Demis Hassabis：AI 的未来、模拟现实、物理学与电子游戏｜Lex Fridman Podcast #475

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 by what classical learning systems might be able to do about even fluids.

Yes, exactly. Fluid dynamics—the Navier–Stokes equations—these are traditionally thought to be very, very difficult, intractable problems to solve on classical systems. They take enormous amounts of compute. Weather prediction systems, these kinds 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 and specular lighting. I love the ones where people have generated videos where there are 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. I know it's so painstakingly hard to build programs that can do that. And yet somehow these systems are reverse-engineering this 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: “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 or 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 that, if you take a step back and look at all the work that we've done, especially with the Alpha projects—I'm thinking AlphaGo, of course, AlphaFold—what they really are is that 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 enumerated all the possibilities, there wouldn't be enough time in the universe.

So you have to do something much smarter, and what we did in both cases was build models of those environments. 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. Somehow, physics solves this problem that we've now also solved computationally.

I think the reason that's possible is that, in nature, natural systems have structure because they were subject to evolutionary processes that shaped 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, crudely stated, anything that can be evolved can be efficiently modeled.

I sometimes call it “survival of the stablest” or something like that, because, of course, there's evolution for living things, but there's also—if you think about geological time—the shape of mountains. That's been shaped by weathering processes over thousands of years. But then you can even take it cosmologically: the orbits of planets, the shapes of asteroids. These have all survived processes that have acted on them many, many times.

If that's true, then there should be some sort of pattern that you can reverse-learn, and a kind of manifold, really, that helps you search for the right solution, the right shape, and actually allows you to predict things about it in an efficient way, because it's not a random pattern. It may not be possible for man-made things or abstract things like factorizing large numbers, because unless there are patterns in the number space—which there might be—but if there aren't and it's uniform, then there's no pattern to learn, no model to learn that will help you search.

So you have to do brute force. In that case, you maybe need a quantum computer, 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. Yeah.

So interesting.

So they can be efficiently rediscovered or recovered, because nature is not random, right? Everything that we see around us, including the elements that are more stable, all of those things, they're subject to some kind of selection pressure.

Do you think, because you're also a fan of theoretical computer science and complexity, that 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, Demis, a new class of systems that could actually be 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=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: 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? 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=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 I think the answer to that is going to be very enlightening.

More specifically, with respect to the P=NP question, some of the stuff we're saying is kind of crazy right now. Just like the Christian Anfinsen Nobel Prize speech—the controversial thing that he said sounded crazy—and then you went and got a Nobel Prize for this with John Jumper and solved the problem.

So let me just stick to P=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 are, 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.

I think it's one of the most interesting questions there is: how far can that paradigm go? I think we've proven, and the AI community in general has proven, 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 or 20 years ago that was decades away, or maybe you would need some sort of quantum machines, 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 neural-network system on top of a classical computer, would be the ultimate expression of that. I think the limit—the bounds—of what that kind of system can do is a very interesting question, and it directly speaks to the P=NP question.

What do you think, again, hypothetically, might be outside of this? Maybe emergent phenomena—if you look at cellular automata, some of them have extremely simple systems, and then some complexity emerges. Maybe that would be outside, or do 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. Most emergent systems, cellular automata, and things like that could be modeled by a classical system. You just sort of do a forward simulation of them, and it would 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 the open questions.

But 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 3—video generation, rendering physics and lighting, 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. In a way, that's what I want to build AGI for: to help us as scientists answer these questions, like P = NP.

Yeah, I think we might be continuously surprised about what is modelable by classical computers. AlphaFold 3, on the interaction side, is surprising—that you can make any kind of progress in that direction. AlphaGenome is surprising: you can map the genetic code to function, kind of playing with emergent phenomena. You think there are so many combinatorial options, and then here you go—you can find the kernel that is efficiently modeled.

Yes. Because there's some structure, some landscape—in the energy landscape, or whatever it is—that you can follow, some gradient you can follow. Of course, what neural networks are very good at is following gradients. 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 may have naively thought about those problems for decades.

If you just enumerate all the possibilities, it looks totally intractable, and there are many, many problems like that. Then you think, well, there are 10^300 possible protein structures; there are 10^170 possible Go positions. All of these are way more than the number of 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. Of course, reality—nature—does do it, right? Proteins do fold. That gives you confidence that there must be—if we understood how physics was doing that, in a sense, and we could mimic or model that process—it should be possible on our classical systems. That's basically what the conjecture is about.

And of course, there are nonlinear dynamical systems—highly nonlinear dynamical systems—everything involving fluids.

Yes.

Right.

You know, I recently had a conversation with Terence Tao, who mathematically contends with a very difficult aspect of systems that have some singularities in them that break the mathematics. 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 by what classical learning systems might be able to do about even fluids.

Yes, exactly. Fluid dynamics and the Navier–Stokes equations are traditionally thought of as very, very difficult, intractable kinds of problems to do on classical systems. They take enormous amounts of compute. Weather prediction systems and these kinds of things all involve fluid dynamics calculations.

But again, if you look at something like Veo 3, our video generation model, it can model liquids quite well—surprisingly well—as well as materials and specular lighting. I love the videos where people have generated 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 so painstakingly hard to build programs that can do that. Yet somehow these systems are reverse-engineering it from just watching YouTube videos.

So presumably, what's happening is that it's extracting some underlying structure around how these materials behave. 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 3. I think a lot of people highlight different aspects, including the comedic stuff, the memes, and all that kind of stuff. 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 3. But exactly the thing you're mentioning is the physics.

Yeah, it's not perfect, but it's pretty damn good. The really interesting scientific question is: what is it understanding about our world in order to be able to do that? The cynical take with diffusion models is that 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 3 understands our world?

I think, to the extent that it can predict the next frames in a coherent way, that is some 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 8 seconds of consistent video that, by eye, at least at a glance, is quite hard to distinguish in terms of what the issues are.

Imagine that in 2 or 3 more years' time. That's the thing I'm thinking about, and how incredible that will look, given where we've come from—the early versions of that 1 or 2 years ago. The rate of progress is incredible. Like you, I think a lot of people love all of the stand-up comedians and the memes that capture a lot of human dynamics very well, including body language. But the thing I'm most impressed with and fascinated by is the physics behavior, the lighting, the materials, and the liquids. It's pretty amazing that it can do that.

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, 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 3 is directly challenging that, it feels like.

Yes, and it's very interesting. Even if you were to ask me 5 or 10 years ago, I would have said—even though I was immersed in all of this—that you probably need to understand intuitive physics. If I push this glass off the table, it may shatter, and the liquid will spill out. We know all of these things, but I thought that—and there were a lot of theories in neuroscience, called action in perception—you need to act in the world to truly perceive it in a deep way.

There were a lot of theories that 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 like you can understand it through passive observation, which is pretty surprising to me. Again, I think that hints at something underlying about the nature of reality, in my opinion, beyond just the cool videos that it generates.

Of course, the next stage is maybe even making those videos interactive, so one can actually step into them and move around in them. That would be really mind-blowing, especially given my games background. So you can imagine that, and then I think we're starting to get toward 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. Of course, that's what you would need for a true AGI system.

I have to talk to you about video games. You were being a bit trolly. I think you're having more and more fun on Twitter, on X, which is great to see. A guy named Jimmy Apples tweeted, “Let me play a video game of my Veo 3 videos already. Google cooked so good—playable world models, when?” And then you quote-tweeted that with, “Now, wouldn't that be something?”

How hard is it to build game worlds with AI? Maybe can you look out into the future of video games, 5 or 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. They were the first major AI systems that I built, and I always wanted to scratch that itch one day and come back to it.

Though I will, I think. I dream about what I would have done back in the ’90s if I’d had access to the kind of AI systems we have today. I think you could build absolutely mind-blowing games.

I think the next stage is that I always used to love making games. All the games I’ve made are open-world games, so they’re games where there’s a simulation, then there are AI characters, and the player interacts with that simulation, and the simulation adapts to the way the player plays.

I always thought they were the coolest games because, in games like Theme Park, which I worked on, everybody’s game experience would be unique to them. You’re co-creating the game. We set up the parameters and the initial conditions, and then you, as the player, are immersed in it and co-create 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. It was always quite difficult to build things like cellular automata—those kinds of classical systems that created some emergent behavior—but they were always a little bit fragile and limited.

Now we’re maybe on the cusp, in the next 5 to 10 years, of having AI systems that can truly create around your imagination. They can dynamically change the story and tell the narrative around you, making it dramatic no matter what you end up choosing. It’s like the ultimate choose-your-own-adventure game.

I think maybe we’re within reach if you think of an interactive version of Veo and then wind that forward 5 to 10 years. Imagine how good it’s going to be.

Yeah. You said a lot of super interesting stuff there. 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. Some games try to do that, to give you choice, but it’s really just an illusion of choice.

Like The Stanley Parable, a game I recently played: there are really just a couple of doors, and it takes you down a narrative. The Stanley Parable is a great video game; I recommend people play it. It kind of, in a meta way, mocks the illusion of choice, and there are philosophical notions of free will and so on.

One of my favorite games in The Elder Scrolls is Daggerfall. I believe they really played with the random generation of the dungeons.

Yeah.

You can step in, and they give you this feeling of an open world. You mentioned interactivity, but you don’t need to interact that much. When you open the door, whatever you see is randomly generated for you.

Yeah, and that’s already an incredible experience because you might be the only person to ever see that.

Yeah, exactly. What you’d like is something a little bit better than just random generation, right? Also, something better than a simple A/B hardcoded choice. That’s not really open-world, right? 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. 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. It’s expensive enough already, the way AAA games are made today. 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 in its early stages that probably still had the best learning AI in it. It was an early reinforcement learning system. You were looking after this mythical creature, growing it, and nurturing it, and depending on how you treated it, it would treat the villagers in that world in the same way. If you were mean to it, it would be mean. If you were good, it would be protective.

