New episode with Dr. Adam Marblestone, CEO and co-founder of Convergent Research. Adam is an extraordinary catalyst and accelerant for science: he has influenced or routed billions in funding, he helped start significant companies and scientific subfields, and he is the world’s go-to person for neuroscience roadmapping. This year, the NSF launched a $1.5B X-Labs initiative inspired by the FRO model Adam pioneered. Though he started with neuroscience, today he helps organize scientific endeavors across many fields, including AI, math, biology, climate, astrophysics, and more.
Adam proposes that fields like neurotech, whole-brain emulation, cryo, and nanotech are severely limited by capital and coordination — not ideas. Connectomics is a clear case. He argues that mapping the brain’s full wiring diagram is the approach most poised-to-scale and still extremely neglected. Costs for a molecularly annotated mouse connectome have come down orders of magnitude, from an estimated $10B to $100M–200M, Adam suggests, with a human connectome perhaps costing around $1B–2B. The cost curve of connectomes is similar to transistors and gene sequencing: once it is low enough, we get extraordinary outcomes for humanity. Near-term applications could pay for it: new drug targets for brain disease, insights for AI development, emulations, and even “control knobs” for mood and focus we haven’t identified yet.
In this episode, we go deep on many topics in neurotech and beyond: what the brain can do that computers still can’t; what neuroscience could teach AI; how you’d actually map a whole mammal brain; how far today’s fly-brain simulations really get toward an upload; what it would take to move mind uploading beyond the fringe or science; where brain-computer interfaces go next; nanotech, reversible cryonics, AI doing its own ML research, and even predictive, agent-based economics. I’m very excited to have Adam on the podcast. Hope you enjoy!
Other links to this episode and references below.
Chapters
00:00:00 Introduction
00:01:40 What are the grand challenges of neurotech?
00:04:50 What the brain does that computers still can’t (on a few bananas a day)
00:14:43 The neuroscience overhang: why brains learn from so little data
00:26:29 How to record and map the brain: DNA “ticker tape,” ultrasound, and the unexplored map
00:40:29 Why connectomics is the area most poised to scale
00:47:54 How whole-brain connectomics works, end to end
00:55:27 Simulating the fly brain: how far does it actually get us?
00:57:51 What would it cost to map a mammal’s brain?
00:59:55 What will be connectomics’ ChatGPT moment?
01:04:35 What a connectome unlocks: disease, drug targets, and the brain’s control knobs
01:14:20 Scaling up to the human brain — a faster timeline than expected
01:17:34 Mapping activity, and what the fly connectome has revealed
01:27:55 What you could build with today’s connectome
01:35:29 Why uploading may be one of civilization’s great cornerstones and transitions
01:42:43 Optimistic visions of a future with uploading
01:49:12 Beyond neurotech: virtual cells, nanotech, and AI-driven science
02:15:29 BCIs today: semi-invasive devices, ultrasound, and what’s on the horizon
02:22:44 Why science is capital-and coordination-limited, not idea-limited
Listen to This Episode on
Links From the Podcast Episode
Guest & Organizations
Research Papers & Technical References
Referenced External Papers & Projects
A drosophila computational brain model reveals sensorimotor processing, Shiu et al. (2024)
ConnectomeBench: can LLMs proofread the connectome?, Brown et al. (2025)
E11 Bio — PRISM technology for self-correcting neuron tracing (2025)
ZAPBench — Zebrafish activity prediction benchmark (Google Research)
Decoupled Neural Interfaces using Synthetic Gradients, Jaderberg et al. (2016)
Books & Media
Juan & Protocol Labs
Transcription
Juan Benet
Our guest today is Adam Marblestone, CEO and co-founder of Convergent Research, where he’s pioneering a new model for accelerating science through focused research organizations. Adam has spent his career at the intersection of neuroscience, AI and biotechnology, with previous roles at Google DeepMind, Kernel, and MIT.
He worked with George Church, Ed Boyden, Konrad Kording, and a whole host of neuroscience luminaries. Adam has developed technologies for large scale brain mapping, connectomics, molecular recording, next generation neural interfaces, and beyond. In addition to neurotech, Adam has become one of the leading voices on how we can build better institutions for R&D and how to tackle the world’s most ambitious scientific challenges.
Welcome, Adam.
Adam Marblestone
Thank you so much. Awesome to be here.
Juan Benet
So let’s dig into the bulk of this conversation is going to be about connectomics, simulations, BCI and so on. You know we can talk about many other fields, but I think this is one where you spend a large fraction of your thinking time on, and it’s a field that we both see as greatly accelerating and with an enormous amount of promise in the coming five, ten, 15 years.
So walk us through what do you see as the grand challenges of neural tech?
Adam Marblestone
I like to think a lot about sort of the tech tree. You know, what are more fundamental capabilities than others? Obviously, intelligence or energy, these are really fundamental capabilities. At a certain point, I started seeing neurotech broadly construed. When I think about neurotech, I think both about direct, real world applications.
And I also think about these science applications that sort of applications in the structure of knowledge. Once we can crack neurotech that it really should be considered to be extremely root level or central in the tech tree. Beyond this kind of applied domain of stuff, we can do with the human brain, with devices or something like that.
So I think of it in maybe my view of neurotech is kind of maximally broad and includes really like the tech for science, which is a lot of also why we’ve been doing the focus research organizations and all that. But at a at a really high level, I think you can bucket. One bucket is sort of understanding the brain in both health and disease.
Sounds like an interesting thing to do. But I mean, if you really think about that, I mean, I think that is a kind of a AGI complete problem, right? That’s not necessarily how we’re mostly making progress on AI today, but understanding the actual algorithms of the human brain is a lot of how people thought we would get to AGI and is sort of clearly an AGI complete problem.
I think it’s also clearly somehow necessary on the path to understanding what consciousness is and more deeply understanding the features of human consciousness and what other types of consciousness is possible. And I think we are still far from that, but clearly that’s needed. Then there’s the sort of in disease part of that sentence where most brain diseases, we just don’t really know still what causes it.
I mean, what why don’t we have a cure for schizophrenia? I think it’s clear that that’s not just a kind of drug development problem in the traditional sense. There are some drug development problems within there, like delivery to the brain or so on, but we don’t actually know really what is different about a schizophrenic brain versus a quote unquote normal brain that then extends into tweaking the brain.
So we think of human nature essentially, or whatever, this distribution of different personalities and drives that people have as this kind of somewhat fixed distribution, in addition to, you know, doing something about really severe mental illnesses and so on, neurological illnesses. I think that we could, if we can understand what are the control knobs on sort of human drives and behaviors in general, how is that steered?
That is both health and sort of kind of the future of the human species in some broader way. Then there is the sort of BCI problem in terms of sort of bandwidth of communication with machines. Also, in terms of having devices that where you can really precisely tweak these types of brain states in real time.
What if I wanted to have control over my mood or my focus, but also a lot of just how do we communicate or even merge with machines? So that’s the sort of whole BCI bucket. I haven’t even talked about certain specific things about the brain. I said as an AGI complete problem, but it’s also enormously energy efficient.
It’s also constructed in a very different way out of biological components. So I really think that there’s a level at which it should be considered to be a root level node in science and technology at the same level as, you know, fundamental physics or energy or AI.
Juan Benet
So I always think that the brain is the most advanced machine that we’ve encountered in the observable universe so far. Now that we can look very far into the observable universe yet, but it clearly is so much more powerful in ways than the bulk of the technology that we’ve been able to build. What are some of the things that the brain can do that our computers still struggle with?
Adam Marblestone
Well, from a sort of AI perspective, it is an agentic system, you know, with continual learning, with world models, you know, that it can query in extremely flexible ways. It is creative, it is social, it is cooperative. These are kind of a who’s who of the grand challenges in AI as well. It runs on sort of a power output of, you know, tens of what’s sort of a few bananas a day or something like that at a scale that is at least comparable or greater to what we see with the largest AI models in terms of computation.
Juan Benet
To me, the brain is doing something really amazing because not only are the neurons dramatically larger, they’re able to run the computations with orders of magnitude less training data and less examples can make very quick judgments and decisions and so on. So there clearly is something about how the brain is computing that is vastly superior to the latest that we’ve been able to build with current artificial neural networks.
Yeah. So what do you think is leading to that difference. And maybe do you anticipate that we will be able to build those kinds of structures into the silicon substrate or.
Adam Marblestone
Yeah, it’s a really good question. I think that it’s an interesting question because there’s sort of different axes or frameworks or dimensions on which you might want to try to do a comparison. And when is it sort of a fair comparison or not. And so I actually think there needs to be a better analysis of exactly kind of the different contributors to these types of efficiency.
On the sort of energy efficiency side, you can point to a number of different aspects, some of which are more directly comparable than others. So you can talk about speed, switching speed, and of course, the neurons are switching at a speed that’s massively lower than what you see with, you know, gigahertz, you know, tens, hundreds of gigahertz switching in computers there, really switching a kind of millisecond timescales rather than kind of nanosecond or sub nanosecond timescales.
So it’s just doing a lot fewer switching operation is one thing. On the other hand, it’s massively parallel. On the other hand, it’s also sparse. You can also talk about the sort of physical.
Juan Benet
Thing between like 3 to 6 orders of magnitude less operations going on each cycle, I guess.
Adam Marblestone
Yeah. So should you think of it more as a sparse memory? People have talked about this for many years. Do you think of it more as a sparse memory versus as a computing substrate? Is it more of a memory lookup? But how is that possible? It’s also obviously running a learning algorithm. It’s also obviously doing real time control.
You can talk about some aspects of the hardware. You know, just if you look at the voltage across some neuronal membrane, you know, talking about tens or hundreds of millivolts as opposed to like volt scale transistors, and that actually with capacitors you can get a sort of v squared difference. So if you can sort of get down closer to the millivolts level with transistors, maybe you can also mimic that just raw hardware.
You can talk about noisy ness and the ability to tolerate noise and imprecision, where each synaptic event is somewhat stochastic. And the receptors that are binding that are sort of these molecular machines that are really maybe close to thermal limits in terms of how precisely they can detect things, just a small number of molecules and a minimum number of molecules and receptors that it needs to be able to use.
And those are stochastic. And so it can use that. So but all of the things I’m talking about there, mixture of hardware algorithm architecture, it’s sort of solving a somewhat different problem. It’s doing it is somewhat different architecture using somewhat different hardware. And so what are the sort of contributors.
And if you were to transfer some of those but not others over, but I suspect that the future of sort of neuromorphic chips that are not biology is very bright, because I do think that you’ll be able to do stochastic switching, and you’ll be able to do lower voltage switching, and you’ll be able to do 3D and you’ll be able to do different architectures and you’ll be able to do different computations or sort of problems that it is solving in principle.
None of those things is something that’s about biology and such a deep way. I mean, if I was to sort of caveat that a little bit, there may be some somewhat special things about it, because it is also self constructing in a certain way. So I could do a computation where I say, you know, if these two neurons are firing together, they will physically form a wire, or even if these are firing together, let’s send a wire over here during development and actually grow a new projection over here.
So you do have something that I think you don’t have on any kind of silicon, where it’s like, I can grow a new wire. I’m not just reweighting or turn on and off, await or sort of reconfigure something like in an FPGA, but like actually physically grow a new wire. So, so the development of the connectome, maybe there’s something hidden in there with a sort of 3D self constructing self-replicating.
Yeah. Human brain messes us up.
Juan Benet
B fab out of Taiwan semi.
Adam Marblestone
Yeah. Does that allow a learning algorithm that is. Or is it really should be thought of a learning and development algorithm. So on these kind of edge cases of how much is happening during prenatal neonatal development. Like maybe there’s some stuff that’s hard to mimic, but I’m not sure how much of the computation.
I suspect that we’ll be able to ultimately mimic it. And the question is, can we understand and partially it what’s an algorithm difference? What’s the hardware difference? What’s a computation difference? It’s all understanding.
Juan Benet
On the kind of growing new wiring together, that reconfiguration of the network at scale across many different substructures in the brain seems like a really powerful algorithm. But potentially, I mean, couldn’t you achieve something like that today in artificial neural networks by just having a lot of connectivity in the graph and then clamping things down, or if edges are just not active?
Adam Marblestone
Yeah, I think that there seem to be effective ways now of doing mixture of experts, effective ways of doing sparsity. And it will only get better in kind of silicon hardware of how can you only use tiny subsets of the weight matrix? There’s still the stuff we say. Maybe. Maybe you have to first load the weight matrix and store it somewhere, and then you have to access different pieces of it.
And the brain doesn’t have the separation between compute and memory, which is people. Something that people talk about and neuromorphic a lot is that it’s not sort of maybe it’s all memory lookups in a certain sense, but it’s also not sort of bottlenecked by memory lookups when it’s doing learning. So there’s there’s a lot.
But I think that in some ways, of course, digital computers have huge advantages. The fact that I can prune an arbitrary weight, this question of the credit assignment problem, the fact that I can run an exact version of backprop, or even better, gradient algorithms, you know, that’s something that the brain has to then work really hard to do when it’s a self constructing from cells and distributed system that’s actually very hard to do.
Juan Benet
And where you primarily have local computations and you have like distributed algorithms that have to kind of synchronize across the entire brain. Yeah, yeah, yeah.
Adam Marblestone
I think it has, you know, some amount of long range transfers and stuff it can do, but it’s a lot of it is sort of everything. I mean, everything in physics is built up locally at some level. But yeah, it really has to build everything out of neurons. I think this is something that is not always appreciated in the AI world, to quite the degree.
It sort of feels like it should be made.
Juan Benet
Do crisply.
Adam Marblestone
Everything. Reward functions, training algorithms. Everything is made of cells. It’s not. You don’t have Python code where you can adjust that. You have to make the entire training process, evaluation process. Any analogies to any of those things have to be done with cells that only know locally that they are themselves and who they’re connected to.
Juan Benet
So the equivalent in artificial neural networks would be everything is made of silicon integrated circuit gates anyway. So like Python is only an abstraction that we’ve built on top. So we might have like a Python equivalent in the brain. That helps.
Adam Marblestone
Yeah, at some level that’s true. And then you have to ask, you know, there’s a lot of great things about the von Neumann architecture being able to say, just give me this half of the weight matrix and send it over here. Transpose the whole thing. Okay, great. Give me number. Give me this element. We can do mathematical operations on it that allow us to kind of move any bit, anywhere in the memory.
That’s something the brain. I don’t think it can move an arbitrary bit to an arbitrary location, you know, based on the architecture.
Juan Benet
Given these order magnitude differences in the computational scale of the brain versus what we can do in silicon systems, that means that there’s a, again, 3 to 6 order of magnitude gap between what artificial neural networks are able to do today versus what they could do if we figure out how to improve them algorithmically.
Could this really be about figuring out how neural circuits learn, or is it some of the loss function type questions, or is it wiring in between? Is it more like an architecture of like how you compose larger neural circuits? Or if you had a guess like today, like what is the opportunity space here? Like if we figure out based on how the brain works, how should we be building artificial neural networks.
Adam Marblestone
It’s a really fascinating, you know, time to be asking this question, because I think in historical times when people have asked this question of computational neuroscientists, you know, they didn’t have any working algorithms, right? They had computational neuroscience models of how neurons and synapses, you know, interact and communicate to some degree.
But they were not achieving something anywhere close to AGI and kind of nobody else’s were either. And so, you know, what are you going to do now? We have at least some working examples of things that really work. So backprop is actually quite a good algorithm. It’s not to say that that’s the only thing that the brain uses, or even it can do something quite as good.
But we didn’t know if that would be sufficient to do anything really significant until whatever, 2012 or something, you know, in terms of real world tasks. And so it’s a really interesting question to sort of now we have a bit of a basis of comparison. We’re still somewhat lacking on ground truth data of what does the brain actually do.
And so I think part of the reason why we’ve had the struggles in sort of neuroscience, really driving AI over the last few years is just neuroscience itself, somewhat slowed down, you know, by comparison. And we’re trying to really accelerate it. But yeah, what are some of the concrete areas where there’s a potential sort of for a neuroscience overhang of some kind?
Juan Benet
Oh, sorry. It was kind of saying before we dig into the neuroscience overhang that there is like an AI overhang relative to how the brain works, and that if we figure out the algorithms that the brain uses now, we can create equivalence for artificial neural networks. Now, that would significantly decrease the cost structure of running AI systems today, right.
So like if you’re spending like billions to tens of billions of dollars in the data centers to train the current generation of models, but you can get a three order of magnitude improvement in like.
Adam Marblestone
Well, the question is how much can we get? That sort of looks like an incremental improvement within the current LLM kind of paradigm. And how much do you have to sort of use neuroscience or something else to get to a new paradigm? And then that has efficiencies. But I think there may be at least some possibility to learn something that is relevant to the current paradigm.
This question about continual learning or data efficient learning. I think of continual learning extremely fast learning and data efficient learning as sort of. Potentially all somewhat the same problem. How do you really learn from a single episode? And of course with in context learning now? This was the whole point of like the GPT three paper with lots and lots of pre-training.
You see some of that behavior of one shot, you know, sort of zero shot or one shot kind of learning. But the brain seems to be able to do that at the level of, of weights, which is something that we don’t have a solution for extremely data efficient or, and or continual learning. And these problems may separate a little bit in the sense that I think that part of why we have so much data efficiency, maybe that we are not trying to learn as general in some ways, a set of information.
We’re not just able unless you’re a savant and you devote a lot of your cortex, you know, to just memorizing, you know, dates of historical events or something. We’re not trying to memorize all that. We are optimized to build models that work for our real world, 3D embodiment, social world to learn from language.
