch of structural advantages.They could last for decades. You can have massive bandwidth. They really enable things that might not be possible any other way. On the other hand, you have totally novel risks to think about. There’s actually a long history of people getting cell transplants into the brain.

And this was one of the things that made us think that this could work, because like, if you go on PubMed, the medical literature database run by the US government, and you search for Parkinson’s cell transplant or Alzheimer’s cell transplant, there’s actually a lot of doctors out there who’ve had a couple of patients who are like, you know what, this patient needs a stem cell injection to the spine or a cell engrafted to the basal ganglia. And those historically have not really helped the diseases. But they did teach us a couple of things.

One, the cells tend to survive and functionally integrate and last decades and nothing bad happens. And so there’s actually a pretty large body of like supporting evidence that even like like even like in some cases stem cells that I don’t think you would have necessarily wanted to inject into the brain that way, turned out to really be pretty safe and pretty well tolerated.

We think that that will turn out to be fine. But it’s definitely going to be systematic from whatever you can do in a dish, whatever you can do in mice, you do in mice or whatever you can do in a rabbit, And then you have to use some monkey studies. Then you go to humans. And certainly anything in humans is much slower and much more methodical than what you do in animal research. But at some point there has to be some first patients. And I think that it’ll probably most likely be stroke patients. There’ll be a population of stroke patients that I think are probably likely to be the first biohybrid recipients. It definitely won’t be next year, but I think it’ll be way less than five years from now.

Juan Benet

From there, what’s the path to like? How long is the trials landscape or something like this?

Max Hodak

It usually takes at least three to four years to get through clinical trial for something like this. Because you’ll spend a year getting set up. You’ll write your protocol, you’ll get IRB approval, you’ll get all your documentation — that takes like at least a year. And then, you’ll enroll your participants. That can take a while, and then you’ll have to follow them for at least 12 months. There’s no way that the FDA is not going to look at your data for less than 12 months, and they’ll probably want two and three year follow ups. But you can do that as you go. And so if you’ve got a two year endpoint and they really want to see a third year on follow up, then you have a year of setup. You have a year where you’re kind of implanting everybody. And then you’ll start hitting the 12 month check and 24 month check points at some rolling window. After that, once you’ve got a suitable number of patients out to two years, that’s probably enough to start thinking about submitting larger filings to the FDA.

And then while you’re going through that process and they’re reviewing it, you’ll start to get your 36 month data. And they’ll look at that as it comes in. After that you can imagine like another year to kind of get an approval and get on market. So that’s at least a four or so year process.

Juan Benet

Plus the sub five years that you mentioned. Now that could be within 10 years this might.

Max Hodak

Yeah, it could be on market. But at the same time we’re going to learn a lot. Going to market is one thing, and the way we actually architected the company allows some of that to be a little more back loaded. Our retinal prosthesis, specifically PRIMA in particular, we hope to be on market with that next summer. That could get delayed. We are going through the regulatory process now.

We don’t have the approval yet, but we think it’s pretty likely that we’ll get approved in the next year. And that alone is a big enough product that it can finance a lot of the rest of the stuff that we want to do. So the company, we’re thinking about how do we get profitable, not just like, oh, we want to raise another $5 billion of venture or something over the next five years.

Juan Benet

And so if you get profitable, what’s the timeline? This seems like one of the fastest biomed oriented companies to become profitable.

Max Hodak

Well, we can say that when we’ve done it. Let me get there first.

Juan Benet

If you can get there.

Max Hodak

Yeah, if you get there. Yeah, exactly. But I mean, I think I don’t know, running a startup is.

Juan Benet

Is there a website tracking these?

Max Hodak

I mean, it’s annoying to look at the software versions of this where it’s like they’re going to.

Juan Benet

Like 100 million in a few weeks.

Max Hodak

Or something. Yeah. It’s like, well, they can’t do that. I don’t know, it’s like the startup.

Juan Benet

Maybe you have a team doing that and then that funds the rest of the program.

Max Hodak

Yeah. Well, we’re grateful for PL support and investment over the years. But yeah the visual prosthesis is really a big enough business. If we don’t mess that up, there can be a multi-billion dollar a year source of cash fully developed. And that buys us a couple things. One, it buys you time to go through the full regulatory process for this — the world’s most complex combination cell therapy device ever made. And it allows us to fund some parallel research along the way that I think will really prove out some of the big concepts, like the missing theory prove that some of these really next generation type applications are possible in animal models and small numbers of humans, separate from a market approval study.

Because you can have these smaller studies where you’ll implant a couple patients like three, five, seven patients. Learn something very targeted, hopefully help them with some disease, but at least get a kind of prove out core elements that are missing. That is very different than a market approval study where you’ve got one thing that you’re going to do very little to tweak, and you implant 50 or 60 patients and you follow them for three years.

Those are very different types of things. Our plan, like the secret master plan here, is bring a breakthrough retinal prosthesis to market. Use the revenue from that to fund the development of the biohybrid core technology uses to prove out big prizes of emerging neurotech and then eventually, of course, translate this to market.

We have this vision of how we think the world is likely to look in the early 2030s. I think 2030 will probably still look a lot like today. I mean, this is also AI and other other things happening, but I think 2035 is likely to look quite a bit different than we might be imagining right now.

Juan Benet

Yeah. At some point we should do that conversation of like, what does 2030 look like and what do we think 2035 is going to look like? Stay tuned. Cool. So first of all, that’s super exciting. Like we’re talking about like if that happens within ten years of now, now having very high bandwidth BCIs that you could then start using to drive computers, interacting with AI, interacting with people and so on.

Max Hodak

Yeah. I mean, your entire experience of reality is rooted in the brain. And so this is also like it’s important to be very sensitive because this is like your sense of identity, your sense of self. Like people do not want you to mess with this. And I think we are very we hear that we were receptive to this.

We want to make sure what the goal is here.

Juan Benet

And at the same time people want to be able to experiment and explore and improve like change their own sense of self. So, just as much as a set of people want to be able to be have what they had in the past restored like they have some disease or some damage or an accident.

They have their sense of self and mobility be restored, there’s a population where a set of people want to be able to enhance their capabilities, like they want to be able to improve their senses, like see a different part of the color spectrum, or being able to add additional senses or interact with their phones or computers in deeper ways. What does what does that look like? What are the kinds of things that you want to be able to do?

Max Hodak

Yeah, some of these things will be possible. I think they’ll also be cultural discussions that happen along the way.

