生命的世界模型：Biohub 的 Alex Rives 谈虚拟细胞、测量与 18 亿美元数据投入

Today, my guest is Alex Rives, formerly co-founder and chief scientist at EvolutionaryScale, 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. Biohub 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 whether I should expect to live much longer than the standard human lifespan. Calling himself a practical person, Alex described not immortality, but a world without cancer, infectious disease, chronic disease, 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 available to us. As a result, we simply don’t have the quality of data needed to represent the fine-grained causal structure that models would presumably be very happy to learn.

Better architectures, built on clever inductive biases, can help make the most of scarce data, as in the thoughtful use of physics in AlphaFold-style models. We do see real progress in next-token modeling of the biological code, but the available data still fails 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 up new modalities, as they did with the laser phase plate, an advance in microscopy that allows scientists to begin seeing directly how individual proteins and other macromolecules interact inside the cell; and by systematically building cell atlases that link molecular and imaging datasets 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 that effectively encodes the physical programs that run life. It starts with a transition function that takes a cell or tissue state and a set of inputs, then predicts a distribution of cell states at the next time step, ultimately building toward what Alex calls a world model of life. This would support the use case I always dream of: 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, and his reaction to my proposal for a million Bryan Johnsons. He also 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 or 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. That surprised me, until I considered that if he and the Biohub team are successful in creating the data layer needed to train highly generalizable bio foundation models, a little bit of timely, relevant personal health context might indeed be all you need.

Alex Rives, head of science at Biohub, welcome to The Cognitive Revolution.

Thank you. Great to be here, Nathan.

I’m excited for this conversation. 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, like 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?

1. Curing All Disease

That’s a great question. How do we define this? I think I’m a practical person, so I think about what’s going to be possible in a realistic timeframe. A good place to set our sights is to say that what we want to do is really live in a world without disease. What would that mean? That would mean a world where there’s no cancer, no infectious disease, and no chronic or degenerative disease—all the kinds of afflictions that have been fundamental to the human condition for a very long time.

I think about that from a fundamental standpoint. How do we build the science that will give us an understanding of biology at a very deep and fundamental level? That will allow us to understand how disease works, identify the core mechanisms, detect disease, and intervene appropriately. There’s also the technology: how can we go beyond the current medical paradigm, the therapeutic paradigm of pills and molecules? Can we go beyond that to far more programmable medicines that will allow us to get to the fundamentals of disease and start to think about reading and writing the programs of biology?

Sometimes when I’ve thought about this in the past, I’ve tried to break this grand challenge down into 2 slightly less grand challenges. One is understanding the causal structure of what’s going on inside our bodies, so we can know what leads to problems and, therefore, what our targets for intervention should be. The other is, having identified a target, figuring out something that can actually have the desired effect. Is that a good taxonomy? How would you break it down further, if you want to, and which parts have the most traction? Where do we need to catch up?

I think that’s really the classic framework of drug development. If you think about how we create a drug today, we think about physiology and disease. We try to identify a cause at the molecular level, because that’s where we can actually intervene with the technology we have. Then we think about finding and designing the intervention. That’s one paradigm, and I think there’s still a lot of progress possible within it. That’s where we are today, of course, and I think AI is giving us new potential to address each part of solving the problem in that way.

But there’s also a more information-theoretic view you might take, where you can think about life as a program. You can think about cells as building living organisms, and about that as a system of rules, of code. What are the programs? How could we read those programs? How could we start to rewrite them? I think thinking in that way opens up a wider aperture for solving these problems.

A big part of where we’re trying to get with Biohub is to think about biology in that way: to create the tools for programmable biology and the models that will allow us to understand biology in that context. We want to understand how the cell works, both down to the molecular level and by zooming out to see the cell in context, and the programs it runs to build a living organism. You can think of disease as part of that program. Disease often unfolds in very regular ways, and if we could really understand that, it would give us fundamentally new tools to think about how to make progress.

I know you guys are working toward virtual cell models. There have been virtual cell models for a while. My general understanding of the state of play is that they don’t work super well. People have advanced different hypotheses for why that might be. The one I’ve found most compelling is that the data we have just isn’t high enough quality, so it may not contain the signals that ML models would really need to learn what’s going on. How would you react to that hypothesis? Why hasn’t everything that’s happened so far taken us further than it has today?

Yeah.