Its making a correct comparison of how data efficient we are, I think is also one of the major kind of outstanding difficult things. It’s not so much that the LLM has pre-training and we have evolution, because I think if you look at how a baby comes out, it takes a really long time to figure out how to bring a spoon to your mouth.
It takes a long time to be able to say any words at all or understand any concepts. So it’s not so much that, you know, oh, we can just kind of say, well, we’re like a few shot learners, but so is Claude because we, you know, we had lots of evolution and Claude had lots of pre-training. There’s something about the within lifetime learning algorithm that we have that is a and this makes the problem maybe easier, more targeted to the sort of things that humans actually have to do.
And that comes into the cost functions and the curriculum. And then there’s also something that may just be objectively better at it, at learning in general, which is this faster, continual learning, which I think we could probably learn a lot from. The hippocampal cortical Consolidation, sort of systems consolidation.
This is something that DeepMind and Demis Hassabis, you know, talked about a lot, you know, even ten years ago, replaying memories. How do you store memories in a way that leads to a lot of generalization and updating the right weights in the cortex? What is the way to sort of access in? Can the hippocampus sort of access into the cortex.
And so these are the ones I need to tweak given this new important memory that I just formed. Can it learn to do that? Is there a learning algorithm that learns how to do fast learning, and do we have that in the brain? So there’s probably something to this continual learning problem that’s like a discrete thing that the brain just generally does better.
There’s also we are somehow learning fewer things in order to get AGI than the pretraining paradigm is. I think we also have in the brain much more powerful reinforcement learning systems than the ones that are being used in post training. In our current paradigm, model based, imagining the future, learning from imagined events, different ways of assigning credit to things over time and space and memory.
Whereas you know, current AI post training that does use RL is like just using very simple versions of RL applied in an enormous scale. But we have large parts of our brain. Specialized systems in our brain, like the hippocampus, are sort of, I think, there to help us with continual learning. Fast learning, fast generalization, fast weights, parts of our brain, amygdala, striatum, etc. are very much there to have a prefrontal cortex to do a much better model based learning.
Juan Benet
What kind of experiments would you design in artificial neural network land to see if we can figure out some of these structures? A lot of what we’re going to talk about is purely on the organic neural network landscape, and how to do all of the mapping and learn what the brain is actually doing. But, you know, there’s a an enormous amount of work being done right now to scale up the just like the raw hardware, energy and systems to power the current infrastructure and maybe a lot of work in the tap out of new chipsets and so on.
That better fit the current paradigm. And surely a range of the labs are running, you know, internal research programs to figure out, like better algorithmic structures. But it seems to me like they’re poking at the layers above the current paradigm and less at the layers below and in the middle of the current paradigm.
And there could be like a range of experiments that are maybe neuroscience inspired that could yield, you know, neuro architectural improvements or finding like better loss functions or whatever. Like in 2026, we should be able to run a large portfolio of experiments.
Adam Marblestone
Yeah. And I think that this gets into the whole issue of neolab and sort of what do they do that’s different? When do you sort of get a certain level of scale. But it’s kind of interesting. I mean, you could say that this points more towards doing more neuroscience, or you could say this points towards doing less in a certain way, or which is that circa 2015, 2016, 2017, there was a lot more exploration of model based RL.
There was a lot more exploration of different ways you could do continual learning. What about recurrent nets instead of transformers, different RL algorithms, and of course the transformer and relatively simple post training and so on is what is working at scale with language and sort of internet scale data.
And that’s scaled. But even using the same GPUs and TPUs and so on. I think there was just a richer research field of what you would call neuro inspired AI. A lot of what was happening is less necessarily you need a new neuroscience experiment. It’s really just that you take neuroscientists and you train them up on AI, and they bring a different set of sort of intuitions or research taste essentially into it.
I remember we didn’t get into it that much in our previous discussion, but this question of sort of, how does the brain do anything in terms of very general learning with credit assignment, you know, adjusting the right neuron to give the right outcome for prediction or for something else that’s even close to what backprop does, right?
What gradients do? Can you actually propagate a gradient through different layers of cortex, areas of cortex or other systems. How does that even work when you don’t have exact asymmetries of weights? And again, this ability to sort of look up any weight in the weight matrix and apply any calculation to that, in principle, that has to be a vectorized calculation or whatever for GPU.
But basically any calculation, there was a bunch of cool work on that that also came from the AI side. I remember a paper on synthetic gradients where instead of having to propagate the exact gradient from layer to layer, you would have part of the network would be making a prediction of what gradient it would be getting and what gradient it should be sending, and then it could just use the predicted gradient instead of using the one it was actually sending.
So you devote some neurons in your local region of cortex, if you will. This is actually a paper, Jaderberg et al. He’s now running Isomorphic Labs. But back in the whatever this was 2016 or 17 era Jaderberg was working on, could you do backprop without having to have weight, you know, perfect symmetry. Could you do a local version?
I mean, if you could scale up something like that even, which doesn’t really have neuroscience, except at a very inspirational level, that somehow the brain does this and there must be some clever thing, even without connectome and so on. I think just slightly shifting the priorities of research from passing ex, you know, next eval to what if we could do this in a local learning way, meta learn to learn, etc.?
You know, these are all things that haven’t been brought to the same scale, but a lot of ideas are out there and even just neuroscience inspired AI. But part of the problem is there are so many such ideas. How do you know? And I think we struggled with that. You didn’t know what was the right memory algorithm, or the right local learning algorithm, or the right reward function.
And so if something works that’s really valuable, you’ve got to scale that. Whereas the brain would provide an additional source of constraint to help narrow down how to use these still scarce resources of GPUs.
Juan Benet
Though I would imagine the budgets of today, both in terms of the improved hardware and the
Juan Benet
vast scale of computing machinery that we now have right at the disposal of all the AI companies. You could run a research program. Many orders of magnitude larger than the 2016 research program exploring a very large design space is happening to some degree. But like for me, it feels like a lot of the attention in AI is to keep squeezing out like the returns of the scaling laws as much as possible, which sure makes sense when it keeps yielding results.
But once you get into like the trillions to soon tens of trillions scale, when you think of all the AI companies put together, you start needing to go back to figuring out if you can just squeeze out returns by finding dramatic efficiencies.
Adam Marblestone
And I should also say that I think there is a bunch of safety implications and questions here where I think that we may have somewhat lucked out with the current paradigm of sort of not exactly alignment by default, but much more ability to get alignment because of the enormous pre-training that these things have been subjected to.
So we can ask it, you know, what? What would be a reasonable thing for someone to do in this situation? Would this make this person upset? Would you go to jail for this? And it’s all in pre-training. And as you get to more powerful, you know, learn end to end RL. I mean, some of the paradigms that were being, let’s say, the AlphaGo style paradigm, I think the AlphaGo style paradigm could be much more dangerous also.
So there’s some question of do we want to be doing that, or do we want to be doing the alignment aspects first? I think there’s there’s a bunch of neuroscience to be done on what is the basis of human morality and what a human’s being aligned to.
Juan Benet
It seems to me that the current paradigm of learning everything about our writings first and develop intelligence on top of that, at least bakes in all of our notions of morality and ethics right into it.
Adam Marblestone
There’s some vector there for us that you can then ask it to collapse onto the vector, as opposed to trying to reason with.
Juan Benet
AlphaStar as it like, is destroying your base, right?
Adam Marblestone
Right. How do you start from AlphaStar and get to morality? Whereas, you know, there’s morality is in there, and I just need to bias it a little bit. So don’t ask it to hack computers. ask it to be helpful and harmless and so on, and then it will collapse onto this cultural concept that we have. So there’s that.
On the other hand, there’s things that would be very beneficial about having ultra low energy, for example. So I think the concentration of power or type issues or, you know, that’s much better if you have lots and lots of low energy chips that can kind of ubiquitously do this. Could have my own AI to counter your.
Yeah. The sort of human society working together is much more powerful than a giant brain of an elephant or something like that. And so maybe not a perfect analogy, but I think there will be pros and cons to this. And yeah, I think when people are talking about we’re going to use lots of our compute and people are very worried about this, probably rightfully so.
You know, instead of serving up models to customers, if they start using their compute internally to do sort of automated research, there’s some things that are exciting and also scary about that. But we’re now at a capability to do that that’s much higher than we were. Now. Imagine automated research on AlphaZero or something like that instead of on just improving Claude.
Like that’s arguably somewhat dangerous. But I do think that there’s a lot you could do, but just by asking it to try to be inspired by the brain, even without knowing more than we know about the brain than we do today. Slightly. Slightly frightening. Yeah, yeah.
Juan Benet
Let’s walk through kind of the approaches to brain recording and mapping. Kind of like, what do we know has worked so far? What are the current approaches that are being tried? You’re one of the people in the world that has the best view into what is happening across the field, and what’s not being tried.
That could be that could be being tried. So, you know, what have we achieved so far in terms of recording the brain and getting full connections and so on?
Adam Marblestone
Great. Yeah. And maybe we can divide. We talk a bit about recording and the dynamics problems sort of, which I think is adjacent to the BCI problem, but not exactly the same. I think asking the question of could I record or monitor in some way the full dynamics of every neuron in a thick, intact brain, let’s say that of a mouse brain.
That’s sort of how I started thinking about that. That has all sorts of scientific value or value as data sets. It has some significant but not perfect overlap with the BCI problem that is largely constrained by what human actually want to put on their head in some sense. But there is overlap, because, you know, a lot of what’s the problem for this recording issue is that the brain is this thick, three dimensional object.
Like, I think if you imagine we live in flatland and everything is either completely optically transparent or just flat, cells are just one layer. I think some people have talked about this thought experiment. Biology would essentially be kind of solved now if we were just kind of this two dimensional, and you could go from the third dimension and say, oh, I want to know what that cell is doing.
Let me put a patch pipette into that cell and stuck out some molecules from that and ask a question about it, or record the activity. Look at which a giant array of cameras, you know, just in parallel. It would be kind of solved if we were completely flat. The basic issue is that we are thick, convoluted in 3D, and that different forms of radiation don’t penetrate even into the center of a mouse brain, which is about a cubic centimeter in size.
And then for the BCI problem, there’s a sort of very related problem, often having to do with the skull, that the same forms of radiation don’t penetrate across the skull. Or if you can even get to superficial layers, it doesn’t penetrate deeply, or it would be damaging or dangerous in some way to try to penetrate that with huge numbers of electrodes.
So if you look in the BCI case, the latest thing that Neuralink did was to have tiny, tiny electrodes. But they go through this dura, which is this protective layer of the brain. But that same issue of being able to penetrate the electrode deep in general without hitting blood vessels and so on, and targeting it, even though you can’t see the end of it, is potentially really powerful for science applications of recording.
So anyway, so in 2013, I started thinking about this, and the reason I had started thinking about this was I was a student in the Church lab, and I was very interested in neuroscience. And what can the Church lab, which is kind of genomics, molecular technology lab do. I was a grad student. I didn’t have a team or a funder or anything, but I was trying to figure out what could a molecular biologist do for this brain recording problem?
And we came to the conclusion that, at least in principle, there was something that you could do, which is that you could have every neuron record its own activity, trace how it’s spiking over time into DNA molecules. And then at the end, you sort of freeze the brain, fix the brain, and you read all the DNA molecules, cell by cell.
So now you really do slice it into flatland, and you pull out this DNA and this DNA, and each one has a recording trace. They’re all temporally locked to each other. This is the idea of a DNA ticker tape for brain recording. And there are lots of practical challenges with this. We did a little bit of very early work on this.
There’s now a bunch of different fields doing different kinds of molecular recording devices for different applications, tracking cell divisions for cancer, all sorts of different things that people are trying to do that are at a slower timescale, usually, than neural activity. There have been variations of rather than recording into DNA, you record into protein filaments and really interesting things happening now.
Can you read it out with a microscope instead of with a DNA sequencer. So this molecular recording field has started to take off. The thing that I took away from it was, gosh, neuroscientists are really not looking at the whole space of possibilities here. They are mostly incentivized to make this electrode a little bit better, or this microscope a little bit better.
And so there’s this whole idea of scientific road mapping. What are all the possible ways you could do neural recording. And we’ve now found that there’s some part of the map which hadn’t really been very much explored, which is the DNA ticker tape or in general, molecular recording device, part of the map.
What are the other unexplored parts of the map? And so we started thinking a lot about this. And I would say that in terms of where we are now, some of those things have started to become more fields or more companies, or partly because of the interest in BCI. So in general, at the time, in 2013, where we were thinking about this, there were not really much in the way of BCI companies at all.
There was the company that did the Utah array, you know, there was fMRI, there was EEG headsets and so on Meg devices. But there was not this flowering of cross-disciplinary deep tech companies Neuralink, Merge Labs, you know, precision pursuing different types of ideas. And I would extend that into others that are now interested in noninvasive methods Open Water, Aleph, some of these others.
But at the time we were thinking a lot about, okay, well, what are the physical modalities that you could use for recording and what types of problems do they have? Would it be possible to somehow make MRI better and record every neurons activity? What about things you can do with light? Optical methods were at that time in 2013, already a workhorse of brain recording with this technology of calcium imaging.
So you’re you have a protein that fluoresce is more or less depending on how much calcium is in the cell. And you basically look at that with a microscope and a camera, and you can see the individual cells as little dots. And those dots have changing intensity with calcium as it flows in when the neuron fires.
So you have this calcium imaging. And people were using things like two photon microscopes, which allow you to see deeper. But it wasn’t at all clear. Like, could you use an optical method to go all the way from one side of the mouse brain or the human brain to another? That was sort of a different field of pure optics that called imaging and scattering media.
And, you know, it turns out that there are things you can do, but I think a lot of those have turned out, and we’re not surprised by this. A lot of things you can do with new modalities require new chips, new hardware, some kind of industrial grade thing beyond just the thing we were piggybacking on already, which was things like CCD cameras, CMOS cameras that of course, neuroscience was already using.
But you need ultrasound on a chip technology to be able to have extremely precise control over the ultrasound wave. You need sort of light on a chip technology where you can alter the wavefront of light at the level of a single wavelength kind of precision, but over a very wide area, and do it very quickly before the light sort of goes out of phase with lasers or scattering.
And so some of these things have progressed. So for example, with Forest Neurotech, which is one of our FROs and now Merge Labs. And now some other companies like Aleph, doing stuff with ultrasound on a chip technology. Ultrasound has this really fantastic advantage and this was something DJ Seo before he was the COO of Neuralink.
He was first of all a coauthor of this road mapping paper that we did. But second of all, more importantly, he had authored a paper I think came out in 2013 as well, called Neural Dust. And one of the things that he pointed out in that paper was that ultrasound, certain wavelengths of ultrasound have very favorable properties.
They can penetrate much deeper into at least the soft tissue of the brain. But those same wavelengths that can penetrate pretty well also are spatially about the same size as a cell. So we’re talking about sort of tens or hundreds of megahertz. Ultrasound frequencies are both penetrate and spatially precise.
So could you use ultrasound for combination of spatial precision and field of view basically or depth. And that is now starting to be embodied with things like forest nanotech merge. And then how do you combine that with molecular transducers of ultrasound. So then that poses a whole engineering problem of how do I make the molecules that are sensitive to ultrasound or the chips, which would have been normal dust, would have been one of them that are sensitive to ultrasound.
With optics, though, we don’t have quite as much progress in the sort of optics on a chip equivalent of what has happened with ultrasound on a chip. In some cases, these are just a few companies have ended up doing these things. So Open Water and Mary Lou Jepsen and some others are in sort of notable company that talks about or has talked about having gigapixel microsecond spatial light modulators where you could steer light, and even if the light is scattering as it propagates into the brain, you could invert that scattering in some way.
That then gets you into combinations of ultrasound and optics. What if you use optics for the contrast but ultrasound for the spatial resolution. And that means how do you filter out the light that is tagged with ultrasound from the light that didn’t see the ultrasound focus, for example? There’s all sorts of interesting work starting to be done on there, but there hasn’t been a large enough community.
There hasn’t been a large enough amount of the underlying foundries and hardware and systems that support controlling these very high degree of freedom radiation or separating out radiation from another in extremely precise ways. So overall, I think that we are still awaiting some of the larger breakthroughs that come from exploiting physics modalities other than electrodes or MRI or sort of simple microscopes.
Gigapixel light modulators will be powerful new ways of separating out frequencies of light. There has been progress. So like from where we were, I think we’re up to being able to do organized projects. I would say focus research, organization size or X lab size projects. And it’s really actually exciting that the National Science Foundation now has a call for these things called X labs, which are sort of like noncommercial, not product commercial, product focused, but sort of organized, focused research organization like teams specifically around new transformative technologies in sensing and imaging.
And I think that we may see things like new ways of using magnetism. Magnetic fields really do penetrate from one end of tissue to the other. Sort of low frequency magnetic fields. They can just shine right through your body. Can we have a sort of magneto genetics? Can we have much more precise ways of picking up magnetic signals with different types of quantum sensors?
So we haven’t even started to get into magnetism. Those combinations of ultrasound and electrical, ultrasound and optical, where there’s some exploration of it. But it’s like nowhere near the scale that you see with even existing BCI companies, let alone, you know, TSMC, you know, GPUs, etc.. So I think that there’s still a very high overhang on the ability to record from large numbers of neurons.
And there has been definitely progress as a technique called light beads microscopy, which is sort of solving the problem that in a one of these two photon microscopes, you’re just kind of you have to scan the spot where the laser is focused across the tissue. And how do you sort of really parallelize the scanning.