Juan Benet

And on that, I think OpenAI doesn’t get enough credit for shipping ChatGPT and having the big AI conversation in public. I think there was a lot of concerns. A lot of the AI community was very divided on whether you should or should not deployand have that conversation in public or not?

But from my perspective, I was always very pro having these kinds of significant changes that are possible appearing early while you can still like figure out what they’re going to look like in the long term and enable the broader groups around the world to come to terms with these kinds of potential futures and chart a path and figure it out together. Where should we go?

Max Hodak

AI is a very cool technology. It is going to enable us to do things that were never possible before. It is increasing the pace of innovation, and of society. But you can have some concern that it is kind of a dehumanizing technology that it could replace humans and have other existential concerns.

But it’s kind of an inherently dual use technology. It is possible that just like intelligence is itself dual use, any differentiated understanding of the universe is dual use. If you’ve got a better understanding of physics than anybody else you can build nuclear weapons, or if a civilization has calculus and one doesn’t, then one is going to be have an advantage.

Any different understanding of reality is dual use. And AI is the ultimate embodiment of that. Now on the other hand, BCI is in some ways a more intrinsically benign technology because somebody has a BCI does not necessarily mean there’s a different category of concern you might have versus like if you heard like somebody has an AGI and they’re the only one with it.

I think a lot of people are like there’s one person with a BCI that isn’t necessarily dangerous in the same category, and so I see this type of middle engineering.

Juan Benet

I mean, it depends, you know.

Max Hodak

Depends on what you do. But in itself, intrinsically, it is without some other thing happening. I think the promise of this technology is to increase and empower human agency. It is like a fundamentally human technology. And I think if you look at the endeavor of lower case science, like, what is the purpose of this?

We’re trying to understand reality. We’re trying to understand the world in order to improve the human condition. We use the knowledge that we get through research in order to improve our lives. Back in the Middle Ages, playing in the forest was dangerous, like a kid got a scratch on the wrong tree and would get infected and they would suddenly die.

And there’s nothing you could do about this, and there’s just this constant jeopardy that life was under, and that was just the human condition.

Juan Benet

It is astonishing. By the way, we do not today at all have an appreciation for how dangerous life was in the past.

Max Hodak

But even so, in the same way, people don’t think maybe I have an undetected cancer right now. That isn’t in the back of the minds of many people. And so this has changed and has certainly gotten better. But the fundamental human condition is still there in that way. So the endeavor of science is to improve the human condition, I see neural engineering as one of a very powerful deep nodes on the tech tree that begins to more dramatically allow you to maximize human agency and improve these things in kind of multiple different ways.

Juan Benet

Yeah. How much do you see this is an imperative for de-risking AI? From my perspective there are massive amounts of AI risks around capability explosion and generating a new age genetic species whether they might develop a degree and level of agency that is very difficult to prove anything about or to have certainty over.

So it can quickly become traditional sci fi, be in deep competition with humanity or in some ways that have been talked about in the last few years, not in harsh competition, just in this weird wau of taking control of the of the world and eroding human agency or control of the future.

Neurotech has seemed to me over the last five, ten years as one of the major avenues that we have to be able to stay ahead or even if not ahead, at least stay competitive in the long term and enable humanity to have a path into the future. How much do you think about that? How much of that is motivating? I am curious.

Max Hodak

There’s a real risk that AI does to us what we did to the chimpanzee 200,000 years ago. Humanity was an undifferentiated primate on the African savanna. And now our closest living relatives live in glass boxes so they don’t go extinct. What was different? Our intelligence. We were more adaptable, and we could collaborate better as a group.

And I think that’s the real thing. The only constant is change, because some people think this started in the 90s with the internet, or maybe in 2000 with the smartphone. This process has really been the story of human civilization since the mid-1800s. I think this really started with the Industrial Revolution, and there was a generation that went from a horse and buggy to seeing a man land on the moon and routine travel.

And I remember ten years ago, we used to joke around Silicon Valley, “We wanted flying cars, and we got 180 characters. What went wrong?” Will we ever see a situation like going from having a horse to the 747? Will we ever see something like that again? And it’s like, yeah, we’re going to see that again.

Juan Benet

We’re finally past the 90s and early 2000s slump. That was a horrible slump.

Max Hodak

I don’t even know if that’s fair, because one of the things that really strikes me about what’s happening in AI is that it feels like it happened as soon as it was possible. It wasn’t like there was a time when all the predicates were there and we were wandering.

But one of the prerequisites for training these AI models is that you need the internet, because you need 10 trillion tokens sitting there in a machine-readable form that you can consume.

Juan Benet

And have some pretty good semi-conductor.

Max Hodak

So you needed semiconductors, which were developed by gaming. Really, it was gaming, crypto, and all kinds of other forces that created the GPU. Those forces were strong and robust, and there were multiple of them. But then you also needed the internet, and you needed humanity to spend 20 years generating tokens, taking thoughts out of their minds and putting them into machines so that you had this machine-readable resource that you could then use to fit things. And really, as soon as you had that, we figured out the LLM.

Juan Benet

Though even without that, with other tokens, you could have the game-playing type of RL model from DeepMind and others. So if we did not have the tokens, we probably still would have been able to go in some direction. I think it’s a lot more constrained by Moore’s Law.

Max Hodak

I think for modern LLMs, the tokens are important. But AlphaGo happened first, so self-play was figured out first.

Juan Benet

And there will be even greater cases of it. I actually think that LLMs right now, and the hunger for improving them, have taken over a lot of the cycles that would have gone into self-play type results like AlphaFold. So I wonder what range of things we would be working on right now if our main modality were not chat and conversation.

Max Hodak

Yeah, I think you can say that Transformers and LLMs became one paradigm that sucked the oxygen out of the room. But I think the reality is that there’s way more oxygen. The total spending, the total amount of talent, and the total number of companies are massively larger than they were five years ago.

Juan Benet

3 or 4 times the entire R&D spend of the US on non-defense.

Max Hodak

Yeah. You could say that maybe we’d be exploring these other architectures if everything weren’t dominated by the LLM. But we are exploring the other architectures, and those efforts are as big now as what people were doing on the scaling hypothesis five years ago. So I think you’re getting all of that.

Juan Benet

Going to be a pretty interesting 2030.

Max Hodak

And BCI, I think, has one interpretation: there are four or five highly funded big companies, and they’re going to drive this. But I think it is going to look similar. In five years, as BCI really begins to translate and more of this is developed, there will be hundreds of companies. There will be a whole ecosystem. None of these things end up with one player in a really robust space.