It was really a reflection of the way you played it. I’ve been working on 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 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. One of the sad aspects of your incredible success in so many domains of science—serious adult stuff—

Yeah.

—is that you might not have time to really create a game. You might end up creating the tooling that others would use to create the game. You have to watch others create the thing you’ve always dreamed of.

Do you think it’s possible that 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?

There are 2 ways to think about that. Maybe with vibe coding, as it gets better, there’s a possibility that one could do that in one’s spare time. 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.

The other thing is maybe a sabbatical after AGI has been safely stewarded into the world and delivered into the world. That, and then working on my physics theory, as we talked about at the beginning. Those would be my 2 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 = NP—or creating a cool video game.

Yeah. Well, they might be related, but in my world they would be related because it would be an open-world simulated game, as realistic as possible. What is the universe? That’s speaking to the same question, right? P = NP. I think all these things are related, at least in my mind.

In a really serious way, video games are sometimes 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 are 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. That’s what human life is. In video games, you can create more sophisticated, more diverse ways of living.

I think so.

I mean, those of us who love games—and I still do—it’s almost like you 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.

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 had never existed before, especially with these open-world games and simulation games where you, as the player, were co-creating the story.

There’s no other entertainment medium where you, as the audience, actually co-create the story. Now, with multiplayer games as well, it can be a very social activity, and you can explore all kinds of interesting worlds in that.

On the other hand, it’s very important to also enjoy and experience the physical world. But the question is then: What is the fundamental nature of reality? What is going to be the difference between these increasingly realistic simulations, multiplayer and emergent worlds, 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 and rigorously, answer the question: Why?

Yeah, 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?

Yes. 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: the simulation, the nature of reality, and what are 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?

My favorite one of all time is Civilization. I have to say that Civilization 1 and Civilization 2 are my favorite games of all time.

I can only assume you’ve avoided the most recent one because if you played it, that would be your sabbatical—you would disappear.

Yes, exactly. These Civilization games take a lot of time, so I have 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. We both love games, and it's interesting that he wrote games as well to start off with. It's probably especially in the era I grew up in, when home computers had just become 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 where I learned all my programming. Of course, it's a very fun thing to program: games.

I think it's a great way to learn programming, and probably still is. I immediately took it in the directions of AI and simulations, so I was able to express my interest in games and my wider scientific interests altogether.

The final thing I think that's great about games is that they fuse artistic design—art—with the most cutting-edge programming. Again, in the '90s, all of the most interesting technical advances were happening in gaming, whether that was AI, graphics, physics engines, or hardware. Even GPUs, of course, were designed for gaming originally.

Everything that was pushing computing forward in the '90s was due to gaming. 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 medium of storytelling. I love that. For me, this sort of multidisciplinary effort is 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: 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 superintelligence systems? For people who don't know, it's kind of—I don't know if it's fair to say—LLM-guided evolution search.

Yeah.

So evolutionary algorithms are doing the search, and LLMs are telling you where to search?

Yes, exactly. LLMs are proposing some possible solutions, and then you use evolutionary computing on top to find some novel part of the search space.

Actually, I think it's an example of very promising directions where you combine LLMs or foundation models with other computational techniques. Evolutionary methods are one, but you could also imagine Monte Carlo tree search—basically, many types of search algorithms or reasoning algorithms—on top of, or using, the foundation models as a basis.

I think there's quite a lot of interesting things to be discovered with these sorts of hybrid systems, let's call them.

But not to romanticize evolution. I'm only human. Do you think there's some value in whatever that mechanism is? We already talked about natural systems. Do you think there's a lot of low-hanging fruit in understanding and being able to model and simulate evolution, and then using whatever we understand about that nature-inspired mechanism to search better and better?

Yes. If you think about breaking down the systems we've built to their really fundamental core, you've got the model of the underlying dynamics of the system. 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.

You can do that in a number of ways. Evolutionary computing is one. With AlphaGo, we used Monte Carlo tree search, and that's what found Move 37—the new, never-before-seen strategy in Go. That's how you can go beyond potentially what is already known.

A 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 and discover something new? Obviously, this is super relevant for scientific discovery, or pushing math, science, and medicine forward, which we want to do with these systems.

You can bolt on some fairly simple search systems on top of these models and get into a new region of space. Of course, you also have to make sure that you're not searching that space totally randomly, because it would be too big. You have to have some objective function that you're trying to optimize and hill-climb toward, and that guides the search.

There are some mechanisms of evolution that are interesting, maybe in the space of programs. The space of programs is an extremely important space because you can probably generalize to everything. For example, mutation is not just Monte Carlo tree search, where it's simply a search.

You could, every once in a while, combine things.

Yeah.

Combine things, alter subcomponents of a thing.

Yes. What evolution is really good at is not just natural selection. It's combining things and building increasingly complex hierarchical systems.

That component is super interesting, especially with AlphaEvolve in the space of programs.

Yeah, exactly. You can get a bit of an extra property out of evolutionary systems, which is that some new emergent capability may come about. Of course, that happened with life.

Interestingly, with naive, traditional evolutionary computing methods, without LLMs and modern AI, the problem was that they could never work out how to evolve new properties—new emergent properties. You always had a subset of the properties that you put into the system. Maybe if we combine them with these foundation models, perhaps we can overcome that limitation.

Obviously, natural evolution clearly did, because it evolved new capabilities, right? From bacteria to where we are now. Clearly, it must be possible with evolutionary systems to generate new patterns, new capabilities, and emergent properties. 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. On my desk at home, most of my time is spent behind those computers just programming. Next to the three screens is a skull of a Tiktaalik, one of the early organisms that crawled out of the water onto land.

I just watch that little guy. Whatever the computational mechanism of evolution is, it's quite incredible. It's truly, truly incredible. Whether that's exactly the thing we need to do our search, never dismiss the power of nature and what it did here.

Yeah. It's amazing, and it's a relatively simple algorithm, effectively, and it can generate all of this immense complexity. Obviously, it's running over 4 billion years of time, but you can think about that as a search process that ran over the physics substrate of the universe for a long amount of computational time. Then it generated all this incredible, rich diversity.

So many questions I want to ask you. 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. For that purpose, on the AI scientist front, just broadly, there's an essay by 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 them is including a superhuman coder and a superhuman AI researcher. In that, there's the term “research taste,” which is really interesting. 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 figure out what the directions are where you want to generate truly novel ideas? 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 do you have the taste to sniff out what the right direction is, what the right experiment is, and what the right question is?

Picking the right question is the hardest part of science. Making the right hypothesis is also extremely difficult. Today's systems definitely can't do that.

I often say it's harder to come up with a really good conjecture than it is to solve it. We may have systems soon that can solve pretty hard conjectures. We solved Math Olympiad problems; with AlphaProof last year, our system got a silver medal in that—really hard problems.

Maybe eventually we'll be able to solve a Millennium Prize–type problem. But could a system have come up with a conjecture worthy of study that someone like Terence Tao would have said, “You know what? That's a really deep question about the nature of mathematics, or the nature of numbers, or the nature of physics”? That is a far harder type of creativity.

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.

Yes.

So, like, it's easy to come up with a thing that's extremely difficult.

Yeah.

It's easy to come up with a thing that's extremely easy. At that very edge—that sweet spot—you're basically advancing the science and splitting the hypothesis space into 2, ideally, right? Whether it's true or not true, you've learned something really useful, and that's hard. Making something that's also falsifiable and within the technologies that you currently have available—that's a very creative process, actually, a highly creative process. 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 2 is super interesting. I've heard you say that there's basically no failure, or that failure is extremely valuable if you construct the questions right, construct the experiments right, and design them right—that failure and success are both useful. So perhaps because it splits the hypothesis space basically in 2, 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. You can learn something equally valuable from an experiment that doesn't work. If you've designed the experiment well and your hypotheses are interesting, it should tell you a lot about where to go next. 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 lie ahead for us to make that happen? We should maybe highlight that AlphaFold—I mean, there are just so many leaps. AlphaFold solved, if it's fair to say, protein folding, and there are so many incredible things we could talk about there, including the open-sourcing of everything you've released. AlphaFold 3 is doing protein, RNA, and DNA interactions, which is super complicated and fascinating. That's amenable to modeling. AlphaGenome predicts how small genetic changes, like single mutations, link to actual function. Those are creeping along toward 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 have these really grand dreams, and then, as you've noticed, try to break them down. It's easy to have a 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?

Virtual Cell, which is what I call the project of modeling a cell, is an idea I've had 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—he founded it—and won the Nobel Prize in 2001. We've been talking about it since the '90s.

I used to come back every 5 years and ask: What would you need to model the full internals of a cell so that you could do experiments on the virtual cell, and those experiments in silico and those predictions would be useful enough to save you a lot of time in the wet lab? That would be the dream. Maybe you could speed up experiments 100-fold by doing most of the search in silico and then doing the validation step in the wet lab.

I've been trying to build these components, AlphaFold being one, that would eventually allow you to model all the interactions—a full simulation of a cell. I'd probably start with a yeast cell, partly because that's what Paul Nurse studied. A yeast cell is a full organism that's a single cell, so it's the simplest kind of single-cell organism. It's not just a cell; it's a full organism, and yeast is very well understood. That would be a good candidate for a fully simulated model.

AlphaFold is the solution to the static picture: What does a protein's 3D structure look like? But we know that all the interesting things in biology happen with the dynamics and the interactions, and that's what AlphaFold 3 is the first step toward modeling. First, pairwise interactions—proteins with proteins, and proteins with RNA and DNA. Then the next step after that would be modeling a whole pathway, maybe the mTOR pathway that's involved in cancer, or something like that. Eventually, you might be able to model a whole cell.

There's another complexity here: stuff in a cell happens at different time scales. Protein folding is superfast. I don't know all the biological mechanisms, but some of them take a long time. Is that another level of complexity? The levels of interaction have different temporal scales that you have to be able to model.