So you’re not limited by sort of moving it from one point to another. It can be limited by the fluorescence lifetime of one thing has to be done fluorescence before you can measure the fluorescence of something else. So you can get fluorescence lifetime limited imaging. So there has been great progress, but it’s still mostly been academic scale efforts.
And so it’s like it’s kind of amazing to imagine. Could you see as much progress in the kind of fundamental capability to record from all the neurons as you’re seeing when we’ve started to industrialize in the last ten years? BCI proper? Yeah. The level of resources that’s applied to it. Neuralink was applied to the science of neural recording.
I mean, I think you would make headway, but these are it’s a little bit slow to date, and I’m optimistic something like X labs can accelerate that.
Juan Benet
But there’s this gigantic acceleration that happens when there is a truly industrial scale lab with a resourcing and team able to do end to end tool building and experiments and so on. And like the very fast paced, startup oriented landscape. And we’ve seen, like you mentioned, BCI, certainly a world of difference today from ten years ago, where now we have drastically better materials.
We have way smaller devices, much higher bandwidths, groups targeting, you know, thousands of channels, groups talking about scaling up to tens of thousands of channels.
Adam Marblestone
You can outsource different parts or source different parts of the supply chain. And sort of you’re not just having to use commercial off the shelf things. I mean, imagine if, you know, James Webb Space Telescope how to just use off the shelf mirrors, we would be in trouble. And so it has to it has to actually really do engineering.
That’s not astronomy, that’s engineering. And similarly this is engineering not neuroscience, but it serves neuroscience.
Juan Benet
The same thing happening in AI. Right? So as soon as the deep learning results showed promise in 2012, 2014, you get this scale of the computational power of different AI systems. You get this inflection point as soon as the graph has an inflection point, when the results start being shown, is downstream of a large amount of capital being poured into the work.
And the thing that’s been at least two major inflection points, 1 in 20 1416 and then another in 2019 2020 as you started getting like the closer to the ChatGPT moment, and I would imagine we’ll end up with another one. You know, like probably the last.
Adam Marblestone
And exactly where will be the first ones? Sort of different subsets of neuroscience. I think these sort of current mini renaissance or mini boom that we’re seeing with BCI companies serving the role as neural recording frontier labs. Neurotechnology. Frontier labs is part of the picture, and I think we also need some other kinds of frontier labs.
So maybe we can get to the connectome. But I think we badly need a circuit mapping frontier lab.
Juan Benet
So the inflection in these graphs represents a deep acceleration of the results that we’re getting out. And it’s all downstream of larger and larger amounts of capital being poured in. So when you’re describing this landscape frontier and recording, it seems to me that there is a huge acceleration that can happen if we can catalyze that scale of capital investment into the R&D of the engineering components.
Adam Marblestone
It’s not to say it’s just a matter of money, but I think that some of these things, it’s a matter of choosing the right problems, money and being organized and doing the right engineering. But that’s way different than needing a kind of Einsteinian breakthrough conceptual breakthrough. I think we can already tell you that if you have the gigapixel microsecond spatial light modulator, we could do things with optics in the brain that are totally different, new kinds of sensors.
So it’s a very good fit for directed research. And directed research is something about the coordination of capital and talent, and it’s less about a completely new idea that we’ve never had.
Juan Benet
Yeah, maybe going into the technology space of recording the dynamics of the brain, what about iconic comics? Seems like that has been scaling the approach of, you know, still like brain slicing and maybe using expansion microscopy and then, yeah, walk us through.
Adam Marblestone
So I will reveal some cards in the sense that I actually think that this is the one that is the most poised to scale. And it’s also extremely important and extremely neglected still that we scaled, as there’s been a lot of fascinating things you can do with recording more neurons. But when you record more neurons, it’s always in the context of a particular task and a particular thing that the animal has learned for a particular period of time.
You’re always choosing a subset of neurons or a species or so with connectome that you have the potential to truly be comprehensive. And I think ultimately all of the function of the brain is encoded in some structural level. You know, you have people that go, you know, fall in a very cold pond or go under anesthesia.
It’s not that brain activity completely stops under anesthesia, but the structure, well, what does evolution do? What does evolution have access to if it wants to create a new intelligence, a new learning machine? I mean, it basically has access to what it has direct access to is the genome. And what can the genome do?
The genome can control what genes are expressed in different conditions and so on with gene regulation. And it can express proteins like proteins that go on the surfaces of cells. One cell can detect another and choose to wire up. So I think that probably the way that evolution specifies or engineers or programs, all of what we’ve been talking about architecture and learning, algorithm and cost function and so on, that’s all programed into particular cell types of neurons that express particular shapes of the neuron that express particular propensities to wire with other neurons based on cell surface proteins, which then once they’ve got that wiring, that then embodies learning rules, because those particular combinations of molecules have both short term dynamics like facilitation or depression of the synapse and also plasticity.
But the evolution is going to program not necessarily just connections, but this kind of this idea of the molecularly annotated connectome. So what are the connections between what types of cells. So what genes are those cells expressing at the cell body. And then what are the particular combinations of molecules that are present at synapses or other kinds of communication points.
Neuro modulatory receptors. That’s a more free space communication gap junctions between cells. But basically if you know not only the connections but you know a select amount of what molecules are where, and you could also include in that things like ion channels are going to determine certain properties of the neurons themselves.
What would.
Juan Benet
Be the corresponding AI parlance description here is like trying to figure out what the weights between it.
Adam Marblestone
Would be. Yeah, and I think it would be the weights. And you will have to do an inference problem.
Juan Benet
If they actually.
Adam Marblestone
Run. Yeah. And to get the actual weights, you’ll have to do as sort of what Konrad calls like a calibration or compiling problem on your other episode and in a paper. But if I give you these molecules at a synapse, that’s not a direct measurement of the weights, but that’s a high dimensional measurement of the state and properties of that particular synapse, which I suspect is highly correlated and predictive of the dynamics, which is really the weight.
If I send a spike on this side, what happens on the other side? One millisecond, ten milliseconds, etc.. After. And also the plasticity rules, because those same molecular machines determine the plasticity and the local learning. So this molecularly annotated connectome, I think it contains sufficient information that with appropriate calibration and other measurements to link this ground truth high dimensional measurement to make a classifier a predictor of what that implies in different conditions.
But it contains all the information that you need to impute architecture, learning rules, cost functions, short term dynamics long term. So it’s basically everything except the input data is I think a bull case. Is that appropriate. Molecularly annotated connectome at least is the full data capture to capture all the information that you could then read it off if you had the appropriate readouts, which maybe you have to learn by some separate additional experiments of if I have this molecule as the synapse, what is then the transfer function in real time.
But those are all learnable measurements. So it’s basically this is the measurement you want to do at scale. And then you want to train some predictors of given that measurement what are other things you know other layers. So and.
Juan Benet
So this molecularly annotated connectome from an engineering perspective this looks like slicing the brain using expansion microscopy to grow the slices to be able to then image them optically, and then using ML to then figure out the wiring and.
Adam Marblestone
All of that. And we can talk about sort of where some of these things stand. That is one way of doing it. And in a sort of roadmap paper with a bunch of people from a similar era of 2013, we have this paper on connectomics, which is sort of what’s the economics or potential scaling behavior? We talk about electron microscopy based connectome.
We don’t really talk about X-rays, which is another way that people are exploring. But there’s electron microscopy based connectome, which is great because the electrons are very high resolution. So you can take very precise images, sort of very dense images, but they’re mostly undifferentiated molecularly, but they can be very dense.
You basically stain with metal broad classes of molecules, lipids and so on. So you can see very clearly boundaries of cells, kind of differentiations of sort of how much overall protein is in different parts of cells, things like that. But you don’t have these very specific molecular, partly because to get that in the optical context, you don’t want to cook the tissue, and electrons have a more propensity to cook the tissue and also preparing them to have electron contrast kind of chemically, it will alter the tissue and potentially hurt some molecules as well.
So with an optical approach, we’re using a light microscope. Classically. The problem is that the wavelength of light is too big to have resolution, that you can resolve the individual neurons and connections. And so back in 2013, we were thinking about even with a regular light microscope, could I target sort of DNA or protein barcodes to just to the synapses and sort of the synapses are then sort of spaced out in space slightly farther than the diffraction limit of light.
So even with low resolution microscopes and maybe some thin slicing, could you get a molecularly annotated connectome and apply different optics tricks and so on, and sort of tricks of making the neurons more distinguishable, identifiable with molecules. So this kind of notion of doing molecular entity connecting even at kind of low spatial resolution, then with expansion microscopy and expansion microscopy works already up to 1000 fold.
You can really get essentially tunable spatial resolution. And so now light microscopes are really on the map for in aggregate. You know, right now there’s not quite as much resolution as electron microscope if you really were slicing. But one of the advantages of that is that when you expand the tissue, it also becomes optically transparent.
It also becomes more permeable to different molecules or tags. And so you can use thicker slices of tissue. So one of the big problems of scaling electron microscopy to the whole mouse brain or whole human brain would be you’re going to lose a slice. It’s going to kind of blow away in the wind or get crushed or something like that, or get distorted.
Whereas if you have thicker sections, that’s a real advantage of the optical approach. You still have to have some sectioning and be able to bridge a little bit. But you have more information rich, thicker sections that I think is going to ultimately make the bridging much more doable. So in short, a lot of the key problems, the cost of the instrument, the lack of molecular annotation, the thin slices, those are mostly gone in the optical approach.
And so it’s very ready to scale. That leaves the AI problem of analyzing the images. Can I actually trace one neuron from another and distinguished segment them and ultimately proofread them. But, you know, maybe one way to point it out. And there’s this great paper from Jeff Brown called Connectome Bench that talks about this.
But the proofreading of these images, that’s ultimately a human computer use task. And for all the things we’ve been talking about, Llms, if there’s one thing that they should be good at, they should be good at automating a task that humans can already do by clicking at a computer, um, on images. Yeah.
Juan Benet
Can you explain sort of how did you like whole brain connectome books end to end?
Adam Marblestone
Yeah. So the sort of near term way of doing it now. And I think that in the future there will be even more advanced ways of doing this. I think you can imagine radically cheap ways of doing it. If you incorporate molecular barcoding, you might even be able to get nondestructive connectome or other things like that.
What we’re talking about, what’s ready to sort of scale. I think the thing that’s ready to scale is to take a brain, let’s say the brain of a mouse, you chemically produce it. So it’s now all the proteins are sort of locked in place. It’s sort of taxidermied, fixed, chemically fixed. You embed it with a bunch of molecular tags and also with a hydrogel that then physically expands.
And what that does is it makes it transparent and it moves molecules apart from each other such that you can then use the fastest, cheapest optical microscopes you can possibly buy or build. Fastest cameras you can possibly obtain to just collect voxels of this thing without any fancy electron microscope.
Just imagine how fast can you take pictures? But they are still 3D pictures because you’ve sliced into some number of sections, which is whatever fits within the working distance of your microscope is how thick the section can be. And so maybe you’re talking about dozens or hundreds of sections for something like a mouse brain, and you image each section, you have to register the sections to each other of where there was a cut between sections.
That cut has to be able to sufficiently limited damage to bridge neurons from one side to the other, or find a way of bridging. And then you have these stacks of 3D images, and they show kind of the shapes of the cells. They show also certain particular molecules, like on the presynaptic side of a synapse versus on the post-synaptic side of a synapse.
There’s also molecular tags that will show up in a different color. And then you have to now stitch this all into a description of the neuron as a connectivity graph with these molecular properties of the particular nodes and connections. And doing that requires a lot of storage, compression, compute, and currently requires a lot of human proofreading.
But I think that the proofreading is solvable by some combination of AI and clever ways of labeling. The neurons and E11 by one of our focus research organizations has published something on showing how you can label the neurons in more clever ways, and other people have shown better algorithms for mimicking the human proofreading behavior.
So then you have your connectome. And again, I think that this object contains much of the information that evolution itself uses to specify what brain is it going to build. You then have to understand what is the function that results from that in different ways, but that you have now digitized the information that is in a brain that determines how it functions.
Juan Benet
So you take a n, goes a brain. You apply this whole process of preparing, then slicing, and then taking pictures of the slices to then produce a three dimensional set of images. Then you annotate the images to then extract the wiring between all of the neurons, ideally with some amount of molecular annotations that can give you more information.
And out pops a file with all of the that just starts listing neurons that are connected to other neurons, and ideally with some amount of the molecular information between all of the synapses. Yeah.
Adam Marblestone
And again, I think that includes what an AI we would call the architecture and the weight matrix trivially. And I think it also if you can interpret this molecular patterns, I think it also contains learning algorithms. I think it also contains most of the cost functions because some of the neurons are sending the cost signals, some of them are sending learning signals, some of them are sending activity signals.
Some of them are sending attention signals on and off of a different region. But it’s all neurons.
Juan Benet
Yes. And it’s all embedded in the wiring and the molecular annotations in that wiring. That’s right. And so this is without any of the dynamics that you were talking about before of like being able to understand the activation pattern. The activation pattern seems to be important for modeling and simulation.
But I guess this is a quick question. But then we.
Adam Marblestone
Can then we can correlate it.
Juan Benet
So actually maybe what do you think right now. Do you think the activation will prove out to be required. Or can you actually get the function out of the structure?
Adam Marblestone
I think you’ll need to do some activity experiments in order to train the translator, or Konrad calls the compiler or calibration to translate between these molecular, annotated, even sort of subunits of neurons. That’s one way to do it. So I think that you need to do additional. I don’t think that just the knowledge that we have of the world now in mathematical calculations, plus the connectome, will give you a full simulation.
I think that you need activity data to get to simulation. You also need activity data to get understanding. But I think that again, with the field of neurosciences exists now, we’re constantly having to pick and choose. It’s sort of the blind men in the elephant situation where I can do an activity experiment and study this, but I don’t understand the circuit that that’s actually and I don’t even know if I know the activity of a neuron.
The likelihood that I also know who it’s connected to is actually really low. IARPA, you know, had this program called MICrONS, where for a cubic millimeter of cortical tissue, it was trying to do calcium recording in the cell bodies of the neurons. So how they’re turning on and off with calcium coupled with not really molecularly annotated, but with some connectome between those neurons and sort of link them.
But that’s like one of very few examples where you comprehensively have both structure and function to relate them. In general, you’re picking and choosing and looking at different dimensions of the black box, as in a very hypothesis driven way. The connectome, I think, is maybe going to be the first commodity that can kind of really scale.
Like you’re getting all the information about the activity is true, but you’re getting all the architectural information and you can compare that across species. Potentially. You can compare that across mouse that went in this maze versus that maze. It’s high dimensional enough. Comprehensive enough measurement that if it could be high throughput, then it’s kind of like the genome doesn’t tell us everything about biology, but we learn a lot by comparing genomes.
RNA sequencing is revolutionizing biology doesn’t tell. RNA doesn’t tell us everything. But single cell RNA sequencing because it’s high throughput and because it’s enough information, not just in a particular RNA, but it’s a genome wide method of all the RNAs. Single cell RNA sequencing is this revolution.
It tells us what all the different types of cells are, because that’s it’s differentiable at that level of comprehensiveness and scale. And I think maybe the first thing in neuroscience really that is particularly the molecular annotated version, that is, we don’t know how to translate it into all details of simulation and function, but it is like a comprehensive reference of each brain for the first comprehensive reference of each brain that we’ve ever had.
And so what’s the implication of that for theory and understanding, for simulation, for using that to as the basis to invent new calibration methods or structure to function translation. So I think it’s like really, really central. Yeah. And it’s something that we should want so much. But you can’t really go through a scientific career saying just I want this.
I wish I would do things differently in neuroscience if I had thousands of connectome. And we have to do the engineering. Yeah.
Juan Benet
Right here there’s like lots of great threads, people. One is great. So you have this file now. How do you simulate it in a way that is effective and like actually gives you an accurate instantiation of the brain in a computer. Another thread is great. So you have a way of producing this file. What is the entire engineering process and what is like the current cost structure of that process?
How expensive is it to get one connectome, and how do we make it cheaper to create like this genomics cost curve or the transformers cost curve to then start being able to get the dozens and hundreds and thousands of connections will allow you to then compare all between these, and then extract learnings from that.
Another third is what has been done so far. Like what connectome do we have to of what kind? What animals, what kind of terms have we gotten to date?
Adam Marblestone
So we basically have from the 1980s in a very painstaking process without any automated, you know, segmentation and proofreading, all sort of manual. We have the C elegans connectome 302 neuron worm. We have a couple of variants of the elegans now of, you know, different bases of different genders, hermaphrodite and other genders that they have in the worm.
A couple of different examples. We have the fly brain. This is kind of the big thing. Recently I think has been really incredible flourishing of research. Having just one full fly connectome, you can start to ask all sorts of questions about simulation, about other things. So we have the fly brain and this was done with electron microscopy.
It was done with millions of dollars of human proofreading. It took many, many years to do one fly brain that is sort of hundreds of thousands of neurons scale, similar in size to a in terms of neurons, roughly to a cubic millimeter of, let’s say, human or mouse cortex. And we have basically more or less, you know, cubic millimeter size, human and mouse, somewhat arbitrary chunks of cortex.
We are about to have a zebrafish connectome soon, where some of the data for it is released, but it’s not fully annotated and proofread yet. But the zebrafish is very small vertebrates, and I think that’s really remarkable. You can read this book, The Brief History of Intelligence. It talks about sort of different evolutionary stages, but a vertebrate has things like reinforcement learning that are architecturally going to be really interesting.
It has something a little bit like kind of like deep learning as sort of sensory hierarchies and multimodal sensory integration. There was just a paper about this, so zebrafish will have a lot. But again that says sort of at this very small scale, um, and I think the next, you know, major thing that has to be done is the first complete mammal brains.