Juan Benet

Let’s talk about company building and broad R&D in driving breakthroughs. I tend to give a talk about fast R&D, and you’re one of the examples I often give. It’s basically lessons from you, Elon, Sam, and Steve.

To set this up, Patrick Collison has this great post called Fast, where he collects a set of examples of people quickly accomplishing ambitious things together.

He has examples like the Alaska Highway, built in 234 days; the Empire State Building, built in 410 days; JavaScript, built in ten days; Unix, the first version, in three weeks; and the iPod, 290 days from concept to shipping to customers. Which is insane, right? A whole device, from Tony Fadell joining Apple to the first iPods shipping to customers, in 290 days, under a year.

Amazing. Now tons of R&D organizations and companies move excruciatingly slower than that. Sometimes it takes many years to get anything out the door. You’ve been able to do many generations of devices, multiple devices, different platforms. You have a fab. There are all kinds of fast R&D approaches that you’ve been able to do. So why does going fast matter?

Max Hodak

For a lot of this type of research-driven innovation, it is basically unknowable how long it’s going to take. You can’t plan out milestones or put it into your task tracker and say, “In December we’ll do this, and by February we’ll do this.” It’s totally different from building many other types of things.

The only thing you can really control is the rate at which you learn. So how long your iteration cycle is has a hugely determinative impact on the outcomes. Because you can’t see years, or even many quarters, into the future, you’re limited to how fast you can learn.

Then we apply resources to things that are making progress and are working. And you reallocate resources away from things that are stuck until there’s either a breakthrough or you have a better idea. So that is a central management mechanism for this type of research.

We learned that there’s really no substitute for vertical integration for much of this stuff. In some sense, it’s actually kind of a market failure. The fact that we need to have a lot of our own capabilities in-house, including, as you mentioned, a fab, but also animal research and a lot of chip design.

In many ways, it would be nicer if there were a robust ecosystem of companies that were really sophisticated and could move fast, and you could just buy from them. Because then we would raise less money, have a smaller org that was easier to manage, be more capital-light, and take less dilution.

And in some other parts of the world, which we should probably be paying attention to, they can do that. You can have these smaller companies that are part of an ecosystem that you can just order this stuff from.

Juan Benet

And it’s kind of like what the US was like.

Max Hodak

What the US was like 50 or 60 years ago. So now I think there have been a couple of really successful examples of vertical integration making a huge difference, with these companies having all of this in-house.

Juan Benet

And probably the first serious example of this was Apple. They were vertically integrated across the entire process of manufacturing, building all the pieces of the product, selling it, and delivering it all the way to the customer.

Max Hodak

Well, Apple famously partnered with Foxconn. So I don’t know if that’s the right example, because they make toys like SMC and Samsung and Foxconn.

Juan Benet

That’s true. So I guess they were still doing a lot of things. The level of integration basically got cranked up a notch, or several notches. Back then, about 20 years ago, people used to say Apple vertically integrated a lot of things. They were still building on top of semiconductors from other companies. Eventually, they developed Apple Silicon.

Max Hodak

Yeah. That’s true.

Juan Benet

They’re still not making their own stuff.

Max Hodak

Yeah. They’re certainly taking advantage of a very sophisticated Asian supply chain. They make their own silicon. The real magic of Apple is the deep integration. They really care about how it works from end to end. But even Apple uses contract manufacturers. They don’t manufacture internally.

Juan Benet

Do you think there’s going to be humanoids in 5 or 10 years?

Max Hodak

I think manufacturing is one of the places where humanoids make the least sense.

Juan Benet

May not humanoids, but just robots?

Max Hodak

Oh yeah, there are tons of robots. Automobile manufacturing lines are incredibly automated.

Juan Benet

But we still haven’t been able to do certain things, like wire harnessing and a bunch of other tasks.

Max Hodak

Yeah, totally. I’m definitely a believer in humanoid robotics. If there’s a meme out there that humanoid robots are stupid and should always be more specialized, I’m skeptical of that. Realistically, the world is built for humanoids, and the more you can fit into that interface, the better.

But in a manufacturing environment, where you’re going to spend $1 billion to build a manufacturing line, there is going to be more specialized equipment. I think people are underestimating how much even relatively basic robotics, with really smart software and pretty good sensors, will be able to do.

So I think that is really software-limited. This is a weirdly contrarian opinion. You say this to some robotics people and they’re like, “Oh no, hands are bad, these sensors are bad, you need more motor bandwidth,” and all this stuff. I don’t know.

I think you could take basically a Roomba with a sufficiently smart model, and it would impress you with what it could do.

Juan Benet

Yeah. A lot of the robot controls problem is just not solved well yet. There are all kinds of easy dexterity-type things that robots just fumble. I think people hide behind “the hardware is not good enough” because the robot software is so bad.

Max Hodak

Yeah, exactly. I think that’s the easiest bet in the world: the models will get better. But on the vertical integration point, there’s one thing I would say. In some sense, it’s kind of a failure. On the other hand, what it enables you to do is innovate more deeply. If you’re limited to things you can piece together from what is commercially available, you will always be somewhat limited in how novel a thing you can build.

Whereas if we can say, “This is the arrangement of matter that we want. I want to put these atoms in exactly these places, and I’m willing to void the warranty on a $2 million fab tool to place those nitrogens exactly where I want them,” that allows deeper innovation. But it’s very capital-intensive.

Juan Benet

And so one part of the approach is how you manage the team. Another part is vertical integration, maybe in the team and how you organize it. You mentioned this kind of control loop of limiting the resources you put into things and trying to maximize the learning rate.

Give us a concrete sense of what that looks like. How fast is fast? What does it mean to drive a project quickly?

Max Hodak

I mean, some of this is getting a sense of what’s fast. I think that’s a totally reasonable question: how do you know? One of the questions I ask people I interview is, “How do you know that you’re good at what you do? What evidence do you have for this?” And it depends on the context.

For some of these things, where it’s unknown how long they can take or should take, I think all you’re really left with is this: at the end of the day, we want to be convinced that there’s nobody else on earth, with our resources, who could have done it faster. In this case, competition can be great, because nobody sets a world record running an open lap, and you don’t really find out what you’re capable of until.

Juan Benet

And as soon as the four-minute mile was broken, a bunch of people did it.

Max Hodak

A bunch of people did it. So you never really find out what you’re capable of unless you feel that pressure to deliver.

Juan Benet

But often a lot of groups don’t describe it at all. You maybe see some of the public timelines and whatnot, but in terms of the whole host of small little projects that a big project breaks down into, how do you keep the internal clock cycle really fast? I’ve just seen so many large organizations start bloating.

