It's hard for us humans to make any kind of clean predictions about highly nonlinear dynamical systems. But again, to your point, we might be very surprised what classical learning systems might be able to do about even fluid.
Yes, exactly. I mean fluid dynamics, Navier-Stokes equations, these are traditionally thought of as very, very difficult intractable problems to do on classical systems. They take enormous amounts of compute, you know, weather prediction systems, these kind of things all involve fluid dynamics calculations. But again, if you look at something like Veo, our video generation model, it can model liquids quite well, surprisingly well, and materials, specular lighting. I love the ones where people have generated videos where there's clear liquids going through hydraulic presses and then being squeezed out. I used to write physics engines and graphics engines in my early days in gaming, and I know it's just so painstakingly hard to build programs that can do that. And yet somehow these systems are reverse engineering from just watching YouTube videos. So presumably what's happening is it's extracting some underlying structure around how these materials behave. So perhaps there is some kind of lower dimensional manifold that can be learned if we actually fully understood what's going on under the hood. That's maybe true of most of reality.
The following is a conversation with Demis Hassabis, his second time on the podcast. He is the leader of Google DeepMind and is now a Nobel Prize winner. Demis is one of the most brilliant and fascinating minds in the world today, working on understanding and building intelligence and exploring the big mysteries of our universe. This was truly an honor and a pleasure for me. This is the Lex Fridman podcast. To support it, please check out our sponsors in the description and consider subscribing to this channel. And now, dear friends, here's Demis Hassabis.
In your Nobel Prize lecture, you propose what I think is a super interesting conjecture that quote any pattern that can be generated or found in nature can be efficiently discovered and modeled by a classical learning algorithm. What kind of patterns of systems might be included in that? Biology, chemistry, physics, maybe cosmology, neuroscience. What are we talking about?
Sure. Well, look, I felt that it's sort of a tradition, I think, of Nobel Prize lectures that you're supposed to be a little bit provocative and I wanted to follow that tradition. What I was talking about there is if you take a step back and you look at all the work that we've done especially with the Alpha projects, so I'm thinking AlphaGo of course, AlphaFold, what they really are is we're building models of very combinatorially high-dimensional spaces that if you tried to brute force a solution, find the best move in Go or find the exact shape of a protein, and if you enumerated all the possibilities, there wouldn't be enough time in the time of the universe. So you have to do something much smarter, and what we did in both cases was build models of those environments. And that guided the search in a smart way and that makes it tractable. So if you think about protein folding, which is obviously a natural system, why should that be possible? How does physics do that? Proteins fold in milliseconds in our bodies. So somehow physics solves this problem that we've now also solved computationally. And I think the reason that's possible is that in nature, natural systems have structure because they were subject to evolutionary processes that shape them. And if that's true, then you can maybe learn what that structure is.
So this perspective, I think, is really interesting. You've hinted at it, which is almost like crudely stated, anything that can be evolved can be efficiently modeled. I think there's some truth to that.
Yeah, I sometimes call it survival of the stablest or something like that because there's evolution for life, living things, but there's also, if you think about geological time, the shape of mountains has been shaped by weathering processes over thousands of years. But then you can even take it cosmological, the orbits of planets, the shapes of asteroids, these have all survived kind of processes that have acted on them many, many times. So if that's true, then there should be some sort of pattern that you can kind of reverse learn and a kind of manifold really that helps you search to the right solution, to the right shape, and actually allow you to predict things about it in an efficient way because it's not a random pattern. So it may not be possible for man-made things or abstract things like factorizing large numbers, because unless there's patterns in the number space, which there might be, but if there's not and it's uniform, then there's no pattern to learn, there's no model to learn that will help you search. So you have to do brute force. So in that case you maybe need a quantum computer or something like this. But most things in nature that we're interested in are not like that. They have structure that evolved for a reason and survived over time. And if that's true, I think that's potentially learnable by a neural network.
It's like nature is doing a search process and it's so fascinating that in that search process it's creating systems that could be efficiently modeled.
That's right. So they can be efficiently rediscovered or recovered because nature is not random. Everything that we see around us, including the elements that are more stable, all of those things are subject to some kind of selection process pressure.
Do you think, because you're also a fan of theoretical computer science and complexity, do you think we can come up with a kind of complexity class, like a complexity zoo type of class, where maybe it's the set of learnable systems, the set of learnable natural systems, LNS? Yeah, this is a new class of systems that could be actually learnable by classical systems in this kind of way. Natural systems that can be modeled efficiently.
Yeah, I mean, I've always been fascinated by the P vs NP question and what is modelable by classical systems, non-quantum systems, Turing machines in effect. And that's exactly what I'm working on actually in my few moments of spare time with a few colleagues, is should there be maybe a new class of problem that is solvable by this type of neural network process and kind of mapped onto these natural systems, the things that exist in physics and have structure. So I think that could be a very interesting new way of thinking about it. And it sort of fits with the way I think about physics in general, which is that I think information is primary. Information is the most fundamental unit of the universe, more fundamental than energy and matter. I think they can all be converted into each other, but I think of the universe as a kind of informational system. So when you think of the universe as an informational system, then the P vs NP question is a physics question.