I’m vastly more interested in complete brains than I am in little chunks, because I don’t think you can understand. Region of the brain without understanding what its inputs and outputs are. It’s the input and output structure to all the other parts of the brain, and also how it compares and contrast with other parts of the brain that will let us interpret, you know, is this neuron actually an attention line or a learning signal, or is this a, you know, training data, or is this something else?
Right. That is all about where it comes from and what it is meant to do inside that little local circuit. So I think that having a whole mammal brain will kind of be the next crucial milestone. Yeah.
Juan Benet
And what’s the scale to get there? Like if you wanted to get a whole mouse connection today, what is the engineering effort for that look like? What does it cost?
Adam Marblestone
It is sort of a time varying estimation of that. So a few years ago I think it was 2024. The Wellcome Trust put out a report, I think it was called Scaling Connectome. And they projected, you know, a year of like $10 billion or something like a 15 year project, basically to do a good job of the mouse brain. Most of that was the part that I’m indicating, I think can make massive progress from some combination of AI and clever labeling of the neurons.
That was presupposing essentially electron microscopy as the way of doing it, but it was like $10 billion or something like that. But most of that being computational, if we assume that we can really drop but not completely eliminate the computational part and switch to things like optical stacks. I think that now you could do in a few years with a focused effort and for sort of depending on whether you have to pay for your computer, whether someone pays for that for you, or depending on various things, sort of on the order of 100 to $200 million for the first mouse sized whole mammal connectome.
Juan Benet
Yeah, expensive, but that’s like a three order of magnitude decrease.
Adam Marblestone
It’s a big decrease. And then you can think about human. And if you were to use sort of that version, you know, human brain is on the order of about a thousand times bigger. So you can then sort of project what it is for human. So then you would have to really optimize the hardware to even get a human within the sort of billion dollar, you know, kind of a few year scale.
And then I think beyond that, that there will be ultimately further optimization of technology. Can you rely more on clever biological labeling, barcoding methods and less on just raw spatial resolution? Number of pixels? Can you be compressing data on the fly, get away with cheaper hardware, cheaper computation by encoding more information into the brain itself to read out?
I think that we can ultimately get. And in extreme case, there have been proposals even for purely barcode based connectome. So just exchanging DNA molecules between neurons and then putting it on a sequencer where I think you can ultimately in the long term, you can have, you know, mouse connectome for only thousands of dollars and human connectome, you know, only millions of dollars.
So, yeah, and.
Juan Benet
We’ve seen this play out in multiple fields, like, again, genomics, AI hardware. This goes back to just straight up Moore’s law. You get the acceleration of returns kicked off. You can write the cost curve down to the point where, like you, a lot of the applications become very possible.
Adam Marblestone
Learning curves, new methods.
Juan Benet
But those usually require some initial short to medium term type applications that prove out the utility that kick off the capital accumulation and deployment to cause the cost curve to improve. So what are the in recent AI terms, that’s like the ChatGPT moment is what triggered the most recent wave of capital deployment.
In computing terms, it’s the early transistors replacing vacuum tubes laid off like the early transistor and then later integrated circuit cost. So what does that and you know, in genomics there was a whole genome project to get one genome. But then a lot of the work and tools ended up building all these cheaper and cheaper sequencers.
So what is that short to medium term. Yeah. Application. And yeah.
Adam Marblestone
What is that. Yeah I think there I think there are a couple possibilities. My preference would be that we have to eat a certain amount of capital to get to roughly the mouse size. The mouse size is also not that different from the size of major human brain regions. Things like the hypothalamus of a brain order of magnitude similar size as an entire mouse brain is sort of a relevant human regions.
So I think if we can get to roughly the sort of cubic centimeter, multi cubic centimeter scale along the way to that, there may be applications, but my preference would be to jump to that scale through capital and engineering once we get there. And then the question is now you have to still scale massively to do lots of mouse guys, human size, have all these learning curves from that point.
So the question is then what are applications at the sort of cubic centimeter or few cubic centimeter multi cubic centimeter scale that are sort of real applications. And I think there’s a few different categories. So one would be can we actually understand something for AI. And I think even looking at the fly connectome, the fly connectome I would say it is both surprisingly interpretable and surprisingly similar.
It doesn’t mean it’s completely interpretable right now or completely simulate all right now, but it’s surprisingly interpretable in the sense that I think that you can see circuits in there for how the fly does particular kinds of reinforcement learning, how it learns differently from tastes versus smells, etc., sort of cost, function, type information.
There is information about how it does sensory processing. There’s actually pretty good simulations now, especially if you can make some assumptions that this is supposed to be a visual system and you sort of do some task optimization of a simulation, and you also look at the constrain it with the connectome, you can have a model of how a fly does something like vision to a significant degree.
Now getting closer to there with how it does motor control, there’s been really interesting interpretable elements of the fly brain from a sort of theoretical neuroscience perspective. There was this idea of a ring attractor. The heading in space is kind of this periodic thing, or you need to represent sort of a point or distribution of points along a kind of a circle to figure out what direction you’re pointing in space.
And there should be some way of building a ring attractor dynamical system out of neurons in the fly. It’s a literal ring. It’s literally a ring of neurons that also does embody this ring attractor concept. So I’m optimistic that there will be interpretable architectural features of whole mammalian scale connectome that could be useful for AI in some of these problems that we’ve been talking about like continual learning, memory consolidation, cost functions.
So there’s AI angle.
Juan Benet
And so for that one it would be pay 100 to 200 million to get a mass connect to them and then study it to potentially get billions and tens of billions in cost reduction for potentially. Yes, AI.
Adam Marblestone
Insights and directions of research and collaborations that could potentially do that.
Juan Benet
That seems like like a straightforward, you know, 10 to 100 x multiple. That seems pretty reasonable for Frontier Labs to spend.
Adam Marblestone
And it’s speculative in some ways, because we haven’t seen that much in the way of reading off of a connectome. And it’s not going to be simply reading it off as one of the things I think of the connectome as a constraint that makes all of neuroscience and all the analysis of all the other experiments in neuroscience go faster.
So you’re really talking not just about sort of here’s my connectome, here’s my algorithm for reading off the AI, you know, improvements from the connectome. It’s really here’s my connectome. Now, any neuroscience question I ever would have wanted to ask about how mammalian brain learns or does anything right now is much faster because of that, much more strongly constrained.
You can interpret and cross compare, potentially use AI agents to cross, compare and interpret. Everything else has been studied about the mouse. We now say, okay, now that we know the connectome, can we derive insights faster?
Juan Benet
So this seems like a really good Christmas gift for next Christmas.
Adam Marblestone
Yeah. So I think I think that this is one.
Juan Benet
That dropped a AlphaFold. So it would be great if another frontier lab drops the entire mass.
Adam Marblestone
We should be doing this as soon as possible. Other types of applications I think there is. When I started talking about the sort of different angles on sort of disease and health from this. So, so one is that some diseases are going to be identifiably deformations or changes of the connectome. You could say here’s a schizophrenia mouse model versus not here’s a chunk of schizophrenic human brain versus not.
Or is there differences in connectivity that tell us what connectivity differences are there, or what molecular differences are there? What genes underlie those molecular differences? These molecular energy connections, you will have some diseases that are like connected, where we can think about this as finding new kinds of drug targets.
Juan Benet
So you’re thinking there sections or whole human brain?
Adam Marblestone
I don’t think you need a whole human brain necessarily to do some of this work. I think it would be helpful. It’s in general, I think it’s helpful to have multi-regional as much of a brain as you can, but then you trade that off against things like mice or other animals are less than perfect models of human disease.
Juan Benet
And so that would be an input into drug development or other types of therapy development.
Adam Marblestone
I think there’s another category of that also, which is not so much there’s a connectopathy. What is the cause of that connectopathy. But here are a bunch of control knobs on the brain that we didn’t know we had. And here’s now an instruction manual for precisely modulating those control knobs. GLP-1, these GLP obesity drugs I mean, a lot of what they are doing, they work on multiple systems.
There’s some effect on the gut, some effect on different parts of the brain. But there’s a kind of weird property where like they’re getting across the blood brain barrier, like heavily in the human hypothalamus and their particular neuron populations in the human hypothalamus that are affecting, for example, the difference between an obesity drug that causes side effects and an obesity drug that doesn’t.
And these neurons that have these neuropeptide receptors in the hypothalamus. So if we actually knew what are all the different things the hypothalamus can control, things like sleep, things that affect mood, fear, aggression. Think of it as a more like wellness or control mapping the control circuitry of mammal brains.
Juan Benet
Yeah. So this would give you a control secretary and map to then have either chemical or drug development targets or potentially BCI targets like could.
Adam Marblestone
You potentially also neuromodulation targets.
Juan Benet
Yeah. Ultrasound to actuate some of those.
Adam Marblestone
Yes. You could potentially do it chemical you know molecular genetic or more real time methods like neuromodulation methods cell type specific or otherwise. If you can figure out which cells, which receptors, what is their placement in circuits. What are the circuits that control different things?
That gives you a set of either drug or modulation or targets.
Juan Benet
So this could be going back to kind of like the end to end. So it would be a few hundred million to a billion to get sections of a human brain generate the circuit diagram and then provide that as an input to BCI or.
Adam Marblestone
Or drug development. What is the catalog of receptors? Even what’s the, you know, GLP-1. That’s a receptor. Yeah. And what are the other receptors that are strategically positioned within different brain circuits that I might want to be targeting, for example, or what are the cell types of cells that I might want to be targeting?
Juan Benet
This seems pretty promising for both of those landscapes. Although those groups are slower than AI, companies are being able to scale up these kinds of engineers.
Adam Marblestone
So that’s that’s ultimately, yeah, sort of biopharma, you know, customer of some kind.
Juan Benet
Of a BCI that has already proved safety. So, for example, if you had like an ultrasound oriented BCI that has proved safety to target any region in the volume of the human brain, then could use these circuit maps to then open up new applications, potentially.
Adam Marblestone
Yes, if you have a platform technology that can target many things already, we have fewer of those than would be ideal, but some of them are being developed. You can even think of other things, other existing neuromodulation. And it may be that some of the mapping information is more from long range projections and sort of tracts of axons that go from one area of the brain to the other, but using similar stacks and microscopes, banks and computer, you know, similar infrastructure to look at neuromodulation targets at a one level scale, larger of sort of axons and axon tracts versus individual synapses.
There’s therapeutically relevant information in there as well for targeting of modulation platforms. You can also go into other types of applications. So you can imagine if your BCI method involves a cell integrating into the brain. Or maybe you have a way of curing stroke or something like that that involves putting in neural stem cells or new neurons.
You want to be able to assay. How are they integrating? Are they forming the correct pattern? If I want to build organoids that are mimics of human brain tissue, I want to be able to compare that organoid to the ground truth connectome that the organoid should have had from a real human specimen. If I want to make a disease model that mimics changes in connections and do that in the organoid and screen, lots of drugs there, I want to make sure that the same changes in connections or cell types or so on or in there.
So mapping has lots of applications at this kind of synaptic up to whole brain scale in this sort of biomedical model development, target development, drug development, type of space tuning. Yeah.
Juan Benet
Great. So that’s two applications so far. Others.
Adam Marblestone
Yeah I think I mean, there are sort of special cases of this where connectome would be part of a larger program. You know, maybe you’re trying to crack the energy efficiency of the brain that we talked about. You want to use Clinic Toolbox as a tool. It’s probably not the only tool you use, but, you know, having that tool available.
Juan Benet
Certainly the price tag, even now, you can imagine if it’s 100 to 200 million for a mouse connectome or 1 billion to 2 billion for a human connection. It strikes me that a third application would be, hey, as a backup option to somebody who is super wealthy but has a terminal disease or whatever, and buys into cryo or uploading is okay, great.
Like be the one of the first few patients.
Adam Marblestone
And yeah, there’s that type of application and there’s.
Juan Benet
The, the wealth that you’re not even going to use anymore because you’re absolutely.
Adam Marblestone
Right. And also just developing the whole basis of what, what uploading would be in the future of the uploading research lab as well, I think is a kind of application as well. And so whether that be for an individual person, I mean, I think people are going to start to have more options. I think real reversible cryo is going to start to be a really interesting option.
I think that might be kind of an option of choice because basically I do, you know, I sort of do reversible cryo wake up for a little bit, get the best chronic toe mix method or uploading method or disease treatment. So I think there’s going to be sort of a maybe there should be a preference for reversible cryo as kind of like the ideal thing for this at some point.
But conceivably you could have someone wanting to scan their own connectome.
Juan Benet
Yeah. Somebody who is on a chronic disease now has a very short time to live. Funds both cryo and uploading. Sure gets cryo first. Gets medical?
Adam Marblestone
Sure, yes.
Juan Benet
Yeah.
Adam Marblestone
Then it’d be.
Juan Benet
Great to upload.
Adam Marblestone
Yeah, yeah. And then? And then try to shorten the time where it makes sense to, uh, whatever therapy includes inclusive of some sort of.
Juan Benet
And this could be, like, an amazing gift to all of humanity, right? Like, imagine as, like your dying legacy. You give humanity the ability.
Adam Marblestone
It’s just an unbelievable gift to humanity in so many, so many ways, including just the basic science. Yeah. And we haven’t even gotten into the value that accelerating neuroscience has for understanding just other pivotal questions consciousness, AI, consciousness. It’s been the only intelligent machine.
It’s also the only machine that has a sort of human like consciousness that we know of. And so there’s just enormous reason to be for society to be sort of doing these hundreds of millions lo billions investment in the very near term. I think it is becoming more of an engineering problem and less of a we have to wait and see how the science shakes out before we could consider actually doing it.
At least whole mammal brain scale. Yeah.
Juan Benet
So how long do you think this projects will take? So you talked a little bit about a cost. How long does it take to get one of these connections, do you think?
Adam Marblestone
Um, these things are sort of integrated because, you know, in theory, you know, you could have one microscope and sort of freeze all the slices that weren’t being used and just very, very slowly. So often when I’m giving these types of numbers, I’m talking about how we do a year or two of to do the imaging.
That’s pretty fast. Eventually it has to be parallelized, but you can sort of say, you know, once you’ve done one, you know, then you’ve made the capital investment in that hardware. So the second one should be a bit cheaper. You’ve also been able to got training data. You can make better segmentation algorithms.
So the second one would.
Juan Benet
Be like a warehouse of, you know, some large number of microscopes.
Adam Marblestone
Scanning the warehouse has a kind of inherent amortization over connectome that has an inherent kind of learning curve within it. And then you’re going to use the information from that to design the best one. How much sensitivity do I really need from my camera? How much speed can I how fast can I really go?
You start with more ground truth methods where you get a little bit more conservative, but there’s going to be optimization factors in the sort of second gen and third gen, just at a kind of very straightforward engineering. And there’s also going to be new concepts like can I use DNA barcodes, protein barcodes.
They’re starting to be work on that. You know, even in the last year there’s a method called connectome seq, which is starting to show progress. And using DNA barcodes for actual synaptic level information, which is the first time it was proposed by 28 or in like 2012, but the first time really seeing synaptic level information being read out is just in the last year or so with barcode methods.
So can we co-op DNA sequencers in the long term. So there’s a lot of potential, but I think that it now is a very good time to do that. Got the learning curve kicked off with sort of 100,000,000 billion sighs investments.
Juan Benet
And yeah it seems to me like finally the cost is amenable to actually being done. The applications are close enough and valuable enough that this can get started. Yeah. And then once you start scaling up to tens to hundreds of connections, a lot of the other applications drop out.
Adam Marblestone
And, and we’ll start to see which applications are the ones that really bear fruit. And once you have that, you can optimize around those applications, too, in all sorts of ways.
Juan Benet
In like a fast paced approach. Suppose that tomorrow you had the 100 to $200 million scale for the mass connectome, and then that fed into 2 to 3 years later, the 1 to 2 billion for a human. Yeah. Then we have questions around like simulation and so on. But that timeline is way faster than I think everybody in the world today.
Yes. Like that would put uploading.
Adam Marblestone
Yeah. Yeah. My, my preference would be to sort of really really nail the mouse, start to get learning curves, then do more on whole human. And while you’re doing the mouse, you start to work on a bunch of other application aspects and a bunch of simulation aspects. I don’t want to overpromise on it, but I think that you can start to have like the shape of the trajectory and like what?
Where do we want to be pointing this learning curve? What do we want to be pointing these engines within a few years?
Juan Benet
To be like possibility does not mean it will happen. So one thing is to predict we will get full contact on X and uploading by date X, right? Which has a lot of like variables around what is likely to happen and who will fund it and what problems you’ll find along the way or whatever, which, you know, it’s a hard question.
Yeah, but even just the idea that it is possible to have a, you know, an Apollo-style program or, you know, OpenAI anthropic style frontier lab approach to do a connectome and uploading FRO or corporate research lab approach and get a full, you know, mouse and then human connectome by 2030 or 2032. Yeah. Seems way faster than what I think most people in tech thought about the possibility space, like most people, would put this out like 15 or 20 years out.
And the senior pilot.
Adam Marblestone
Within enough effort, I think you could do major progress in that timescale. Definitely a whole mouse, quite possibly whole human in that kind of yeah, five ish, five, six year time scale. Um, with enough effort and focus and learning and so on. And, you know, at a large scale. And I think one thing that’s a little bit tricky.
Interesting is that if you’re believing AI 2027 timelines, that’s still too long, right? That’s right. If you’re if you’re but if you’re thinking AI 2040 manifesto just came out and, you know, maybe in the AI 2040 world, actually this is one of the most useful things you could be doing. Yeah.