That would be hard. You'd probably need several simulated systems that can interact at these different temporal dynamics, or at least maybe it's a hierarchical system, so you can jump up and down the different temporal stages.

Can you avoid simulating, for example, the quantum-mechanical aspects of any of this? One of the challenges here is to avoid overmodeling. You want to skip ahead and just model the really high-level things that get you a good estimate of what's going to happen.

You've 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 at that then captures the dynamics that you're interested in? 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. Of course, that's where AlphaFold kicks in.

That would be the basis, and then you'd build these higher-level simulations that take those as building blocks. Then you get the emergent behavior.

Apologies for the naive questions ahead of time, but do you think we'll be able to simulate and model the origin of life—being able to simulate the first living organism from nonliving matter, 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 Ten Great Inventions of Evolution*. I think it's fantastic, and it also speaks to what the Great Filter might be: Is it prior to us, or is it ahead of us?

I think they're most likely in the past, if you read that book, because of how unlikely it was to have any life at all. Then going from single-cell to multicellular seems like an unbelievably big jump that took, I think, around a billion years on Earth. It shows you how hard it was.

Bacteria were super happy for a very long time.

A very long time before they captured mitochondria somehow, right? I don't see why AI couldn't help with that through some kind of simulation. Again, it's 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 are some initial conditions. Can you generate something that looks like a cell?

Perhaps that would be a next stage after the Virtual Cell project: How could something like that actually 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. Ultimately, I think what we'll figure out is that there's a continuum. There's no such thing as a line between nonliving and living. But if we can make that rigorous, yes.

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 around the origin of life—from nonliving to living—and see that it's not a line, but a continuum that connects physics, chemistry, and biology, then 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 kinds of questions. I don't really understand why the average person doesn't worry about this stuff more. How can we not have a good definition of life, of living and nonliving, and of the nature of time, let alone consciousness and gravity and all these things?

It's just—and quantum mechanics' weirdness. To me, it's always been screaming in my face the whole time, and it's getting louder. It's like, what is going on here? I mean that in the deepest sense, in 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 them with microscopes and take them apart almost down to the atomic level, and yet we still can't answer clearly, in a simple way, that question of how to define living.

Yeah, it's kind of amazing. Living, you can kind of talk your way out of thinking about, but consciousness—we have this very obviously subjective conscious experience. We're at the center of our own world, and it feels like something. Then how are you not screaming—

At the mystery of it all. I mean, really, humans have been contending with the mystery of the world around them for a long time. There are a lot of mysteries, like what's up with the sun and the rain? What's that about? Last year we had a lot of rain, and this year we don't have rain. What did we do wrong? Humans have been asking that question for a long time.

Exactly. So we're quite—I guess we've developed a lot of mechanisms to cope with these deep mysteries that we can't fully understand. We can see them, but we can't fully understand them, and we have to just get on with daily life. We keep ourselves busy, right, in a way. Do we keep ourselves distracted? I mean, weather is one of the most important questions in human history. That's still the go-to small-talk direction: the weather.

Especially in England.

And then there's the weather, which is famously an extremely difficult system to model, and even that system Google DeepMind has made progress on. Yes, we've created the best weather-prediction systems in the world with our WeatherNext system, and they're better than traditional fluid-dynamics systems, which are usually calculated on massive supercomputers and take days to calculate.

Again, it's interesting that those kinds of dynamics can be modeled even though they're very complicated, almost bordering on chaotic systems in some cases. A lot of the interesting aspects of that can be modeled by these neural-network systems, including, very recently, cyclone prediction—where the paths of hurricanes might go. Of course, that's super useful and super important for the world, and it's super important to do that very timely and very 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—I need to talk to them more—is that they're extremely tech-savvy. What they have to do is use models to predict where the storm is. They're crazy enough to go into the eye of the storm and—

In order to protect your life and predict where the extreme events are going to be, they have to have increasingly sophisticated models of weather.

Yeah. Yeah, it's a beautiful balance of being in it as living organisms and at the cutting edge of science. So they actually might be using a DeepMind system. So that's—

Yeah, they hopefully are, and I'd love to join them on one of those chases. They look amazing, right? To actually experience it one time.

Exactly. And then also to experience the correct prediction of where something will come and how it's going to evolve. It's incredible.

Yeah.

You've estimated that we'll have AGI by 2030, so there are interesting questions around that. How will we actually know that we got there? And what might be the Move 37 of AGI?

My estimate is sort of a 50% chance in the next 5 years, so by 2030, let's say. I think there's a good chance that could happen. Part of it is: What is your definition of AGI? Of course, 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 universal Turing machines, approximately. Of course, we've created incredible modern civilization with our minds, so that also speaks to how general the brain is. For us to know we have a true AGI, we would have to make sure that it has all those capabilities. It isn't kind of 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 do you test that? One way would be a brute-force test of tens of thousands of cognitive tasks that we know humans can do, and maybe also make the system available to a few hundred of the world's top experts—the towering figures of each subject area. Give them a month or two and see if they can find an obvious flaw in the system. If they can't, then I think you can be pretty confident that we have a fully general system.

Maybe to push back a little bit, it seems like humans are really incredible at taking intelligence for granted as it improves across all domains.

Mhm.

Like 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 in on the limitations. Well, right there—you know, I consider myself, first of all, human.

Yeah.

Second, I identify as human. Some people listen to me talk and they're like, “That guy is not good at talking—the stuttering.” So even humans have obvious limits across domains, even just outside of mathematics and physics and so on. 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, holy, this is—” I think exactly. So I think there's the sort of blanket testing to just make sure you've got the consistency, but I think there are the sort of lighthouse moments, like the Move 37, that I would be looking for.

One would be inventing a new conjecture or a new hypothesis about physics, like Einstein did. So maybe you could even run the back test of that very rigorously: Have a cutoff of knowledge at 1900, and then give the system everything that was 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 one 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, as aesthetically beautiful, and as elegant as Go? Those are the sorts of things I would be looking out for, and probably a system being able to do several of those things for it to be very general—not just one domain.

I think those would be the signs I would be looking for that we've got a system that's AGI-level. Then maybe to fill that out, you would also check the consistency, make sure there's no holes in that system either.

Yeah. Something like a new conjecture or scientific discovery—that would be a cool feeling. Yeah, that would be amazing. So it's not just helping us do that, but actually coming up with something brand new.

And you would be in the room for that. So it would probably be 2 or 3 months before announcing it.

Mhm.

And you would just be sitting there trying not to tweet—

Something like that. Exactly. It's like, “What is this amazing new physics idea?”

And then we would probably check it with world experts in that domain, validate it, and kind of go through its workings. I guess 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? Might miss it?

It may be pretty complicated. The analogy I give there is that I don't think it will be totally mysterious to the best human scientists, but it may be a bit like, for example, in chess. If I was to talk to Garry Kasparov or Magnus Carlsen and play a game with them, and they made a brilliant move, I might not be able to come up with that move, but they could explain afterward why that move made sense. We would understand it to some degree—not to the level they do, but to some degree—if they were good at explaining.

That's 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.

But I wonder—maybe you can educate me on Go. I wonder if there are moves for Magnus or Garry where, at first, they'll dismiss them as a bad move.

Yeah, sure. It could be. But afterward they'll figure out with their intuition why this works. Then, empirically, one of the great things about games is that you can use them as a sort of scientific test: Do you win the game or not? That tells you, okay, that move in the end was good. That strategy was good.

And then you can go back and analyze that and explain, even to yourself a little bit more, why—explore around it. That's how chess analysis and things like that work. So perhaps that's why my brain works like that, because I've been doing that since I was 4, and you're trained—it's sort of hardcore training in that way.

But even now, when I generate code, there is this kind of nuanced, fascinating contention happening where I might at first identify a set of generated code as incorrect in some interesting, nuanced ways. But then I always have to ask the question: Is there a deeper insight here—that I'm the one who's incorrect?

And that's something that, as the systems get more and more intelligent, you're going to have to contend with. It's like, what do you— is this a bug or a feature of what you just came up with?

Yeah. They're going to be pretty complicated to do. But of course, you can imagine AI systems that are producing that code, or whatever that is, and then human programmers looking at it, but also not unaided—they'll have the help of AI tools as well. So it's going to be kind of interesting—maybe different AI tools from the ones that generated it, more monitoring tools than the ones that generated it.

So if we look at an AGI system—sorry to bring it back up—but AlphaEvolve is super cool. AlphaEvolve enables, on the programming side, something like recursive self-improvement, potentially. What can you imagine that AGI system—maybe not the first version, but a few versions beyond that—actually looking like? What does that actually look like? Do you think it would be simple? Do you think it'll be something like a self-improving program, a simple one?

I mean, potentially, that's possible. I would say I'm not sure it's even desirable, because that's a kind of hard takeoff scenario. But these current systems, like AlphaEvolve, have a human in the loop deciding on various things. They're separate hybrid systems that interact. One could imagine eventually doing that end to end. I don't see why that wouldn't be possible, but right now I think the systems are not good enough to do that in terms of coming up with the architecture of the code.

And again, it's a little bit connected to this idea of coming up with a new conjectural hypothesis. They're good if you give them very specific instructions about what you're trying to do. But if you give them a very vague, high-level instruction, that wouldn't work currently. I think that's related to this idea of, “Invent a game as good as Go,” right? Imagine that was the prompt. That's pretty underspecified. And so the current systems wouldn't know, I think, what to do with that, how to narrow that down to something tractable.

And I think there's a similar—look, “Just make a better version of yourself” is too unconstrained. But we've done it, as you know, with AlphaEvolve, with things like faster matrix multiplication. When you hone it down to the very specific thing you want, it's very good at incrementally improving that. But at the moment, these are more like incremental improvements, small iterations, whereas if you wanted a big leap in understanding, you need a much larger advance.