This is what happened to all of the major government contractors that got really slow, and the major tech companies. How do you avoid this kind of ossification or this bloat?

Max Hodak

Yeah. Small teams of people doing things. All the prizes are for making things that work. So you want small teams of people working with their own hands. There are not a lot of other managers. There are not people sitting around whose job is just paperwork, unless it is quality and regulatory, where the job is literally paperwork, which is also important.

But when we interview, people ask, what do you look for? And I think there are two things that I look for, and they are not the hard skills. The technical team they are interviewing for will determine if they meet the threshold of hard skills. But even then, that can often be taught.

The two things that I care about the most, which I have found are the least teachable, are an intrinsic sense of urgency. Does it matter to you that this stuff happens sooner? And then judgment. Do you make good decisions?

My central measure of capacity is: how successful have you been in making your life look like you wish it were, whatever that means? Because people have different interests and different ambitions, and some people want maximum free time to spend with their kids.

Other people want to be at the very top of their field and are willing to make trade-offs for that. But whatever it is, how good are you at making your life look like you wish it were? And by the time you have gotten to your mid-career, there should be visible evidence of this. And so the sense of judgment.

Juan Benet

Why instill a sense of urgency when you can just hire for it.

Max Hodak

Yeah. And that is tough to teach. Sometimes you’ve had really smart people who were just in the wrong environment or had a different cultural example, but they knew that they wanted something faster. But if someone is happy on island time, then we’re not going to convince them otherwise.

Juan Benet

Yeah. Do you do a Netflix-style approach, where you get very clear about that really quickly and early? How do you approach that in the interview process to clarify to candidates the clock speed that you expect in the organization?

Max Hodak

I don’t know that there’s a way. It’s definitely a thing that gets talked about in every interview. How would this person fit in here? I don’t know that there’s a bright-line rule or guideline that you rely on, but it’s something that everybody is looking for.

Juan Benet

And how does the team embody this? I think historically, lots of organizations have degraded in part because the team, if the incentives are not set up right or if they do not have this very strong intrinsic motivation, lets everything start taking a little bit longer and a little bit longer.

Max Hodak

You’ve got to keep an eye on this. Certainly, keeping an eye out for evaporative cooling is important. A players hire A players. B players hire C players. It’s like a massive explosion. Then your best people get disenfranchised, they start to leave, and then you’ve got a real problem. So being very paranoid about that is important.

Juan Benet

Can you clarify that a bit? The “A players hire A players” idea comes from early Silicon Valley perspectives and so on. I’ve often thought that model of just these three buckets is a vastly oversimplistic view of a deep exponential, where there’s an enormous ladder of skills.

I think there’s something qualitatively different. You could have people who are at a deep mastery level in some field, and yet they don’t have the quality that the Silicon Valley “A player” idea meant, and that maybe you’re getting at with the sense of urgency and so on.

How do you define what an A player looks like when they’re very early in their career, or when they don’t yet have the degree of accomplishments that you can clearly look at from a LinkedIn profile perspective and say, “A player through and through”?

Max Hodak

Yeah. I think one early marker of this is how intentional you have been. Even early in your career, you have gotten through 20-something years of life, or 18 or so. How intentional were you through that?

One of the things I like to do is start with: where did you grow up? Where did you go to high school? Where did you go to college? What did you do after that? The thing I’m looking at is: did you make decisions, or were you just kind of swept along? Many people are just swept along.

So the first thing I look for in these junior candidates is whether they have made intentional decisions. Did they end up in a place they meant to be?

Juan Benet

Do you think it’s a personality trait, or is it learned? Have you encountered people who were maybe swept along for a while and then started taking agency, or vice versa? Maybe they had a lot of agency and then that sort of broke down.

Max Hodak

I mean, you see all kinds of patterns, but you’re looking for high-agency people. And you’re totally right. The very best people are much more dramatically effective than the average person.

Juan Benet

I think this is really not well appreciated by people globally, especially even in deep knowledge-work domains where the leverage is so significant. In Silicon Valley, people talk about it in terms of 10x, 100x, 1000x types of things, but these are deep exponentials. You’re getting into the 10,000 or 100,000 level of impact, where one person, in the same year of time, will be able to have an incredible degree of leverage over the same problem space. And it really does break down to typing on a keyboard. It’s the same mechanical range.

Max Hodak

But it’s not that they’re producing more lines of code or more PCBs. So in that sense, it’s tough to think about it in terms of 10x or 100x or whatever, because it’s not like they often just work longer hours. It’s not that they’re moving faster in that sense. It’s that the things they do work, unblock things, and there’s no amount of time others could have spent on it where they would have found as good a solution.

And I think one of the biggest mistakes really smart engineers make is highly optimizing a thing that just shouldn’t exist in the first place. And a famous saying out there now is, “The best part is no part.” And this is definitely true. I think the best systems are very simple. Those are also the most reliable. The closer our products get to a single block of covalently bonded matter, the higher performance, the lower power, and the better they’ll be. And this means that you have to find ways to really simplify it to whatever the minimum physical thing can be.

And that simplicity is not trivial. So when you think about someone being ten times as effective, that doesn’t mean they are doing ten times as much. In fact, they might write half as much code or a third as much. The thing they come up with might be fundamentally simpler, and in that sense, more powerful.

Juan Benet

You emphasize judgment and good judgment. I have also found that this is extremely critical. You kind of described interviewing for it and looking at people’s judgment and decision-making across their life. What are some of the ways it shows up when running a team? What does good judgment look like in a team?

Max Hodak

Do the things you propose tend to work? Do the ideas you have tend to end up working?

We give resources and responsibility to people when, if I see them get excited about something, I think, “This is probably going to work, and it’s probably going to be cool.” And you also know when people have pushed for things that turned out to be dead ends or did not work. So you know it when you see it.

Juan Benet

You encourage exploration on things where you might legitimately have to pursue a whole bunch of dead ends before you find the right one.

Max Hodak

I mean, you shouldn’t really need to pursue that many dead ends, or when you do, it’s fairly intentional. Sometimes, in practice, it’s not like this whole idea of, “Oh, you throw away 4,000 prototypes.” It doesn’t really work that way, at least in the stuff I’ve seen.

But there might be a thing where you’re like, for example, on the retina, we knew that there would be some way to get a visual signal into the brain past the photoreceptors. For that, you’ve got this two-by-two matrix of bipolar cells, optic nerve cells, optogenetic, electrical. You could extend that. You could do genetic. You could do other things.