2. Why Virtual Cells Fall Short

The cell is incredibly complex, so I think you have to start there. It’s maybe the right mesoscale for thinking about biology. The cell is constituted by an incredibly complex, dynamic collection of matter, where events down to the level of individual atoms and interactions between molecules determine what happens at the larger scale of the cell.

You can think about it from the bottom up, starting at atomic resolution and the basic physical and chemical principles. You can also think about the cell as embedded within a living system: within tissue, within physiology, within an organism. Cells execute complex programs to sense and respond to their environment, communicate with each other, and do incredibly complicated, dynamic things.

It’s an immense challenge, and I think there are a number of limitations. One is the models themselves: having the right modeling approach and paradigm. We have the transformer in other areas of AI, but it doesn’t have the kind of multiscale structure that could represent those different levels of complexity. So there’s a question of what the right modeling paradigm is.

But even more fundamentally, there’s the question of observation, observability, and measurement. There’s so much about the cell that we just can’t measure. Sometimes we can’t see it at all, and in many cases, generating data at scale would be really hard. Many advances in science and biology come from having a new technology that changes what we can measure and observe. New powers to see and observe biology can really open up a new understanding of it.

I think there are 3 parts to solving this problem: the modeling part; generating data at scale that reveals the information architecture and programs of the cell in their statistical structure; and the tools we have to observe the cell and measure the things that matter. I think you need to see progress on all 3. The models being called virtual cells today, I just don’t think meet the definition of a model that could actually simulate the cell in a meaningful way and serve as an oracle for experimental biology.

Let’s do one more beat on how you imagine modeling evolving. My naive outside view, based on all the success we’ve seen in AI and language modeling, is that if we had enough good data, today’s techniques would probably work. Maybe you’d disagree with that. If so, what sort of future model do you envision? How do you think we encode the right inductive priors? How much of an actual Rube Goldberg machine is the model? Is it trying to match up the actual cell with all the little proteins bopping around and interacting with each other, and modeling all those microstates? Or would it be a naive black box that says, “I have a cell, I perturb it—what’s the next state?” More of an input-output model, without worrying about all the detailed mechanisms inside. Where do you think we land on modeling when this really starts to work?

3. A World Model of Life

That’s a great question. I should start by saying that, in the limit of infinite data and compute, if you’re thinking in limits, all model architectures probably end up being the same. They’re just function approximators. If you have the right inputs and outputs, and you learn a very good general function approximator, then, in those limits, the model architecture itself falls away.

But in biology, we’re very far from that point, with some exceptions. With protein biology, for example, we have gene sequencing at scale, so we can get closer and closer to that limit of infinite data. But for cellular biology, we’re just so far from that right now. Historically, when you’re far from that limit, inductive bias becomes really important.

When I think about where we ultimately want to get, we want a model of the cell. As I was saying before, it has complexity at the molecular level, and the model would have a representation of that. Imagine being able to project out the state of the cell down to atomically resolved structures of its components. It would also be able to model the cell at the level of its programs, allowing development and interactions within tissue to unfold, and telling you what would happen. In that sense, you could think of it as a world model—a world model of life, a world model of biology.

I think we’re very far from that. What I think is possible now is an information-theoretic model of the cell. There are programs that run life, and they’re encoded in the genome. Historically, there have been 2 dual ways to think about biology. One is a bottom-up, first-principles approach, trying to build it from basic chemical and physical principles. But the other view that’s been really productive is to think about the emergent, informational model of the cell: the cell as an information-processing machine with a genome that encodes the programs that run life.

I think that’s the level of understanding of the cell that could be attainable via AI in the next 5 years. We’re in the age of information theory at scale. We have incredibly powerful methods that perform pretraining and use self-supervised learning on vast datasets. From statistical patterns that link inputs to outputs, they’re able to capture high-fidelity representations of the underlying latent variables driving the relationship between input and output.

You can imagine generating data on the cell that could reveal how information at the genetic, transcriptomic, and molecular protein levels links to the programs and responses of the cell. Those could be observed through different molecular readouts, such as transcriptomics, gene sequencing, and epigenetic measurements, but also through imaging and simply being able to see how the cell responds.

I think you’re exactly right to think of one of the first things we can do as learning this transition function that takes us from a state and a set of inputs to the next state. It’s a probabilistic function, not deterministic. I think the cell will transition to a distribution of possible downstream states. It’s a very complex problem to learn, but I think we now have the technologies that will make it possible to create that kind of model.

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 of 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.