Juan Benet
Or if you land somewhere in between where.
Adam Marblestone
2043, it still has extremely accelerated progress. It’s just.
Juan Benet
I don’t know, like based on like the last five years, 2040 feels so far away.
Adam Marblestone
Yeah. 2040 in that scenario is like when all of society is like run by the ASI. You allow it to go off all the way, but you know, they’re 2029 to 2030. You know, you’re already you’re just not letting RSI go full on until a little later. But so even slightly more modest but still extremely high speed AI timeline, it seems like it should be one of the things you should be doing in late 2020 or early 2030s at like a massive, massive scale.
Juan Benet
Yeah, yeah, humanity would be on a dramatically safer, better path if we greatly accelerated connectomics right now. Do you think that by default, these cost curves keep coming down on their own, or does it really require the focus efforts? Right. Because in some cases, AI cost curve came down as a consequence of just the normal computing.
Adam Marblestone
And I think there are pieces of it that are happening more autonomously. We can ride the wave. So this question of whether computer use agents can just knock out proofreading, for example. Yeah. Normally when you’re doing this research on automated proofreading you’re talking about is really dependent on how large of a data set you have.
Diversity of data sets ground truth human proofreading to train on. But I think it’s possible that the computational side gets significantly boosted, kind of regardless of what we do by AI. But I think that one of the things that’s most important is to just start to know the unknown unknowns or sort of know what we don’t know.
I mean, we couldn’t have predicted that the fly brain ring attractor is an actual ring, you know, without actually looking at those neurons. And for the next thing, maybe the hippocampus has a really great way of indexing into the cortex that solves continual learning. That may be an identifiable connectome pattern.
We won’t know until we know. So I think kicking off the value creation is really important. I think kicking off. There are still lots of subtle problems. This question of sort of bridging slices, this question of what molecular annotations do you need to do different things? The question is.
Juan Benet
About the engineering.
Adam Marblestone
And how interpretable and how similar it is. You want to start getting this mammal scale data to get insight. How many new cell types do you discover in the hypothalamus the different connectivity patterns and receptors because of this. So you want to start getting the learning curve at a field level, going by doing these large data sets of entire mammal scale.
Juan Benet
And certainly all the engineering for the preparing of the brain and yeah, slicing.
Adam Marblestone
And you won’t know exactly how good a camera you can use until you try it with a slightly better camera. And then you can kind of downsample from that. You know, how much spatial resolution do you really need?
Juan Benet
But presumably the simulations will get a lot easier. And for example, today trying to do we haven’t talked about the simulation landscape yet, But like trying to build simulations of a connectome today is dramatically easier thanks to cloud code codecs and cloud code.
Adam Marblestone
It’s a great time in the cloud code to do neuroscience.
Juan Benet
It’s like trying to do this like three years ago it would have been super difficult.
Adam Marblestone
Absolutely.
Juan Benet
Even just like a year ago.
Adam Marblestone
What can cloud code do with a connectome? Plus all of the existing neuroscience literature, all the existing AI literature.
Juan Benet
What simulations have been done so far because I keep hearing very different statements and claims about them. So yeah, you know, like, what have we gotten to work with C elegans? What have we gotten to work with the fly?
Adam Marblestone
C elegans I think has been a bit of a disappointment. And granted, we don’t have a full molecular annotated connectome elegans. We don’t have, let alone a molecular annotated connection with a sort of structure, function, calibration or compiling that Konrad talked about. And we’re only starting to have a functional ground truth for C elegans, where you can really record all the neurons at once.
So depending on how you think of it, C elegans has been a little bit disappointing. I don’t think we have simulation of class, even though we have lots of attempts at simulating it in a different kinds, the sort of more academic or sort of small open source projects, but we don’t have something that’s like going to do really surprising out of distribution stuff that C elegans does, and it has all sorts of weird behaviors.
It has some amount of learning, has some amount of memory. Some of those things only turn on when it’s around a particular kind of fungus or particular locust or whatever. Right? That has all sorts of weird behavior that’s very multiplexed and crammed into this tiny system. So C elegans a bit of a disappointment.
But I also think that my particular perspective on simulation is that it’s not completely separable from reverse engineering, or it isn’t principle possible to do a completely raw. I’m going to characterize mapping between molecules and function for every kind of synapse that there is exhaustively, and then build it up.
The high dimensional description you have is every single sub compartment of cell for which you have a structure, function, translation, and how that interacts. That is a little bit what Konrad is talking about, about the bottom up neuroscience or compiler. I think that is possible, but is going to be going to be a bit non robust.
It’s going to be very tricky. You know self-driving cars, it’s the last 1% is you know every nine is the same difficulty etc.. It’s going to be just a really big slog. And so I think it will be much less of a slog if we can get some understanding algorithmically of what is the algorithm like if we want to do a simulation of cortex.
That simulation should be running a learning algorithm. That’s the function of cortex as a learning algorithm. Probably maybe some combination of memory and a learning algorithm, a few other things. But you want to have some algorithmic understanding of what you’re trying to reproduce. If you’re doing simulating the ring attractor of the fly, it should be able to navigate based on that, right?
So it’s not just pure bottom up. You want to get some understanding. And so for C elegans, I’m not too worried about the fact that we don’t have a super great simulation description of C elegans, in part because I think C elegans is this really weird creature. It’s more like an ASIC application specific integrated circuit than it is like a CPU or a GPU or an FPGA.
It’s just putting all sorts of stuff into this kind of very kluge together. It has to do all sorts of things and one very tiny number of 302 neurons. When you start to get to the fly, I think you start to see brain systems, you start to see algorithms, you start to see structure, function, correspondences that make more sense.
And that’s really exciting. You can say, what is the visual system of a fly? And then can we start simulating that? What is the motor system of like? Can we start simulating that? And there are different approaches people are now taking. One of them is to just neglect the fact we don’t have a sort of full molecular annotated connectome, etc., but just give kind of reasonably okay models of neurons take the connectome.
We have excitatory and inhibitory distinction that you have from sort of the genetics of what you know about the fly, different neuron types. And just like try to boot it up. And so this is what Philip Shiu did. Ian has a post about this issue. And colleagues have a nature paper about this. And I would say it is not an upload of the fly, but it is maybe doing more than naively.
It’s a compared to what question, because it’s doing more than one might naively have assumed. If I just took the simplest model of neurons, I have kind of not that much effort to boot it up.
Juan Benet
We don’t think of upload as a binary thing, and we think of it as a self-driving car analogy, like I think with like a bunch of different levels. It’s like the first I don’t know if it’s the first night, maybe it’s not even a nine, but it’s like one step into it.
Adam Marblestone
Yeah. It’s like, you know, we thought it might be this totally unsolvable problem to just get it to drive straight down the street and track the yellow line.
Juan Benet
Kind of like DARPA Grand.
Adam Marblestone
Challenge contract the yellow lines. It still has lots of things it can’t do. It can’t recognize people, but a relatively simple prototype, you know, not even hasn’t even been a derby. There should be a DARPA Grand Challenge for this, right? It hasn’t even been that level of effort. But with much less than that level of effort, the first person who tried something in the DARPA Grand Challenge has actually got it, you know, kind of following yellow lines or something.
Juan Benet
So what does that simulation capable of? And this is like a leaky integrated.
Adam Marblestone
Yeah. Yeah. Simple model of neurons. Yeah. So if I recall correctly, it can simulate certain neurons that say sort of you should reach for food. There’s sort of food nearby you that sort of propagates through the brain. And then it sort of preferentially chooses the sort of motor neurons that would be causing the sort of proboscis to sort of then reach toward that food.
So you have a sort of a basin of attraction of sort of relatively easily getting a not crazy sensory motor transformation. Now, there’s lots of other things in the middle of that. It’s not necessarily doing correctly. It doesn’t, I think, reproduce these ring attractor stuff. I don’t think it internal state.
How rewarding was that behavior? Learning it doesn’t recapitulate all that, but it also just doesn’t have a complete epileptic fit. It actually has some sensory neurons that activate. There’s sort of a cascade and some appropriate, relatively appropriate motor neurons seem to activate. So long story, you could spend hours talking about sort of different versions of what that could be or isn’t.
A bunch of stuff with how that then controls a simulated fly in a simulated environment that’s a little bit more wishy washy, because some of that comes from your training with reinforcement learning. The body model and the body model sort of already knows how to do a bunch of coordination of different motor neurons and spinal neurons to sort of do something.
The uploaded fly is kind of I think of it more as like they’re demonstrating a software platform in which the simulated fly kind of sits, and it can be a kind of development platform, but it’s not like an upload that does a lot of the things. But it’s still, I would.
Juan Benet
Say it’s somewhere in the continuum.
Adam Marblestone
On the spectrum of if you pulled neuroscientists in, you know, 2015 or something, and you’re like, yo, what’s the probability that we get the fly connectome that we know the excitatory inhibitory, that we boot the thing up and that we get anything that isn’t just complete nonsense? I think they would have given many neuroscience would give a low probability.
You’re crazy. This bottom up approach will never work. You don’t understand the principles. You don’t understand the biophysics.
Juan Benet
I mean, this is the same thing that happened with neural nets, right? Where people made all kinds of negative predictions of what was going to be possible and the goalposts had to keep changing.
Adam Marblestone
AlexNet was not.
Juan Benet
Still moving today. Right? Like, like there’s still like AGI deniers at scale today.
Adam Marblestone
But it was surprisingly good. AlexNet was surprisingly good for computer vision of that time, using a method that was surprising that that would be that good. And it’s yeah, it’s not AGI, but it’s compared to what came before or what you might have naively expected. It’s a promising signal that you can do something.
Now, that model is these integrated fire neurons they don’t use. There’s lots of sort of free parameters. What should they be? The actual strengths of those synapses, other properties of the neurons. So the other approach that people start to do is they sort of combine some sort of simple models, like that kind of initialized kind of with the connectome or either initialized with connect, and were sort of constrained to keep only the weights that the connectome has positive, you know, still be positive or something as you train it.
But now then you sort of do a combination of connectome constrained and then sort of task optimized. So I assume that this thing is a visual system and sort of the output. I need to start to see differentiation of different orientations of motion or stuff like that. Now I do backprop to actually adjust some of the details of the synapses, but in a connectome constrained way.
Now, can I then predict properties and so evolution optimize that and maybe in a way other than backprop. But can I start with the connectome information plus task optimization. And to do task optimization you need to know a little bit about what the task is, which means you have to do some sort of reverse engineering or functional description.
So for a human cortex, the task is like learning and memory over a lifetime for arbitrary data streams or something like that. And so it will be more you know, it will be it’s not necessarily that will just pop out, but it’s a difficult task. For some brain areas, we don’t really understand the task. If we understood the task, we would probably be able to simulate it much better.
So you can do some combination of unconstrained and task optimize. You can do other things like if you have activity mapping, the task can predict the next video frames of the activity, essentially recording from some, you know, number of seconds in the past.
Juan Benet
And how would we get activity mapping today?
Adam Marblestone
It depends on what you’re doing. So for something like a zebrafish, which is optically transparent, you can really do this much more for the fly. It’s actually been somewhat less. The fly isn’t optically transparent. And actually the activity recording methods in the fly are slightly worse compared to the connectome.
So actually, Janelia, which, you know, led a lot of the work on the fly connectome, has just pivoted to work, spend a huge amount of their effort. Howard Hughes Medical Institute in Virginia. Everybody has to work on this transparent vertebrate called Danionella now, on the basis that I think they think they’ll be able to relatively efficiently get connectome based activation happens and they can get the activation patterns very comprehensively because it’s transparent, even the adult phase of the fish.
And so that’s a vertebrate where you could potentially get full connect on full activity molecular annotations. They have to do a bunch of work to this or a bunch of Janelia is now working on this. I think their timeline is not like, you know, 2030 is AGI or something, but it’s really starting to focus on this question of how do you do structure to function, optimize everything you’re doing for a comprehensive every neuron.
Juan Benet
Yeah. And so that would be like structure. Once that connectome drops then we’ll have potentially some amount of activity mapping. And then we’ll be able to like do better.
Adam Marblestone
And they’ll be doing activity as well. Yeah.
Juan Benet
But going back to the flight, like what other things have people done with the because we have that connectome. Right.
Adam Marblestone
So yeah.
Juan Benet
We don’t have the full molecularly annotated connectome. So we don’t have a lot of the information that’s required. But we have a little bit like we have a version of it. Yeah.
Adam Marblestone
So there’s also a lot of more hypothesis driven work of sort of theorizing. So they figured out things like I was talking about the ring for head direction. Well, that ring has to be then if the animal is going to learn to navigate, maybe I want to navigate towards purple things because in my environment, purple things, you know, have come with a beautiful flower that has lots of nectar that I want to go and get.
I have to learn that the purple is not a bad thing. It’s something good. One of the things you have to be able to do. First of all, you have to be able to associate those with sort of valence or actions, but you also have to be able to figure out, sort of make a memory of an environment like the purple things are over here.
When I see these neurons in my visual system, I’m probably heading in this direction. If I’m heading in this direction, it’s possible that my visual neurons are going to be going towards purple. So relating to sort of which direction is in space with what is it heading toward and what is it going to be seeing.
Juan Benet
They have analytical exploration of the connectome to try to.
Adam Marblestone
Yeah. It’s like, what are some neurons that go between this head direction ring and the visual rings? Yeah. Okay. That’s interesting. They’re inhibitory neurons, but somehow they have to learn to do this type of association between one and the other. We don’t know any plasticity mechanism that does inhibitory plasticity between this and this to link, you know, activity here with this visual, you know, activity there.
Well, they discovered a new plasticity mechanism based on that. And how did they discover the new plasticity mechanism that we have to look at this neuron when it is in this condition of getting this and this, and they wouldn’t have had that hypothesis without the connectome to ever look at that. But it turns out this is the neuron, and this is the plasticity mechanism that lets you link, you know, vision with head direction.
And they use the connectome as a guide to say that’s the next.
Juan Benet
And that’s kind of like a good learning and result that comes from studying the connectome. But it sounds like there’s no simulation piece there.
Adam Marblestone
It’s not a direct simulation, but I think that these are all on a continuum. Because why, you know, why can’t we simulate that in the Phil Shiu? You know, leaky integrate-and-fire model. Well, I didn’t know it had this type of inhibitory plasticity I see. Okay, cool. Well, now that I know that. So I think that reverse engineering and uploading, simulating it’s an iterative process.
And you sort of say, well, what are the things I don’t either understand or I can’t simulate. Use the connectome as a guide to sort of focus your energy to find missing parameters, and you can do any experiment. It doesn’t have to be a fully automated algorithm to find those missing parameters. Think of it as a robotic lab, but right now it’s a human lab.
But imagine an automated lab that says, I can’t simulate this. There are no papers about this. I better go measure the activity of this and try to figure out what’s going on or something.
Juan Benet
Right. Do you think there’s a space of things for people to try with the existing connectome that’s there? Like that could be.
Adam Marblestone
There’s still a lot. There’s still.
Juan Benet
A cure. Software only approaches, whether it’s analysis or actual simulations could yield really strong results.
Adam Marblestone
Like what would I think? I think it would be stronger results. If you can close the loop with experimental predictions at different levels of scale, not just I ran the simulation and it did this or it did that, but I can actually go in and say, well, I predict that this neuron should do this and actually measure the neuron.
So I think with something like Danionella especially, you can imagine a sort of unified platform where they do all sorts of behaviors, all sorts of things, where they do all the activity, all the connecting.
Juan Benet
But that’s again, yeah, expensive like basically requires a lab and people and so on. And we can.
Adam Marblestone
Do.
Juan Benet
Constrained because today we have the flight connectome the software package. Yeah. And we have tens of thousands.
Adam Marblestone
Yeah. There are things you can.
Juan Benet
Try, hundreds of thousands of people who are very capable in this landscape equipped with cloud code and codex.
Adam Marblestone
I think that there is.
Juan Benet
Something on building something.
Adam Marblestone
Like concretely, I think something that you could do is another thing that’s happening with AI is sort of these AI literature analysis agents, Google co-scientist, or um, Robin from Edison Scientific. Some of those suggest hypotheses. Some of those just do they sort of orchestrate computational studies.
Right. And so given the fly connectome as a guide, I could say, well, look, I can’t do any experiments. Sorry, but I can go and look up the gene expression profile of these neurons from another study. And I can sort of have a hypothesis that I think these neurons are all doing the similar type of thing. Well, does that mean they’re going to have also similar gene expression.
So you can imagine a sort of in silico and it’s not purely in silico, um, just looking at a physical system, but it’s also looking at the neuroscience literature. And so you’re sort of doing connectome guided hypothesis search through the scientific literature, where your validation method is putting together neural data sets that have been collected about the fruit fly.
We do know a lot. Even other data sets Janelia has about fly genetics and neuron activity and what proteins they express and so on. And so can you say, okay, give me a better version of the fish simulation that takes into account all the RNA sequencing data that we have, you know, etc., and try to run that. I think we need a better benchmark suite and test bed is my simulation.
In fact, recapitulating the ring attractor is in fact recapitulating the right visual properties. What does it mean to get it right? So I think maybe it would start by building in silico benchmarks. I’m not sure exactly what is the best benchmark we could build now this purely in silico, but yeah, I think you could have cloud code do a lot.
Juan Benet
I wonder if having like a grand challenge type thing with like fly worlds where people can try lots of different simulations and submit their own thing and create like benchmark?
Adam Marblestone
Yeah, you have to create a good benchmark where you have.
Juan Benet
Like this simulation is able to you.
