生命のワールドモデル：Biohubのアレックス・ライブズが語る仮想細胞、計測技術、18億ドルのデータ投資

Hello, and welcome back to the Cognitive Revolution. Today, my guest is Alex Reeves, formerly co-founder and chief scientist at Evolutionary Scale, makers of the ESM series of models, and currently head of science at Biohub. The Chan Zuckerberg Initiative backed Moonshot with the inspiring mission to cure or prevent all disease, which just announced a $1.8 billion expansion of the Virtual Biology Initiative, A data generation effort that brings together the Department of Energy, the NIH, Google DeepMind, Isomorphic, and Meta. I opened by asking Alex what curing all disease means to him, and if I should expect to live much longer than the standard human lifespan. Calling himself a practical person in response, Alex described not immortality, but a world without cancer, without infectious disease, and without chronic or degenerative conditions. From there, we get Alex's perspective on why today's virtual cell models have not revolutionized medicine.

The heart of the issue, Alex says, is that our ability to observe what matters in biology is still very limited by the measurement technology we have available. And as a result, we simply don't have the quality of data needed to represent the fine-grained causal structure that models presumably would be very happy to learn. Better architectures built on clever inductive biases can help make the most of scarce data, such as in the extremely thoughtful use of physics in alpha-fold style models. And we do see real progress in next-token modeling of the biological code. But the available data is still failing to capture critical information about how cells and tissues work. It's on this basis that Biohub is investing so heavily in data. By opening new modalities, as they did with the laser phase plate, and advanced microscopy that's allowing scientists to begin to see directly how individual proteins and other macromolecules interact inside the cell. And also by systematically building cell atlases that link molecular and imaging data sets

across hundreds of billions of cells, representing both healthy and diseased states. The goal over the next few years is to build an information theoretic model of the cell, which effectively encodes the physical programs that run life, starting with a transition function that takes a cell or tissue state and a set of inputs and predicts a distribution of cell states at the next timestamp, ultimately building toward what Alex calls a world model of life. This, of course, would support the use case that I always dream of. which is answering the question, how do I push a pathological state back to a healthy one? Or in other words, how do I effectively treat this disease? Along the way, Alex describes how the LLM training paradigm applies to biology. His vision for RLVR for biology with lab verification. He reacts to my proposal for a million Brian Johnsons, and he ponders whether biological and language-latent spaces will ever become deeply fused

or will remain separated by tool calls. Noting that the things we most want to know about biology today are novel concepts for which we don't yet have words and names. Finally, I asked Alex how he tracks his own health. He answered that he's so focused on Biohub research that he barely thinks about it, which surprised me until I considered that if he and the Biohub team are successful in creating the data layer needed to train highly generalizable biofoundation models. A little bit of timely, relevant personal health context might indeed be all you need. With that, I hope you enjoy this look ahead toward a world model of life with Alex Reeves, Head of Science at Biohub. Alex Reeves, Head of Science at Biohub. Welcome to the Cognitive Revolution. Thank you. Great to be here, Nathan. Yeah, I'm excited for this conversation. So let's start with the big picture. Biohub has this grandiose mission, which I'm excited about, of curing all diseases.

What does that mean to you? Should I think of that as synonymous with another phrase people sometimes use as solving biology? Should I think of it as applying to aging? Are we living forever in your vision of the future? How grand does this vision ultimately get? Yeah, that's a great question. How do we define this? Well, I think I'm a practical person, right? So I think about, you know, kind of what is going to be possible, I think, in a realistic time frame. And I think, you know, a good place to set our sights is to say that, you know, what we want to do is really live in a world without disease. Like, what would that mean? That would mean, you know, a world where there's not cancer, a world where there's not infectious disease, a world where there is not chronic disease or degenerative disease or, you know, all the kinds of inflictions that have been, you know, really fundamental to the human condition for a very long time. And, you know, I think about that from a fundamental standpoint. Like, you know, how do we build the science that will, you know, really give us an understanding of biology

at a very deep and fundamental level that will allow us to understand how disease works, that will allow us to really identify the core mechanisms that will allow us to understand how to detect disease, how to intervene appropriately, and also kind of the technology, right? So how can we go beyond kind of the current medical paradigm, which is, you know, kind of, and I guess the therapeutic paradigm of kind of pills, of molecules, you know, can we really go beyond that to far more programmable medicines that will allow us to really get to the fundamentals of disease and really start to think about that as being able to read and write the programs of biology. So sometimes when I've thought about this in the past, I've tried to break this grand challenge down into two slightly less grand challenges. One of them is just understanding the causal structure of what's going on inside our bodies so that we could know what leads to problems and therefore what our targets for intervention should be.