4. Building the Biology Data Layer

Yeah, we’re incredibly excited about this. In April, we announced the Virtual Biology Initiative with a group of leading scientific institutions and consortia, and really said that we’re going to go out and build the datasets to enable virtual biology broadly. The key goal to start is to have this kind of universal or general model of cellular biology. We’ve been working over the summer, and we announced, I guess it was just the day before yesterday, a flagship partnership with the Department of Energy, the NIH, Google DeepMind, Isomorphic Labs, and Meta to come together and coordinate an effort to generate this data.

We’re collaborating with a broad scientific community. We’ve put out a call, inviting the worldwide scientific community to be a part of this and for other funders to join it. There’s a core group focused on how we can have a more coordinated effort across science to do this. The team is convening to create that longer-term roadmap.

Some things that are really clear are going to come out of this: a dramatic expansion of these contextual maps of cellular biology. That includes things like spatial biology, where you’re able to see the cell in the context of tissue and surrounding cells, link molecular-level measurements to imaging, and do that at the scale of hundreds of billions or more cells. The goal is to move faster and create the first datasets that will allow us to start training these contextual information-theoretic models of the cell.

In the context of self-driving cars, there’s this longstanding debate about fancy, more expensive LiDAR-type sensors versus just cheap cameras, and how much it matters that the data will be richer versus being able to 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 being able to look inside the cell at higher resolution and even start to see macromolecules in position, and to some extent how they’re interacting with each other. How do you see this in biology? On the one hand, we can do super-expensive stuff, but not as much of it. Alternatively, what are the cheaper techniques, and how do you think about 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 Müller at Berkeley imagined it many decades ago, and Biohub started investing in it years ago. Basically, what it is is phase contrast for electron microscopy, and it’s truly high technology. It’s the most powerful standing-wave laser in the world: a beam of light that interferes with the electron beam to shift its phase so that you can have phase contrast in electron microscopy.

The big application is cryo-electron tomography, which is basically sectioning the cell and using electron microscopy to see all the proteins and other macromolecules inside the cell in their living context. The contrast is really critical there because the observations are incredibly noisy, and that contrast allows you to start to make out the individual molecules and understand what they are and how they’re interacting. I think it’s just going to open up a totally new understanding of the dynamic structural biology of the living cell.

This gets back to one of the core principles of Biohub: the power of new technologies for measurement and observation to change how we understand the world and the science we can do. So many historic paradigm shifts in science have really come from new technologies for measurement. Biology has incredible parallels to other domains where we’re developing AI, but there are also things that make biology fundamentally different.

I think the biggest difference is the limits in our ability to observe, see, and measure what matters. So much of biology has been driven by the development of new tools to do that. 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. That’s really how we think about this problem: 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 happening inside them, down at the most fundamental levels?

5. Connecting Biological Scales

Tell me more about that, because I think a big question I had as I was preparing for this is: To what degree does the unit of a cell, or even maybe a few cells, tell us what we really need to know? Obviously, there are quite a few orders of magnitude from cells to tissues to systems to whole organisms. I could imagine that “everything’s connected” would be one story, and maybe one random cell in my body, because it’s all so integrated and information is flowing everywhere, has enough information to make a good prediction about my overall well-being. Or maybe it doesn’t, and probably sometimes it varies.

How do you think about bridging these scales, and what new techniques do we need? Do we have the techniques we need to do that, or do we need new ones? I was thinking about things like micro-biopsies, or what it would look like to have a million Bryan Johnsons measured in all these different ways, with all the data linked. How do you think about that challenge, and what techniques do you think are most promising to get over those humps?

Yeah, it’s a really great question. I think it really gets to what is so hard and difficult about this problem. It’s really a multiscale problem, and complexity down to the molecular level can influence events at the macroscale, right? A single point mutation in a gene can cause disease. Ultimately, you’re 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. In the near future, identifying the entire system down to the molecular level is not possible. Maybe someday something like that will be possible. But today, we can look at tissue and at cells within tissue, so we can see that mesoscale of cellular biology.

What I do think is that we can use these informational principles to bridge the gap. I think you have to be clear: when I talk about the cell and a model of the cell, I don’t mean the cell in isolation. Biologists often think about the cell in isolation and in culture systems. They think about these kinds of laboratory proxies for in situ biology. But a real, complete model of the cell would have to include the—

Right.

—the understanding of the cell in its context in these living systems, because living systems are created by interactions between cells. In a certain sense, that means the interfaces they have with each other, the input-output relations, and the transitions in state that they can undergo. I think you can learn a lot at this mesoscale that will be really informative to the complete picture of the system.