Adam Marblestone
Have the ground truth to correspond to that benchmark. Yeah. One thing is that that makes this tricky is that one way to hit that benchmark is just do a ton of RL. So just measure lots of fly behavior. I can give some sort of black box policy that sort of behaves like the fly. Yeah. And there’s. Yeah.
Juan Benet
So you’d have to like look into what exactly.
Adam Marblestone
Yeah. So, so it has to have sort of what’s the sort of held out set of this. And you know what. Yeah. Like fixing versus learning. So I think is a non-trivial problem. And therefore my like in some ways I prefer to like let’s go and try to understand a mammal to let’s understand the differences across species.
Let’s do this Danionella thing, simulated zebrafish, etc.. But I think that there is still certainly something that computational neuroscientists or neuroscientists will have really interesting stuff to do with the existing fly connectome for years.
Juan Benet
Yeah, yeah, yeah. Because, um, the reason I’m pushing on this is while the set of people working on the zebrafish and the mass connectome are doing that. Yeah, there’s a whole host of people who could be squeezing out returns out of the existing connectome that we have purely in software.
Adam Marblestone
Yeah. And also improving some of the methods. So, you know, I mentioned this connectome bench paper can llms proofread the connectome. They could also be experimenting on upstream image segmentation proofreading. For that, we also need better data sets and benchmarks. But they could be contributing to reducing the cost of the method as well.
And as we start to get better data sets, I think Google actually did something called ZAPBench zebrafish activity prediction Bench, where they did release activity maps of the whole zebrafish brain. Again, because it’s transparent and you can see is it better to predict based on pixels of video, or is it better to segment the cell bodies and then just treat them as nodes in a graph and predict based on that?
What if I use other parts of the cell that have little bits of calcium over there? What if I now constrain that with the connectome? What if I now make assumptions based on the connectome and molecular annotations and cell types? How good can I get at predicting the next seconds of neural activity? How good is it with more black box versus more connectome constrained?
And you want to see a crossover where the black box stops being the right way, and you’re actually measuring enough dimensions that you’re getting at the how, as opposed to just the recapitulating the particular time series that you are measuring, you know.
Juan Benet
So suppose that we do all of this and we succeed and we get connectomes. We’re able to simulate them. We’re able to get all of the short term medium term returns that we talked about in terms of better understanding the brain and being able to treat diseases much better and heal a certain set of conditions, or start improving our own control over our own bodies.
Then we open the door to uploading and kind of digital life. I think the world doesn’t quite understand the possibility space in that trajectory. When humanity can start accessing a form of life that is digital. You’ve obviously thought a lot about this. A lot of people in the field have thought about this as well.
Certainly tons of sci fi authors have discussed it, but I think it is still one of those things that is not well understood even in the tech community. And so I think it would be really valuable to explore the idea space of like what sort of becomes possible and interesting and valuable and concretely like, kind of like what are paths to get there?
Like, what does this look like mid to long term? When you think about uploading, what do you think is like the positive outlook, the positive vision?
Adam Marblestone
Yeah. I mean.
Juan Benet
What becomes possible?
Adam Marblestone
Yeah. And again, you know, sort of no promises in terms of near-term product caveats of how difficult it may be. But thinking about what this could ultimately be. I mean, I think it has to be considered as one of the kind of great civilizational, you know, cornerstones or transitions. It’s not just this kind of weird sci fi thing that, you know, life extension oriented billionaires or something are interested in.
That is often the sci fi portrayal, is often the individual person you know, for their own personal kind of life extension or something like that gets this thing. I think it has to be seen as this kind of on a par with AGI kind of level of civilizational technology, and maybe we can go into like some of what it would solve if you had this.
So one that people don’t often think about is like biosecurity, right. There’s a huge going to be, you know, trillions of dollars of investment in biosecurity and preventing infectious diseases. Right. Well, you know, with a digital system, you can have provably secure software, and you can have cryptographic keys that you can’t access in firewall.
And so, yes, there is such a thing as computer viruses, but I think uploading solves biosecurity. It solves long range space travel, it solves a certain form of aligned AGI, it solves life extension and aging and death. It potentially would allow new forms of self. So people obsess a lot about oh, is the copy really you?
Well, I don’t know. You know, you’re changing moment to moment. This sort of philosophy of personal identity is a complicated topic. But maybe you make some copy of you. And now can I actually download something that that copy learned. So now me actually we sort of end up splitting and then merging again.
Right. So it can sort of totally change personal identity. It doesn’t mean you have to be always part of a hive mind. But you know, if you want to eat at this restaurant and I want to eat at this restaurant, you know, let’s just split and then, you know, we can compare which one we liked, and then we can have both memories.
You know, we can remember both the Chinese food and the Indian food in our future self that is now come back. Yeah. What else does it solve? I mean biosecurity, AGI, transportation, longevity, extending the self in consciousness, expanding consciousness states.
Juan Benet
Yeah, I know a lot of people in.
Adam Marblestone
BCI in a certain sense. If you had uploading BCI, you know, becomes a much more trivial problem. So it’s sort of like if you actually had like this idea of substrate independent minds that is on a par with the computer or AGI or the steam engine or what, you know, all, you know, any of these types of things like by far.
So it’s kind of this like fundamental node of like the next century, like something big is going to be related to this question. I think that we actually do probably need more sci fi on this, in part because I think that many of the sci fi treatments of it I mean, it’s the classic thing about sci fi. You sort of assume one thing changes, but you don’t assume lots of other things changes.
But I think often the sci fi portrayal is actually kind of an undesirable and also sort of very, almost scientifically epistemic risky version that they’re proposing. So first of all, it sort of only works on kind of dead people, right? If that makes sense. You sort of have to slice and scan having to first die.
And then it raises this philosophical question about continuity, but it’s also just kind of unattractive as a product. It’s not necessarily really verifiable because you sort of first die, and then you have your simulation, and you don’t get to actually check if you like your simulation. Right. That’s kind of a problem people worry a lot about.
You know what? If someone can make unauthorized copies of you? What if we have this Malthusian situation? Robin Hanson sort of has this whole book about this. What if we have a malthusian situation where it’s just all copies and we have sort of subsistence labor? Are you being asked to do these random computational tasks to survive?
Juan Benet
It seems very expensive to simulate an entire person.
Adam Marblestone
First of all, it seems unnecessarily expensive when we can just run into ChatGPT, you know, optimized inference, you know? You know, I don’t I don’t.
Juan Benet
Really buy the.
Adam Marblestone
I don’t necessarily buy it. But also another reason I don’t necessarily buy it is like, you know, it would be a malthusian outcome if everyone had access to my bank account too. But I have a cryptographic key that doesn’t in a password, maybe a YubiKey, you know, that says you can’t access my bank account.
And so it’s also not just getting unauthorized copies of my bank account either. You could do things in just an ecosystem around in uploading existed. You would have some cybersecurity around uploading. People say, well, I wouldn’t be embodied. I would be in the cloud. Right? But that’s not at all obvious, because in this type of world, you could print a neuromorphic chip that is your body.
So you could actually be very much in body. You could even potentially print it out of biological neurons. Then your connectome is basically just like a backup store. But always when you want to make one, you’re always printing it rather than just running it. And so you could have take.
Juan Benet
Advantage of that cheap 30 watt.
Adam Marblestone
Compute power, because maybe you learn how to do that. So I don’t think assuming that it’s running on like some tech companies, the servers, you know, is the way to think about this. I think you’re figuring out how to make it substrate independent, which you could then print something. It could have a body, it could people worry about it not running in real time, but you could make it run in real time if you have the right neuromorphic and hardware and so on.
And so there’s sort of there’s yeah.
Juan Benet
Likely faster time, like real time just starts going.
Adam Marblestone
In and you could potentially run them faster. Although that’s also maybe concerning because now people are going in the real world, are going slower than you are, but you could sort of run it at a speed where you could interact in real time with the real world, and you could have a body. So I think that disembodied, you know, unverifiable, insecure, other people have control.
You have to die to do it. That’s kind of an unappealing product. And so no wonder it’s kind of relegated to this kind of fringe of imagination. But if you instead say, well, we have sort of substrate independent sort of Android bodies or something like that.
Juan Benet
There are some.
Adam Marblestone
More obvious you have to.
Juan Benet
Have it. But talk about like all the positive outlooks, like we are Legion, we are. Bob is a good sci fi story. We’re like a, you know, a person. Without spoiling anything, you know, a person signs up for froyo, immediately dies, and then wakes up and is put on a random probe and sent out to help humanity. And, you know, hilarity ensues.
Lots of like, amazing adventures, but it has like a good depiction of what it would be like to be in this kind of like uploaded society and landscape and whatnot. And that’s like one, one version, right? Like, I.
Adam Marblestone
Think we.
Juan Benet
Need more.
Adam Marblestone
We need more like that. It’s a very general capacity or capability that we’re talking about at being able to replicate and understand and also be understanding the brain. So, you know, in Pantheon, you know, they struggle to crack integrity, right. But I think that’s a very real possibility. We might actually be able to simulate the zebrafish for 20s, but not figure out how it learns or doesn’t remember it or so, and then lose integrity because it didn’t learn anything or something like that, or learn the wrong thing.
We have to have some understanding of neuroscience. But I also think the technology stack that leads to any kind of work on uploading is also the missing bottlenecks and technology stack. You know, if you do uploading, you sort of get a lot of neuroscience.
Juan Benet
So an atom’s request for sci fi. What are some like some of the things that you would love to see depicted or like, you know, substrate example. Did you give examples of like the negatives? What are the positives to explore?
Adam Marblestone
Yeah, I mean, what if you could split with yourself, you know, I’m stay here and my other self goes to Mars and then we can compare notes and choose to merge again, I think.
Juan Benet
And how identity and relationships and society shifts. Yeah. With that, like.
Adam Marblestone
Um, yeah. One thing that I.
Juan Benet
Think like like a vision of the future where, you know, humanity has this different modality or like each individual human has a different modality and you have the ability to like split and then merge and potentially also exchange thoughts in the same degree with other people. Like not just your copies of you, but like.
Adam Marblestone
It’s all a sort of telepathy or potentially solves a kind of downloading of thoughts, which is a little different than what people typically think of BCI in terms of real time bandwidth. Real time bandwidth matters, but it also matters kind of transmit this most important memory that I have to you, even if it takes me hours to transmit, that it’s not so much that I have to immediately, like control the keyboard or something.
It’s like, can I transmit more important information over a longer or deeper information over a longer period of time? They also need to be more BCI sci fi one. Just comment on all this. I think that the consciousness thing is a big issue and there’s disagreement about this, but personally, I don’t think it’s totally obvious that if I did run a simulation on a von Neumann Turing computer server, that that would necessarily have the same consciousness as the 3D neuromorphic real time chip.
That’s a very different physical phenomenon than the 3D neuromorphic chip. So that is not what people you know are a strict computational functionalist would say. But that could also.
Juan Benet
This question is probably an opener for like many hours of conversation that in the future we will.
Adam Marblestone
Have maybe beyond the scope. But I think one thing about it is maybe I can run a digital simulation. It actually be kind of great if I could run digital simulations that weren’t conscious right before I make my new neuromorphic chip of me, tweak my hypothalamus or something, or erase this memory, I could run in simulation and I not have to worry about that guy.
That’s just a simulation. It doesn’t have actual consciousness, but I can still use it to explore the forks of my brain or changes or so on, and then download it to the actually conscious one that actually matters. So this distinction matters a lot.
Juan Benet
It would definitely be very useful ethical point. If these are separable.
Adam Marblestone
It would be fantastic.
Juan Benet
I think way less ethically fraught to worry about simulations.
Adam Marblestone
Yeah. And this issue of making lots of copies in the cloud. Okay. But they’re not conscious. Fine. Don’t worry about that. The problem is when you print one. Yeah. Um, so there’s this whole space.
Juan Benet
Yeah. One thing to explore for sci fi is, like the different degrees of consciousness that appear with different types of hardware, or interesting questions open up when technology in the society has the ability to, like, produce non-conscious versions. And what other kind of ways in which humanity and society can get a lot better that could be explored here.
Like, I think certainly with the medical, time travel or death things, which I think features a lot in the bio longevity space, still kind of like imagines this kind of very constrained world because it’s sort of like physically bounded. Once you jump into uploading, you can have, you know, like versions of society and families that, like, have this very different landscape where you can build and maintain relationships into like millennia or like billions six years into the future.
And like that just starts looking like a very different type of society. And when people think about what are the really fulfilling things in their life, it usually comes down to their relationships. And there’s only like all the work achievements and so on. But most kind of studies around fulfillment tend to connect to the relationships humans form and the fulfillment they get from their family, their children, their parents, their friends, their significant others, and so on.
And if you have like a greatly expanded horizon of what both the types of relationships you could have or the number of them or the or you just have a lot more time with people to explore the universe together. And I think, like sci fi said, talk about that. I think could be super interesting. And one.
Adam Marblestone
Of. Yeah, no, absolutely. I like, um, The City and the Stars by Arthur C. Clarke, where people sort of get woken up and awoken at different eras of their, like, city people, the inhabitants of the city sort of go sleep mode and then come back and there’s just a lot and I think it reflects, yeah, the mind is just such a fundamental thing and you’re overpromising and, you know, it’s not realistic or you sort of have these not very appealing concepts of this, but yeah, what would be both realistic and appealing and ambitious?
Juan Benet
Are there any good explorations of what it would look like to expand, like what your mind itself can do in a functional way? This would be like growing the capability space of what your mind is capable of thinking about or doing, or being able to explore different types of mind states. Certainly a lot of people do all kinds of activities, whether it’s like physical activity or playing sports or playing games or different kinds of like recreational drugs or psychedelics to explore many different kinds of mind states.
When you’re talking about building a digital mind, you then can explore all of those potential states with very few, if none of the repercussions. Or you have the ability to like snapshot. And even beyond that, you have the ability to explore a totally untapped opportunity space that we just.
Adam Marblestone
Truly we’re truly sort of beyond the limit of. I think it would be great for there to be more sci fi. But I think at this point now, you know, Claude is a very different kind of mind than we are. I mean, the sort of mind space. And then you ask about how does sort of consciousness tap into this? What would it mean to be conscious of another sensory modality?
These are lots of questions that also come up with BCI, but it’s kind of extremely unpredictable by us biological humans right now. So. But it opens up this vast exploration space.
Juan Benet
If you zoom out from our current, you know, 2026 landscape, and you think about it in terms of the perspectives of people over hundreds or thousands of years, like this kind of quest of understanding ourselves and introspection and exploring our own minds and the minds of other people and relationships is like what a huge part of the human story is all about.
And it seems like this is like a major leap into a greater version and a greater possibility space there.
Adam Marblestone
And with all the things we’re saying, I think that you also weigh against that. That’s very similar technology stack as all the other stuff we were talking about, like, you know, peptide drugs and neuro AI. And it’s really just like we are massively under investing in the project of understanding the human brain.
Yeah, and the brains of all species.
Juan Benet
Yeah, yeah. Maybe since we’re talking about all this kind of optimistic visions for the future, separate from all of the connectomics and neuroscience and neurotech landscape, there’s a whole range of other technologies that are advancing tremendously right now. And, like, you have a fantastic view into a range of them.
What are some of the other areas that you’re really excited about that you think are like 10 to 20 year time scale can open up like our horizon? It’s like, what are some of the things that you look forward to unfolding over the next few years?
Adam Marblestone
Um, you know, virtual cells are all the rage, but I think that if we usually what that means right now is predicting the effects of perturbations to sort of which RNAs a cell expresses gene expression of a cell at the RNA level as a function of kind of genetic perturbations that you might make, and so on. That’s kind of what people mean by a virtual cell is a model that can predict lots of perturbation experiments, and that’s useful for things like what cell can I differentiate the stem cell into, or what’s wrong with this immune cell?
Could I get back into a young state or versus an old state. But we could have like really real virtual cells with much more comprehensive measurement, you know, nanoscale resolution, small molecules and also extremely powerful AI, really. How does things physically move and change within the cell, but also an AI perspective on this of sort of a universal latent variable model of cells and tissues and bodies, a little bit different than a brain emulation, but kind of a slightly different set of questions is less about what your memories are or these learning algorithms or stuff, but it’s still about state transitions.
I’m actually really excited about vitrification and cryo that we started talking about. That’s one that’s sort of was taken off the menu for a few decades, but I think is a kind of a very tractable physical problem of how do you cool something without ice crystal formation when you have various affordances, microwaves and nanoparticles and blood vessels and chemicals and so on.
To do that, I’m really excited about sort of automated reasoning, what AI will do in math and in physics. So we don’t I mean, we don’t know how good it will be as a neuroscientist or so on, but I think that it has a really strong potential to be amazing as a mathematical physicist. Provably correct math, formal math.
What breakthroughs will we see with new kinds of mathematical proof, new ways of storing knowledge? I’m really interested in what AI can let us do in sort of societal discourse new democratic mechanisms, new mechanisms of coordination, exchanging information, AI agents that act on your behalf as advocates.
So can we change the sort of social structure of society? Can we predict the economy with AI agents and have a real, real predictive version of economics? Fabrication, nanotechnology, although a lot of it will be just kind of biotech things. But I think there’s also the possibility of, of like true nanotechnology.
Atom by atom or molecule by molecule. You know.
Juan Benet
What are the bottlenecks.
Adam Marblestone
In construction?
Juan Benet
This is a field that you and I are both very interested in. Yeah. And we both think is very heavily bottlenecked. And maybe the world isn’t paying nearly as much attention to it as is.