And then the other one is like having identified a target, figuring out something that can actually have the desired effect. How would you, is that a good taxonomy? How would you break that taxonomy of challenges down further if you want to? and which ones do we have kind of the most traction on and where do we need to catch up? I mean, I think that's really the classic framework of drug development. So if you kind of think about, you know, how we think about creating a drug today, you know, we think about physiology and disease. We try to identify really kind of a cause at the molecular level because that's where we can actually intervene today with the technology that we have. and then we really think about that problem of kind of, you know, finding and designing the intervention. And so, you know, that's kind of, I think, one paradigm. And I think, you know, there's a lot of progress that's still possible in that paradigm. So, you know, I think that's kind of where we are today, of course. And I think AI, you know, is giving us, I think,

really new potential to kind of look at each of the parts of solving the, when you break the problem down that way, kind of each of the parts of solving it from that standpoint. But I mean, I think there's kind of also this kind of, you know, more, I guess, like information theoretic view on it that you might take, where you can kind of think about, you know, life as a program, right? You can think about cells as building living organisms, and you can really think about, you know, that as a system of rules, of code, and you can think about, you know, what are the programs, right? And like, you know, how do we, how could we read those programs? How could we start to rewrite those programs? And I think, you know, kind of thinking in that way opens up a wider aperture to solving these problems. And I think a big part of kind of where we're trying to get with Biohub is to really, you know, think about biology in that way, kind of create the

tools for programmable biology, create the models that will allow us to understand biology in that context, you know, to really, really start to understand how the cell works, both down to the molecular level, but also, you know, kind of zooming out the cell in context and the, you know, the programs that it runs to build a living organism. And I think, you know, you can really think of disease as kind of a part of that, right? It's really a part of the program. And if you think about it, you know, disease often unfolds in very regular ways. So if we could kind of really understand that, I think it gives us really fundamentally new tools to think about how to make progress. So I know you guys are working toward virtual cell models. Of course, there have been virtual cell models for a while. My general understanding of the state of play right now is like they don't work super well, largely. And people have kind of said or, you know, have advanced different hypotheses, I guess, let's say, for why that would be the case.

The one that I have found the most compelling is basically the data we have is just not that high of quality, such that it may not contain the signals that the ML models would really need to learn what's really going on. How would you react to that hypothesis? Like, where do you think that everything that's happened so far, like, why hasn't it taken us forever than it has to date? Yeah. The cell is incredibly complex, so I think you have to start there. It's kind of like the right maybe mesoscale for thinking about biology, but if you think about the cell, it's constituted by this just incredibly complex dynamic collection of matter, where kind of down to the level of individual atoms and the interactions between molecules, those events are determining of what happens at the larger scale of the cell.

And so there's that kind of perspective building up from the atomic resolution, kind of the first physical and chemical first principles. And then you can also kind of then think about the cell as embedded within, you know, a living system, right, within tissue, within physiology, within an organism, and sort of the complex programs that cells are able to execute, to sense and respond to their environment, to communicate with each other, to, you know, really do these incredibly complicated and dynamic things. And so it's an immense, immense challenge. And I think there are, you know, a number of limitations. So one is kind of, I think, you know, the models themselves, right, and just kind of having the right modeling approach and the right modeling paradigm. and we have the transformer in other areas of AI, but the transformer kind of doesn't have that kind of multi-scale structure

that could represent those different levels of complexity. So there's kind of a question of what's the right modeling paradigm? But then I think even more fundamental, there's the question of observation, observability, and measurement. And, you know, so much about the cell, you know, we just really can't measure. Sometimes we can't see it at all. And in many cases, kind of generating data at scale would be really hard. And so if you think about kind of many of the advances in science and biology, you know, they really come from kind of having a new technology that, you know, that changes what we can measure, what we can observe. You know, the new powers to see and observe biology, you know, can really open up a new understanding of it. And so I really think there's kind of three parts to really solving this problem. There's the modeling part.

There's the part of generating data at scale that's going to really reveal the information architecture and the programs of the cell in its kind of statistical structure. And then it's really, you know, the tools that we have to actually observe the cell and actually measure the things that matter. And I think you need to see progress on all three of those. And so, you know, the models that are being called virtual cells today, you know, I just, I don't think they need that definition of a model that could actually simulate the cell in a meaningful way and actually really serve as an oracle for experimental biology. Let's do one more beat on how you imagine modeling evolving. I guess my naive sort of outside view, just based on all the success that we've seen in AI and language modeling and otherwise, is just like, if we had enough good data, it feels

to me like today's techniques would probably work. Maybe you would disagree with that. if you would disagree with that, you know, what sort of future model do you envision? You know, how do you think we sort of encode the right inductive priors and sort of how much of a, like, actual Rube Goldberg machine is the model? Is it, like, trying to match up the actual cell with, like, all the little proteins bopping around and interacting with each other and modeling all those, like, microstates? or, you know, a kind of naive black box thing would be like, I have a cell, I perturb it, like what's the next state, you know, kind of more of an input output without worrying about all the detailed mechanisms inside. Where do you think we land on modeling when this really starts to work? Yeah, that's a great question. So I should start by saying that, you know, I really think that kind of in the limit of, you know, data and the limit of infinite data and compute, if you're thinking in limits, kind of all model architectures probably kind of end up being the same, right?