But for us, we were like, you’ve got a two-by-two matrix, and we’re going to do a survey of all four quadrants. And we’re going to exhaustively explore this space so we understand the trade-offs of all four quadrants. And then we’re going to narrow in on the things that we think are worth exploring further.

And so sometimes you want to do a parameter sweep of what the space is, and you can understand the space in some cases almost exhaustively. But from that, you then want to narrow this down. And if you’re constantly doing a lot of effort, if by six months into a project it’s a big setback, that doesn’t actually happen that often.

Juan Benet

Tell us a bit about how you manage time in the company. How do you set goals, timelines, and deadlines?

Max Hodak

It’s tough to set arbitrary deadlines. Smart engineers and scientists do not really like arbitrary deadlines. If I come along and say, “You’ve got to do this by this date,” they ask, “Why?” And if I say, “Because I said so,” that is not convincing. That does not work very well.

Now, in some cases, you have exogenous deadlines. For example, there could be a big opportunity where, if you have something ready by then, you know you will be able to get people to check it out. Or there is some other dependency where, if you are not ready by then, something else will be blocked that could have proceeded.

Nobody wants to be the bottleneck. They all want to hold up their end of the bargain and have their parts ready by the time it needs to be integrated. So you have these mutual bottlenecks, where the chip people do not want to block things, the animal people do not want to block things, the pro people do not want to block things, and the cell people do not want to block things.

So they have a sense of how it comes together. In practice, there are trade-offs you can make in each of these so that you can ship something, and then you can make each part better over time. You just keep iterating. It is a high rate of iteration.

The first version is not going to be perfect, but you want to get to some threshold of performance so that you can start making those trade-offs and say, “Okay, this feature is not important.”

You can have a higher noise floor in this chip, or you can have a smaller number of channels, or you could have a cell that does not have some particular phenotype, or some thin film where we are going to go with two layers instead of four layers right now, because that is the trade-off to keep the thing moving. And then we are going to get that back on the schedule next time.

Juan Benet

But it requires a great manager to set those concrete external goals.

Max Hodak

This is what I do all day. This is what the leadership here does.

Juan Benet

Because I think if left to their own devices, teams will just expand the timeline and erode time. And then things will just take a lot longer. And there will always be really good reasons for doing it.

Max Hodak

I mean, at the team level, people still want to ship things, and they want to get things out, and they are able to do that in a decoupled way. But as part of the culture, you need to have this sense of paranoia. We do not know how long it will take. We do not know if we will need to respin a chip. We do not know if we will need to do another surgery, or if we are going to need to generate more evidence. And at some point, you do eventually run into these absolute limits. You can only raise some amount of money at some price, on some terms, and then you are out of money and everybody goes home.

Juan Benet

You thought you were going to be able to raise money in two years, and then the economy changed and you couldn’t.

Max Hodak

And so you have to have this paranoia that you do not know exactly how long this is going to take. Because of that, you have to go as fast as you can, because even that might not be enough. But at the end of it, you want to know there was nothing else you could have done.

Juan Benet

But there seems to be a quality in really great founders: they are able to pick the right threshold for maintaining this extremely fast pace and externalizing the deadline.

For example, one of the famous stories on this is the iPhone keynote, the iPhone launch keynote. It is this magical moment in Silicon Valley history that people reference a ton.

Many people know the iPhone was built in around two years, roughly from deciding to do it to shipping the thing, or from deciding to do it to announcing it in a keynote. And Steve gets onstage at the keynote, describes the price and the shipping date, and the manufacturing team learns about this at the keynote.

So they have no idea that they need to ship this device in six months, and the cost structure they need to fit. Of course, they had probably ballparked that, but there was not an agreement or a plan. And now they have to scramble and do this within six months.

Max Hodak

And I think part of what made Steve somebody like that is knowing what is possible. In some sense, he is making it up, but part of what makes him so special is that he knew what the bounds were, that they could do it, that they could be pushed to do it, and then he was right. And I think that is tough to know.

I think lots of other founders then model on that and are like, “Well, he was a dick, so I can be a dick.” And it’s like, you’re missing the point. It’s not about that. The thing that mattered was the revealed judgment. He was right about these things.

And the being a dick part, it wasn’t a feature. He was successful despite that.

Juan Benet

Yeah. I’ve been wanting to write an essay called Cargo Cult Founders, where they learn all the wrong lessons from great founders and forget to learn the really critical ones.

Max Hodak

Yeah. People who have worked with me know I’m not perfect. You care a lot, you want to move fast, and it’s a high-stakes, high-stress environment. But at the end of the day, it’s not about that. Those things are still to be overcome in these incredibly stressful, incredibly high-stakes situations.

Juan Benet

There’s a deep, unfortunate thing about being a very strong founder, which is that the demands of the technology and the product and the market and the people and the team massively compress the timeline in which you have to care about a range of things. So you can try extremely hard not to be a dick about a bunch of things and yet come off like one at various points in time.

So this is a hard balance. Many people do it way better than others. Of course, Steve certainly had lots of examples where he just did not need to be the kind of jerk that he was, and many people do it better. But at the end of the day, the time you have to be human with other people is compressed.

Max Hodak

Yeah. This is part of the trade-off of getting to work on the critical path of civilization. The reality is, the stuff that gets done here matters, and it changes the world. It’s a privilege to get to work on it. And it’s a lifestyle choice that is not for everybody.

Juan Benet

Yeah, I think that’s something Silicon Valley is probably a lot better about now than it was in the past. People are much more aware and self-selecting into a lot of this.

And I think this is where the Netflix culture of, you know, “good performance gets your severance package” type of mentality comes in, of saying, “No, no, no. We’re really trying to build an athlete-level team here, and we’re trying to win the Olympics of business.”

And that looks a certain way. You have to self-select into that if that’s what you want to do. Amazing. Great. If that’s not what you want to do, that’s okay. But this is just not the right place.

Max Hodak

Totally. Yeah.

Juan Benet

Deadlines, though, you don’t have the ability to externalize as many deadlines because you have to be a lot more private about everything. How do you set deadlines like that? Are you able to commit the team down a path to force things to happen?

Max Hodak

It depends. There are always other things that we’re interacting with. Like I said, there are these mutual interdependencies where you want things to come together at some iteration cadence. People do not want to let their teammates down, so they are going to make sure that they do not block them and make that other work a waste.

But then there are other things, like tape-out dates. There are shuttle runs, and that is a TSMC date. If you miss it, then you are going to wait two months. So you are going to make it. There are other things like that that you get attached to.

Juan Benet

You also use these internal demos to create a clock cycle for the company, where you demo certain things on a cadence and things have to be good for that demo.

