哈萨比斯把 AGI 闹钟拨到 2030：领先五年是远见，领先五十年是灾难

Demis, thank you so much. Exciting to be here. Thanks, everyone, for coming. It's great to be here. We're so honored to have you at our chocolate factory. Yes, I just heard about that. Looking forward to the chocolate afterwards. Excellent. Well, Demis, we're going to jump right in. We have one of the OGs in every way, original thinkers, founders, visionaries in all things AI, a true believer, true scientist in Demis. We're going to spend the beginning of the conversation about the early days, then early days of DeepMind, and then we'll get into the science and open up the room for some questions. So let's jump right in. So Demis, you were a chess prodigy. You also were a founder of a gaming company. You're a neuroscientist. And you're the founder of DeepMind, now leader of a really big consequential company. Those seem like pretty unrelated pieces, but you've said that there's a common thread. Can you walk us through that? There is a common thread, and maybe I made it into a common thread. So it could be a post hoc sort of shaping. But I wanted to do AI for a long time. So I kind of decided it was the most important thing I could possibly, and most interesting thing I could do in my teenage years. And then I picked things to study or do that I felt eventually would help me build a company like DeepMind. So I had that as a plan from about 15, 16 years old. I had a detour into games because actually that was in the 90s. That's where all the most cutting edge technology was being done. Obviously, not just in AI, but in graphics, especially, including hardware. Of course, the GPUs we all use today, they were designed for graphics engines. And we used the, I mean, I was using the very first GPUs back in the day in the late 90s. So there was a lot of really cutting edge technology. And then all the games I made, including all the games I did for Bullfrog, but also my own company, Elixir Studios, all involved AI as the main gameplay component. So probably the most well-known game I made was Theme Park when I was about 17. And that was a simulation of an amusement park. And thousands of little people came into your theme park and played on your rides and decided what to buy from the shop. So there was a whole kind of economics AI model underneath. And it was one of the first games of its type, along with Sim City. And when I saw it sold, you know, 10 million plus copies. And when I saw how delighted people were to interact with the AI, you know, that was one of the things that made me think about spending my whole career on it. And then, of course, neuroscience is to get inspiration from how the brain works and different algorithmic ideas from that. And then just bringing all those different things together for the start of DeepMind, when the timing we felt was right. And then, of course, we use games as the early proving ground for our AI ideas. So we've got a room full of founders here. And you can relate because you're a founder not just once, but twice. Take us back to the first time, Elixir Studios. What was that like? It's not the startup that you're most known for, but it was one that you had incredible success with. How did you lead that and what did it teach you about building? Well, look, we started, I started Elixir Studios straight out of college. And I was lucky enough to work at Bullfrog Productions, which, for those of you who know games, it was a kind of legendary game studio in the early days of the game industry, probably the best one in the UK, in Europe. And I wanted to do something that combined, pushed AI. So effectively, I was funding AI back in those days through the backdoor, through games development. And then pushed the forefront of that and combine it with cutting edge creativity. And I think that's still relevant today with the way we do our blue sky research. But maybe the biggest lesson I learned was you want to be five years ahead of your time, not 50 years ahead. So we tried to do a game called Republic at Elixir Studios, which simulated a whole country. And then the idea of the game is you could sort of overthrow, I think there was a dictator in charge of the country in any number of different ways. And we basically simulated living, breathing cities. And this is, bearing in mind, this is like the late 90s on a Pentium. So we had to get all the graphics and all the AI for a million people working on a home PC at the time. So it was a little bit ambitious. And maybe it was too ambitious. And it caused some issues. And I took that lesson with me of like, you want to be ahead of your time. You don't want to be, obviously, when it's obvious to everyone, it's too late. But if you're 50 years ahead, then there's probably no way you can get it to be successful. All right. So speaking of not being too far ahead of your time, it was 2009 and you decided there would be AGI. Yes. Maybe it was only, you know, 10 years ahead of our time that time. Better than 50 years. So tell us about, again, room full of founders here. Tell us about 09. How did you convince the first few brilliant talents? Because you pulled in really high caliber employees, early team members. How do you convince them to believe in what seemed like total sci-fi at the time? Well, there was some interesting threads that I think we picked up on. I think we thought we were five years ahead, but maybe we were more