That's right. And it's a question that can help us actually solve the entirety of this whole thing going on.
Yeah, I think it's one of the most fundamental questions actually if you think of physics as informational, and the answer to that I think is going to be very enlightening.
More specific to the P vs NP question. This again, some of the stuff we're saying is kind of crazy right now. Just like the thing Edison Nobel Prize speech controversial thing that he said sounded crazy, and then you went and got a Nobel Prize for this with John Jumper, solved the problem. So let me just stick to the P vs NP. Do you think there's something in this thing we're talking about that could be shown if you can do something like polynomial time or constant time compute ahead of time and construct this gigantic model, then you can solve some of these extremely difficult problems in a theoretical computer science kind of way?
Yeah, I think that there are actually a huge class of problems that could be couched in this way. The way we did AlphaGo and the way we did AlphaFold, where you model what the dynamics of the system is, the properties of that system, the environment that you're trying to understand, and then that makes the search for the solution or the prediction of the next step efficient, basically polynomial time, so tractable by a classical system, which a neural network is, it runs on normal computers, classical computers, Turing machines in effect. And I think it's one of the most interesting questions there is, is how far can that paradigm go? I think we've proven and the AI community in general that classical systems, Turing machines, can go a lot further than we previously thought. They can do things like model the structures of proteins and play Go to better than world champion level. A lot of people would have thought maybe 10, 20 years ago that was decades away or maybe you would need some sort of quantum systems to be able to do things like protein folding. And so I think we haven't really even scratched the surface yet of what classical systems could do. And of course AGI being built on a neural network system on top of a classical computer would be the ultimate expression of that. And I think the limit, the bounds of that kind of system, what it can do, is a very interesting question and directly speaks to the P vs NP question.
What do you think, again hypothetical, might be outside of this? Maybe emergent phenomena, like if you look at cellular automata, some of the extremely simple systems and then some complexity emerges. Maybe that would be outside, or even would you guess even that might be amenable to efficient modeling by a classical machine?
Yeah, I think those systems would be right on the boundary. I think most emergent systems, cellular automata things, could be modelable by a classical system. You just sort of do a forward simulation of it and it probably be efficient enough. Of course, there's the question of things like chaotic systems where the initial conditions really matter and then you get to some uncorrelated end state. Those could be difficult to model. So I think these are kind of the open questions. But I think when you step back and look at what we've done with the systems and the problems that we've solved, and then you look at things like Veo on video generation, sort of rendering physics and lighting and things like that, really core fundamental things in physics, it's pretty interesting. I think it's telling us something quite fundamental about how the universe is structured in my opinion. So in a way, that's what I want to build AGI for, is to help us as scientists answer these questions like P vs NP.
Yeah, I think we might be continuously surprised about what is modelable by classical computers. I mean AlphaFold 3 on the interaction side is surprising that you can make any kind of progress on that direction. AlphaGenome is surprising that you can map the genetic code to the function, kind of playing with the emergent kind of phenomena. You think there's so many combinatorial options and then here you go, you can find the kernel that is efficiently modeled.
Yes. Because there's some structure, there's some landscape, you know, in the energy landscape or whatever it is that you can follow some gradient. And of course what neural networks are very good at is following gradients. And so if there's one to follow and you can specify the objective function correctly, you don't have to deal with all that complexity, which I think is how we maybe have naively thought about it for decades. Those problems, if you just enumerate all the possibilities, it looks totally intractable, and there's many, many problems like that. And then you think, well, it's like 10 to the 300 possible protein structures, it's 10 to the 170 possible Go positions, all of these are way more than atoms in the universe. So how could one possibly find the right solution or predict the next step? But it turns out that it is possible. And of course, reality, nature does do it, right? Proteins do fold. So that gives you confidence that there must be, if we understood how physics was doing that, in a sense, and then we could mimic that process, model that process, it should be possible on our classical systems. That is basically what the conjecture is about.
And of course there's nonlinear dynamical systems, highly nonlinear dynamical systems, everything involving fluid.
Yes. Right. You know, I recently had a conversation with Terrence Tao who mathematically contends with a very difficult aspect of systems that have some singularities in them that break the mathematics, and it's just hard for us humans to make any kind of clean predictions about highly nonlinear dynamical systems. But again, to your point, we might be very surprised what classical learning systems might be able to do about even fluid. Yes, exactly. I mean fluid dynamics, Navier-Stokes equations, these are traditionally thought of as very, very difficult intractable problems to do on classical systems. They take enormous amounts of compute, you know, weather prediction systems, these kind of things all involve fluid dynamics calculations. But again, if you look at something like Veo, our video generation model, it can model liquids quite well, surprisingly well, and materials, specular lighting. I love the ones where people have generated videos where there's clear liquids going through hydraulic presses and then being squeezed out. I used to write physics engines and graphics engines in my early days in gaming, and I know it's just so painstakingly hard to build programs that can do that. And yet somehow these systems are reverse engineering from just watching YouTube videos. So presumably what's happening is it's extracting some underlying structure around how these materials behave. So perhaps there is some kind of lower dimensional manifold that can be learned if we actually fully understood what's going on under the hood. That's maybe true of most of reality.