But you have to think about the different levels of complexity. You can think about the complexity of the cell in isolation, and that’s going to be really valuable. You can learn a great deal about cellular biology from a cell in culture and from transformed cells. They may be very different from the states of cells in situ in tissue, but they still have that core machinery and those core programs, right? You can learn a lot about them, even if you’re looking at the cell through a poor proxy like a culture system.

But we can learn even more if we start to look at the cell at that next level of context, where we can see it in tissue or in systems. We’re really working to address those gaps. At Biohub, we have an imaging institute that’s building technologies at each of these levels of complexity. The laser phase plate is opening up an atomic-resolution view, but you can only see part of the cell that way. It’s not going to give you a picture of an entire tissue.

We’re building technologies that allow us to see fields of cells, perturb those cells, understand their responses, and measure those responses statistically across many cells in culture. We’re also developing whole-organism-level imaging, which is really incredible. We do a lot of work with zebrafish, which are transparent, so you can image them and actually see every single cell. We have microscopes that allow us to track a developing zebrafish embryo over time and see and track every single cell as the system goes from one cell all the way to a fully developed organism.

We have systems where you can label different cell types—for example, all the immune cells. You can look inside a living fish and track all the immune cells, seeing how they move, traffic, and interact with each other and the surrounding tissue. I think it’s possible to get information at each of these different scales. Then what you really need to do is build a model that can integrate across the hierarchies of biological complexity.

Do these model organisms—I'm interested to hear your perspective on the role they will play. Obviously, the old joke is that 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.

You look at these headline stats about how much DNA we share with mice and how much of the core mechanisms of cells are conserved, and it’s an awful lot, right? So it seems like treatments we develop to treat mouse cancer or mouse Alzheimer’s, or whatever, should work in humans more often than they seem to. I’m a little confused by that gap and how you think about closing it.

Is it that we somehow need linked datasets between model organisms and humans? Because the other story I could imagine telling is that, in the end, we’re just going to need a ton of human data, and there’s no amount of zebrafish data that’s ultimately going to add up to the level of insight we really want about humans to get us where we want to go. What’s your take on all that?

Yeah. I think it’s important to start by recognizing that most of the fundamental advances in our molecular understanding of disease, and of the biological pathways that are core to it, have come from experiments in model systems and model organisms. Even experiments in yeast, the simplest eukaryote, have revealed really important mechanisms relevant to human biology and human disease. Model organisms have played a fundamental role in the development of molecular biology and in our understanding of biology.

But you’re right to point out that they’re not perfect proxies by any means. Overly relying on them can lead to 2 things. You can think this is a cure to cancer, but it only works in mice. You can also miss a mechanism that’s relevant in humans because you don’t have a good proxy for it in mice. I think you can make that mistake in both ways.

There’s no simple solution to this problem. Biology is constantly pushing for better and better experimental proxies for human biology, and I think the latest advance there is organoids, which allow you to recapitulate more and more of the complexity of tissues in a culture system. But I think this is also where AI can play a role.

I come from a background of thinking about protein biology and building models of proteins. One thing that’s really interesting to observe is that ESM and AlphaFold did not come from training a model just on human proteins. They came from training a model across all of evolution. What that says is that there’s information across evolution that’s relevant to making predictions about and understanding human biology. It’s not just relevant; it’s essential. If you train a model just on human proteins, you wouldn’t get an accurate predictive model.

There’s something really powerful in this idea of generalization and transfer. So much of biology is conserved across all of life, across all the different domains of life. And as you get closer and closer to humans—even if you just think about vertebrates—so much of the core molecular machinery is conserved: the transcription factors, the way they interact with enhancers and promoters. That language is preserved.

What’s happened is that the programs have been rewired. The programs are completely rewired, but they’re built with the same building blocks. I think you can look broadly, and there’s a lot to learn from all of these different systems. It’s very different to think about building a model that’s going to capture, integrate, and learn from all of that information versus trying to create a perfect experimental proxy, which just isn’t going to be possible. But I think there’s power in trying to build a predictive model that can use all of that information to make far better predictions for human biology.