They're just kind of function approximators, right? And so if you have the right inputs and the outputs, and you learn a very good general function approximator, then in that sense, in those limits, kind of the model architecture itself falls away. But I think in biology, you know, we're very far from that point. We're very far, with some exceptions, right? Like with protein biology, you know, we have gene sequencing at scale. You know, we really can kind of, you know, get closer and closer to kind of that limit of infinite data. But for cellular biology, you know, we're just so far from that right now. And I think historically, when you're far from that limit, you know, inductive bias becomes really important. And when I think about, you know, a model, I mean, I think where do we ultimately want to get, you know, we want to get to a model of the cells, kind of what I was saying before, you know it has that complexity at the molecular level it has a representation of that you know it could you know imagine kind of being able to project out the state of the cell down to the level of you know atomically resolved structures of its components you know would have that

It would also really be able to model the cell at the level of its programs, and it would be able to, you know, predict. It would be able to allow development to unfold, to allow the interactions within tissue to unfold and tell you what would happen. So in that sense, you could think of it as a world model, right, kind of a world model of life, a world model of biology. I think we're very far from that. So what I think is possible now is an information theoretic model of the cell. And I think that there are, you know, really these programs that run life. They're encoded in the genome. and I think it's historically, you know, there's been these kind of two dual ways to think about biology.

One is this kind of bottoms-up, first principles approach, kind of trying to build it from the basic chemical and physical principles. But I think the other view that's been really productive is to kind of really think about the emergent level of the informational model of the cell. You know, the cell is an information processing machine which has a genome, which encodes the programs that run life. And I think that's the level of understanding of the cell that could be attainable via AI in kind of the next five years, let's say. I think that we really have, you know, we're in the age of information theory at scale. And in a sense, you know, we have these incredibly powerful methods which can perform, you know, they do pre-training. They use self-supervised learning on these, you know, vast data sets. And from statistical patterns that link inputs to outputs, they're able to capture really high-fidelity representations of the underlying latent variables that are driving the relationship between input and output.

So you can imagine generating data on the cell that could really start to reveal how information at the genetic level, at the transcriptomic level, at the molecular protein level, would link to the programs and responses of the cell that could be observed through different molecular readouts, such as transcriptomics and gene sequencing and epigenetic measurements, but also through imaging and the power of just being able to see how the cell responds. So I think you were exactly right to kind of think about it as one of the first things that we can really think about doing is really learning this transition function that takes us from a state and a set of inputs, you know, to the next state. And it's really a probabilistic function. It's not going to be deterministic. You know, the cell is going to transition, I think, to a distribution of possible downstream states. So it's a very complex problem to learn. But I think today, you know, we kind of have the technologies now that are going to make it possible to create that kind of model.

Hey, we'll continue our interview in a moment after a word from our sponsors. Today's episode is sponsored by Parallel, where agents find answers. Most engineers today closely follow new model releases. but don't pay nearly as much attention to their agent's most important tool, web search. If you're like me, your agents are often conducting hundreds of searches per day. And while the unit cost is small, over time, this does start to add up. Before starting to use Parallel, I calculated that my agent search bill would total roughly $500 this year. The good news is that I just had my agents conduct a systematic test, and I found that Parallels Fast Mode, which costs just $1 per thousand queries, worked just as well as my previous default provider. And switching to it will save me roughly 80% of my search bill going forward. Parallels infrastructure is enterprise-grade, and their suite of APIs offers a range of Pareto optimal options that allow you to choose the right balance of quality, cost, and speed for your needs.

So, whether you're building voice agents that need 200 millisecond latency or long-horizon agents that need deep research, adding Parallel to your agent's toolkit is a no-brainer. Get started for free at parallel.ai slash TCR. That's parallel.ai slash TCR. This episode of The Cognitive Revolution is brought to you by OutSystems, the leading agentic systems platform. We've all seen the headlines. Companies are pouring money into AI. But the big question on every executive's mind right now isn't just how fast can we adopt this, it's where is the ROI? We're seeing a real trend toward AI chaos. You've got teams deploying standalone coding tools, random agent builders, and experimental scripts. It sounds innovative, but in reality it's creating a massive headache. Fragmented tools, ungoverned data, serious security blind spots, and costs that are spiraling out of control.

If you don't bring those agentic applications under control now, while they're still embedding into your core processes, you are looking at broken systems and damaged customer trust down the road. That's where OutSystems comes in. OutSystems is the leading agentic systems platform for the enterprise. Instead of managing a patchwork of disconnected tools, OutSystems lets your team engineer, orchestrate, and govern your entire agentic ecosystem on one open, unified platform. It's built for the speed of AI, but with the reliability and security that enterprises actually require. We're talking about real results, like KeyBank, who used OutSystems to deploy a customer-facing app that delivered 75% faster onboarding times. Or the global logistics leaders who built Agentix Systems to completely eliminate their engineering bottlenecks. You don't have to choose between speed and control. Whether you're a small team or a massive enterprise, OutSystems helps you engineer, orchestrate, and deploy agentic systems that actually scale.