Yeah. I've been continuously surprised by this aspect of Veo. I think a lot of people highlight different aspects, including the comedic and the meme and all that kind of stuff, and then the ultra realistic ability to capture humans in a really nice way that's compelling and feels close to reality, and then combine that with native audio. All of those are marvelous things about Veo. But the exactly the thing you're mentioning, which is the physics. It's not perfect, but it's pretty damn good. And then the really interesting scientific question is what is it understanding about our world in order to be able to do that? Because the cynical take with diffusion models, there's no way it understands anything, but it seems, I mean, I don't think you can generate that kind of video without understanding. And then our own philosophical notion of what it means to understand is brought to the surface. To what degree do you think Veo understands our world?
I think to the extent that it can predict the next frames in a coherent way, that is a form of understanding, right? Not in the anthropomorphic version, it's not some kind of deep philosophical understanding of what's going on. I don't think these systems have that, but they certainly have modeled enough of the dynamics, put it that way, that they can pretty accurately generate whatever it is, 8 seconds of consistent video that by eye at least, at a glance, is quite hard to distinguish what the issues are. And imagine that in two or three more years time. That's the thing I'm thinking about, how incredible that will look given where we've come from, the early versions of that one or two years ago. And so the rate of progress is incredible. And I think, like you, a lot of people love all the standup comedians and that actually captures a lot of human dynamics very well and body language, but actually the thing I'm most impressed with and fascinated by is the physics behavior, the lighting and materials and liquids. And it's pretty amazing that it can do that. And I think that shows that it has some notion of at least intuitive physics, how things are supposed to work intuitively, maybe the way that a human child would understand physics, as opposed to a PhD student really being able to unpack all the equations. It's more of an intuitive physics understanding.
Well, that intuitive physics understanding, that's the base layer, that's the thing people sometimes call common sense. It really understands something. I think that really surprised a lot of people. It blows my mind that I just didn't think it would be possible to generate that level of realism without understanding. There's this notion that you can only understand the physical world by having an embodied AI system, a robot that interacts with that world. That's the only way to construct an understanding of that world. But Veo is directly challenging that.
Yes, and it's very interesting. Even if you were to ask me 5, 10 years ago, I would have said, even though I was immersed in all of this, I would have said, well, yeah, you probably need to understand intuitive physics, you know, like if I push this off the table, this glass will maybe shatter and the liquid will spill out. So we know all of these things, but I thought that, and there's a lot of theories in neuroscience called action in perception, where you need to act in the world to really truly perceive it in a deep way. And there was a lot of theories about you need embodied intelligence or robotics or something, or maybe at least simulated action, so that you would understand things like intuitive physics. But it seems that you can understand it through passive observation, which is pretty surprising to me, and again I think hints at something underlying about the nature of reality in my opinion, beyond just the cool videos that it generates. And of course the next stage is maybe even making those videos interactive, so one can actually step into them and move around them, which would be really mind-blowing especially given my games background. So you can imagine, and then I think we're starting to get towards what I would call a world model, a model of how the world works, the mechanics of the world, the physics of the world, and the things in that world. And of course that's what you would need for a true AGI system.
I have to talk to you about video games. So, you were being a bit trolly. I think you're having more and more fun on Twitter on X, which is great to see. So, a guy named Jimmy Apples tweeted, "Let me play a video game of my Veo videos already. Google cooked so good, playable world models when spelled W-E-N question mark." And then you quote tweeted that with, "Now, wouldn't that be something?" So, how hard is it to build game worlds with AI? Maybe can you look out into the future of video games 5, 10 years out? What do you think that looks like?
Well, games were my first love really, and doing AI for games was the first thing I did professionally in my teenage years and was the first major AI systems that I built. And I always want to scratch that itch one day and come back to that. And I will do, I think. And I think I'd sort of dream about what would I have done back in the '90s if I'd had access to the kind of AI systems we have today. And I think you could build absolutely mind-blowing games. And I think the next stage is, I always used to love making all the games I've made are open world games. So they're games where there's a simulation and then there's AI characters and then the player interacts with that simulation and the simulation adapts to the way the player plays. And I always thought they were the coolest games because games like Theme Park that I worked on, everybody's game experience would be unique to them, because you're kind of co-creating the game. We set up the parameters, we set up initial conditions, and then you as the player immersed in it, and then you are co-creating it with the simulation. But of course, it's very hard to program open world games. You've got to be able to create content whichever direction the player goes in and you want it to be compelling no matter what the player chooses. And so it was always quite difficult to build things like cellular automata type of those kind of classical systems which created some emergent behavior, but they're always a little bit fragile, a little bit limited. Now we're maybe on the cusp in the next few years, 5, 10 years, of having AI systems that can truly create around your imagination, can dynamically change the story and the narrative around and make it dramatic no matter what you end up choosing. So it's like the ultimate choose your own adventure sort of game. And I think maybe we're within reach if you think of a kind of interactive version of Veo, and then wind that forward 5 to 10 years and imagine how good it's going to be.