Yeah. But it could also be—sort of to push back against the hard takeoff scenario—just a sequence of incremental improvements, like matrix multiplication. It has to sit there for days thinking about how to incrementally improve a thing, and it does so recursively. As you do more and more improvement, it'll slow down, so there'll be a path to AGI. It won't be like a hard takeoff; it'll be a gradual improvement over time.

Yes. If it was just incremental improvements, that's how it would look. So the question is: Could it come up with a new leap, like the Transformer architecture? Could it have done that back in 2017, when we did it at Google Brain? It's not clear that these systems—something like AlphaEvolve—wouldn't be able to make such a big leap. For sure, these systems are good. We have systems, I think, that can do incremental hill climbing. And that's a kind of bigger question: Is that all that's needed from here, or do we actually need 1 or 2 more big breakthroughs?

And can the same kind of systems provide the breakthroughs as well? So make it a bunch of S-curves: incremental improvement, but also every once in a while, leaps.

Yeah. I don't think anyone has systems that have shown unequivocally those big leaps, right? We have a lot of systems that do the hill climbing of the S-curve that you're currently on.

Yeah. And Move 37 would be a leap.

Yeah. I think it would be a leap.

Do you think the scaling laws are holding strong on pre-training, post-training, and test-time compute? On the flip side of that, do you anticipate AI progress hitting a wall?

We certainly feel there's a lot more room just in scaling, actually, across all steps: pre-training, post-training, and inference time. There are sort of 3 scalings happening concurrently. And again, it's about how innovative you can be. We pride ourselves on having the broadest and deepest research bench. We have amazing, incredible researchers and people like Noam Shazeer, who came up with Transformers, and David Silver, who led the AlphaGo project, and so on.

That research base means that if some new breakthrough is required, like AlphaGo or Transformers, I would back us to be the place that does that. So I'm actually quite like it when the terrain gets harder, right? Because then it veers more from just engineering to true research, or research plus engineering, and that's our sweet spot. I think that's harder—it's harder to invent things than to fast-follow.

So we don't know. I would say it's kind of 50/50 whether new things are needed or whether scaling the existing stuff is going to be enough. And so, in true empirical fashion, we're pushing both of those as hard as possible: the new blue-sky ideas—and maybe about half our resources are on that—and scaling to the max the current capabilities. We're still seeing some fantastic progress on each different version of Gemini.

That's interesting, the way you put it, in terms of the deep bench: If progress toward AGI is more than just scaling compute—the engineering side of the problem—and is more on the scientific side, where breakthroughs are needed, then you feel confident that Google DeepMind is well positioned to kick ass in that domain?

Well, I mean, if you look at the history of the last decade or 15 years, it's been—I don't know—maybe 80% or 90% of the breakthroughs that underpin the modern AI field today were from, originally, Google Brain, Google Research, and DeepMind. So, yeah, I would back that to continue, hopefully.

On the data side, are you concerned about running out of high-quality data, especially high-quality human data?

I'm not very worried about that, partly because I think there's enough data, and it's been proven to get the systems to be pretty good. This goes back to simulations again: If you have enough data to make simulations, you can create more synthetic data that's from the right distribution. Obviously, that's the key. You need enough real-world data in order to be able to create those kinds of generative data generators, and I think we're at that step at the moment.

Yeah, you've done a lot of incredible stuff on the side of science and biology, doing a lot with not so much data.

Yeah.

I mean, it's still a lot of data, but I guess it's enough to kick that off.

Exactly, yeah.

How crucial is the scaling of compute to building AGI? This is an engineering question. It's almost a geopolitical question, because integrated into that are supply chains and energy, a thing that you care a lot about—potentially fusion. So, innovating on the side of energy as well, do you think we're going to keep scaling compute?

I think so, for several reasons. I think compute—there's the amount of compute you have for training. Often it needs to be colocated, so even bandwidth constraints between data centers can affect that. There are additional constraints even there, and that's important for training. Obviously, you want to train the largest models you can, but there's also the fact that AI systems are now in products and being used by billions of people around the world, so you need a ton of inference compute.

And then, on top of that, there's the thinking models—the new paradigm of the last year—where they get smarter the longer the inference time you give them at test time. So all of those things need a lot of compute, and I don't really see that slowing down. As AI systems become better, they'll become more useful, and there'll be more demand for them.

So both from the training side—the training side is actually only just one part of that—it may even become the smaller part of what's needed in the overall compute that's required.

Yeah, that's one sort of almost memey kind of thing, which is the success and the incredible aspects of Veo 3. People kind of make fun of it: The more successful it becomes, you know, the more the servers are sweating.

Exactly. We did a little video of the servers frying eggs and things. That's right, and we're going to have to figure out how to do that. There's a lot of interesting hardware innovations that we do. As you know, we have our own TPU line, and we're looking at inference-only things—inference-only chips—and how we can make those more efficient.

We're also very interested in building AI systems that help with energy usage: data center energy, making the cooling systems efficient, grid optimization, and then eventually things like helping with plasma containment in fusion reactors. We've done lots of work on that with Commonwealth Fusion Systems, and one could imagine reactor design.

Then material design, I think, is one of the most exciting areas. New types of solar material, solar-panel material, room-temperature superconductors have always been on my list of dream breakthroughs, as well as optimal batteries. I think a solution to any one of those things would be absolutely revolutionary for climate and energy usage, and we're probably close, again in the next 5 years, to having AI systems that can materially help with those problems.

If you were to bet—sorry for the ridiculous question—but what is the main source of energy in 20, 30, 40 years? Do you think it's going to be nuclear fusion?

I think fusion and solar are the 2 that I would bet on. Solar—I mean, it's the fusion reactor in the sky, of course. I think the problem there is batteries and transmission. As well as more and more efficient solar material, perhaps eventually in space, these kind of Dyson sphere-type ideas, and fusion—I think fusion is definitely doable.

It seems that if we have the right design of reactor and can control the plasma fast enough and so on, both of those things will actually get solved. We'll probably have at least those 2 as the primary sources of renewable, clean, almost-free, or perhaps free, energy.

What a time to be alive. If I traveled into the future with you 100 years from now, how much would you be surprised if we'd passed a Type I Kardashev-scale civilization?

I would not be that surprised if there was a 100-year timescale from here. I think it's pretty clear: if we crack the energy problems in one of the ways we've just discussed—fusion or very efficient solar—then, if energy is kind of free and renewable and clean, that solves a whole bunch of other problems.

For example, the water-access problem goes away because you can just use desalination. We have the technology; it's just too expensive, so only fairly wealthy countries like Singapore and Israel actually use it. But if it was cheap, then all countries that have a coast could use it.

Also, you'd have unlimited rocket fuel. You could just separate seawater into hydrogen and oxygen using energy, and that's rocket fuel. Combined with Elon's amazing self-landing rockets, it could be like a bus service to space. So that opens up incredible new resources and domains.

Asteroid mining, I think, will become a thing, and maximizing human flourishing to the stars. That's what I dream about as well: Carl Sagan's idea of bringing consciousness to the universe, waking up the universe. I think human civilization will do that in the fullness of time if we get AI right and crack some of these problems with it.

I wonder what it would look like if you were just a tourist flying through space. You would probably notice Earth because, if you solved the energy problem, you'd see a lot of space rockets, probably. So it would be like traffic here in London.

But in space.

Just a lot of rockets.

And then you would probably see floating in space some kind of source of energy, like solar.

Yeah.

Potentially. So Earth would just look, on the surface, more technological, and then you'd use the power of that energy to preserve the natural—

Yes.

—like the rainforest and all that kind of stuff, because for the first time in human history we wouldn't be resource-constrained. I think that could be an amazing new era for humanity, where it's not zero-sum, right? I have this land; you don't have it. Or if the tigers have their forest, then the local villagers can't—what are they going to use? I think that this will help a lot.

No, it won't solve all problems because there are still other human foibles that will exist, but it will at least remove one of the big vectors, which is scarcity of resources, including land, more materials, and energy. I sometimes call it—and others call it—this kind of radical abundance era, where there's plenty of resources to go around. But, of course, the next big question is making sure that it's fairly shared and everyone in society benefits from that.

There's something about human nature where I go—it's like Borat: “My neighbor.” We do start conflicts. That's why games, throughout history—as I'm learning more and more, even in ancient history—serve the purpose of pushing people away from war, actually hot war.

Maybe we can figure out increasingly sophisticated video games that pull us, that scratch the itch of conflict, whatever that is about us, our human nature, and then avoid the actual hot wars that would come with increasingly sophisticated technologies. We're now long past the stage where the weapons we're able to create can actually just destroy all of human civilization, so that's no longer a great way to start with your neighbor. It's better to play a game of chess.

Or football. Yeah.

And I think that's what modern sport is. I love football, watching it, and I used to play it a lot as well. It's very visceral and tribal, and I think it channels a lot of those energies into a way which I think is a kind of human need: to belong to some group, but in a fun, healthy, non-destructive way—a constructive thing.

Going back to games again, I think one reason they're so great for kids to play, things like chess, is that they're great little microcosm simulations of the world. They're simulations of the world, too. They're simplified versions of some real-world situation, whether it's poker, Go, chess, or Diplomacy—different aspects of the real world.

They allow you to practice at them, too. How many times do you get to practice a massive decision moment in your life? What job to take, what university to go to—you get maybe a dozen or so key decisions that one has to make, and you've got to make those as best as you can. Games are a kind of safe, repeatable environment where you can get better at your decision-making process. They may have this additional benefit of channeling some energies into more creative and constructive pursuits.

Well, I think it's also really important to practice losing and winning, right?

Losing is really important. That's why I love games. That's why I love even things like Brazilian jiu-jitsu.

Yeah.

You can get your ass kicked in a safe environment over and over. It reminds you about physics, about the way the world works, about how sometimes you lose and sometimes you win. You can still be friends with everybody. But that feeling of losing is a weird one for us humans to really make sense of. It's just part of life. That is a fundamental part of life: losing.