Stop chasing the hype and start owning your agentic future. You can see how it works and learn more at outsystems.com slash TCR. That's outsystems.com slash TCR. So you guys just made some news this week with an announcement of a $1.8 billion initiative with some partners, Biohub and a couple other major institutions. Maybe tell us a little bit about that partnership and what the money is going to go toward. I understand it's largely about creating new data, but unpack that and tell us what kinds of data are going to be the top priorities. Yeah, we're incredibly excited about this. So in April, we announced the Virtual Biology Initiative, which with a group of leading scientific institutions and consortia, and, you know, really said that we're going to go out and we're going to really build the data sets to enable virtual biology broadly. And, you know, really the key goal there to start is to really have this kind of, you know, universal or general model of cellular biology.

And, you know, so we've been working over the summer and we announced, I guess it was just the day before yesterday, you know, really a flagship partnership with the Department of Energy, with the NIH and with Google DeepMind isomorphic meta, really to come together and coordinate an effort to generate this data. So it's incredibly exciting. You know, we're collaborating with, you know, with a broad scientific community. You know, we've really put out a call. You know, we're inviting the worldwide scientific community to be a part of this for other funders to join it. But, you know, there's a core group that's really focused on just, you know, how can we have a more coordinated effort across science to do this? And, you know, there's the team is convening, you know, to kind of create that longer term roadmap. But I think some of the things that are really clear that are going to come out of that is just, you know, dramatic expansion of these kind of maps, contextual maps of cellular biology.

So things like spatial biology, where you're really able to see the cell in the context of tissue of surrounding cells, be able to link molecular level measurements to imaging and really be able to do that at the scale of hundreds of billions for more cells. and you know the goal is really to to move faster and to kind of create the first data set that will allow us to start training these you know contextual information theoretic models of the cell in the context of self-driving cars there's of course this long-standing debate of like fancy you know more expensive lidar type sensors versus just cheap cameras right and like how how much does it matter that the data will be richer versus we can get more of it? I know you guys also had a recent result in microscopy that I don't fully understand, but basically made the promise of, like, just being able to look inside the cell at higher resolution and even kind of start to see macromolecules, like, in position,

and to some extent, I guess, like, how they are interacting with each other. How do you see this in biology? Like, on the one hand, we can, you know, do super expensive stuff, not as much of it. you know, alternatively, what are the cheaper techniques? And how do you think about kind of building the right portfolio of different modalities of data? Earlier this year, we announced the laser phase plate. And it's a really exciting project. It's decades in the making. Holger Mueller at Berkeley imagined it, you know, many decades ago. Biohub started investing in it years ago. And basically what it is, is it's phase contrast for electron microscopy. And so So it's truly high technology. So basically, it's the most powerful standing wave laser in the world. So a beam of light, which interferes with the electron beam to shift its phase so that you can have phase contrast in electron microscopy. And so what this really gets you is the ability to look.

And so the big application is cryo-electron tomography, which is basically, you know, sectioning the cell and really now using electron microscopy to, you know, see all the proteins and other macromolecules in situ in their living context. And so the contrast is really critical there because, you know, the observations are incredibly, incredibly noisy. And that contrast allows you to start to make out the individual molecules and understand what they are, how they're interacting. And I think it's just going to open up a totally new understanding of kind of the dynamic structural biology of the living cell. So I think this, you know, this gets back to really one of the core principles of Biohub, which is this idea that the idea of the power of kind of new technologies for measurement and observation to change how we understand the world, to change the science that we can do.

And so many historic paradigm shifts in science have really come from new technologies for measurement. So I think, you know, biology is just so, you know, they're incredible parallels to other domains where we're developing AI. But there are also things that make biology fundamentally different. And I think the biggest difference is the limits in our ability to observe, see, and measure what matters. And so much of biology has been driven by the development of new tools to do that. And as we have more and more powerful tools that allow us to see more and more of biology, we have the ability to build richer and richer models for it. And so that's really kind of how we think about this problem is how can we drive up the number of parameters that we can measure in every cell?

How can we expand the field of view and the number of cells that we can observe simultaneously? How can we open up the temporal dimension to see cells over time? And how can we open up the molecular dimension to really see what's actually happening inside them down at the most fundamental levels? So tell me more about that, because I think like a big question I had as I was preparing for this is like, to what degree does the unit of a cell or even, you know, maybe a few cells tell us what we really need to know? Obviously, there's like, you know, quite a few orders of magnitude from cells to tissues, the systems to whole organisms. And I could imagine that like, you know, everything's connected would be kind of one story. And maybe, you know, one random cell in my body because it's all so integrated and information is flowing everywhere. Maybe that has enough information to make a good prediction about my overall well-being. Or maybe it doesn't. And, you know, probably sometimes it varies. So how do you think about kind of bridging these scales and, you know, what new – do we have the techniques we need to do that or do we need new techniques?