Yeah. So you said a lot of super interesting stuff there. So one, the open world built into that is a deep personalization the way you've described it. So it's not just that it's open world like you can open any door and there'll be something there. It's that the choice of which door you open in an unconstrained way defines the worlds you see. So some games try to do that to give you choice.
Yes. But it's really just an illusion of choice because the only, like Stanley Parable game I recently played, it's really there's a couple of doors and it really just takes you down a narrative.
Stanley Parable is a great video game, I recommend people play that. It kind of in a meta way mocks the illusion of choice and there's philosophical notions of free will and so on. But I do like one of my favorite games of Elder Scrolls is Daggerfall. I believe that they really played with random generation of the dungeons.
Yeah. Of you can step in and they give you this feeling of an open world and you mentioned interactivity. You don't need to interact that much. You just when you open the door, whatever you see is randomly generated for you. And that's already an incredible experience because you might be the only person to ever see that.
Exactly. And so but what you'd like is a little bit better than just sort of a random generation, right?
So you'd like, and also better than a simple A/B hardcoded choice, right? That's not really open world, as you say, it's just giving you the illusion of choice. What you want to be able to do is potentially anything in that game environment. And I think the only way you can do that is to have generated systems, systems that will generate that on the fly. Of course, you can't create infinite amounts of game assets, right? It's expensive enough already how AAA games are made today. And that was obvious to us back in the '90s when I was working on all these games. I think maybe Black & White was the game that I worked on, early stages of that, that had the still probably the best AI learning AI in it. It was an early reinforcement learning system that you were looking after this mythical creature and growing it and nurturing it, and depending how you treated it, it would treat the villagers in that world in the same way. So if you were mean to it, it would be mean. If you were good, it would be protective. And so it was really a reflection of the way you played it. So actually all of the, I've been working on sort of simulations and AI through the medium of games at the beginning of my career, and really the whole of what I do today is still a follow on from those early more hardcoded ways of doing the AI to now fully general learning systems that are trying to achieve the same thing.
Yeah, it's been interesting, hilarious, and fun to watch you and Elon obviously itching to create games because you're both gamers. And one of the sad aspects of your incredible success in so many domains of science, like serious adult stuff, that you might not have time to really create a game. You might end up creating the tooling that others would create the game. You have to watch others create the thing you've always dreamed of. Do you think it's possible you can somehow in your extremely busy schedule actually find time to create something like Black & White, some actual video game where you could make the childhood dream become reality?
You know, there's two ways to think about that. One is maybe with vibe coding as it gets better, there's a possibility that I could do that actually in my spare time. So I'm quite excited about that as that would be my project if I got the time to do some vibe coding. I'm actually itching to do that. And then the other thing is, maybe it's a sabbatical after AGI has been safely stewarded into the world and delivered into the world. And then working on my physics theory as we talked about at the beginning. Those would be my two post-AGI projects, let's call it that way.
I would love to see which game post-AGI you choose. Solving the problem that some of the smartest people in human history contended with, P vs NP, or creating a cool video game.
Well, but they might, in my world they'd be related because it would be an open world simulated game as realistic as possible. So, what is the universe? That's speaking to the same question, right? P vs NP. I think all these things are related, at least in my mind.
I mean in a really serious way, like video games sometimes are looked down upon as just this fun side activity, but especially as AI does more and more of the difficult boring tasks, something we in the modern world call work. Video games is the thing in which we may find meaning, in which we may find what to do with our time. You could create incredibly rich, meaningful experiences. Like that's what human life is. And then in video games, you can create more sophisticated, more diverse ways of living, right?
I think so. I mean, those of us who love games, and I still do, it's almost can let your imagination run wild. I used to love games and working on games so much because it's the fusion, especially in the '90s and early 2000s, the sort of golden era maybe the '80s of the games industry, and it was all being discovered, new genres were being discovered. We weren't just making games, we felt we were creating a new entertainment medium that never existed before, especially with these open world games and simulation games where you were co-creating, you as the player were co-creating the story. There's no other media, entertainment media, where you do that, where you as the audience actually co-create the story. And of course now with multiplayer games as well, it can be a very social activity and can explore all kinds of interesting worlds in that. But on the other hand, it's very important to also enjoy and experience the physical world. But the question is then, I think we're going to have to kind of confront the question again of what is the fundamental nature of reality? What is going to be the difference between these increasingly realistic simulations and multiplayer ones and emergent ones, and what we do in the real world?
Yeah, there's clearly a huge amount of value to experiencing the real world, nature. There's also a huge amount of value in experiencing other humans directly in person the way we're sitting here today. But we need to really scientifically rigorously answer the question why. And which aspect of that can be mapped into the virtual world?
Exactly. It's not enough to say, "Yeah, you should go touch grass and hang out in nature." It's like, why exactly is that valuable? And I guess that's maybe the thing that's been haunting me, obsessing me from the beginning of my career. If you think about all the different things I've done, they're all related in that way. This simulation, nature of reality, and what is the bounds of what can be modeled.
Sorry for the ridiculous question, but so far, what is the greatest video game of all time? What's up there?
Well, my favorite one of all time is Civilization. I have to say that, Civilization 1 and Civilization 2. My favorite games of all time.
I can only assume you've avoided the most recent one because it would probably, you would disappear.