like 10. But it was, you know, deep learning had just been invented by Jeff Hinton and colleagues sort of in academia. But almost no one had really realized it was a big deal. We knew a lot about reinforcement learning and we felt there was huge progress to be made by combining those two techniques, which almost had never been mixed together, really. Certainly not at any kind of anything other than toy problems in the academic subjects. They were quite two quite siloed parts of AI. Then we could see the compute that GPUs at the time were going to be really useful. Of course, we use TPUs now. But the accelerated computing industry was was was going to be very helpful. And then we also felt at the end of my PhD in postdoc and some of the other people I got together were computational neuroscientists that we had enough ideas and principles from the brain that could be useful, including the idea that reinforcement learning could eventually scale to AGI. So so we felt we had these ingredients and we almost felt like we were keepers of a secret because no one either in academia industry really believed that any big progress was possible. In fact, a lot of the people in academia used to roll out that, you know, literally roll their eyes up at us when we were sort of suggested we would work on AGI or strong AI. It was sometimes called at the time because it was like, well, we know this doesn't work. So, you know, everyone tried it in the 90s and I did my postdoc at MIT, which was the which was the sort of center point for expert systems and first of all, the logic language systems. I mean, it seems amazing to think that now, but I was already feeling like that was a CAIC then. But they they you know, that's still how it was done, both in Cambridge in the UK and also in MIT, these big centers of traditional AI. And it felt like but actually that convinced me even more that we were onto something because at least if we were going to fail, we would fail in a different way than people had failed, you know, to get to AGI in the 90s. So that felt like it was worth doing no matter what, even if obviously it was research, we didn't know for sure it would be successful. But at least we would we would fail in an original way if it didn't work. Was there any common sticking point in that early belief? Was there something that you had to prove either to yourself or to your early followers to get them on board? Well, it was I mean, we had put it this way. I would have been spent my life on A.I. No matter what happened. So as it's turned out, it's gone. You know, it's sort of gone on the absolutely amazing side of the optimistic side of what we thought still actually within what we were predicting in 2010. We thought it would be a 20 year mission. And I think we're basically exactly on track as a field for that. And obviously we played our part in that. But even if it hadn't transpired that way and it was still now a niche subject, that's what that's what I would still be doing, because I felt it was the most important technology ever. If it was obvious to me, you know, our original state mission statement at DeepMind was step one, solve, solve intelligence. I build AGI step to use it to solve everything else. So it was always I always thought it was the most important technology that could ever be invented, but also the most interesting one. So as a tool for science, as an interesting artifact in itself, and actually as one of the best ways to understand our own minds, you know, like the nature of consciousness, dreaming, creativity, all of these questions I had as a neuroscientist. I felt one of the things that was missing was an analysis tool like A.I., but also a comparison that you could do sort of a controlled experiment, study and compare two different systems against each other. Let's talk about A.I. for science. You've been early to that. You've been a believer and you've been really a purist about this. This is the driving mission. What about the way you set up DeepMind and set the culture has positioned it to be on the constant forefront of A.I. for science? Well, that was the ultimate goal, at least for me, my personal passion. There's one. There's my own drive to build A.I., which was to advance science and medicine and our understanding of the world. It's my expression of that mission was to sort of do it in a meta way, right? Build the build the ultimate tool and then come back when that was ready and use it to make breakthroughs in science. Things like AlphaFold that we've done and I think many more things. So we've always had that at the heart of what we've been trying to do at DeepMind. So actually we've had an A.I. for science group division led by Pushme Koli that has existed for nearly a decade now. Actually, pretty much the day after we got back from Seoul and the AlphaGo match, which is sort of 10 years to the month now, is when we started, formally started the A.I. for science efforts. Because I was waiting for the algorithms to be powerful enough and the ideas to be general enough. And for me, you know, Cracking Go was that point, that time that we thought, OK, now we're ready to really apply these ideas to important real world problems, starting with these big scientific challenges. So we've always had that in mind as the most beneficial use of A.I., like what could be better than using it to, you know, cure diseases and