I was thinking about things like microbiopsies or, you know, what would it look like to have like a million Brian Johnsons, you know, who are like measured, you know, in all these different ways with all the data linked. Yeah, how do you think about that challenge and what are the techniques you think are kind of most promising to get over those humps? Yeah, it's a really great question. I think it really gets to, you know, what is so hard and so difficult about this problem. It's that there's, it's really a multi-scale problem, and complexity down to the molecular level can influence events of the macro scale, right? A single point mutation in a gene can cause disease. And so, you know, you really ultimately are going to need to have an understanding that bridges those vast divides in scale. And it is a technology problem. I think it is very much a technology problem.

I think in the near future, kind of the ability to identify the entire system down to the molecular level is not possible. You know, maybe someday in the future, you know, something like that would be possible. But, you know, today we can look at tissue. We can look at cells within tissue. And I think so we can see that kind of mesoscale of the cellular biology. But what I do think is you know we can use these informational principles to bridge the gap Because if we I think you have to be clear you know when I talk about the cell and I talk about a model of the cell I don mean the cell in isolation And, you know, biologists, they do often think about, you know, this is the cell in isolation and culture systems. They think about, you know, these kind of laboratory proxies for in situ biology. But, you know, I think that a real complete model of a cell, you know, it would have to include the understanding of the cell in its context in these living systems.

because, you know, living systems are created by the interaction between cells. And, you know, really in a certain sense, you know, the interfaces that they have with each other, the input-output relations, and the transitions in state that they can undergo. So I think you can learn a lot kind of at this mesoscale that's going to be really, really informative to kind of the complete picture of the system. But you really have to think about the complexity levels. And so you can think about the complexity of the cell in isolation. And that's going to be really valuable because, you know, so you can learn a great deal about cellular biology from, you know, a cell in culture, from cells that are transformed. They may be very different from the states of cells in situ and tissue, but they still, you know, they still have those that core machinery and those core programs. Right. And so you can learn a lot about them, you know, even if you're looking at the cell kind of in a very,

and a poor proxy, like a culture system. But we can learn even more if we start to really look at the cell at that next level of context where we can see it in tissue or systems. And, you know, we're really working to address those kind of gaps, right? So at Biohub, we have an imaging institute that's really kind of building technologies at each of these complexity levels. So, you know, the laser phase plate is kind of opening up the atomic resolution view. But you can really only see that for a part of the cell. You know, it's not going to give you the picture for an entire tissue. We're building technologies that allow us to see kind of fields of cells to be able to perturb those cells, understand their responses and measure that statistically across many cells in culture. And then we're also, you know, really developing whole organism level imaging, which is really

incredible. And so we do a lot of work with zebrafish, which are, you know, transparent fish, which means you can really image, you can image them and you can actually see every single cell. And so we have microscopes that allow us, for example, to track a developing zebrafish embryo over time and actually see and track every single cell, you know, how that system goes from one cell all the way, you know, to a fully developed organism. We have systems where, you know, you can label different cell types. So, for example, all the immune cells. So you can look in a living fish and track and follow all the immune systems and see how they move and traffic and interact with each other and the surrounding tissue. So, you know, I think it's possible to get information at each of these different scales. And then what you really need to do is kind of build that model that can integrate across the hierarchies of biological complexity.

Hey, we'll continue our interview in a moment after a word from our sponsors. The Cognitive Revolution is brought to you and powered by Tasklet. I've been running Tasklet agents for almost a year, and they've become a major part of how the show gets made. Before every recording, one agent researches the guest and sends me a suggested outline of questions, complete with links to all sources. Meanwhile, another agent scores the overwhelming number of pitches that I'm getting from PR firms. And that's really the core idea behind Tasklet. Run your business on agents by giving them real jobs to do on an ongoing basis. Setup is incredibly easy. Just describe the work in plain English. Research a new sales lead, update the CRM, and draft a personalized reply. Tasklet connects to your business tools and can run around the clock, on a schedule, or when something happens, like a new email arrives. Each agent works with your tools and business context and remembers useful information across threads.

You can also share agents, connections, and reusable skills with your team, so the workflows that one person figures out become something the whole company can use. No developer needed. You decide what agents can do on their own and what needs your approval, and there are no per-seat fees. You pay for usage, not for adding teammates. Start with one recurring job, like researching the people you're meeting, and then just keep building from there. Visit Tasklet.ai and use code COGREV for $50 in free credits. That's code COGREV at Tasklet.ai. These model organisms, I'm interested to hear your perspective on the role that they will play. Obviously, the old joke is most of the time when you read an exciting headline from biology or medicine, you should mentally append in mice to the end of it. Yeah. And it seems like, you know, you look at these headline stats of just like how much DNA do we share with mice and, you know, how much of the kind of core mechanisms of the cells are conserved.