Yes, exactly. They take a lot of time, these Civilization games. So I got to be careful with them.
Fun question. You and Elon seem to be somehow solid gamers. Is there a connection between being great at gaming and being great leaders of AI companies?
I don't know. It's an interesting one. I mean, we both love games and it's interesting he wrote games as well to start off with. It's probably especially in the era I grew up in where home computers just became a thing in the late '80s and '90s, especially in the UK. I had a Spectrum and then a Commodore Amiga 500, which is my favorite computer ever, and that's why I learned all my programming. And of course it's a very fun thing to program is to program games. So I think it's a great way to learn programming, probably still is. And then of course I immediately took it in directions of AI and simulations, so I was able to express my interest in games and my sort of wider scientific interests all together. And then the final thing I think that's great about games is it fuses artistic design, art, with the most cutting edge programming. So again, in the '90s, all of the most interesting technical advances were happening in gaming, whether that was AI, graphics, physics engines, hardware, even GPUs of course were designed for gaming originally. So everything that was pushing computing forward in the '90s was due to gaming. So interestingly, that was where the forefront of research was going on and it was this incredible fusion with art, graphics, but also music and just the whole new media of storytelling. And I love that. For me, this sort of multidisciplinary effort is again something I've enjoyed my whole life.
I have to ask you, I almost forgot about one of the many, and I would say one of the most incredible things recently that somehow didn't yet get enough attention is AlphaEvolve. We talked about evolution a little bit, but it's the Google DeepMind system that evolves algorithms. Are these kinds of evolution-like techniques promising as a component of future super intelligence systems? So for people who don't know, it's kind of, I don't know if it's fair to say it's LLM guided evolution search.
Yeah. So evolutionary algorithms are doing the search and LLMs are telling you where. Yes. Exactly. So LLMs are kind of proposing some possible solutions and then you use evolutionary computing on top to find some novel part of the search space. So actually I think it's an example of very promising directions where you combine LLMs or foundation models with other computational techniques. Evolutionary methods is one, but you could also imagine Monte Carlo search, basically many types of search algorithms or reasoning algorithms on top of or using the foundation models as a basis. So I actually think there's quite a lot of interesting things to be discovered probably with these sort of hybrid systems, let's call them.
But not to romanticize evolution. I'm only human. But you think there's some value in whatever that mechanism is because we already talked about natural systems. Do you think where there's a lot of low-hanging fruit of us understanding, being able to model, being able to simulate evolution, and then using that, whatever we understand about that nature inspired mechanism, to then do search better and better and better?
Yes. So if you think about again breaking down the sort of systems we've built to their really fundamental core, you've got the model of the underlying dynamics of the system. And then if you want to discover something new, something novel that hasn't been seen before, then you need some kind of search process on top to take you to a novel region of the search space. And you can do that in a number of ways. Evolutionary computing is one. With AlphaGo we just use Monte Carlo tree search, right? And that's what found move 37, the new kind of never seen before strategy in Go. And so that's how you can go beyond potentially what is already known. So the model can model everything that you currently know about, all the data that you currently have, but then how do you go beyond that? That starts to speak about the ideas of creativity. How can these systems create something new, discover something new? Obviously this is super relevant for scientific discovery or pushing medicine forward, which we want to do with these systems. And you can actually bolt on some fairly simple search systems on top of these models and get you into a new region of space. Of course, you also have to make sure that you're not searching that space totally randomly. It would be too big. So you have to have some objective function that you're trying to optimize and hill climb towards, and that guides that search.
But there's some mechanism of evolution that are interesting maybe in the space of programs. But then the space of programs is an extremely important space because you can probably generalize to everything. For example, mutation. So it's not just Monte Carlo tree search where it's like a search. You could every once in a while combine things, alter subcomponents of a thing.
Yes. So then you know what evolution is really good at is not just the natural selection. It's combining things and building increasingly complex hierarchical systems. So that component is super interesting, especially with AlphaEvolve in the space of programs.
Exactly. So there's a, you can get a bit of an extra property out of evolutionary systems which is some new emergent capability may come about, but of course like happened with life.
Interestingly, with naive traditional evolutionary computing methods without LLMs and the modern AI, the problem with them was they were very well studied in the '90s and early 2000s and some promising results, but the problem was they could never work out how to evolve new properties, new emergent properties. You always had a sort of subset of the properties that you put into the system. But maybe if we combine them with these foundation models, perhaps we can overcome that limitation. Obviously natural evolution clearly did because it did evolve new capabilities, right? From bacteria to where we are now. So clearly it must be possible with evolutionary systems to generate new patterns, going back to the first thing we talked about, and new capabilities and emergent properties. And maybe we're on the cusp of discovering how to do that.
Yeah, listen, AlphaEvolve is one of the coolest things I've ever seen. I've on my desk at home, most of my time is spent behind the computers just programming. And next to the three screens is a skull of a Tiktaalik, which is one of the early organisms that crawled out of the water onto land. And I just kind of watch that little guy. It's like, whatever the computation mechanism of evolution is, it's quite incredible. It's truly, truly incredible. Now whether that's exactly the thing we need to do to do our search, but never dismiss the power of nature what it did here.