Adam Marblestone
Yeah, I think we think in a similar way about this as sort of what happens to a field when you have the ChatGPT moment or when you have the even the AlexNet moment, when you have the Neuralink moment. And we’re talking about sort of for BCI, where you have a frontier lab equivalent gets created with all the engineering and operations and capital, what are fields that are kind of ripe for that transformation?
And it’s not every single one. I’m not sure that we have a quantum gravity frontier lab. I think that that may come from just brilliant people diversifying the set of minds, thinking different thoughts about science. Those are more like people getting individual fellowships and stuff like that to affect quantum gravity.
But I think just like these neuroscience measurement problems and modeling problems, I think that nanotech is a one in which coordinated Research, not necessarily immediately a frontier lab, but sort of DARPA style research. Bigger lists, bigger milestones could go a long way. There starting to be some evidence of this.
I mean, one sort of path goes through sort of protein engineering, molecular machines, DNA nanotechnology, sort of making a molecular 3D printer to try to get molecular machines that we can build and self-assemble, to have enough mechanical kind of precision that they can control something as precise as sort of their own assembly or covalent bonds.
Even the other thing they can sort of control covalent bonds is the scanning tunneling microscope is sort of this big machine where you’re relying on very precise mechanics and stages and feedback control and electronics and measurement to manipulate individual atoms on surfaces. You know, for a long time, this was kind of something people did.
IBM did a bunch of work on this manipulating atoms, but it kind of fizzled out, um, in the sort of early 2000. There wasn’t a coordinated research vision of how to make that, do 3D structures or make that do real machines that you can build with atoms that way, then, you know. In Canada, this company called CBN Nanotechnologies, Canadian Banknote Company.
Long story that I don’t fully understand, but Canadian Banknote Company internalized a bunch of AFM STM researchers with a much more ambitious sci fi kind of vision of Meccano synthesis, and they have not solved it. But they have made more, way more progress in the surrounding academic fields of STM and chemistry and so on.
On the whole, in terms of, you know, multi-step reactions and kind of how do you have a molecular tool that you can attach to something attached and remove different kinds of atoms, attach an atom to something where you’ve already attached an atom. So that’s showing some value of a sort of secretive Canadian company.
But in this category of coordinated, goal driven, well-capitalized research program versus in the sort of distributed academic model making more progress, I would say we haven’t yet seen that really be applied to the biomolecular protein engineering molecular machine. Make a molecular 3D printer.
So if you had DARPA program level of effort. Now a days on molecular 3D printers or if you have a DARPA program level of effort on STM Meccano synthesis. Yeah, I think you start to see we’re not yet at the AlexNet moment equivalent for either of those, but we might be in the kind of 1980s, you know, kind of Geoff Hinton showing that these nets are kind of doing some cool demos kind of phase, but we’re not at the phase of it radically outperforming other types of methods for any real world application.
For example, what.
Juan Benet
Do you think it takes 30 years, or is that timeline actually way more compressed because.
Adam Marblestone
Well, I think there’s self accelerating feedback loops. Like once you get that Alex, in that moment, you know, all bets are off.
Juan Benet
But from the 80s Hinton.
Adam Marblestone
Oh I don’t think it needs to. I don’t think it needs to take from 1980 to 2012. No, I don’t think so. I think that DARPA sometimes runs a series of programs, a series of 4 or 3, 4 or 5 year programs. You know, by the time you get to program three of those, you know, if you were doing in a really coordinated fashion, you could see really interesting demos.
You know, I think the quantum computing field is an interesting example. I think of quantum computing is a very healthy field where actually quantum computing has a lot of the things that a nanotechnology field, or for that matter, neuroscience field would need. It has national lab centers. It has big, well-capitalized companies.
It has extremely diverse, deep talent base of physicists and engineers. It has Google working on it. Google getting Nobel prizes. You know, in quantum, if we apply the sort of used quantum computing as a model, it’s also very open field. Everything is on arxiv.org. Lots of underlying both theory and experiment really progressing in lockstep.
It’s okay to be a pure theorist in quantum computing and do really productive things. It’s often been less okay to be a pure theorist in neuroscience or nanotechnology. So I think if, you know, if you had the level of affordances in many of these fields, what the bio nanotechnology, neuroscience that quantum computing has as a field, we would be making way faster progress.
And that’s a very exciting thing. Yeah. Because it’s not an impossible thing to reconfigure the capital.
Juan Benet
You know what I think is the capital scale needed? Like, if you had a nanotech frontier lab type thing to make truly meaningful progress quickly. Like, is this in the billions scale or like tens or hundreds of billions?
Adam Marblestone
I find it hard to estimate those things until I see the AlexNet moment equivalent. So I tend naturally to think in sort of what a group of 30 people or 50 people could do over five years, which again, is much, much bigger on any one project than what a distributed set of grad students, postdocs, professors, small companies are doing adjacent but not perfectly goal aligned things.
So I sort of tend to think of a single focused research organization size project. And then what are the sort of 3 or 4 FROs? What are the 3 or 4 DARPA programs when you sort of start to add those things up? The things I feel more confident in envisioning this as a chance of giving us the AlexNet style moment, what, five DARPA programs?
Okay, that’s $250 million or something, or $300 million, right? But if.
Juan Benet
It was.
Adam Marblestone
A super FRO, you know.
Juan Benet
Then that’s like 20 years.
Adam Marblestone
Though. If it has to be sequential, but some of it can be parallel. Yeah.
Juan Benet
And so part of why pushing that is like my sense is that if you unlock pieces of the nanotech promise, you can then start printing chips better in 2026. The dominant economic constraint that is driving all capital and all geopolitics is the fabric of our computing infrastructure. Yeah. So if you figure out a better and cheaper way of growing computers like.
Adam Marblestone
That’s right. Yeah. And I think that it is one of the challenges. I mean, I thought about this actually, before going way back when 2009, 1011 and before I was even working on this George Church molecular ticker tape stuff, I was thinking about, could you use DNA nanotechnology as an alternative way to fab chips?
And there’s lots of attractive reasons why you might. I mean, a DNA origami, you know, you can have sort of small single digit nanometer sized pixels, but there’s not just, you know, metal semiconductor insulator, different levels of doping. It’s like I could put whatever nanoparticle I wanted there, whatever protein I wanted.
They’re small molecules that they can convert themselves in some chemical process into crystals. So you could have a kind of bio chip that has very addressable many more ranges of materials, and then some of this sort of ramified into work in the Boyden lab and then by Dan Oran. And this company irradiate was sort of running expansion microscopy in reverse where you expand the hydrogel, you then you use photo patterning to put lots of different, diverse materials, and then you shrink it.
And so this was sort of this problem of like, how do you have a totally alternative stack for like a much more diverse material set for chips? I still think that’s pretty cool, but I think one of the things that I sort of at some extremely emotional level or something I realize is like it is really hard to compete with Intel and TSMC.
It is trillions of dollars of optimization. Yes. Not necessarily Intel. Is this raw new ideas and innovation. But trillions is always optimizing learning curves, supply chains, everything. And so and the idea that we were riding.
Juan Benet
On 50 years of that.
Adam Marblestone
Yeah. And so the idea that we were going to have a kind of oh, we’ll just like, do research and demonstrate a bio chip. And then we will, like, compete with Intel. Like that wasn’t really.
Juan Benet
If you show that it’s possible and it has the right. Think of it more like maybe like the storage industry where like, you went from old magnetic drives and things like tape, then hard drives, then SSDs. And I do see it happening.
Adam Marblestone
In quantum by the way.
Juan Benet
And so if you showed the new technology, then immediately Intel and others jump into the new technology and drive it. So you don’t have to like yes.
Adam Marblestone
Yeah. That’s right, that’s right. You don’t have to do it all yourself. Yeah. So you have to you have to have a demo, but you have to have a demo. That is where they justify that in some sense. That same question of optimization, where it’s very.
Juan Benet
Clear that like the orders of magnitude will check out better in this new approach.
Adam Marblestone
There’s some there’s a more general thing, which I think that I think general purpose fabrication technologies have a difficult thing because for any one thing I want to make, like you can talk about, nanotechnology is going to let us make better drug delivery vehicles and better quantum computers and better classical computers and you name it.
But for any one of those things, if it’s actually important to humanity, they spend a lot of time optimizing an incumbent that is made with a special purpose fabrication technology, lipid nanoparticles and all sorts of things. And billions of dollars spent on that for something that’s really important.
There will be a special purpose method of doing it. So you’re saying I’m going to have a general purpose manufacturing method, general purpose technology that is better at all these things. And I’m not going to just do a specialized, narrow application that’s going to be a real nano assembler kind of thing.
Then it’s going to start out for many years being worse at every particular application. And so it’s a little bit of a public goods kind of thing. Obviously the military, you know, Cold War sort of subsidize a lot of early work on microchips, Bell Labs, monopoly, etc.. And so I do think that it is a particularly hard and slow process to bootstrap progress in nanotechnology.
It has been disappointingly slow in a certain sense for like 30 plus years, um, 40 years. So the way I think about it is, what if we weren’t not as ambitious as to say it has to compete Intel and get self-sustaining revenue in the way that we’ve had this great success with GPT and everything. These most valuable companies now are, you know, this was just pure research.
You know, ten years ago. Well, if you don’t assume that, but you just have to have a functioning research field, the research field has to be as good a research field as roughly, quantum computing is a research field. The field has just to be as good a research field as metal organic framework chemistry, or just you name your thing as good a research field.
I think that should be an achievable problem. And that’s a few DARPA programs away.
Juan Benet
Yeah, I would imagine that as AI automated research loops keep growing, we’ll likely see a frontier emerging here where.
Adam Marblestone
I do think that so particularly good, this kind of chemistry materials, nanotechnology stuff is a particularly good case for automated labs, automated research. If I want to, I can give me a huge training data set by running quantum chemistry calculations in a computer only because we understand physics and chemistry pretty well.
So you know how great of a foundation model can I make for quantum chemistry? Okay, great. I want to run an experiment. I can do that in a self-contained lab. I don’t need to go get access to a precious human that has late stage Alzheimer’s disease. With this genetics that I can’t run as trial on and won’t sign up for my trial, I just need a lab.
And so this sort of I do think that, you know, if you were to try to run a kind of AI driven research loop. Yeah. Gosh, you know, deterministic construction kind of chemical Legos. I mean, yeah, I bet the AI could actually run that lab way better than humans at certain point. And so I think that.
Juan Benet
That’s not a far.
Adam Marblestone
Away. Right? Yeah. I think that this is a field that could have some AI acceleration in a way that isn’t totally it’s not right.
Juan Benet
But currently people are applying this type of AI search for drug development and drug discovery. It’s not that far away to then start doing it for molecular anomaly precise manufacturing type systems.
Adam Marblestone
Right. And people are doing it for materials. People are looking for superconductors, those kinds of things. But a lot of that is bottlenecked on this question of fabricating the thing. And so if you apply your AI directly to being better at fabricating the physical thing. That would be a pretty great, you know, AI automation loop and doesn’t require patient recruitment and regulatory and whatever else you might need.
Juan Benet
And this is how you get like the really fast takeoff loops though.
Adam Marblestone
I think there’s a possibility that the CBN nanotech you know is very AI driven. Could be you know faster.
Juan Benet
So yeah there is a funny observation of why in our timeline we didn’t get nanotech until like post AI because if we had gotten nanotech first, the very early pretty dumb models would have exploited it and then wiped out the species.
Adam Marblestone
And so that’s funny. Yeah, I always.
Juan Benet
Like a safer timeline because nanotech is.
Adam Marblestone
Like, yeah, maybe this is anthropic selection or something. But basically, I always assumed reading as a kid that lots of smart people are working on AI. Presumably the problem is more like we don’t have the right types of computers. Fabrication. I thought that nanotech would come first. It seemed like a very concrete thing to do, and something a little bit like that happened that we got GPUs and then we got the AI.
Progress. But I thought we were gonna have nanotech progress first. And that definitely is not what happened. I think nanotech is like way behind, but then it may be one that is more poised for AI acceleration than some. Yeah, yeah.
Juan Benet
Yeah, yeah. We might see over, you know, maybe not in the next three years, but the subsequent 5 to 7 we should see like some crazy acceleration here in construction fabs materials all drug discovery.
Adam Marblestone
I think it’s possible if people work on it. Yeah. In the right ways.
Juan Benet
Are you talking about AI automated research loops. Like how are you seeing that develop today. What are some of the most promising things. What would you point people towards doing? What are like other problems that you’re interested in people trying or, you know, some opportunity space that you’re like, oh man.
Like in the next 6 to 24 months?
Adam Marblestone
Like it depends a lot on the space, I guess. I think that there’s still this question that’s still like, not clear to me how hard it is of sort of when do we get like very automated. Let’s just talk about sort of very much preclinical kind of bio research. You’ve got ginkgo. You’ve got some amount of cloud labs.
There’s more work going on this. You’ve got people working on humanoid robots. You’ve got people working on robots with vision and cameras on the lab bench. You’ve got people working on wearing head cams, you know, to capture biological protocols. It still feels like I haven’t seen the AlexNet moment of flexible automation of bio research.
I think you have it in limited settings, pipetting robots and stuff like that. But in sort of, you know, if you’re starting an automated lab company, your biggest customer is like maybe pharma companies that have a very well defined set of workflows that don’t require as much flexibility. If you’re an AI model company, you have this model that can, in principle, kind of do anything.
But, you know, do I really need it to help me do my mini prep or do it for me or something if it’s a routine task, and then if it’s a very advanced task, maybe it’s kind of I’m the one postdoc in the lab who’s very good at that task. And so, you know, I kind of stick to that myself, or it would be a lot of upfront work to do that.
So I think when do we get to the sort of ubiquitous. All of bio is sort of truly cloud based. You sort of have cloud code for all of bio. And what are the things? I mean, there’s sort of different layers of kind of a routing layer of I want to do things X. So therefore outsource this to these ten different Cros and send it back to me.
That feels like cloud code is close to being able to do that kind of thing, you know, those types of models. But of course, there’s big biosecurity questions about this. And maybe rightfully, there’s only a limited amount of exploration of this that the model companies are really allowing. I think there’s big questions about, like whether we get these sort of automation loops to what degree and how and in what fields of sort of of bio is like real like closed loop automation, but lots of other questions of what do we get more efficiency, do get more efficient, you know, AI generated hypotheses about drug targets.
Do we get lots of sort of specific things happening? But sort of the thing we were talking about, if you sort of have a warehouse and it sort of goes and it gets better at scanning, probe atom manipulation or something like that. I think we’re maybe a little further from that, from the very open ended set of things that biologists do.
Juan Benet
We’re seeing significant acceleration in math. Yeah, already. What about like, physics? Chemistry?
Adam Marblestone
I am pretty optimistic about physics, but it’s not necessarily from like a full automation loop with math. There are certain versions where you can imagine a full automation loop, particularly with formally verified math. You can debate this, but you have in silico verification with physics. It’s like, I don’t know what the right theory of quantum gravity is just based on a pure in silico graph.
You know, I think something that may happen, though, is that we vastly expand the sort of physics workforce. So like, imagine if you’re just a really great person and you have really interesting ideas about cosmology, but you’re kind of like a philosophy major with some math background, and you’ve read some books and you know, well, now, if I can have a model and I can use formal methods and so on to verify the math in that model.
So now everybody. You’re sort of de bottleneck ING the world to contribute. And so you’re going to get a lot of slop physics, but you’re also going to get some people that have like this idea that wasn’t possible for them to pursue because they didn’t make it through the like, series of three postdocs and, you know, Stanford and the Perimeter Institute and so on, that you need to get an academic position to spend decades working on.
Juan Benet
And presumably, we should have pretty automated ways of filtering. A lot of the slop to.
Adam Marblestone
You can probably filter. There may also just be social processes, better forms of sort of peer review. There might be things that people can do in groups. But yeah, I’m wondering whether right now it’s kind of there’s some amateur physicists out there and it’s kind of a little bit of a joke. You know, you’ve got AI psychosis if you think you solve quantum gravity.
But I think I do wonder whether until.
Juan Benet
You actually solve quantum.
Adam Marblestone
Gravity, till one of those people actually kind of like does or something or like, is there is an idea that people haven’t seen before. And so physics, you know, and then what is the sort of experimental loop that you close on physics. Well, do you have to just build a much bigger particle accelerator? Well, maybe in some cases, but I actually think that astronomy is one of the also best laboratories of physics, right?
Can you predict properties of galaxy formation or stuff that you can see with space telescopes? So if you use cosmology and astronomy as your sort of experimental, vast experimental data, AI can sort of comb through huge amounts of telescope data, design, better optics, all sorts of things. And then you have, like everybody in the world that sort of can conceptually understand what they’re trying to do in physics, but might not be able to run the math calculations or get the academic position, can sort of participate in physics that you might get some acceleration through that mechanism.
That’s speculative, but I think that we may like it would be really interesting to see the first like true amateur quantum gravity person. We’re getting closer to it with math, I think we’re not quite there. I still I think the people who are making a lot of progress with AI models as math are also kind of pretty good, at least math like students, and would be on track to doing really rigorous, tasteful work in math otherwise.
But when are we going to get the first, you know? Yeah. I mean, or another way to say it would be imagine, you know, you had Max Planck as a high school student or something, right? But he has access to this. What? You know, what would happen, right? How many other Max Plancks were there that didn’t go through that path?
Right. Yeah. But had the same potential as Max Planck as a teenager with models.
Juan Benet
So you mentioned predictive economics model. That’s an area that I’ve also been super interested in. And to me, it seems like economics is unfortunately very siloed as a field in a way that, you know, when you think of physics, computer science, biology, you have this enormous explosion of possibility in great part because you can do a lot of experiments and you have like tractable questions you can ask and answer and build on.