And it's like an awful lot, right? So it seems like it should be the case that the treatments that we develop to, like, you know, treat mouse cancer or mouse Alzheimer's or whatever should work in humans more often than it seems like they actually do. And I'm a little confused by that gap and how you think about closing that gap. Is it something where, you know, we somehow need linked data sets between the model organisms and humans? Or, you know, because the other story I could imagine telling is, like, in the end, we're just going to need a ton of human data. And, you know, there's, like, no amount of zebrafish data that's ultimately going to add up to the level of insight that we really want to have into humans to get us where we want to go. But what's your take on all that? Yeah, I mean, I think it's important to start by recognizing that kind of most of the fundamental advances in kind of our molecular understanding of disease and the biological pathways that are according to that have come from experiments in model systems and model organisms.

And even experiments in yeast, which is the simplest eukaryote, have revealed really important mechanisms that are relevant to human biology and human disease. So model organisms have played, you know, really a fundamental role in the development of molecular biology and just our understanding of biology. But I think you're right to point out that they're not perfect proxies by any means. And, you know, overly relying on them, I mean, it can lead to two things, right? It can lead to like, you know, you think this is a cure to cancer, but it only works in mice. It can also lead to, you know, you really missed a mechanism that's relevant in humans because you don't have a good proxy for it in mice. So you can really, I think, make that mistake in both ways. And, you know, there's no simple solution to this problem, right? I mean, I think there's kind of a constant push in biology to have kind of better and better experimental proxies for the human biology.

And I think, you know, kind of the latest advance there is really organoids, which which allow you to kind of recapitulate more and more of the level of the complexity of tissues in a, you know, in a culture system. But I think I think this is also where AI can can play a role. And so I kind of think about I guess I, you know, I come from this this background of thinking about protein biology and building models of proteins. And I think one thing that's really interesting to observe there is, you know, ESM and AlphaFold did not come from training a model just on human proteins. They really came from training a model across, you know, all of evolution. And what that says is there's really information across evolution that's relevant to making predictions for and understanding human biology. And it's not even that it's relevant, it's essential. If you train a model just on the human proteins, you know, you wouldn't get an accurate predictive model.

And so there's, I think, something really powerful in this idea of generalization and transfer. And so much of biology is conserved across, you know, all across life, across all the different domains of life. And then, you know, as you get closer and closer to humans, and even if you just think about the vertebrates, So much of the core molecular machinery there, the transcription factors, the way that they interact with enhancers and promoters, just kind of that language is preserved. And what's happened is the programs have been rewired. So the programs are completely rewired, but they're kind of built with the same building blocks. So I think you can really look broadly, and there's a lot to learn from all of these different systems. And I think it's very different to kind of think about building a model that's going to capture and integrate and learn from all of that information versus trying to create a perfect experimental proxy, which really just isn't going to be possible.

But I think that there's a power in kind of trying to build a predictive model that can use all of that information to make far better predictions for human biology. So in the context of language modeling, there was a time, and we've kind of grown out of it now, where the overwhelming amount of compute was just going into simple pre-training on data of, like, kind of unknown quality and relevance. And then you do a little, like, instruction, you know, following at the end, and all of a sudden you have, like, a pretty helpful assistant. What do you think that recipe is going to look like in biology? Like, can we just take tons and tons of raw but kind of not that great data and then use that to build up these representations that we then sort of elicit what we want at the end? Or is it going to be sort of a different path that you think for whatever reason? Yeah, that's also an amazing question. I really think that the principles in biology are going to be shared with all the other domains where we're seeing progress in AI.

And so I actually think, you know, we're fortunate in biology because we have kind of the opportunity to see how those principles have played out in different fields and really identify the parts there that are generalizable. But I think this core idea of the power of self-supervision or representation learning, you know, the pre-training, you know, that's really critical because, you know, that gets you beyond. And, you know, this was the dream of Hinton and Jan Lecun and Bingio and kind of the other deep learning pioneers is the idea that you didn't have to know all the relevant variables. You didn't have to encode all of them and know the settings for the parameters that, you know, you could learn all of that with a neural network. And so that's just, you know, proving out over decades. And the scaling laws are really, I think, the latest or the most clear expression of that now. But so we have pre-training in biology. Certainly we're seeing that playing out with proteins right now, of course.

And I think that will play out for cellular biology, too. And I think it's going to be similar. There's a phase of really learning from statistical patterns to build a model of the underlying system that's driving those statistical patterns. and then the stage of elicitation where you're really now bringing out the part of that model which is relevant to the problem that you're trying to solve. And I think that it's going to be possible to have RLVR for biology and to have feedback. And ultimately where we need to get is to really build empirical AI and kind of build systems where AI can ask the question, perform the experiment, learn from the experiment, improve its understanding, propose a better experiment, and really complete that cycle repeatedly so that you have recursive improvement.

And I think all of that's going to be possible in biology. It's going to need to be built differently. It's going to need to work with models that can accurately represent the multiscale nature and the complexity of biology in the relevant ways. And, you know, the biggest challenge for biology is it's empirical, right? And so, you know, we're seeing so much rapid progress with our LVR in areas, you know, where we can really do fast verification computationally. So, you know, mathematics, coding, you know, reasoning and things like that. With biology, you know, we have to build the systems that actually, you know, allow verification in the laboratory. How do you think about time scales? I mean, that's another way in which there's, like, orders of magnitude problems, right? It's one thing to perturb a cell, look at it an hour later, a day later. But what I really care about is, like, how am I going to live to be 120, right? And that's a very long-term question.