Yeah. And it's amazing, which is a relatively simple algorithm effectively, and it can generate all of this immense complexity emerges, obviously running over 4 billion years of time. But it's, you can think about that as again a search process that ran over the physics substrate of the universe for a long amount of computational time, but then it generated all this incredible rich diversity.
So many questions I want to ask you. But one, you do have a dream. One of the natural systems you want to try to model is a cell. Yes, that's a beautiful dream. I could ask you about that. I also just for that purpose on the AI scientist front just broadly. So there's an essay from Daniel Kokotajlo, Scott Alexander, and others that outlines steps along the way to get to ASI and has a lot of interesting ideas in it. One of which is including a superhuman coder and a superhuman AI researcher, and in that there's a term of research taste that's really interesting. So in everything you've seen, do you think it's possible for AI systems to have research taste to help you in the way that AI co-scientist does, to help steer human brilliant scientists, and then potentially by itself to figure out what are the directions where you want to generate truly novel ideas? Because that seems to be a really important component of how to do great science.
Yeah, I think that's going to be one of the hardest things to mimic or model, this idea of taste or judgment. I think that's what separates the great scientists from the good scientists. All professional scientists are good technically, right? Otherwise they wouldn't have made it that far in academia and things like that. But then do you have the taste to sort of sniff out what the right direction is, what the right experiment is, what the right question is? So picking the right question is the hardest part of science. And making the right hypothesis. And that's what today's systems definitely can't do. So I often say it's harder to come up with a conjecture, a really good conjecture, than it is to solve it. So we may have systems soon that can solve pretty hard conjectures. I mean, Math Olympiad problems, where AlphaProof last year got a silver medal in those really hard problems. Maybe eventually we'll be able to solve a Millennium Prize kind of problem. But could a system have come up with a conjecture worthy of study that someone like Terrence Tao would have gone, you know what, that's a really deep question about the nature of maths or the nature of numbers or the nature of physics? And that is far harder type of creativity, and we don't really, systems clearly can't do that, and we're not quite sure what that mechanism would be. This kind of leap of imagination like Einstein had when he came up with special relativity and then general relativity with the knowledge he had at the time.
As for conjecture, you want to come up with a thing that's interesting and amenable to proof. So it's easy to come up with a thing that's extremely difficult. It's easy to come up with a thing that's extremely easy. At that very edge, that sweet spot, of basically advancing the science and splitting the hypothesis space into two ideally, right? Whether if it's true or not true, you've learned something really useful.
And making something that's also falsifiable and within the technologies that you currently have available. So it's a very creative process actually, highly creative process, that I think just a kind of naive search on top of a model won't be enough for that.
Okay. The idea of splitting the hypothesis space in two is super interesting. So I've heard you say that there's basically no failure, or failure is extremely valuable if it's done, if you construct the questions right, if you construct the experiments right, if you design them right, that failure and success are both useful.
So perhaps because it splits the hypothesis space in two, it's like a binary search. That's right. So when you do real blue sky research, there's no such thing as failure really, as long as you're picking experiments and hypotheses that meaningfully split the hypothesis space. So you learn something, you can learn something equally valuable from an experiment that doesn't work. That should tell you, if you've designed the experiment well and your hypotheses are interesting, it should tell you a lot about where to go next. And then you're effectively doing a search process and using that information in very helpful ways.
So to go to your dream of modeling a cell, what are the big challenges that lay ahead for us to make that happen? We should maybe highlight that Alpha, I mean there's just so many leaps. So AlphaFold solved, if it's fair to say, protein folding, and there's so many incredible things we could talk about there including the open sourcing of everything you've released. AlphaFold 3 is doing protein, RNA, DNA interactions, which is super complicated and fascinating. That's amenable to modeling. AlphaGenome predicts how small genetic changes, like if we think about single mutations, how they link to actual function. So those are, it seems like it's creeping along to much more complicated things like a cell, but a cell has a lot of really complicated components.
Yeah. So what I've tried to do throughout my career is I have these really grand dreams and then I try to break them down. It's easy to have a kind of crazy ambitious dream, but the trick is how do you break it down into manageable achievable interim steps that are meaningful and useful in their own right. And so Virtual Cell, which is what I call the project of modeling a cell, I've had this idea of wanting to do that for maybe more like 25 years. I used to talk with Paul Nurse, who is a bit of a mentor of mine in biology, he runs the Francis Crick Institute and won the Nobel Prize in 2001. We've been talking about it since before the '90s, and I used to come back to it every 5 years, like what would you need to model the full internals of a cell so that you could do experiments on the virtual cell and those predictions would be useful for you to save you a lot of time in the wet lab? That would be the dream. Maybe you could 100x speed up experiments by doing most of it in silico, the search in silico, and then you do the validation step in the wet lab. That's the dream. And so maybe now finally, I was trying to build these components, AlphaFold being one, that would allow you eventually to model the full simulation of a cell. And I'd probably start with a yeast cell, partly because that's what Paul Nurse studied, because a yeast cell is like a full organism that's a single cell, so it's the kind of simplest single cell organism. And it's not just a cell, it's a full organism. And yeast is very well understood. So that would be a good candidate for a full simulated model. Now AlphaFold is the solution to the static picture of what does a protein look like, the 3D structure. But we know that biology, all the interesting things happen with the dynamics, the interactions. And that's what AlphaFold 3 is, the first step towards modeling those interactions. So first of all pairwise, proteins with proteins, proteins with RNA and DNA. But then the next step after that would be modeling maybe a whole pathway, maybe like the p53 pathway that's involved in cancer or something like this. And then eventually you might be able to model a whole cell.