Economics seems extremely limited by the fact that the experimental landscape of economics usually implies geopolitical disasters and the like come up with a new economic theory, and then the only way to try it is convince, like millions of people to believe it. That’s right. And then you enter into like some global war landscape.
And, you know, 20th century was an exercise in different economic theories, fighting each other through.
Adam Marblestone
Three instances of what happens when there’s a global, you know, market crash or whatever it is. You have only a few instances of that.
Juan Benet
Yeah, exactly. Yeah. But now we’re entering the landscape of being able to do large scale, computable simulations to then actually be able to test computational economics type of landscape. Yes. I’m curious if you’ve seen any serious groups or individuals.
Adam Marblestone
There are certainly people who want to do it. I mean, so Doyne Farmer was one of the original people that did sort of the Santa Fe Institute, and he’s most famous for a sort of very statistical approach where he does sort of quantitative trading kind of stuff with early machine learning stuff. Um, you know, but with what he was doing with Santa Fe, she was more about agent based models and sort of complexity.
And, you know, he has a book about this. He has research about this where he’s trying to do agent based models. And he says also part of it is collecting the data, right? Is like you have to calibrate the parameters to try to match it to the real world. You know, what is the probability that someone does participate in the Covid lockdown but doesn’t, you know, go to the grocery store and but does whatever, you know do something else leaves their job but still doesn’t go to the grocery store or whatever.
Um.
Juan Benet
How high fidelity does your simulation need to be.
Adam Marblestone
To.
Juan Benet
Calibrate correctly?
Adam Marblestone
Yeah. Or if you want to try to make a prediction about the labor market, you know, what are all the different kinds of jobs that people have right now? Do economists have to go collect data and find out about this? So there’s an empirical component to sort of matching it to the real world. But yeah, maybe I think we’re going to have just like so much better agent based models of everything that are possible with this.
And we can have a synthetic human population also for social science. So, you know, in short, I don’t know where it goes, but I think it should be tried to we should be retrying sort of agent based models and economics that is otherwise very out of fashion. I mean, Larry Summers has an endorsement of Doyne Farmer’s book, right?
But he’s only one of very few researchers working on this. Yeah.
Juan Benet
Now, kind of like zooming into, you know, we’re in the middle of 2026, massive amount of acceleration and a ton of fronts. A lot can be achieved. I don’t know, six, 12, 24 months, even seven months ago. This is like the days before olden era, you know. Before. Like the really good clot code? Yeah, kind of.
Adam Marblestone
Crazy.
Juan Benet
Like software engineering. So what are some of the things that you think we might start seeing happen, or that you would point people towards doing or trying? In the short term?
Adam Marblestone
I find it very hard to really predict this. I think that related to a bunch of things that we talked about, you know, right now in the sort of AI for science kind of space. There is these couple of different things that people do. Let’s say you want to make a simulation of a fly brain using the existing connectome literature, or you want to make a virtual cell model.
What you’re typically doing is one of two things. You’re either sort of having this kind of general language agent model where you’re like, come up with a hypothesis for me or like, tell me, what’s the interesting papers in this area? You’re sort of doing working with the natural language interface on kind of mostly over the literature, and only secondarily a little bit over some data analysis that the agents might run.
And you’re sort of asking in an English language question, that’s one kind of AI for science, which is kind of the LLM based version. And then there’s kind of a version that is the AlphaFold kind of version, where you have a bunch of machine learning researchers that spend a lot of time to crack protein folding.
Now, I think we are getting close to, like, clogged code. Write me the code that cracks protein folding. Wow. And I’m not I’m not saying we’re there yet, but I think in terms of the paradigm, instead of saying, Claude, tell me what the literature says about the fly, or I’m going to use Claude to help me as a machine learning trained researcher, you know, model the fly with a that’s my thesis project or something is to model the fly.
We could say, you know, cloud code, do the ML research within an empirical focus domain for virtual brains or virtual cells or so on. And so, you know, how much of the research that is simulating a zebrafish or something like that is going to be not only the code, but the choice of ML model and training environment and setup and testing and so on.
How much of that would then be
Adam Marblestone
coming out of the AI, reducing the scarcity of ML research essentially, in the bio world or the chemistry world, that might be a near term thing. That would be pretty cool to see. Or it’s like, hey, look, I don’t really know how to make an agent based model of the economy. I’m not an economist. I’m not an agent based modeler.
Oh,
Adam Marblestone
yeah, it’s just coding. Yeah.
Juan Benet
One of the things we haven’t talked about is just BCIs in general. There’s an enormous amount of progress that the field is seeing over the last ten years. A number of devices are reaching the end of their trials are about to go into market. We’re going to start seeing a number of conditions start being treated and potentially cured with these devices.
And then, you know, kind of consumer applications are not that far behind. Maybe it’s a 1 or 2 generations away. How do you see the BCI landscape evolving? Like you helped shape a lot of it, and you have seen it develop now and are thinking about its future. What are you most excited about in the field and whether invasive or noninvasive.
Obviously, you’re very interested in ultrasound or like you place a very significant bet on ultrasound that bore out very well.
Adam Marblestone
I think that there’s lots of potential right now. I still think it’s a really hard field to go into. There’s a lot we don’t understand about the brain diseases that we might treat with neuromodulation. And there’s always been this kind of chicken and egg problem. On the neuromodulation more neuromodulation focused end of BCI.
You sort of need a very rich data generating platform in the brain to be able to explore rapidly enough this design space of different closed loop treatments, exactly what millimeter of the brain in combination with which others, with which stimulation pattern matters for this person’s treatment resistant depression or something like that.
There’s been these amazing results, but sometimes hard to reproduce, sometimes hard to know what’s going on and how generalizable it is. Can you base a company off of that? And then you need to spend hundreds of millions of dollars to make the data generating platform. That, in principle, would do that.
I think that there are things, including the ultrasound, that make that more plausible. But you can imagine you can write at a given location in the brain, and at least with ultrasound, you can sort of see what happens everywhere. Whereas previously, if you’re writing somewhere, you can kind of see what happens at that location or a few other select locations, but that might have been the wrong locations or even stimulating there might have interfered with your being able to measure nearby.
So ultrasound lets you do that. It also lets you do the opposite. It’s like you can control writing and then maybe do a more precise electrophysiological type measurement. So there’s a bunch of things to do scientifically with ultrasound. Some of the chicken and egg problems of going beyond just kind of motor control or communication that people have the typical BCS, there’s also a ton going on with making things into sort of a semi invasive category.
And my question about Neuralink, always in the beginning was little wires. How deep can you go? What kind of coverage and with what trade off with sort of invasiveness and safety? I mean, the most recent thing that they put out was these tiny electrodes that penetrate through the dura, a sort of protective layer of the brain.
I think that’s like very much the right thing to be focusing on. It’s a very hard thing to be focusing. I think it requires sort of imaging through the dura to see that you’re not hitting a blood vessel. That’s kind of combining technologies that aren’t just kind of brute force less stick wires, and you’re combining that with imaging, semi invasive ultrasound also kind of above the dura, things like unfolding an ECoG array from a tiny hole over larger areas.
And I think you’ve covered in some other episodes, I think that these semi invasive space is starting to flourish. And that was always one of the most important things to do is figure out a kind of right trade off of sort of resolution and invasiveness. I also think there’s a ton on the horizon, and there kind of always has been.
But we need these like coordinated research programs, midsize focus projects, some public goods in the field on the noninvasive side, magnetoencephalography. Nobody’s been able to do that outside of a shielded room. But there are potential ways you can do that with different kinds of magnetic sensing, different kinds of magnetic sensors that, with some of the types of sensors use, saturate the Earth’s magnetic fields.
And if you kind of move your head in the Earth’s magnetic field, all the dynamic range is gone. So even if it was pretty sensitive. You don’t have any dynamic range. Or the ones that have lots of dynamic range are not very sensitive. So you can’t really see neural signals. So magnetic I think should still be on the horizon of like long range research programs.
These combinations of ultrasound optics, AI reconstructions, you can do a lot more even with noninvasive ultrasound with better algorithms. So I think that we are really what.
Juan Benet
Do you think of like the results on, on um, being able to do like fully noninvasive ultrasound and kind of like what, what resolution do you think?
Adam Marblestone
I think it’s coming. Uh, I think stimulation and calibrating stimulation with anatomy, you know, with the sort of essentially fully ultrasound, you know, very programable stack is like probably coming, I think, for functional ultrasound and the kind of single neuron level or those types of things you’re still probably talking about semi invasive ultrasound.
There are groups that are working on trying to push fully noninvasive ultrasound toward functional dynamic measures.
Juan Benet
Do you think that CTRL-labs results on the wristband where like as soon as they got enough training data, they were able to squeeze out a lot of signal.
Adam Marblestone
Yeah, they’re squeezing out the person to person variability. They’re squeezing out what is their exact right way to analyze it. And yeah, I mean, I think that that’s sort of a scaling law for decoding, for sort of decoding, although I think it was never a question of whether you had the underlying biological signal.
It was a question of whether there’s so much variability you couldn’t interpret it. So I guess I think that’s a scaling law for decoding rather than a scaling law for like physical sensitivity or something like that, although I haven’t looked that closely at it. But I think that scaling law may be interesting for applications of BCI.
Right. So can I use it to improve AI models? Can I use it for very complex closed loop control of my own brain or mood state or different things? Can I find the right targeting for neuromodulation? There may be scaling laws for those things that have to do with interpreting or decoding. You know, whether you got from like neural foundation models.
So I think that we’re still at the beginning of BCI. There is a lot that can be done. It still is difficult to go after it in a medical and business sense. And I really like really pushing on semi invasiveness as the sort of core principle or non invasiveness as the core principle. Um, and on the science of the brain.
So these things all reinforce these. I’m very bullish on the science and technology. I’m very uncertain on sort of business models and timeframes. And when the economy at large gets this as a big sector or something like that. But um, I think we’ve crossed this point where it was sort of very artisanal and kind of people don’t know what to do till you have like really focused efforts and they’re kind of starting to target some of the right things.
And, um, we can get more, uh, with more FRO-like projects. It’s in a much better state than it was ten years ago. It’s still not quite at the ChatGPT moment, but I think we’re getting a lot closer.
Juan Benet
Yeah, yeah. What’s kind of like your optimistic stance over the next few five, ten years coming up, like the progress we’ve made across the grand challenges of neural tech. And maybe as we unlock the recursive AI, AI for science loops, what’s kind of like the optimistic vision you’re pushing for?
Adam Marblestone
I think that, you know, as you push these things more, the bottlenecks become clearer and we get closer to having the sort of frontier lab level of effort for the really core, you know, bottlenecks in these different fields. So I’m, I’m optimistic that we start removing these kind of oh, but if we could only collect the data, it’s just a really hard thing to go collect the data, but nobody’s going to do it.
I think that we are getting much closer with Froze with AI, recognizing that missing data is a huge issue, to sort of yeah, having the right scientific data for complex systems. I mean, that’s been a dream for a long time, and I think it will still take a huge amount of work to do it. But like, I feel that the stars are starting to align to do more like the right coordinated approaches.
Still areas where there’s not even anything close to a frontier lab. Like the thing we talked about with nanotechnology. Again, there isn’t the kind of connectomics frontier lab really, as of now, even it sounds like Janelia might be becoming a Danionella with structure function, frontier lab, which is exciting.
Juan Benet
Do you think the capital outlook is much better now than like ten years ago? Like, certainly FROs exist now. I did helped shape the in my sense. You inspired the recent move in the US to like do the X Labs.
Adam Marblestone
I think it’s much better. I think it’s dramatically better, both because of ultra long range deep tech companies and FROs and a kind of increasing just knowledge flowing and people sort of thinking about what are the really important things for each institution to focus on, as well as a bunch of technological breakthroughs.
You know, I think some of the new bio institutes would have been hard to do without technologies like single cell sequencing to sort of justify a new institute or things like that. So there’s a lot happening. I hope that things like connectome and mapping will be similar in justifying a new generation of institute scale, company scale efforts.
But yeah, I think the capital landscape is much better, but still to me still feels plodding slow. You know, we spent the whole conversation talking about all these opportunities and to actually raise FRO-sized or larger than FRO size, let alone frontier lab sized efforts on these is still like a multi year long effort by a relatively restricted set of people that has the right networks focus.
Juan Benet
Access to me is totally insane, right? Like when no truly intelligent, self-respecting civilization would make it so difficult.
Adam Marblestone
These are the scientific public goods in some sense, or the scientific public goods, or adjacent to scientific public goods, because they’re particularly hard to sort of monetize. And but they’re also super fundamental. I am very excited about this NSF labs that you mentioned, and sort of governments starting to institutionalize things closer to the FRO model or similar types of models.
I’m very excited about new ARPA agencies. I’m very excited about ultra long range platform companies and this whole category of neuro tech company that I think sort of Neuralink was a prototype for. We still needed in more areas. It still needs to be easier. Yeah.
Juan Benet
My sense of capital at scale, especially in public markets, is that it is a lot more rational now than it was ten years ago. Ten years ago you couldn’t expect the public markets to actually bet well on any of the technology landscape. And it looks like right now that’s interesting. Public markets are actually moving quite quickly to readjust landscapes on at least the short to medium term outlook for some of the high tech.
Adam Marblestone
What has happened with AI is truly remarkable. I think you can imagine a world where.
Juan Benet
Or even correctly valuing space now.
Adam Marblestone
Yeah.
Juan Benet
Which, you know, for, for like ten years, it was seen as this insane thing that it was never going to work.
Adam Marblestone
You can weirdly, I still feel like you can sort of imagine a world where that didn’t happen. You know, DeepMind never got acquired. You know, OpenAI never got enough capital to do ChatGPT. Anthropic struggled to raise its first round. Yeah. You know, oh, you’re just a copycat of OpenAI or something. You know, what’s the application of this?
I think you can imagine a world where it’s way worse. And I think it might have been way worse in the early 2000.
Juan Benet
Yeah, it definitely.
Adam Marblestone
Was, you know, and I think it was and I think the institutional diversity of scientific organizations and scientific funders is growing. So I’m overall very optimistic. It still feels on a day to day basis. Like a very overly slow process to me. There’s also lots of questions about how do we prioritize steer safety.
Juan Benet
The value of death still exists.
Adam Marblestone
Yeah. So there’s also bottlenecks. And so there’s still like a lot of issues. But I do feel that, you know, 2026 is way different than 2016. And, you know, all of those are just enormously better than, you know, 2011 or when we started talking. Yeah. So yeah.
Juan Benet
I think looking back, we’re finally getting out of this super slow period. I think we were in, you know, when we started talking about accelerating science in 2012, 2013, I’ve come to kind of reflect back that that was the last legs of this super intense slowdown from 1980 through 19. Yeah.
Adam Marblestone
If I had to imagine, I think I might put some sort of like peak dysfunction or something, kind of, even though there was also lots of cool stuff going on. The first quantum computing research and, you know, Janelia and things like that, other institutes getting started and really cool things happening, obviously.
And Google, you know, all sorts of interesting things, but there’s sort of a maybe peak dysfunction sort of around that same period of maybe around 2008 or 5 or so. And then of course, there’s the 2008 crash. And so when, when we were early sort of students or entrepreneurs, it was like, oh, man. Like the capital is the problem, and it’s not going.
Juan Benet
To all of the right.
Adam Marblestone
Things. Yeah. And I still think it’s kind of if you look at this conversation, most of what we’re talking about is sort of capital or sort of capital plus coordination limited versus like totally new idea limited. But yeah, we have, you know, hundreds of millions into, um, real long range companies and some of these areas and FROs.
Juan Benet
Yeah, it’s much better and much faster now. It’s still very far away from what I would consider actually a high functioning society or civilization. Yeah. Like from my perspective, to really be unlocking the returns from science and technology development the world should be spending on, you know, double digit percent of GDP on R&D.
And it is maybe like single digit of GDP. Right. And so like ideally would be something like 40 to 60% of all of the capital expenditure would be going into unlocking the scientific and technological frontier. Because doing that is what enables the standard of living for everybody to get dramatically better.
And so we’re just so far away from that.
Adam Marblestone
We really are. And people, individual scientists, entrepreneurs still face tremendous uncertainty of if I take this path, pursue this goal. Am I going to just sort of end up not getting funded to do this? That’s still like a huge dominant factor.
Juan Benet
What I tend to think about it is like, you know, when you think about the most successful technology companies, they tend to generate a revenue and then immediately plow most of that back into R&D for the next generation? We are not doing this at a civilizational scale. Yeah, like civilization only. Imagine if you had a technology company, and a technology company was only reinvesting like 1 to 5% of its right, of its.
Adam Marblestone
Revenue scale, right? Yeah, yeah. It’s actually amazing how much, you know, research things like meta and stuff. They actually do. I mean, they do a lot of research. It’s not necessarily fundamental public goods for understanding the brain or something, but it’s like tech companies do a decent amount of research.
And society as a whole should do more.
Juan Benet
Yeah, yeah. Hey, thank you so much for talking to all of this. Thank you so much for all the work that you do accelerating all these fields. You’ve had a your hand in really improving the outcomes and odds for lots of people out there. I think in ten, 20 years, we will look back and trace all of the projects that you either help start or fund or initially think about.
So yeah, thank you for everything.
Adam Marblestone
You do very kind and you’ve been a part of it, you know, for 15 years now. It’s kind of crazy to think about. It’s great to get to kind of take a time point and reminisce on where we are now. Yeah.
Juan Benet
So thanks so much for being here.
Adam Marblestone
Thanks a lot.
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.