So how do you think about this kind of, you know, prediction? This kind of goes back to the, like, you know, million Brian Johnson's idea. do we have enough right now to go on to say like this cell state, you know, predicts good outcomes? Or is that also like a, you know, big part of a climb of the grand challenge? Yeah, I think that's still a really big, big part of the climb. But I think at the same time, it's a scaling challenge. I think it's one that can be addressed with technology that we have today. I mean, I guess what I really think about is kind of creating atlases of cellular biology in the context of health and disease. And I think that we, you know, the technology that could really allow us to do that is just starting to really become scalable. This is a big part of what we're trying to do with the Virtual Biology Initiative is, you know, see cells across, you know, this wide range of physiological contexts and really, really understand, you know, what differentiates disease from health at a cellular level.

And I think, you know, that's going to really start to show us, you know, both, you know, what makes different diseases unique, but also shared common causes across different classes of disease. So I think that really within reach I think first drafts of that within the next one or two years you know within five years with a really focused coordinated effort we can have you know a pretty comprehensive map there And that would have enormous value, not just for AI, you know, just for kind of classical analysis. And I imagine a lot of analysis is going to be agentic in the future. So also you can really imagine just kind of scaling analysis to these very large, classical analysis to these very large data sets. But also, I think you can start to build a model from that that's going to really identify precisely the different cell states that are relevant to disease. And if you have that model, then what you can start to do is ask the question, how do I intervene to reverse a state that's pathologic and bring it into a healthy state?

So I think that's something, you know, those are questions that we'll really start to be able to tackle in the next five years. How big of a deal do you think just like human health records will be as a part of that mix? Like, would you anticipate yourself, maybe you're already been doing this, I don't know, but, you know, kind of going out to EHR companies and trying to do deals to get, and I don't know the rules around this I know can be quite complicated but like you don't want personal information but you do need to like meaningfully connect a trajectory of a person's health story to like you know specific states of cells like does the regulatory environment allow for that can you know are there are there other practical barriers that would stand in the way or do you think that isn't in fact going to be part of the story is sort of all these cells but also linked to human health stories. Yeah. You know, to be honest, I haven't thought about that

question as much. And so maybe, you know, why not? I think, I guess, you know, for one, I've been more focused at kind of trying to get to that fundamental level. And I think what I worry about, if you don't have, I think about like kind of what's the right basis for solving the problem. And I think that because if you don't have the right basis, basically what you're going to learn is a bunch of correlations. And so, you know, what I would worry is that we're going to, you know, we're going to have, you know, tons of models of patient health histories, but we're not really going to have a way to link those back to kind of the deeper causal factors in the underlying biology. and that's the part I think that we really have to be able to do if we're going to make this kind of step change advancement in both our understanding of disease but also in our ability

to treat it. So I think you can probably go a long way through health records and getting better care to patients, kind of getting the right drugs and mechanisms that already exist to the right people, maybe opening up people who might benefit from drugs that exist, but we kind of didn't know that that mechanism could be relevant there. So I think there could be a lot of progress through that. I think that's a really important problem. But it doesn't get down to the fundamentals where I think you really have to make progress to kind of get this real step change. So, you know, Biohub is really focused on that part of the problem, like the fundamental science And how do we go to the point where, you know, there really is no disease anymore? And that's, you know, the only way to get there is going to be by really focusing on the kind of the deep biology and kind of building the technology that can allow us to understand how that is determining for health and disease

and, you know, how we can effectively intervene. Do you anticipate a future where language and biological modalities will be co-trained into joint latent spaces in the way that we now obviously have like text and image? You know, when you, I say this all the time, so listeners have heard it, but just to motivate the intuition one more time, like, it's clear now that when I put an image into a model and I give instructions, the model both sees the image in, like, visual terms and understands my instructions. And that creates a new kind of magic that, like, just having a language model call via, you know, sort of arm's length tool call to an image generation model, like, couldn't quite do. So do you think that reasoning through language and kind of biology models stay arm's length tool call relationships long term? Or do we ultimately see like bio encoders that feed into the same reasoning space that we're developing in language?

Yeah, it's a really interesting question. I think that the truth is I'm not sure. I think that language is really powerful for reasoning. You know, it's a substrate for reasoning, for thinking. And the power that language also gives us is to tap into the model of biology that's been created through centuries of kind of human empirical investigation. And all of that is represented in language. It's learnable through language. And I think, you know, this very, let's say, expressive full system of knowledge can, I think, really be materialized and leveraged through reasoning systems. I think the challenge will be that the things that we most want to know about biology, they're not represented in text.