Also there's another complexity here, that stuff in a cell happens at different time scales. Is that tricky? Like protein folding is super fast, some of them take a long time. So that's an additional level. The levels of interaction have a different temporal scale that you have to be able to model.
So that would be hard. So you'd probably need several simulated systems that can interact at these different temporal dynamics, or at least maybe it's like a hierarchical system. So you can jump up and down the different temporal stages. So one of the challenges here is not to avoid simulating, for example, the quantum mechanical aspects of any of this. You want to not over-model, you can skip ahead to just model the really high-level things that get you a really good estimate of what's going to happen. So you got to make a decision when you're modeling any natural system, what is the cutoff level of the granularity that you're going to model it to that then captures the dynamics that you're interested in. So probably for a cell I would hope that would be the protein level, and that one wouldn't have to go down to the atomic level. So of course that's where AlphaFold kicks in. So that would be kind of the basis, and then you'd build these higher level simulations that take those as building blocks and then you get the emergent behavior.
Apologize for the bad questions ahead of time, but will we be able to simulate and model the origin of life? So being able to simulate the first from non-living organisms, the birth of a living organism.
I think that's one of the deepest and most fascinating questions. I love that area of biology. There's a great book by Nick Lane, one of the top experts in this area, called "The 10 Great Inventions of Evolution." I think it's fantastic, and it also speaks to what the great filters might be, whether they are prior or ahead of us. I think they're most likely in the past, if you read that book, of how unlikely it is to have any life at all, and then single cell to multicell seems an unbelievably big jump that took like a billion years on Earth to do. So it shows you how hard it was. Bacteria were super happy for a very long time before they captured mitochondria somehow. I don't see why not, why AI couldn't help with that, some kind of simulation. Again, it's a bit of a search process through a combinatorial space. Here's all the chemical soup that you start with, the primordial soup that maybe was on Earth near these hot vents. Here's some initial conditions. Can you generate something that looks like a cell? So perhaps that would be a next stage after the Virtual Cell project, is how could you actually have something like that emerge from the chemical soup?
Well, I would love it if there was a move 37 for the origin of life.
Yeah, I think that's one of the great mysteries. I think ultimately what we will figure out is there's a continuum. There's no such thing as a line between non-living and living. But if we can make that rigorous, that the very thing from the Big Bang to today has been the same process. If we can break down that wall that we've constructed in our minds of the actual origin from non-living to living, and it's not a line, that it's a continuum that connects physics and chemistry and biology. There's no line. I mean, this is my whole reason why I've worked on AI and AGI my whole life, because I think it can be the ultimate tool to help us answer these kind of questions. And I don't really understand why the average person doesn't think about this stuff more. How can we not have a good definition of life and non-living, and the nature of time, and let alone consciousness and gravity and all these things, and quantum mechanics weirdness. It's just, to me, I've always had this sort of screaming at me in my face the whole time, and it's getting louder. It's like, how, what is going on here? I mean that in the deepest sense, the nature of reality, which has to be the ultimate question that would answer all of these things. It's sort of crazy if you think about it. We can stare at each other and all these living things all the time. We can inspect it with microscopes and take it apart almost down to the atomic level, and yet we still can't answer that clearly in a simple way, that question of how do you define living?
Yeah, it's kind of amazing. Living, you can kind of talk your way out of thinking about, but like consciousness, like we have this very obviously subjective conscious experience.
We're at the center of our own world and it feels like something. How are you not screaming at the mystery of it all?
But really, humans have been contending with the mystery of the world around them for a long time. There's a lot of mysteries, like what's up with the sun and the rain? What's that about? And then last year we had a lot of rain and this year we don't have rain. Like, what did we do wrong? Humans have been asking that question for a long time.
Exactly. So we've developed a lot of mechanisms to cope with these deep mysteries that we can see but can't fully understand, and we have to just get on with daily life. Do we keep ourselves distracted?
Weather is one of the most important questions of human history. It's the go-to small talk direction, especially in England, which is famously an extremely difficult system to model. Even that system, Google DeepMind has made progress on. We've created the best weather prediction systems in the world, better than traditional fluid dynamics systems that usually take days to calculate on massive supercomputers. We've managed to model a lot of the weather dynamics with neural network systems with our GraphCast system. It's interesting that those kinds of dynamics can be modeled even though they're very complicated, almost bordering on chaotic systems. A lot of the interesting aspects can be modeled by these neural network systems, including very recently we had cyclone prediction of where paths of hurricanes might go. Of course, super useful and important for the world, and it's super important to do that very timely and quickly as well as accurately. I think it's a very promising direction for simulating and running forward predictions of very complicated real-world systems.