give us healthier lifespans and to help with medicine, followed obviously by other really important areas like material science and the environment and energy and these kinds of topics, which I think A.I. is also going to be have a huge part to play in the next few years. And how does A.I. break through in biology? You're deeply involved with isomorphic. This is an area of deep passion. You have been a purist on the potential of A.I. to cure diseases from the very beginning. When do we have the type of moment that we've had in language and coding, but in biology? Yeah, well, I mean, I'd argue we've already had one of those moments with alpha folds. So, you know, it's a 50 year grand challenge protein folding and the 3D structure proteins is incredibly important thing to know about. If you want to design medicines or if you want to understand biology, of course, it's only one part of the drug discovery process. It's an important part, but it's only one part. So isomorphic labs, which is our, you know, latest spin out, having a lot of fun running that as well, is to build adjacent technologies in more biochemistry and chemistry space that can actually design the compounds automatically to kind of fit and bind to the right part of the protein. So we now know the protein, the shape of the protein. We know that what's on the surface of the protein and what we have to target. But now we've got to build the right compounds that, of course, bind strongly to where you want it to bind on the target of interest, but doesn't bind to anything else, ideally, because that would be a toxic side effect. So the dream is to do almost all the exploration, which is 99% of the work and the time in silico and then save the wet lab step just for the validation step. Right. So that would be, you know, I think if we can do that and I think we can get there in the next few years, I think we could reduce drug discovery times instead of that for taking like, you know, an average of 10 years down to months, maybe even weeks and perhaps even days one day. And then I think then all disease could be in reach. And I think things like personalized medicine will become possible, you know, like personalized variations off of base medicines. So I think the whole of the whole medical area, drug discovery areas is going to be revolutionized in the next in the next few years. Brilliant. You talked a lot about AI for science. Do you think that at some point, AI will create new sciences, a la Industrial Revolution and thermodynamics? Will there be something new taught fundamentally in our education system? And if so, what would it be like? Well, I think there's several things that along those lines that I think is going to happen. So first of all, the understanding and the analysis of AI systems themselves, I think, is going to become a whole science, a kind of engineering science. These incredible, incredibly interesting artifacts that we are building and they're incredibly complex as well as complex. Eventually they'll be as complex as the human mind and the brain. And so they'll need to be studied so we can understand fully way beyond where we are today, how these systems work. So I think there's a whole kind of field. Mechinturps is part of that. But there's a lot more, I think, that we can do to analyze these systems. So that would be a science. But I think also, AI itself will maybe unlock new sciences, which is maybe what you're getting at. The one I'm particularly excited about is AI for simulations. So I love simulations. All the games I wrote not only had AI, but they were simulations. And I think simulations is the way we can address some of the what we maybe think of social sciences like economics and other more humanistic subjects, because it's very difficult to do control studies. And, you know, why aren't they just sciences like physics today? Because the problem is there are emergent systems just like biology, actually. And it's very hard to do repeated controlled experiments. You know, if you raise interest rates by half a percent, you have to do it in the real world and then see what happens. You can have theories, but you can't run it thousands of times. But if you could simulate things really accurately, then maybe there's sort of new sciences to be done where you can sort of rigorously sample from a very accurate simulator. And then I think that will allow us to make much better decisions in these today, very uncertain domains. What will it take to get to those extremely accurate simulations? World models, what kind of science is necessary and engineering? Yeah. Well, look, I mean, I'm thinking a lot about that in we do a ton of that work, like learning simulators, basically what it would be. So, you know, these are in domains where you can't we don't know the mathematics of it well enough or it's perhaps too complex. We can't just write we can't just write a directly down a special case simulator. It's just not accurate enough, doesn't capture all the variables. We're doing that. We've done it with weather. We have the most accurate kind of weather simulator in the world, Weather Next. And it's far faster than what the meteorologist use. Can you change the weather yet? We can't. We can't. No. And I'm not sure that would be a good idea. But the first step is to understand it better. And but then even biology, you know, we're working on a kind of what I