Max Hodak

Yeah, exactly. We’ve done these about twice a year since the beginning of the company. We do not really have board meetings, but we invite the investors, we invite the whole company, and then we organize an hour-and-a-half presentation of what our progress has been in the last four to six months.

And that is as much for internal purposes as it is for updating our investors. Every once in a while, the story is complicated. There are a lot of moving parts. You should organize your thoughts and show what you have to show for yourself.

And for some of the employees, especially with how we have been growing, they do not see the whole story put together end to end because they are focusing on their part. They know that they have been working on the lifetime testing of implant packaging, or they have been working on new chips, or they have been working on some other part of it.

Then they see the whole story together. That is a morale-building experience for the team. It allows us to compact the story and make sure that what we are doing is sensible. And because there are these spinning blades of death every six months, it creates this forcing function of, “We’re going to have something to show for ourselves.”

Juan Benet

I love it. It’s like a whole-company board meeting structure, because board meetings have famously been used as a forcing function for the executive team to actually think through things carefully and report to a set of stakeholders who then get to reason about things and so on. So many teams use board meetings as a forcing function for high-quality thinking across the board.

But that often tends to happen only with a small fraction of the company or stakeholders. And if you do these larger demos with the whole company and the entire stakeholder set, that gives a shot in the arm to the whole group.

Max Hodak

Yeah. Getting the whole organization aligned is one of the central challenges. The idea of the company can be perfectly formed in my mind, but if that is not conveyed to 200 people who are touching it, then it does not really matter.

We have been fortunate that our governance is relatively simple, and we can act with high conviction very quickly. But then aligning the rest of the company to that, I think one of the lessons I learned early in my career was that these companies are really human organizations. You can usually get the technology to work, but aligning and motivating hundreds of humans to do anything in the same direction is not trivial.

Juan Benet

Yeah. How do you think that is going to start changing as AI systems get good enough to start planning and managing entire swaths of knowledge work, and then eventually teams?

Max Hodak

Yeah, I don’t know. Ironically, the company and our science are kind of architected around this. We’ve got this big internal software tool that we run most of the company through. So basically all of the information that the company generates is in this internal tool called Helix.

All of our purchasing, all the animals, all the parts, all of the meetings, it’s a system we built in-house. This is kind of a contrarian bet. Early on, I heard all these stories and saw these things secondhand about the power of internal tools at Facebook, YC, and others. And I’ve used a lot of ERP systems and other tools like that, and nobody likes their ERP install.

Juan Benet

Something like a common denominator for a user.

Max Hodak

Yeah, exactly. It’s this category of software that seems to resist generic solutions. But when one company grows around a piece of software like that, it can be very powerful. And because it’s all in one place, you can expose a lot of this to AI agents.

So, for example, all of the meeting notes, there’s a note taker that joins many internal meetings, takes notes, gets them into Helix, and then people can chat with it and be like, “Hey, what’s the status of this project?” And it can read all the meeting notes and give you a summary.

Juan Benet

Can you connect it to project management software and then figure out what?

Max Hodak

I’m sure that down the road, there are some parts that are really sophisticated. Every dollar the company spends flows through this thing. On the other hand, there’s an endless wish list of features that are not a high priority to add.

Juan Benet

Automatically schedule a new fab run?

Max Hodak

Someday, yeah. I joke that the next CEO of science will be an AGI, and I’ll know that we’re getting there when all of the control surfaces and all the information required to run the company are there. And I’m just hitting “accept, accept, accept, accept.” By 2035, I’ll be able to delegate that, and then I’ll get to go back and do the fun stuff.

Juan Benet

Let’s talk a bit about you as an individual. How did you grow up and become the modern Max Hodak? When did you first get into science and tech? What inspired you? I want to get your story.

Max Hodak

Yeah. I started programming when I was really little. My parents told me that I sat on the floor of a bookstore and cried until they bought me a “Learn QBasic” book or something like that. From the time I was a kid through being a teenager, the compiler was not especially concerned with how old I was. I was into science fiction.

Juan Benet

When you say you started programming really little, what years was that?

Max Hodak

I don’t know exactly. I think I started when I was as little as 5 or 6, and I became a good programmer when I was an early teenager. My dad knew how to program. He had an undergrad degree in aerospace engineering. My parents were kind of in business, but I grew up building model rockets and having exposure to STEM.

One of the big inspirations was definitely the movie The Matrix. I think one of my personal ambitions is to disappear into full VR, never to be found. Atoms are really annoying to work with. The speed of light is low. Earth is small. In the world of bits, it can be whatever we want, and I think there is something very inspiring about that.

That was one of many potential missions, but through that and other things, I got really interested in the brain. And I spent a lot of my teenage years reading about and learning about the brain.

Juan Benet

Through like textbooks or sci-fi.

Max Hodak

Textbook, sci-fi. All of the above. I remember when I was in high school, I discovered one book that stood out to me called The Biophysics of Computation by Christof Koch. The brain is extraordinarily cool, and the idea of being able to engineer that is such a transcendent goal that, if you can really do that... Sometimes I think my life would be easier if I had gone into AI instead of biotech, but let people figure that out. BCI remains to be solved. I think there are some important things there.

Juan Benet

We need to run a portfolio approach here.

Max Hodak

But I went to the college that I did. I have an undergrad degree in biomedical engineering from Duke, and one of the reasons I went there was because, at the time, one of the best labs in the world doing brain-computer interfacing in monkeys was there, Miguel Nicolelis’s lab.

So I talked my way into that lab freshman year. When I showed up at Duke as a freshman, I got asked for an advisor. They asked, “Is there any researcher that you want to work with?” And I told them that I wanted to work for Miguel. I was basically rejected. They were like, “No, no, no. He doesn’t take undergrads. Only grad students and postdocs. He’s in the medical center.”

Then I figured out that there was a chemistry course, a seminar in chemistry, that would place you in a lab, and I used that as a backdoor to sneak in. They took me as part of this chemistry seminar. It was not a chemistry lab, but I got in through that. And that was really where most of my education in college happened. I spent most of my time working in that lab.

That was also where I met a bunch of people who would later become some of the Neuralink co-founders. Tim Hanson was a grad student and then a postdoc in the lab, as was Joey O’Doherty, who is now one of the senior BCI engineers at Neuralink.

In fact, I remember that, a decade plus later, this would become the Neuralink surgical robot. I remember the first time Tim had this idea for building a neural sewing machine, and he ordered this old CNC machine, a pick-and-place machine, or a CNC, whatever it was. It was this extraordinarily busted thing, available for $1 on eBay plus $400 in shipping, that showed up and that, over the following 18 months, he turned into a very early prototype of this neurosurgical implantation robot.