Yeah. And I think in martial arts, as I understand it, but also in things like chess, at least the way I took it, it's a lot to do with self-improvement and self-knowledge. You know that, okay, I did this thing. It's not really about beating the other person; it's about maximizing your own potential.

If you do it in a healthy way, you learn to use victories and losses in a way. Don't get carried away with victory and think you're the best in the world. The losses keep you humble, always knowing there's something more to learn. There's always a bigger expert who can mentor you. I think you learn that in martial arts, and I think that's also the way that I was trained in chess.

In the same way, it can be very hardcore and very important, and of course you want to win, but you also need to learn how to deal with setbacks in a healthy way and wire that feeling you have when you lose something into a constructive thought: “Next time I'm going to improve this,” or “get better at this.” There is a source of happiness, a source of meaning, in that improvement step. It's not about the winning or losing.

Yes. The mastery. There's nothing more satisfying, in a way, than saying, “Oh, wow. This thing I couldn't do before, now I can.” Again, games and physical sports and mental sports are ways of measuring. They're beautiful because you can measure that progress.

Yeah. I mean, this is why I love role-playing games: the number going up on my skill tree. Literally, that is a source of meaning for us humans.

Yeah, we're quite addicted to these numbers going up, and maybe that's why we made games like that, because obviously that's something we're quite—We're hill-climbing systems ourselves, right?

Yeah, it would be quite sad if we didn't have any mechanism.

Color belts. We do this everywhere, right?

I don't want to dismiss that there is a source of deep meaning for us as humans.

One of the incredible stories on the business and leadership side is what Google has done over the past year. I think it's fair to say that Google was losing on the LLM product side a year ago with Gemini 1.5, and now it's winning with Gemini 2.5. You took the helm and led this effort. What did it take to go from, quote-unquote, losing to quote-unquote, winning in the span of a year?

Yeah. Well, firstly, it's an absolutely incredible team that we have, led by Koray, Jeff Dean, and Oriol, and the amazing team we have on Gemini. They're absolutely world-class. You can't do it without the best talent, and of course, we have a lot of great compute as well.

But then it's the research culture we've created. Basically, it's about bringing together the different groups in Google. There was the Google Brain world-class team and then the old DeepMind, pulling together all the best people and the best ideas and gathering around to make the absolute greatest system we could. It's hard, but we're all very competitive, and we love research. This is so fun to do.

It's great to see our trajectory. It wasn't a given, but we're very pleased with where we are, and the rate of progress is the most important thing. If you look at where we've come from 2 years ago to 1 year ago to now, I think what we call relentless progress, along with relentless shipping of that progress, is being very successful.

It's unbelievably competitive—the whole space, the whole AI space—with some of the greatest entrepreneurs, leaders, and companies in the world all competing now because everyone's realized how important AI is. It's been very pleasing for us to see that progress.

Google's a gigantic company. Can you speak to one of the natural things that happens in that case, which is the bureaucracy that emerges? You want to be careful, but there are meetings and managers. What are some of the challenges, from a leadership perspective, in breaking through that in order to, as you said, ship? The number of Gemini-related products that have been shipped over the past year is just insane.

Right, it is. Exactly. That's what relentlessness looks like. I think it's a question of how any big company ends up having a lot of layers of management and things like that. That's sort of the nature of how it works.

But I still operate—and I was always operating with old DeepMind—as a startup. It's still a large one, but we still act like a startup today with Google DeepMind, acting with the decisiveness and energy that you get from the best smaller organizations. We try to get the best of both worlds, where we have these incredible surfaces with billions of users, incredible products that we can power up with our AI and our research. That's amazing.

There are very few places in the world where you can do incredible, world-class research on the one hand and then plug it in and improve billions of people's lives the next day. That's a pretty amazing combination.

We're continually fighting and cutting away bureaucracy to allow the research culture and the relentless shipping culture to flourish. I think we've got a pretty good balance, whilst being responsible with it, as you have to be as a large company, and also with the number of huge product surfaces that we have.

A funny thing you mentioned about the surface of the billions: I had a conversation with a brilliant guy here at the British Museum called Irving Finkel. He's a world expert at cuneiform, which is an ancient writing system on tablets. He doesn't know about ChatGPT or Gemini. He doesn't even know anything about AI.

But his first encounter with this AI—

Is AI Mode on Google. Yes.

He's like, "Is that what you're talking about? This AI Mode?" It's just a reminder that there's a large part of the world that doesn't know about this AI thing.

Yeah, I know. It's funny because if you live on X and Twitter—and I mean, at least my feed—it's all AI. There are certain places in the Valley and certain pockets where everyone's thinking about AI, but a lot of the normal world hasn't come across it yet.

But that's a great responsibility: their first interaction, on the grand scale of rural India or anywhere across the world. You get to—

Right, and we want it to be as good as possible. In a lot of cases, it's just under the hood, powering and making something like Maps or Search work better. Ideally, for a lot of those people, it should just be seamless. It's just new technology that makes their lives more productive and helps them.

A bunch of folks on the Gemini product and engineering teams have spoken extremely highly of you on another dimension that I almost didn't even expect, because I kind of think of you as the deep scientist who cares about these big research and scientific questions. But they also said you're a great product guy—how to create a thing that a lot of people would use and enjoy using. Can you maybe speak to what it takes to create an AI-based product that a lot of people would enjoy using?

Yeah. Well, that comes back to my game design days, where I used to design games for millions of gamers. People would forget about that. I've had experience with cutting-edge technology in products. That was how games were in the '90s.

I love the combination of cutting-edge research and then applying it in a product to power a new experience. I think it's the same skill of imagining what it would be like to use it viscerally and having good taste. Coming back to earlier, the same thing that's useful in science can also be useful in product design.

I've always been a sort of multidisciplinary person, so I don't really see the boundaries between arts and sciences or product and research. It's a continuum for me. I only like working on products that are cutting-edge. If they were just run-of-the-mill products, I wouldn't be excited about having cutting-edge technology under the hood.

It requires invention, creativity, and capability.

What are some specific things you've learned about when, even on the LLM side, you're interacting with Gemini and you're like, "This doesn't feel like the layout, the interface"? Maybe the trade-off between the latency—how to present it to the user, how long to wait, and how that waiting is shown—or the reasoning capabilities.

There are some interesting things because, like you said, it's the very cutting edge. We don't know how to present it, how to present it correctly. Are there some specific things you've learned?

I mean, it's such a fast-evolving space. We're evaluating this all the time, but where we are today is that you want to continually simplify things, whether that's the interface or all the interactions you build on top of the model. You kind of want to get out of the way of the model.

The model train is coming down the track, and it's improving unbelievably fast—this relentless progress we talked about earlier. You look at 2.5 versus 1.5, and it's just a gigantic improvement. We expect that again for future versions.

The models are becoming more capable. The interesting thing about the design space in today's world, these AI-first products, is that you've got to design not for what the technology can do today, but for what it will be able to do in a year's time. You actually have to be a very technical product person because you've got to have a good intuition and feel for, "Okay, that thing that I'm dreaming about now can't be done today, but is the research track on schedule to basically intercept that in 6 months or a year's time?"

You've got to anticipate where this highly changing technology is going. In addition, new capabilities are coming online all the time that you didn't realize were possible before, which can allow Deep Research to work. Now we've got video generation—what do we do with that?

With this multimodal stuff, one question I have is: Is it really going to be the current UI that we have today—these text-box chats? It seems very unlikely, given these super-multimodal systems. Shouldn't it be something more like Minority Report, where you're sort of vibing with it in a collaborative way? It seems very restricted today.

I think we'll look back on today's interfaces, products, and systems as quite archaic, maybe in just a couple of years. I think there's a lot of space for innovation to happen on the product side as well as the research side.

We were talking offline about this keyboard. The open question is how, when, and how much we'll move to audio as the primary way of interacting with the machines around us versus typing stuff.

Yeah, I mean, typing is a very low-bandwidth way of doing things, even if you're a very fast typer. I think we're going to have to start utilizing other devices, whether that's smart glasses, audio earbuds, and eventually maybe some sort of neural device, where we can increase the input and output bandwidth to something like 100× what it is today.

I think the underappreciated art form is interface design.

But I think you cannot unlock the power of the intelligence of a system if you don't have the right interface. The interface is really the way you unlock its power. It's such an interesting question of how to do that.

You would think getting out of the way is a real art form.

Yes. You know, it's the sort of thing that I guess Steve Jobs always talked about, right? It's simplicity, beauty, and elegance that we want, right? And we're not there. Nobody's there yet, in my opinion, and that's what I would like us to get to. Again, it sort of speaks to Go, right? As a game, it's the most elegant, beautiful game. Can you make an interface as beautiful as that?

Actually, I think we're going to enter an era of AI-generated interfaces that are probably personalized to you, so they fit your aesthetic, your feel, and the way that your brain works. The AI kind of generates that depending on the task. That feels like that's probably the direction we'll end up in.

Because some people are power users and they want every single parameter on screen, everything—perhaps me with keyboard-based navigation. I like to have shortcuts for everything. And some people like the minimalism.

Just hide all of that complexity.

Exactly. Yeah. Well, I'm glad you have a Steve Jobs mode in you as well. This is great: Einstein mode, Steve Jobs mode.

All right, let me try to trick you into answering a question. When will Gemini 3 come out? Is it before or after GTA 6? The world waits for both. And what does it take to go from 2.5 to 3.0? Because it seems like there have been a lot of releases of 2.5 that are already leaps in performance.

So what does it even mean to go to a new version? Is it about performance? Is it about a completely different flavor of an experience?

Yeah. Well, the way it works with our different version numbers is that we try to collect everything together. Maybe it takes roughly 6 months or something to do a new kind of full run and the full productization of a new version. During that time, lots of new, interesting research iterations and ideas come up, and we sort of collect them all together.

You could imagine the last 6 months' worth of interesting ideas on the architecture front. Maybe it's on the data front. It's many different possible things, and we package that all up, test which ones are likely to be useful for the next iteration, and then bundle that all together. Then we start the new giant hero training run, and of course that gets monitored.