call a virtual cell. So a hugely dynamical emergent system. And I think biology is is perfect. Sort of machine learning is perfect description language for biology in the same way maths is for physics, because I think in biology and in lots of these natural systems, you have loads of weak signals, weak correlations, tons of data, far too much that any human mind can analyze. But there are connections and correlations and interesting causalities within that mass of data. So I think it's sort of it's always struck me that machine learning is the perfect tool to describe those kinds of systems where until today, you know, mathematics hasn't been able to do that either because we can't manage it as top mathematicians because too complex or the expressive power of maths is not enough for to understand these sort of highly emergent dynamical systems. Is it also because of the messiness and stochastic nature? Yeah, sure. And I mean, eventually you could, by the way, once you learn these simulators, it may be there's another, this is maybe another branch of new branch of science, you could maybe extract some equations from the once you have the simulator. So you have this sort of implicit simulator or intuitive simulator, and then maybe you could extract explicit equations from that, partly because you could also sample it as many times as you want. As fundamental as Maxwell's or something like that. Maybe. If, if I don't know if that exists for such emergent systems. But if they do exist, I don't see why we won't be able to find them in this with this, with this methods. That would be amazing. You've talked about this theory that the basic building block of everything in the universe could be information-like. This is more theoretical. How do you think about it that? And what does that mean for a traditional classical Turing computer? Well, look, I think you can, of course, all the famous, you know, equals MC squared and all the stuff Einstein did and energy and matter are kind of equivalent. But I actually think information has a kind of equivalency in the same way. So you can think of, you know, the organization of matter and structure and especially things like biology that are resisting entropy as basically information processing systems at their heart. So I think one can convert all of those three kind of quantities into each other. But I have this feeling information is most fundamental. So it's a little bit the opposite way around to the classic physicist thought in the 1920s and things where, you know, it's sort of energy and matter primary. I actually think it's a better way to understand the world, the universe is to think about it as information first. And if that's true, and I think there's quite a lot of evidence for that, then, of course, AI is even more sort of profound in a sense than we think. And it's already pretty profound because it's also about organizing information and understanding information and constructing informational objects. So AI, in my opinion, is all about information processing. So I think there's something sort of very deeply connected with these different areas if you think of it through the lens of information processing as the primary way to think about it. And do you think a classical Turing machine will be able to compute everything? Well, I sometimes think, you know, I sometimes sort of think about what we're doing and refer to ourselves as Turing's champion because Turing machines, I think Alan Turing is one of my all time, you know, favorite scientific heroes. I think what he did obviously laid the foundations for the computers for computer science, but also AI. And I think it's one of the most profound results ever is the Turing machine result. You know, everything that is computable can be computed by a relatively simple description of a machine. So I think our brains are likely to be approximate Turing machines. And I think it's interesting to think about the connection between Turing machines and quantum computers and quantum systems. But I think at least what we've shown with things like AlphaGo and especially AlphaFold is that a classical Turing machine, obviously in the guise of a modern neural network, it can model what was thought to be in the case of protein folding. It's a quantum system, you know, it's a quantum system, you know, at some sense, it's very, you know, it's dealing with very small particles. And one might think you'd have to take into account all the quantum effects of the water bonds and all sorts of things. But it turns out you can get to an approximate optimal sort of solution on a classical system. So it may turn out that a lot of things that we think that would need a quantum system to model or run might be modelable on a classical system if thought about in the right way. So you've talked about AI consistently as a tool, like a telescope or a microscope, astrolabe through the centuries. But when you think about a machine that can model almost anything, let's say it can even model quantum systems like you pointed out, when does it stop becoming a tool? And will that ever happen? Well, my strong feeling is we should, in this sort of mission and journey to build AGI, those of us on that journey, many of the people in this room, you know, I feel like it would be best to build a tool first, an incredibly intelligent and useful and precise tool, and then cross the next