Then they ended up moving out to San Francisco and becoming postdocs for Philip Sabes, who was later part of the founding team. But that lab was a formative experience.

Juan Benet

There are these very special labs, sometimes at universities, that end up shaping the people who then go on to create the whole field. This has happened a few times in computer science. For example, a lot of graphics came out of one lab at the beginning of the field, and AI came out of one or two labs.

Max Hodak

And Duke ’08 to ’12, there’s a lot of high-profile alumni there, but it is really overrepresented in Silicon Valley to some degree. Like Frederson, Byers, Zac Parrott, me, and a handful of others. I don’t know what was in the water there. Harvard, Stanford, yeah, that makes sense. But then there’s this whole group, especially in biotech, from Duke from that era that ended up being super overrepresented.

Juan Benet

How much did you learn at the university from the university itself, versus just getting connected to the relevant people in the relevant lab?

Max Hodak

I was not really into school. For my last two years of college, I actually ended up working out in Silicon Valley and commuting to college because I wanted to keep working in the lab, and that was interesting research. But I kind of showed up to exams and turned in problem sets and otherwise did not spend that much time thinking about the coursework.

Each semester, I would stuff all of the labs into either the first half or the back half of the week and spend the rest of the time in California. There isn’t really advice in that.

Juan Benet

I think it’s just a trade of how people go through college and so on. We both know a number of people who chose not to go down the university path and now do incredibly breakthrough R&D work in a range of places and are effectively self-taught, or they found ways of learning the relevant material.

Max Hodak

I did really feel like I needed to finish the undergrad degree because I’m enough of a dropout without having a PhD. Someday, if the stuff that we’re doing really works well, I’ll fix that at some point. But most of my education happened outside the classroom.

Juan Benet

Yeah. I think in your case, in neuro, you can’t be a university dropout. You have to be a PhD dropout.

Max Hodak

Something like that.

Juan Benet

What about after university? From there, I think you went to Transcriptic.

Max Hodak

Yeah. So when I graduated from college, I thought what I wanted to do was start a BCI company. But this was 2012, and I thought it would take at least $100 million. And this technology was at least ten years away from clinical translation, and I did not have $100 million.

So the choice was basically: go to grad school, in which case I could spend 6 or 7 years in grad school, and after that be kind of a postdoc who still did not have access to $100 million that nobody really cared about. Or I could move out to Silicon Valley, start a different company that I thought was more accessible with the kind of resources that I could raise at the time, and really learn how to build companies.

And there was another idea that I had that I thought was pretty good. In addition to working in the Nicolelis lab, I also spent a little bit of time in a synthetic biology lab where I had the experience of going into the lab to press a button on a machine every three hours for three days to take a growth curve with this bacteria that I was engineering.

And I was like, there’s got to be a better way.

Juan Benet

Yeah, it is astonishing.

Max Hodak

And to do that, you also needed all this lab space and millions of dollars of equipment. This was around the time that cloud computing was really starting to happen. And so the idea felt very obvious at the time, which was, what if instead of having your own lab, you had a cloud lab?

And that was not something you could simulate. You needed a physical laboratory with a bunch of robotics. The idea was that we’d have a set of APIs that scientists could use over the internet to run experiments. And one of the things that made me think this was possible was that I’d figured out by this point that you could buy lab equipment surprisingly cheaply at auction.

And so for the ten or so million dollars that was accessible, that I could come out and raise from Google Ventures and some others, you could get started on that. And so that became Transcriptic. And from 2012 to the end of 2016, I was founder and CEO of Transcriptic. We built a small business there.

Juan Benet

It was automating all of the flows of a specific synthetic biology process.

Max Hodak

Some parts of cell and molecular biology, primarily molecular biology.

Juan Benet

And this is automating it with robots. How much of it was software around existing machines versus robots on top of the machines?

Max Hodak

It was existing machines and custom robotic arms. We built custom refrigerators, freezers, incubators, and custom transport systems. The standard unit of device here is this thing called the microplate. So we made custom robots for moving microplates between devices. And all of the software is custom.

Juan Benet

By the way, this is kind of what TSMC and others look like today, with the wafers.

Max Hodak

Exactly. So Transcriptic ultimately did not revolutionize its industry. I think we made some key mistakes there. That was definitely on hard mode.

Early on, we were selling to academics because I was a student. That was what I knew. I knew these academic labs. I was not connected to big drug companies.

But automation in biology is good if you need to run an assay 10,000 times, or 100 times. It can do that quite well. But biology is still really pretty finicky. It is not the case that you can just write your protocol out as Python and then run that and have it work the first time. There is a lot of high-touch interaction there.

And with academic customers, if you gave them the option that their protocol could be twice as complicated, but they would save 20% per sample, they would do that every time. And that just turns out to not really be a good fit for this type of automation.

So a couple of years in, we had really figured out that where the business was, was in serving pharma and conventional drug discovery. And we eventually got some big contracts there. We ended up with this huge contract to run this $100 million Eli Lilly facility in San Diego, and that business grew to pretty substantial revenue.

That year was my last year as CEO, when I stepped down to co-found Neuralink. And ultimately, I do not think it lived up to its potential. There are elements of that business that I think should be tried again at some point.

But still, to this day, I think one of the structural things that I got wrong there, or that just was not in my worldview at the time and is only now becoming possible, was that biology does not happen in milliliters. It barely even happens in microliters. Biology happens in picoliters and nanoliters.

And microfluidics has been around for a long time. People have talked about building a lab on a chip for a long time. And there are lots of lab-on-a-chip single-purpose things. There are lots of ASICs, but there is no lab-on-a-chip CPU. And that has been a very resistant problem that a lot of people have thought about.

But things like modern DNA sequencers, and lots of other things like that, run on microfluidics. And whenever you can scale down, whenever you can scale down your automation to deal with biology on its length and time scales, things will get more predictable.

So I would approach it very differently if I were to do it today.

Juan Benet

As you mentioned, you have to step down from transcriptic.

Max Hodak

Yeah. In the summer of 2016, I got introduced to Elon, who already had the name Neuralink in mind. He knew he wanted to start this and was very concerned about what he saw on the horizon with AI, which I think has proved to be remarkably prescient.

And I think he deserves a lot of credit for creating the modern instantiation of this field. Without his bet on BCI, there certainly would not be the ecosystem in industry that exists today.