6. Training Models in Biology

In the context of language modeling, there was a time—and we’ve kind of grown out of it now—when the overwhelming amount of compute was going into simple pretraining on data of unknown quality and relevance. Then you’d do a little instruction following at the end, and all of a sudden you’d have a pretty helpful assistant. What do you think that recipe is going to look like in biology? Can we take tons and tons of raw but not-that-great data and use it to build up these representations, then elicit what we want at the end? Or do you think it’s going to be a different path for some reason?

Yeah, that’s an amazing question. I really think the principles in biology are going to be shared with all the other domains where we’re seeing progress in AI. We’re fortunate in biology because we have the opportunity to see how those principles have played out in different fields and identify the parts that are generalizable. I think this core idea of the power of self-supervision or representation learning—the pretraining—is really critical, because that gets you beyond needing to know all the relevant variables.

That was the dream of Hinton, Yann LeCun, Bengio, and the other deep learning pioneers: you didn’t have to know all the relevant variables, encode them, and know the settings for the parameters. You could learn all of that with a neural network. That’s been proving out over decades, and scaling laws are, I think, the clearest expression of that now. We have pretraining in biology. We’re certainly seeing that play out with proteins right now, and I think it will play out for cellular biology too.

I think it’s going to be similar. There’s a phase of learning from statistical patterns to build a model of the underlying system that’s driving those patterns, and then a stage of elicitation, where you bring out the part of that model that’s relevant to the problem you’re trying to solve. I think it’s going to be possible to have RLVR for biology and to have feedback. Ultimately, we need to build empirical AI and systems where AI can ask a question, perform the experiment, learn from it, improve its understanding, propose a better experiment, and complete that cycle repeatedly, so that we have recursive improvement.

I think all of that is going to be possible in biology. It’s going to need to be built differently. We’re going to need models that can accurately represent the multiscale nature and complexity of biology in the relevant ways. The biggest challenge for biology is that it’s empirical. We’re seeing rapid progress with RLVR in areas where we can do fast verification computationally, like mathematics, coding, and reasoning. With biology, we’re going to have to build systems that allow us to verify things in the laboratory.

How do you think about timescales? That’s another way in which there are orders-of-magnitude problems, right? It’s one thing to perturb a cell and look at it an hour later or a day later, but what I really care about is how I’m going to live to be 120. That’s a very long-term question. How do you think about this kind of prediction? This goes back to the idea of a million Bryan Johnsons. Do we have enough right now to say, “This cell state predicts good outcomes,” or is that also a big part of the climb toward the grand challenge?

Yeah, I think that’s still a really big part of the climb. But at the same time, I think it’s a scaling challenge, and one that can be addressed with technology we have today. What I think about is creating atlases of cellular biology in the context of health and disease. The technology that could allow us to do that is just starting to become scalable.

This is a big part of what we’re trying to do with the Virtual Biology Initiative: see cells across a wide range of physiological contexts and really understand what differentiates disease from health at a cellular level. That’s going to start to show us both what makes different diseases unique and the common causes shared across different classes of disease. I think that’s within reach. We’ll have first drafts of that within the next 1 or 2 years, and within 5 years, with a focused, coordinated effort, we can have a pretty comprehensive map.

That would have enormous value, not just for AI, but for classical analysis too. I imagine a lot of analysis is going to be agentic in the future, so you can imagine scaling classical analysis to these very large datasets. You can also start to build a model from that, one that will precisely identify the different cell states relevant to disease. With that model, you can start to ask: How do I intervene to reverse a pathological state and bring it into a healthy state? I think those are questions we’ll really start to be able to tackle in the next 5 years.

How big of a deal do you think human health records will be as part of that mix? Would you anticipate going to EHR companies and trying to make deals, maybe you’re already doing this, to get access to that data? I know the rules around this can be quite complicated, but you don’t want personal information, and you do need to meaningfully connect a person’s health trajectory to specific states of cells. Does the regulatory environment allow for that? Are there other practical barriers that would stand in the way? Or do you think that is going to be part of the story, with all these cells linked to human health stories?

Yeah, to be honest, I haven’t thought about that question as much, so maybe—why not? I’ve been more focused on trying to get to that fundamental level. I think about what the right basis is for solving the problem. If you don’t have the right basis, what you’re going to learn is a bunch of correlations. What I’d worry about is that we’ll have tons of models of patient health histories, but no way to link those back to the deeper causal factors in the underlying biology. That’s the part I think we really have to be able to do if we’re going to make this kind of step-change advancement in our understanding of disease and our ability to treat it.