sort of rubric. That's already profound enough and has, you know, of course, the tool could start becoming more and more autonomous and agent like that we're all seeing we're in the midst of that the agent era now. But then there's a further step of like, you know, does it have agency? Is it conscious? These sorts of questions, which are also going to be questions we're going to need to address. But I would, I would recommend we do that as a second step, perhaps using the tool in the first step to help us with those next profound questions. And ideally, also, we could understand our own brain and minds better and define things like consciousness a lot more precisely than we can today. Do you have estimations of what that definition of consciousness might look like? No, I mean, I haven't got much to add beyond that thousands of years of philosophy hasn't said already. But I mean, it's very clear to me that it's obvious some components are going to be needed. They're probably necessary, but not sufficient. Things like self awareness and, you know, the idea of self and other some kind of continuity over time. So some of these things are clearly needed for anything that might look like consciousness. But I mean, obviously, it's an open question as to as to what the full definition is. And I've talked to many of the great philosophers about that. Daniel Dennett, obviously, sadly passed away recently. But we had a long conversation a few years back about this. And I think, you know, one of the issues is how does a system behave? Does it behave like a conscious system? So that's you know, you could argue some of the AI systems might end up being able to do that as they get close to AGI. But then there's still the question of why do you you know, why do we think each other are conscious? One is the way we're behaving. We're behaving like conscious beings. But the other thing is we're running on the same substrate. So I think if both those things are true, then it's parsimonious to imagine you're you're experiencing the same thing I'm experiencing, which is why we don't have that debate about, you know, normally about each other conscious. But I think we'll obviously we'll never have the substrate equivalence with an artificial system. So there'll always be a I think it will be hard to to completely close that gap. So you can look at it behaviorally. But but what about experientially? There are probably some ways to do that post AGI, but it's a bit about out of scope today, even for AI for science discussion. Brilliant. So we're going to open the room to questions in just a moment. Get your questions ready. But you brought up philosophers. You've mentioned Kant and Spinoza as two as your two of your favorite philosophers. Kant is this, you know, deontological, highly duty driven philosopher. Spinoza almost has this deterministic view of the universe. How do you kind of connect those two beliefs and where where is your thinking of how the world works? Well, the reason I like those two, they stuck out for me is that I think Kant, when I was doing my PhD in neuroscience, you know, his his sort of statements about the mind creates reality. Right. I think that's basically true. And and so another reason to study the mind, right, and how the brain works. And I'm interested ultimately in the nature of reality. So we have to understand how the mind is interpreting that. And so I think that's for me what I took from Kant. And then Spinoza is more about the you could almost call spiritual dimension of like, well, if you're trying to understand the universe using science, in my case, as the tool, you're sort of understanding some deep mystery about how the universe works. Right. In a in a really kind of deep way. And that's what I feel we're doing and I'm doing when, you know, I do my science and we work on A.R. and we're building these tools is somehow we're kind of reading the language of of the universe. Beautiful. What a beautiful way to say what you do every day. Demis, scientist, orator and philosopher. We will before we wrap do a couple of rapid fire questions. OK. Thank you for finishing. He's not seeing these yet. Over under on distribution year of AGI. Oh, wow. Or reject premise of question. No, 2030. I've been pretty consistent about that. OK, 2030. Yeah. Must read book, poem, or paper for when we achieve AGI. Oh, wow. For when we achieve, once we achieve it. Well, my favorite book is The Fabric of Reality by David Deutsch. So I think that still holds. I'd hope to answer the questions in that book with the AGI. That's my post-AGI work. Brilliant. Proudest moments so far in DeepMind? Oh, wow. We've been lucky to have a lot. I mean, probably AlphaFold. OK. Yeah. Now a couple games questions. If you were engaged in a high-stake strategy game, turn-based strategy game. OK. Civ, Polytopia, serious games. Right. And you could select one scientist from history. We're thinking the Einsteins, the Turing's, Newtons. Oh, wow. Who would you select to be on your team? On my team? On your team. Oh, gosh. Probably Von Neumann, I think. Great choice. Yeah, you want a game theorist, I think. And I think he's the best. Makes sense. Yeah. That's a real intelligence. I think he'd be a good teammate. Yeah. All right. Yeah. Well, Demis, you do it all. Thank you so much for being with us. Great. Please join me in taking the time. Thank you. Thanks, everybody.