Juan Benet

Let’s end on this: what advice would you have for yourself coming out of high school or college, or maybe not necessarily yourself, just people right now staring into 2030 and 2035 coming up? What would you recommend brilliant people out there be focused on?

Max Hodak

Well, I think one thing to be very careful of is how social media and mobile devices have atrophied attention. I think there is a much more pernicious danger in copy-pasting from ChatGPT atrophying reasoning ability. So you have to keep thinking for yourself.

In fact, I think this is one of the central things that makes having any interesting career really hard. If you cheat on a test, then your grade will trend toward the average of the class. And certainly for startups, that would be failure. So you have to do better than that. And the only way to do better than that is to really think for yourself.

It is actually quite hard to do that because, by its nature, you will have people telling you otherwise, including objectively successful, smart people who you think are capable of giving you all this input. And nevertheless, you must pick things that make sense to you, even when you are alone in it. Because you might be wrong, but you cannot beat that, ultimately.

And as soon as you start delegating that, I mean, you can, in your judgment, delegate the judgment, but you still have to take all-cause responsibility for the results.

I cannot tell you how many times I have seen something where there is a consensus view, or literally everybody has some perspective on something, and it does not feel right to me. But then you are like, should I trust that? Because I do not think my judgment is perfect. So how do I know?

But ultimately, you just have to do things that make sense to you, and then it will get revealed how good your judgment is. That is the only way to get differentiated results.

And I worry as we start offloading more. Like we were talking about earlier, if you have your phone with you, your memory gets worse and your attention gets worse. And now you have this possibility of offloading thinking, so unless you want to be totally dependent on ChatGPT or Claude or Gemini, I think that is a real danger.

That is one of the AGI takeover scenarios that does not get talked about as much. It is not that there are robots out hunting humans, or even a single big agent that is clearly running the world. It is just that all of the humans have totally delegated their thinking to the machine, because it is producing better decisions than they used to have, and then suddenly you are kind of being puppets.

Juan Benet

Well, hopefully you’ll fix that by getting the biohybrid in place, and then other next-generation devices.

Max Hodak

So you have to be really thoughtful about how you integrate that with your own decision-making and your own thinking.

Juan Benet

When people are just starting out, it is sometimes hard to remember how hard certain things were, in terms of how you build your first project, your first attempt at a company, or raise your first money.

Max Hodak

Like, totally. I remember this clearly because I’d go ask experienced entrepreneurs, “How do you raise money?” or “How do you organize a team?” And they’d say, “Well, you need to have a really clear vision, and you need to have a good culture.”

And I’m like, okay, but what does that mean? How do I raise $1 million? “Having a good culture” is not practical advice. There are tactics that you can learn, and there are various things that are practical.

But then later on, you realize that really was the important thing. You actually did need it. Like you’re saying, how do you make sure that you’re always moving as fast as you can, that you have these short time constants, and that you don’t just bleed out by a death of a thousand cuts, these small slips that people don’t take seriously? The answer to that is culture.

And in some sense, these are the answers. But you also need the more tactical specifics.

Juan Benet

Yeah. There’s a great meme about how to draw an owl, which starts with two circles.

Max Hodak

And then you draw.

Juan Benet

You draw the rest of the fucking owl. So a lot of the advice is you start with a two circles, right?

Max Hodak

And so the way that I’ve characterized it, the apprenticeship that I did at the company I was at before Science was invaluable, because there aren’t really generic principles of startup advice where, if you hear the right thousand words, you can turn the crank.

A career is a long series of decisions, and you want to be able to look at those fact patterns and make local decisions that are best for those fact patterns. And even if this looks inconsistent over time, if it is a situation that is similar, but the facts look slightly different, you should feel empowered to make a different decision.

But the thing that you want is for your filters to be well tuned so that, when you’re faced with these things, you make high-judgment decisions. And that is, I think, a different lens than asking what the key tactics are. You just need to be good at making decisions.

And the best way that I’ve seen for that is, this is really an apprenticeship. It’s tough. You can’t really learn it through school. I tried to do it on my own, and that was very difficult. Humans are very mimetic, and I think the PayPal experience, for a whole cluster of Silicon Valley, was that they figured it out and they did it. And that was a really powerful founding moment, the culture that came out of that. And then I, like many others, ultimately got some of that imparted by doing the apprenticeship. But these cultures are oral traditions.

Juan Benet

And also, in practice, it’s kind of like a Jedi apprenticeship situation where you have to join the group and then go through it.

Max Hodak

And they’re tough. I had the experience where it was your best friend’s birthday, and you had been making plans for it, and then you get the text: “The plane leaves in 45 minutes. Get there.”

And the only way to learn these things is to be trying to make these decisions with jeopardy and stakes attached, looking forward with uncertainty, when it matters. And you go through a bunch of reps of that, and then you find out if you can learn it or not.

Juan Benet

Do you have a young Max style apprentice?

Max Hodak

Well, we’re fortunate to work with some really talented teammates. We still have a lot to prove. I do not want to sit here and say I’ve worked for very successful people, I have friends who are very successful, and we still have a long way to go. I think we should acknowledge that.

But yes, we’re fortunate to work with some really talented teammates, including others who I hope receive the oral tradition.

Juan Benet

Yeah. Let’s finish with any recommendations for very inspiring sci-fi, people to read, books, or maybe underrated sci-fi.

Max Hodak

Pantheon is great. It’s an animated TV show. Highly recommended. I joke that it’s actually a horror story about the extinguishing of consciousness in the universe, but people should go watch it.

Juan Benet

Is that a spoiler?

Max Hodak

I don’t know, maybe a very mild one. The Culture series. I think if you want to talk about people saying, “Oh, we want to build an optimistic future,” the Culture series probably best embodies that. And it’s kind of unique among science fiction in that it deals with artificial intelligence.

AI is kind of a tricky topic because as soon as you have it, it’s kind of like time travel. The story is only about AI or only about time travel. And the Culture paints a picture of a far-future humanity that is fully elaborated in these ways and is, in some ways, a utopia, but in other ways very complicated.

So many of the stories are told on the periphery of the civilization, where they interact with others and face these complex moral and ethical dilemmas. And you see how those get navigated even in a basically utopian society.

I also just love The Expanse and hope very much that Mr. Bezos decides to make the last two seasons.

Juan Benet

Yeah, that’d be great.

Max Hodak

Yeah.

Juan Benet

Hey, thank you so much for doing this. It’s great chatting.

Max Hodak

Thanks for having me.

Disclaimer: https://bit.ly/PodcastDisclaimer