At the end of the pretraining, there's all the post-training. There are many different ways of doing that, different ways of patching it, so there's a whole experimental phase there where you can also get a lot of gains. That's where you see the version numbers usually referring to the base model, the pretrained model. The interim versions of 2.5, the different sizes and little additions, are often patches or post-training ideas that can be done afterward off the same basic architecture.

On top of that, we also have different sizes—Pro, Flash, and Flash-Lite—that are often distilled from the biggest ones, the Flash model from the Pro model. That means we have a range of different choices if you're a developer: do you want to prioritize performance, speed, and cost?

We like to think of this Pareto frontier. On the one hand, the y-axis is performance, and the x-axis is cost or latency and speed, basically. We have models that completely define the frontier. Whatever trade-off you want as an individual user or as a developer, you should find one of our models that satisfies that constraint.

So behind the version changes, there is a big hero run. And then there's an insane complexity of productization. Then there's the distillation of the different sizes along that Pareto frontier. And then, with each step you take, you realize there might be a cool product. There are side quests.

Yes.

Exactly. But then you also don't want to take too many side quests, because then you have a million versions of a million products and it's very unclear. But you also get super excited because it's super cool. How do you even look at Veo? How does it fit into the bigger thing?

Exactly. Exactly. And then you're constantly in this process of converging upstream, we call it—ideas from the product surfaces or from the post-training, and even further downstream than that, you kind of upstream that into the core model training for the next run. Then the main model, the main Gemini track, becomes more and more general, and eventually AGI.

One hero run at a time.

Yes. Exactly. A few hero runs later.

Yeah. So sometimes when you release these new versions—or every version, really—are benchmarks productive or counterproductive for showing the performance of a model?

You need them, but it's important that you don't overfit to them, right? They shouldn't be the be-all and end-all. There's LMArena, or it used to be called Chatbot Arena; that's one of them that turned out sort of organically to be one of the main ways people like to test these systems, at least the chatbots.

Obviously, there are loads of academic benchmarks that test mathematics and coding ability, general language ability, science ability, and so on. Then we have our own internal benchmarks that we care about.

It's a kind of multi-objective optimization problem, right? You don't want to be good at just one thing. We're trying to build general systems that are good across the board, and you try and make no-regret improvements, where you're improving coding but it doesn't reduce your performance in other areas, right?

That's the hard part, because you could put more coding data in, or you could put more gaming data in, but then does it make your language system or your translation systems and other things that you care about worse? You've got to continually monitor this increasingly larger suite of benchmarks.

Also, when you put these models into products, you care about direct usage and the direct stats and signals that you're getting from the end users, whether they're coders or the average person using the chat interfaces.

Yeah. Because ultimately you want to measure usefulness, but it's so hard to convert that into a number, right? It's really vibe-based benchmarks across a large number of users, and it's hard to know. I think it would be terrifying to me: you have a much smarter model, but it's just something vibe-based; it's not quite working. That's so scary, because everything you just said—it has to be smart and useful across so many domains.

You get super excited because it's suddenly solving programming problems you've never been able to solve before, but now it's crappy poetry or something, and it's just—I don't know, that's stressful. That's so difficult to balance, because you can't really trust the benchmarks. You really have to trust the end users.

Yeah. And then other things that are even more esoteric come into play, like the style, the persona of the system. Is it verbose? Is it succinct? Is it humorous? Different people like different things.

It's very interesting—it's almost like cutting-edge psychology research or personality research. I used to do that in my PhD: the Five-Factor Model. What do we actually want our assistants to be like? Different people will like different things as well.

These are all new problems in the product space that I don't think have ever really been tackled before, but we're going to rapidly have to deal with them now. I think it's a super fascinating space, developing the character of the thing, and in so doing, it puts a mirror to ourselves: what are the kinds of things that we like?

Prompt engineering allows you to control a lot of those elements, but can the product make it easier for you to control the different flavors of those experiences, the different characters that you interact with?

Yeah, exactly. So what's the probability of Google DeepMind winning?

Well, I don't see it as winning. I think we need to think—winning is the wrong way to look at it, given how important and consequential what it is we're building.

Funnily enough, I don't try to view it like a game or competition, even though that's a lot of my mindset. It's about, in my view, all of us—those of us at the leading edge—having a responsibility to steward this unbelievable technology that could be used for incredible good but also has risks, and steward it safely into the world for the benefit of humanity.

That's always what I've dreamed about, and what we've always tried to do. I hope that's what eventually the community, maybe the international community, will rally around when it becomes obvious that, as we get closer and closer to AGI, that's what's needed.

I agree with you. I think that's beautifully put. You've said that you talk to and are on good terms with the leads of some of these labs as the competition heats up. How hard is it to maintain those relationships?

It's been okay so far.

I try to pride myself on being collaborative. Research is a collaborative endeavor. Science is a collaborative endeavor. Right? It's all good for humanity in the end if you cure incredible, terrible diseases and come up with an incredible cure. This is a net win for humanity, and the same with energy.

All of the things that I'm interested in helping solve with AI, I just want that technology to exist in the world and be used for the right things, with the productivity benefits being shared for the benefit of everyone. So, I try to maintain good relations with all the leading lab people. They are very interesting characters, many of them, as you might expect.

I'm on good terms, I hope, with pretty much all of them. I think that's going to be important when things get even more serious than they are now: that there are those communication channels. That's what will facilitate cooperation or collaboration if that's what's required, especially on things like safety.

Yeah, I hope there's some collaboration on stuff that's less high-stakes and, in so doing, serves as a mechanism for maintaining friendships and relationships. For example, I think the internet would love it if you and Elon somehow collaborated on creating a video game. That kind of thing, I think, enables camaraderie and keeps you two on good terms. Also, you two are legit gamers, so it's just fun to create.

Yeah, that would be awesome. We've talked about that in the past, and it may be a cool thing that we can do. I agree with you. It'd be nice to have side projects in a way where one can just lean into the collaboration aspect of it, and it's a win-win for both sides. It kind of builds up that collaborative muscle.

I see the scientific endeavor as that kind of side project for humanity, and I think Google DeepMind has been really pushing that. I would love to see other labs do more scientific stuff and then collaborate, because it just seems like it's easier to collaborate on the big scientific questions.

I agree, and I would love to see a lot of the other labs talk about science. But I think we're really the only ones using AI for science and doing that, and that's why projects like AlphaFold are so important to me and, I think, to our mission: to show how AI can be clearly used in a very concrete way for the benefit of humanity.

We also spun out companies like Isomorphic Labs off the back of AlphaFold to do drug discovery, and it's going really well. You can think of building additional AlphaFold-type systems to go into chemistry and help accelerate drug design. Those are the examples I think we need to show, and society needs to understand what AI can bring: these huge benefits.

Well, from the bottom of my heart, thank you for pushing the scientific efforts forward with rigor, with fun, with humility—all of it. I just love to see that you're still talking about P = NP. I mean, it's just incredible. I love it.

There has been seemingly a war for talent. Some of it is a meme; I don't know. What do you think about Meta buying up talent with huge salaries and the heating up of this battle for talent? I should say that I think a lot of people see DeepMind as a really great place to do cutting-edge work for the reasons that you've outlined. There's this vibrant scientific culture.

Yeah. Well, of course, there's a strategy that Meta is taking right now. From my perspective, at least, I think the people who are real believers in the mission of AGI and what it can do, and who understand the real consequences—both good and bad—from that and what that responsibility entails, are mostly doing it to be, like myself, on the frontier of that research. They can help influence the way that goes and steward that technology safely into the world.

Meta right now is not at the frontier. Maybe they'll manage to get back on there. It's probably rational what they're doing from their perspective because they're behind and they need to do something, but I think there are more important things than just money.

Of course, one has to pay people their market rates, and all of these things continue to go up. I was expecting this because more and more people, including leaders of companies, are finally realizing what I've always known for 30-plus years now: AGI is probably the most important technology that's ever going to be invented. So, in some senses, it's rational to be doing that.

But I also think there's a much bigger question. People in AI these days are very well paid. I remember when we were starting out back in 2010, I didn't even pay myself for a couple of years because there wasn't enough money. We couldn't raise any money.

These days, interns are being paid the amount that we raised as our entire first seed round, so it's pretty funny. I remember the days when I used to have to work for free and almost pay my own way to do an internship. Right now, it's all the other way around, but that's just how it is. It's the new world.

We've been discussing what happens post-AGI, when energy systems are solved and so on. What is money even going to mean? What is the economy going to mean? We're going to have much bigger issues to work through: How does the economy function in that world? So, I think salaries are a little bit of a side issue today.

Yeah, when you're facing such gigantic consequences and gigantic, fascinating scientific questions, which may be only a few years away, on the practical, pragmatic side, if we zoom in on jobs, we can look at programmers. It seems like AI systems are currently doing incredibly well in programming and increasingly so.

A lot of people who program for a living and love programming are worried they will lose their jobs. How worried should they be, do you think? What's the right way to adjust to the new reality and ensure that you survive and thrive as a human in the programming world?

Well, it's interesting that programming—and, again, this is counterintuitive to what we thought years ago—some of the skills that we think of as harder skills have turned out to maybe be the easier ones for various reasons. Coding and math are easier because you can create a lot of synthetic data and verify whether that data is correct. Because of that nature, it's easier to make things like synthetic data to train from.

It's also an area, of course, we're all interested in because, as programmers, we want it to help us get faster at it and more productive. So, for the next era—the next 5–10 years—I think what we're going to find is that people who embrace these technologies become almost at one with them. Whether that's in the creative industries or the technical industries, they will become superhumanly productive, I think.

The great programmers will be even better. They'll be 10 times what they are today because they'll be able to use their skills to utilize the tools to the maximum and exploit them to the maximum. I think that's what we're going to see in the next decade. That's going to cause quite a lot of change, and a lot of people will benefit from that.