I should mention that I got a chance in Texas to meet a community of folks called the storm chasers. What's really incredible about them is they're extremely tech-savvy because they have to use models to predict where the storm is. It's this beautiful mix of being crazy enough to go into the eye of the storm, and in order to protect your life and predict where extreme events are going to be, they have to have increasingly sophisticated models of weather.
Yeah, it's a beautiful balance of being in it as living organisms and the cutting edge of science. They might actually be using a DeepMind system.
Hopefully they are. I'd love to join them on one of those chases. They look amazing, to actually experience it and also experience the correct prediction of where something will come and how it's going to evolve.
You've estimated that we'll have AGI by 2030. There are interesting questions around that. How will we actually know that we got there? Maybe the move 37 of AGI. My estimate is sort of 50% chance in the next 5 years, so by 2030. I think there's a good chance that could happen. Part of it is what is your definition of AGI? People are arguing about that now. Mine's quite a high bar and always has been: can we match the cognitive functions that the brain has? We know our brains are pretty much general Turing machines, approximate. We've created incredible modern civilization with our minds, which speaks to how general the brain is. For us to know we have a true AGI, we would have to make sure it has all those capabilities. It isn't a jagged intelligence where some things it's really good at, like today's systems, but other things it's really flawed at. That's what we currently have with today's systems; they're not consistent. You'd want that consistency of intelligence across the board. Then we have some missing capabilities, like true invention capabilities and creativity that we were talking about earlier. You'd want to see those. How you test that? I think you just test it. One way would be a brute force test of tens of thousands of cognitive tasks that we know humans can do. Maybe also make the system available to a few hundred of the world's top experts, the towering experts in each subject area, and see if they can find an obvious flaw in the system. If they can't, then I think you can be pretty confident we have a fully general system.
To push back a little bit, it seems like humans are really incredible at taking intelligence for granted as it improves across all domains. You mentioned Terence Tao. These brilliant experts might quickly, in a span of weeks, take for granted all the incredible things it can do and then focus on the flaws. I consider myself first of all human. I identify as human. Some people listen to me talk and think that guy is not good at talking, the stuttering. Even humans have obvious limits across domains, even just outside of mathematics and physics. I wonder if it will take something like a move 37 on the positive side versus a barrage of 10,000 cognitive tasks where it would be one or two where it's like, yes, this is it.
I think exactly. There's the blanket testing to make sure you've got the consistency, but there are also lighthouse moments like the move 37 that I would be looking for. One would be inventing a new conjecture or new hypothesis about physics, like Einstein did. You could run a back test of that very rigorously: have a cutoff of knowledge of 1900, give the system everything written up to 1900, and see if it could come up with special relativity and general relativity like Einstein did. That would be an interesting test. Another would be: can it invent a game like Go? Not just come up with move 37, a new strategy, but can it invent a game that's as deep, aesthetically beautiful, and elegant as Go? Those are the sorts of things I would be looking out for. Probably a system being able to do several of those things for it to be very general, not just one domain. I think that would be the signs that we've got a system at AGI level. Then to fill that out, you would also check the consistency, make sure there are no holes in that system.
Something like a new conjecture or scientific discovery. That would be a cool feeling.
That would be amazing. Not just helping us do that, but actually coming up with something brand new. You would be in the room for that. It would be probably 2 or 3 months before announcing it. You would just be sitting there trying not to tweet something like, 'What is this amazing new physics idea?' Then we would check it with world experts in that domain, validate it, go through its workings. It would be explaining its workings too. It would be an amazing moment.
Do you worry that we as humans, even expert humans like you, might miss it? It might be pretty complicated.
I don't think it will be totally mysterious to the best human scientists. It may be a bit like in chess: if I talk to Garry Kasparov or Magnus Carlsen and they make a brilliant move, I might not be able to come up with that move, but they could explain afterwards why that move made sense. We would understand it to some degree, not to the level they do, but if they were good at explaining, which is actually part of intelligence too, being able to explain in a simple way what you're thinking about. I think that would be very possible for the best human scientists.
I wonder if there are moves in Go where even experts like Lee Sedol or Ke Jie at first dismiss it as a bad move, but then afterwards figure out with their intuition why it works. Empirically, the nice thing about games is you can test it: does it make you win the game or not? That tells you that move was good, that strategy was good. Then you can go back and analyze it and explain to yourself a little bit more why. That's how chess analysis works. Perhaps that's why my brain works like that, because I've been doing that since I was four. Even now when I generate code, there's a nuanced contention happening where I might at first identify a set of generated code as incorrect in some nuanced way, but then I always have to ask: is there a deeper insight here that I'm the one who's incorrect? As the systems get more intelligent, you're going to have to contend with that. Is this a bug or a feature of what you just came up with?
Yeah, and they're going to be pretty complicated. But you can imagine AI systems producing that code, and human programmers looking at it, but also aided with the help of AI tools. So it's going to be an interesting situation where different AI tools, maybe monitoring tools, are used to check the ones that generated it.
AlphaDev enables on the programming side something like recursive self-improvement. What does that AGI system, maybe not the first version but a few versions beyond, actually look like? Do you think it would be simple, like a self-improving program?