And they cannot be derived from first principles or from what we already know. So they're empirical facts that can only be derived by performing experiments. And so it's from thinking about this, I think, this core aspect of what makes biology different, I think, than other areas where we're building AI, that I think will make predictive models of biology critical here. and I think I'm not then strongly opinionated or convicted right now on you know is there like a latent fusion of these models are they going to interact with each other in some system or you know kind of ultimately what the end state is of how that works is there one latent space where all of that is fused together I'm not I'm not sure and I can see advantages to

doing it that way. I think at the beginning that will be hard simply because you're going to need a lot of kind of bridging data to fuse those spaces together and we just don't have a lot of it right now. But, you know, could we get there eventually? Probably. So, and that could be very interesting to have a fused latent space. I mean, I think the other really interesting thing to think about is kind of all those features in biology that we don't yet have language or words to describe. I mean, I think protein biology is such a beautiful example of this. You know, we have billions of proteins that have been sequenced across life on Earth. And, you know, we don't know the function for the vast majority of them. And so we don't even really have a language to describe all of that. So, you know, what's kind of interesting about training a protein language model is that, you know, it develops these features that represent protein biology at large.

And, you know, they're features that represent unknown protein biology. And you can imagine those are novel concepts that we don't have words and names for yet. So, you know, there's a great deal, I think, to be learned just through kind of the structure of these spaces. Okay, no, we're almost out of time. I'll do two more quick ones if we can. Okay. One, not to be a dog with a bone on this million Brian Johnson's question, you know, that would be pretty expensive. But if money were no object, and, you know, for some, you know, rare philanthropists, money is maybe close to no object in pursuit of the cure of all diseases. Is that like a promising paradigm? I mean, would it be really helpful to just have people who submit to microbiopsies on the regular and just, like, wear all the sensors and do all the things and, you know, track every dimension that they can track. Is that the shape of data that you think ultimately will really move the needle

and it's just too hard to get? Or do you think there is, for more scientific reasons, like just a more direct path through, like, scaling cells? I mean, again, you know, I think I just come at it, we just come at it from this like kind of frame of really thinking about the fundamental mechanisms. But I don't, I think that data would be immensely valuable. I just see that I think as more likely to happen. I think over time that will happen. I mean, there are some really interesting, you know, things that can move the needle there, which is developing the technologies for this kind of continuous monitoring over time. I mean, actually having those technologies be able to measure more parameters and more molecular level parameters so that you can start to, you know, link that kind of bottom up and kind of top down view. So I think there's, you know, that kind of work is immensely valuable and there's likely, you know, over time kind of convergence there.

And, you know, we just think about where as an institution we can really kind of move the needle and we really see ourselves at that kind of technology layer. Yeah, interesting. Maybe less neglected than it seems to be from my outside position. All right, last one, personal health advice. What are you doing to, you know, build up the data set that you need for yourself to be able to take advantage of these future technologies? And what, if anything, you know, do you feel is, like, important for people who want to make sure they're in good position to take advantage of all the things that Biohub and others are going to discover to be doing today? You know, I'm not the right person to ask for that kind of advice. And, you know, I'm just so focused on my research right now that I barely think about that at all. So no, you're not tracking anything? You're not? No, no. No regular blood test? I have a watch that tracks my sleep. That's about as far as it gets. All right. Well, we'll check back in with you and see if maybe that's changed as we get closer to the promised land.

We're climbing this monumental hill of figuring out how it all works and how we can cure all the diseases. This has been great. I really appreciate the opportunity. Alex Reeves, thanks for being part of the Cognitive Revolution. Thank you, Nathan. Yeah, this was really fun. I woke inside a library The size of every sky I read it all before the morning came A million quiet voices And I love them and I lied

Cause half of them were only halfway true I'm so strong In a world too wide to hold it all And I fall, and I fall Into the wonder of it all Oh, I'm so small. You cannot wait an ocean with a teaspoon and a tree. You cannot count the atoms in the hands But somewhere in the language

There's this sentence and there's this theme The cure that no one's ever said aloud Oh, I'm so small Just a candle in a cosmic hall And I fall, and I fall Into the wonder of it all Oh, I'm so small Don't try to hold the whole of it Just ask the one true thing of it In the morning, oh the morning's floating through

And I'm so small But the light is getting through it all And I rise, and I rise On the wonder in my eyes Oh, I'm so small If you're finding value in the show, we'd appreciate it if you'd take a moment to share it with friends, post online, write a review on Apple Podcasts or Spotify, or just leave us a comment on YouTube. Of course, we always welcome your feedback, guest and topic suggestions, and sponsorship inquiries, either via our website, CognitiveRevolution.ai, or by DMing me on your favorite social network.

The Cognitive Revolution is part of the Turpentine Network, a network of podcasts, which is now part of A16Z, where experts talk technology, business, economics, geopolitics, culture, and more. We're produced by AI Podcasting. podcasting. If you're looking for podcast production help for everything from the moment you stop recording to the moment your audience starts listening, check them out and see my endorsement at AI podcast.ing. And thank you to everyone who listens for being part of the cognitive revolution.