One example of that is if coding becomes easier, it becomes available to many more creatives to do more. But I think the top programmers will still have huge advantages in terms of specifying—going back to specifying what the architecture should be—the questions they should ask, how to guide these coding assistants in a way that's useful, and checking whether the code they produce is good. So, I think there's plenty of headroom there for the foreseeable future, over the next few years.

So, I think there are several interesting things there. One is that there's a lot of incentive to just get better and better at consistently using these tools. They're riding the wave of the improving models.

Yes.

Versus competing against them. But sadly, that's the nature of life on Earth. There could be a huge amount of value to certain kinds of programming at the cutting edge and less value to other kinds.

For example, frontend web design might be more amenable, as you mentioned, to generation by AI systems. Maybe game engine design, or something like this, or backend design, or guiding systems in high-performance situations—high-performance programming-type design decisions—might be extremely valuable. It will shift where the humans are needed most, and that's scary for people to adjust.

I think that's right. Anytime there's a lot of disruption and change—and we've had this, not just this time, but many times in human history, with the internet, mobile, and, before that, obviously, the Industrial Revolution—it's going to be one of those eras where there will be a lot of change.

I think there will be new jobs we can't even imagine today, just like the internet created new jobs. Those people with the right skill sets to ride that wave will become incredibly valuable. People may have to relearn or adapt their current skills a bit.

The thing that's going to be harder to deal with this time around is that I think what we're going to see is probably 10 times the impact the Industrial Revolution had, but 10 times faster as well.

Right? So instead of 100 years, it takes 10 years. And so that's going to make it like 100x the impact and the speed combined. So that's what I think is going to make it more difficult for society to deal with. There's a lot to think through, and I think we need to be discussing that right now.

I encourage top economists in the world and philosophers to start thinking about how society is going to be affected by this and what we should do, including things like universal basic provision or something like that, where a lot of the increased productivity gets shared out and distributed to society, maybe in the form of services and other things. If you want more than that, you still go and get some incredibly rare skills and things like that and make yourself unique. But there's a basic provision that's provided.

And if you think of government as a technology, there are also interesting questions—not just in economics, but in politics. How do you design a system that's responding to the rapidly changing times such that you can represent the different pain that people feel from different groups? How do you reallocate resources in a way that addresses that pain and represents the hopes, pain, and fears of different people in a way that doesn't lead to division?

Politicians are often really good at fueling division and using that to get elected—othering the other and then saying that's bad based on that. I think that's often counterproductive to leveraging rapidly changing technology to help the world flourish. So we almost need to improve our political systems rapidly as well, if you think of them as a technology.

Definitely. I think we'll need new governance structures and institutions, probably to help with this transition. Political philosophy and political science are going to be key to that. But I think the number-one thing, first of all, is to create more abundance of resources, right?

That's the number-one thing: increase productivity, get more resources, and maybe eventually get out of the zero-sum situation. Then the second question is how to use those resources and distribute them, but you can't do that without having that abundance first. You mentioned to me the book The Maniac by Benjamin Labatut, a book, first of all, about von Neumann. There's a biography about him. Strange, yeah.

It's unclear how much is fiction and how much is reality, but I think the central figure is John von Neumann. I would say it's a haunting and beautiful exploration of madness and genius, and of the double-edged sword of discovery.

For people who don't know, John von Neumann is a kind of legendary mind. He contributed to quantum mechanics, was on the Manhattan Project, and is widely considered to be the father of, or a pioneer of, the modern computer and AI, and so on. Many people say he's one of the smartest humans ever, so it's fascinating.

What's also fascinating is that he saw nuclear science and physics become the atomic bomb. He got to see ideas become a thing that had a huge amount of impact on the world. He also foresaw the same thing for computing.

Yeah.

And that's the, again, beautiful and haunting aspect of the book. Taking a leap forward and looking at this—at least at AlphaGo and AlphaZero—as a big moment where maybe John von Neumann's thinking was brought to reality. So I guess the question is: What do you think, if you got to hang out with John von Neumann now, what would he say about what's going on?

Well, that would be an amazing experience. He's a fantastic mind, and I also love the place where he spent a lot of his time at Princeton, the Institute for Advanced Study—a very special place for thinking. It's amazing how much of a polymath he was and the range of things he helped invent, including, of course, the von Neumann architecture that all modern computers are based on.

He had amazing foresight. I think he would have loved where we are today, and he would have really enjoyed AlphaGo—you know, games; he also did game theory. I think he foresaw a lot of what would happen with machine-learning systems that are kind of grown. I think he called it “grown” rather than programmed. I'm not sure; maybe he wouldn't even be that surprised at the fruition of what I think he already foresaw in the 1950s.

I wonder what advice he would give. You got to see the building of the atomic bomb with the Manhattan Project. I'm sure there's interesting stuff that maybe is not talked about enough—some bureaucratic aspect, perhaps the influence of politicians, maybe not enough picking up the phone and talking to people who are called enemies by those politicians. There might be some deep wisdom that we just may have lost from that time, actually.

Yeah, I'm sure. I'm sure there is. We've studied it—I read a lot of books chronicling that time as well—and some brilliant people were involved. I agree with you. I think maybe there needs to be more dialogue and understanding. I hope we can learn from those times.

I think the difference here is that AI has so many uses. It's a multi-use technology. Obviously, we're trying to do things like solve all diseases, help with energy, and address scarcity—these incredible things. This is why all of us, myself included, started on this journey 30-plus years ago.

But of course there are risks too, and my guess is that von Neumann foresaw both. I think he said to his wife that computers would be even more impactful in the world. As we just discussed, I think that's right. I think it's going to be at least 10 times the impact of the Industrial Revolution, so I think he was right. I imagine he would have been fascinated by where we are now.

And I think one of the—maybe you can correct me—but one of the takeaways from the book is that reason, as said in the book, “mad dreams of reason,” is not enough for guiding humanity as we build these super-powerful technologies. There's something else.

I mean, there's also a religious component. Whatever God or whatever religion gives it to you, it pulls at us—something in the human spirit that raw, cold reason doesn't give us.

And I agree with that. I think we need to approach it with whatever you want to call it: a spiritual dimension or a humanist dimension. It doesn't have to do with religion, right? But this idea of a soul, what makes us human, this spark that we have, perhaps has to do with consciousness when we finally understand that. I think that has to be at the heart of the endeavor.

Technology, I've always seen technology as the enabler—the tools that enable us to flourish and to understand more about the world. I'm sort of with Feynman on this, and he used to always talk about science and art being companions, right? You can understand it from both sides: the beauty of a flower, how beautiful it is, and also why the colors of the flower evolved like that. That just makes it more beautiful—the intrinsic beauty of the flower.

Maybe in the Renaissance, the great discoverers, people like da Vinci, didn't see any difference between science and art, and perhaps religion. Everything was just part of being human and being inspired by the world around us. That's the philosophy I try to take.

One of my favorite philosophers is Spinoza, and I think he combined all of that very well. This idea of trying to understand the universe and understanding our place in it was his way of understanding religion, and I think that's quite beautiful. For me, all of these things are related and interrelated: technology and what it means to be human.

It's very important, though, that we remember that when we're immersed in the technology and the research. I think a lot of researchers that I see in our field are a little bit too narrow and only understand the technology. I think that's also why it's important for this to be debated by society at large.

I'm very supportive of things like the AI summits that will happen, and of governments understanding it. I think that's one good thing about the chatbot era and the product era of AI: everyday people can actually feel and interact with cutting-edge AI and feel it for themselves.

Yeah. Because they force the technologists to have the human conversation. For sure, that's the hopeful aspect of it. Like you said, it's a dual-use technology that we're forcefully integrating all of humanity into through the discussion about AI.

Ultimately, AI and AGI will be used for the things that states use technologies for, which is conflict and so on. The more we integrate humans into this picture by having chats with them, the more it will guide society. Society will be able to adapt to these technologies, like we've always done in the past with the incredible technologies we've invented.

Do you think there will be something like a Manhattan Project, where there will be an escalation of the power of this technology, and states, in their old way of thinking, will try to use it as weapons technology? Will there be this kind of escalation?

I hope not. I think that would be very dangerous to do and also not the right use of the technology. I hope we'll end up with something more collaborative, if needed, more like a CERN project, where it's research-focused and the best minds in the world come together to carefully complete the final steps and make sure it's responsibly done before deploying it to the world. We'll see.

It's difficult with the current geopolitical climate to see cooperation, but things can change. I think, at least on the scientific level, it's important for researchers to keep in touch and keep close to each other on those kinds of topics.

Yeah. I personally believe that, on the education side and the immigration side, it would be great if, in both directions, people from the West immigrated to China and people from China immigrated back. There is some family, human aspect of people just intermixing, and thereby those ties grow strong.

So you can't sort of divide against each other in this old-school way of thinking. Multicultural, multidisciplinary research teams working on scientific questions—that's the hope. Don't let the world leaders who are warmongers divide us. I think science is ultimately a really beautiful connector.

Yeah. Science has always been quite a collaborative endeavor, and scientists know that it's a collective endeavor as well. We can all learn from each other, so perhaps it could be a vector to get a bit of cooperation.

What's your ridiculous question? What's your p(doom)—the probability that human civilization destroys itself?

Well, look, I don't have a p(doom) number. The reason I don't is because I think it would imply a level of precision that isn't there. I don't know how people are getting their p(doom) numbers. I think it's a little bit of a ridiculous notion.

What I would say is that it's definitely nonzero and probably non-negligible. That in itself is pretty sobering, and my view is that it's just hugely uncertain. We don't know what these technologies are going to be able to do, how fast they're going to take off, or how controllable they're going to be.

Some things may turn out to be, hopefully, way easier than we thought, but there may be some really hard problems that are harder than we guess today. Under those conditions of a lot of uncertainty but huge stakes both ways, we could, on the one hand, solve all diseases, energy problems, and the scarcity problem, and then travel to the stars and conquer the stars, with maximum human flourishing. On the other hand, there are these p(doom) scenarios.