You can probably go a long way with health records and getting better care to patients: getting the right existing drugs and mechanisms to the right people, and maybe opening up access for people who might benefit from drugs that exist, but where we didn’t know that the mechanism could be relevant. I think there could be a lot of progress through that. It’s a really important problem, but it doesn’t get down to the fundamentals where I think you have to make progress to get this real step change.

Biohub is really focused on that part of the problem: the fundamental science, and how we get to the point where there really is no disease anymore. The only way to get there is by focusing on the deep biology and building the technology that can allow us to understand how it determines health and disease, and how we can effectively intervene.

7. Fusing Language and Biology

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 with text and image? I say this all the time, so listeners have heard it, but to motivate the intuition one more time: it’s clear now that when I put an image into a model and give instructions, the model both sees the image in visual terms and understands my instructions. That creates a new kind of magic that just having a language model call an image generation model via an arm’s-length tool call couldn’t quite do. Do you think reasoning through language and biology models will stay in arm’s-length tool-call relationships long term, or do we ultimately see bio encoders that feed into the same reasoning space we’re developing in language?

Yeah. That’s a really interesting question. I think the truth is, I’m not sure. Language is really powerful for reasoning. It’s a substrate for reasoning, for thinking. The power that language also gives us is to tap into the model of biology that’s been created through centuries of human empirical investigation. All of that is represented in language; it’s learnable through language. I think this very expressive, full system of knowledge can really be materialized and leveraged through reasoning systems.

I think the challenge will be that the things we most want to know about biology aren’t represented in text, and they cannot be derived from first principles or from what we already know. They’re empirical facts that can only be derived by performing experiments. Thinking about this core aspect of what makes biology different from other areas where we’re building AI is what makes predictive models of biology critical here.

I’m not strongly opinionated or convicted right now about whether there’s a latent fusion of these models, whether they’re going to interact with each other in some system, or ultimately what the end state is for how that works. Is there one latent space where all of that is fused together? 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 bridging data to fuse those spaces together, and we just don’t have a lot of it right now. Could we get there eventually? Probably. It could be very interesting to have a fused latent space.

I think the other really interesting thing to think about is all those features in biology that we don’t yet have language or words to describe. Protein biology is such a beautiful example of this. We have billions of proteins that have been sequenced across life on Earth, and we don’t know the function for the vast majority of them. We don’t even really have a language to describe all of that.

What’s interesting about training a protein language model is that it develops features that represent protein biology at large. There are features that represent unknown protein biology, and you can imagine those are novel concepts that we don’t have words or names for yet. There’s a great deal to be learned just through the structure of these spaces.

Okay. We’re almost out of time. I’ll do 2 more quick ones, if we can.

Okay.

One, not to be a dog with a bone on this million Bryan Johnsons question, but that would be pretty expensive. If money were no object—and for some rare philanthropist, money is maybe close to no object in pursuit of the cure for all diseases—is that a promising paradigm? Would it be really helpful to just have people submit to micro-biopsies on the regular, wear all the sensors, do all the things, and track every dimension they can track? Is that the shape of data that you think will ultimately really move the needle, and it’s just too hard to get? Or do you think there’s, for more scientific reasons, a more direct path through scaling cells?

I think we just come at it from this frame of really thinking about the fundamental mechanisms. I think that data would be immensely valuable. I just see that as more likely to happen. I think over time that will happen.

Mm-hmm.

There are some really interesting things that can move the needle there, which is developing the technologies for this kind of continuous monitoring over time, and having those technologies be able to measure more parameters and more molecular-level parameters, so that you can start to link that bottom-up and top-down view. I think that kind of work is immensely valuable, and there’s likely to be convergence there over time. We just think about where, as an institution, we can really move the needle, and we really see ourselves at that technology layer.

Yeah, interesting. Maybe it’s less neglected than it seems to me from my outside position. All right, last one: personal health advice. What are you doing to build up the data set you need for yourself to be able to take advantage of these future technologies? And what, if anything, do you feel is important for people who want to make sure they’re in a good position to take advantage of all the things Biohub and others are going to discover to be doing today?

I’m not the right person to ask for that kind of advice. I’m just so focused on my research right now that I barely think about that at all.

So you’re not tracking anything? You’re not—

No. No.

—getting regular blood tests?

I have a watch that tracks my sleep. That’s about as far as it gets.

All right. We’ll check back in with you and see if maybe that’s changed as we get closer to the promised land and 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 Rives, thank you for being part of The Cognitive Revolution.

Thank you, Nathan. Yeah, this was really fun.