Given the uncertainty around it and the importance of it, it's clear to me that the only rational, sensible approach is to proceed with cautious optimism. We want the outcome, and we want the benefits, of course, and all of the amazing things that AI can bring.

I would be really worried for humanity if, given the other challenges that we have—climate, disease, aging, resources, all of that—I didn't know something like AI was coming down the line. How would we solve all those other problems? I think it's hard.

It could be amazingly transformative for good, but on the other hand, there are these risks that we know are there but can't quite quantify. The best thing to do is to use the scientific method, do more research to try to more precisely define those risks, and of course address them. I think that's what we're doing.

I think there probably needs to be 10 times more effort on that than there is now, as we're getting closer and closer to the AGI line.

What would be the source of worry for you more? Would it be human-caused or AI- or AGI-caused—humans abusing that technology versus AGI itself through mechanisms that you've spoken about, which are fascinating, like deception or this kind of thing getting better and better and better secretly, and then...?

I think they operate over different time scales, and they're equally important to address. There's just the common garden-variety of bad actors using new technology—in this case, a general-purpose technology—and repurposing it for harmful ends. That's a huge risk, and I think it has a lot of complications because generally I'm in huge favor of open science and open source.

In fact, we did it with all our science projects, like AlphaFold and all of those things, for the benefit of the scientific community. How does one restrict bad actors' access to these powerful systems, whether they're individuals or even rogue states, but enable access at the same time to good actors to maximally build on top of them? It's a pretty tricky problem, and I've not heard a clear solution to it.

So there's the bad-actor use-case problem, and then there's obviously, as the systems become more agentic and closer to AGI and more autonomous, how do we ensure the guardrails, that they stick to what we want them to do, and that they remain under our control?

Yeah. I tend to worry more about the humans, so the bad actors. It could be, in part, how you don't put destructive technology in the hands of bad actors, but in another part, from a geopolitical technology perspective, how do you reduce the number of bad actors in the world? That's also an interesting human problem.

Yeah, it's a hard problem. Look, we can maybe also use the technology itself to help with early warning on some of the bad-actor use cases, whether that's bio or nuclear or whatever it is. AI could potentially be helpful there, as long as the AI that you're using is itself reliable.

So it's a sort of interlocking problem, and that's what makes it very tricky. Again, it may require some agreement internationally, at least between China and the U.S., on some basic standards.

I have to ask you about the book The Maniac. There's this hand-of-God moment—Lee Sedol's move 78—that perhaps was the last time a human made a move of pure human genius and beat AlphaGo, or broke its brain, if—sorry to anthropomorphize—but it's an interesting moment, because I think in so many domains it will keep happening.

Yeah, it's a special moment. It was great for Lee Sedol, and I think, in a way, they were kind of inspiring each other. We, as a team, were inspired by Lee Sedol's brilliance and nobleness, and then maybe he got inspired by what AlphaGo was doing to conjure this incredible, inspirational moment.

It's all captured very well in the documentary about it. I think that'll continue in many domains, at least for the foreseeable future, with humans bringing in their ingenuity and asking the right question, let's say, and then utilizing these tools in a way that cracks a problem.

Yeah. As AI becomes smarter and smarter, one of the interesting questions we can ask ourselves is: What makes humans special? It does feel—I'm perhaps biased—that we humans are deeply special. I don't know if it's our intelligence. It could be something else, that other thing that's outside the mad dreams of reason.

I think that's what I've always imagined when I was a kid and starting on this journey. I was, of course, fascinated by things like consciousness. I did a neuroscience PhD to look at how the brain works, especially imagination and memory. I focused on the hippocampus, and it's going to be interesting.

I always thought the best way, of course—one can philosophize about it and have thought experiments, and maybe even do actual experiments like you do in neuroscience on real brains—but in the end, I always imagined that building AI, a kind of intelligent artifact, and then comparing that to the human mind and seeing what the differences were would be the best way to uncover what's special about the human mind, if indeed there is anything special.

I suspect there probably is, but it's going to be hard to define. I think this journey we're on will help us understand that and define it. There may be a difference between the carbon-based substrates that we are and silicon ones when they process information.

One of the best definitions I like of consciousness is: It's what information feels like when we process it. It isn't a very helpful scientific explanation, but I think it's an interesting, intuitive one. This scientific journey we're on will, I think, help uncover that mystery.

Yeah. “What I cannot create, I do not understand.” That's from somebody you deeply admire, Richard Feynman, as you mentioned. You also reach for Wigner's dreams of universality that he saw in constrained domains, but also broadly, generally, in mathematics and so on. There are so many aspects on which you're pushing toward—

Not to start trouble at the end, but Roger Penrose.

Yes.

Okay. Do you think consciousness—there's this hard problem of consciousness, how information feels—do you think consciousness, first of all, is a computation? If it is information processing, like you said everything is, is it something that could be modeled by a classical computer, or is it quantum-mechanical in nature?

Well, look, Penrose is an amazing thinker, one of the greatest of the modern era, and we've had a lot of discussions about this.

Of course, we cordially disagree, which is that I feel he collaborated with a lot of good neuroscientists to see if he could find mechanisms for quantum mechanical behavior in the brain. To my knowledge, they haven't found anything convincing yet. My bet is that it's mostly classical computing that's going on in the brain, which suggests that all the phenomena are modelable or mimicable by a classical computer. But we'll see.

There may be these final mysterious things about the feeling of consciousness—the qualia, these kinds of things that philosophers debate—where it's unique to the substrate. We may even come toward understanding that if we do things like Neuralink and have neural interfaces to AI systems, which I think we probably will eventually, maybe to keep up with the AI systems. We might actually be able to feel for ourselves what it's like to compute on silicon, right? So I think it's going to be interesting.

I had a debate once with the late Daniel Dennett about why we think each other are conscious. There are 2 reasons. One is that you're exhibiting the same behavior that I am, so behaviorally you seem like a conscious being if I am. But the second thing, which is often overlooked, is that we're running on the same substrate. So if you're behaving in the same way and we're running on the same substrate, it's most parsimonious to assume you're feeling the same experience that I'm feeling.

But with an AI that's on silicon, we won't be able to rely on the second part. Even if it exhibits the first part—the behavior looks like the behavior of a conscious being—it might even claim that it is, but we wouldn't know how it actually felt. It probably couldn't know what we felt, at least in the first stages. Maybe when we get to superintelligence and the technologies that it builds, perhaps we'll be able to bridge that.

No, I mean, that's a huge test for radical empathy: to empathize with a different substrate.

Right, exactly. We never had to confront that before.

Yeah. So maybe through brain-computer interfaces, we'd be able to truly empathize with what it feels like to be a computer, for information to be computed not on a carbon-based system. That's deeply—some people think about that with plants and other life forms, which have a different, though similar, substrate, but are sufficiently far enough on the evolutionary tree that it requires a radical empathy. But to do that with a computer—

Look, there are animal studies on this. Higher animals like killer whales, dolphins, dogs, monkeys, and elephants have some aspects, certainly, of consciousness, right? Even though they might not be that smart in an IQ sense, we can already empathize with that. Maybe even some of our systems one day—we built this thing called DolphinGemma, where one version of our system was trained on dolphin and whale sounds. Maybe we'll be able to build an interpreter or translator at some point, which should be pretty cool.

What gives you hope for the future of human civilization?

What gives me hope is our almost limitless ingenuity, first of all. I think the best of us and the best human minds are incredible. I love meeting and watching any human who's at the top of their game, whether that's in sport, science, or art. It's just nothing more wonderful than that, seeing them in their element and flow.

I think it's almost limitless. Our brains are general systems, intelligent systems, so I think it's almost limitless what we can potentially do with them. Then the other thing is our extreme adaptability. I think it's going to be okay, even though there's going to be a lot of change.

But look where we are now with our effectively hunter-gatherer brains. How is it that we can cope with the modern world? Flying on planes, doing podcasts, playing computer games, and using virtual simulations. Given that our brains were developed for hunting buffalo, society's already adapted to this mind-blowing AI technology we have today. It's like, "Oh, I talk to chatbots." It's totally fine.

And it's very possible that this very podcast activity, which I'm here for, will be completely replaced by AI. I'm very replaceable, and I'm waiting for—

Not to the level that you can do it, Lex, so don't think—

Thank you. That's what we humans do to each other: we compliment.

All right. And I'm deeply grateful for us humans to have this infinite capacity for curiosity, adaptability, like you said, and also compassion and the ability to love.

Exactly. All of those human—

All the things that are deeply human.

Well, this is a huge honor, Demis. You're one of the truly special humans in the world. Thank you so much for doing what you do and for talking today.

Well, thank you very much, Lex.

I got a note on May 21 this year that said, "Hi, Lex. 20 years ago today, David Foster Wallace delivered his famous This Is Water speech at Kenyon College. What do you think of this speech?"

First, I think this is probably one of the greatest and most unique commencement speeches ever given. Of course, I have many favorites, including the one by Steve Jobs. David Foster Wallace is one of my favorite writers and one of my favorite humans. There's a tragic honesty to his work, and it always felt as if he was engaging in a constant battle with his own mind. His writing was kind of his notes from the front lines of that battle.

Now, onto the speech. Let me quote some parts. There's, of course, the parable of the fish and the water, which goes:

"There are these 2 young fish swimming along, and they happen to meet an older fish swimming the other way who nod

It only adds. All right, back to David Foster Wallace's speech. He has a great story in there that I particularly enjoy. It goes:

“There are these 2 guys sitting together in a bar in the remote Alaskan wilderness. One of the guys is religious. The other is an atheist, and the 2 are arguing about the existence of God with that special intensity that comes after about the 4th beer.

“And the atheist says, ‘Look, it's not like I don't have actual reasons for not believing in God. It's not like I haven't ever experimented with the whole God-and-prayer thing. Just