Founder Stories: Demis Hassabis, AI to Accelerate Scientific Discovery

Ladies and gentlemen, a very warm welcome to this evening's lecture here in the splendour of the Sheldonian Theatre. It is hosted by Oxford University's Institute for Ethics in AI and is part of the Obert C. Tanner Lectures on Artificial Intelligence and Human Values. My name is Nigel Shadbolt, Principal of Jesus College. I'm also a Professor of Computer Science here in Oxford and chair the Institute Steering Group. It was my privilege to help set up the Institute, which brings together world-leading philosophers and other experts in the humanities with the researchers, developers, and users of AI. The Director of the Institute is Professor John Tsoules, and its ultimate home will be the Stephen A. Schwarzman Centre for the Humanities. Whose construction is soon to start. In recent years, AI has gone from strength to strength. It's now ubiquitous in our phones, the games we play, in our cars, our drug discovery companies, the search engines we use, and the translation tools we depend on. Much of that is down to a new generation of AI methods and techniques that are powered by modern machine learning. Learning algorithms, great swathes of data, and the prodigious power of modern-day computing hardware. Some of AI's most dramatic recent accomplishments owe a great deal to our speaker here with us this evening and the company he co-founded. Demis Hassabis, CEO and co-founder of DeepMind, one of the world's leading AI research companies. Demis's own career and intellectual journey is an extraordinary one. A chess prodigy, hugely successful computer games developer with a double first in computer science from Cambridge. Demis has always been fascinated by the human brain, understanding how it gives rise to intelligence. After the success of his games companies, he went on to a PhD in cognitive neuroscience at UCL, followed by a Henry Wellcome postdoctoral research fellowship at the Gatsby Computational Neuroscience Unit also at UCL. His papers in cognitive neuroscience investigated imagination, memory, and amnesia, and appeared in leading journals such as Nature and Science. He combined his interest in computing and neuroscience with the formation of DeepMind in 2010. Its compelling ambition to solve intelligence and then use intelligence to solve everything else. He and his team used games as the context in which to test new ideas, about how to build AI systems using machine learning methods inspired by neuroscience. First arcade games, and then famously Go. A previous talk here in the Sheldonian in February 2016 prefigured AlphaGo winning 4-1 against former world champion Lee Sedol just a month later. Games have proven to be a great training ground for developing and testing AI algorithms. But the aim of DeepMind has always been to build general learning systems ultimately capable of solving important problems in the real world. DeepMind's AlphaFold system is a solution to the 50-year grand challenge of protein structure prediction, culminating in the release of the most accurate and complete picture of the human proteome. A core aim for the Institute for Ethics in AI is to bring together world-leading academics and the practitioners at the cutting edge of AI development. Tonight, we will hear firsthand experience of AI's enormous potential to accelerate scientific discovery, experience which will inform our research and thinking about the critical ethical considerations that must be considered by policymakers and technical developers of AI. Demis has predicted that artificial intelligence will be one of the most beneficial technologies ever, but that significant ethical issues remain. Please join me in welcoming Demis Hassabis to deliver for tonight's Tanner Lecture, Using AI to Accelerate Scientific Discovery.

Thank you, thank you, Sir Nigel, for such a great introduction, and it's a real pleasure to be back here in Oxford in the Sheldonian and giving the Tanner Lecture. It's a real honor. So, what I'm going to talk about today is using AI to accelerate scientific discovery. And in fact, as you'll see throughout my talk, this was my original motivation and has always been my motivation behind spending my entire career on trying to make AI a reality. And I'm going to talk a lot about some of our most recent advances actually now coming to fruition over the, especially the last year or two of using AI to crack difficult scientific problems. But I'm also going to talk about the lead-up to there and how I think about the games work we did originally and the foundational work we did originally. And last time I talked here was just before the AlphaGo match in Korea, so that was kind of a major moment for us, and how in the last even 5, 6 years things have progressed enormously. So just to sort of talk a little bit about what our vision was behind DeepMind back in 2010. It's quite hard to remember the state of AI back in 2010 because today, as, as Nigel was saying, that, you know, AI is ubiquitous all around us. It's one of the biggest buzzwords in industry. It's sort of hard to remember just 12 years ago almost nobody was talking about AI, I would say. And it was almost impossible to actually get funding in the private sector for AI at all. And we have many funny stories back in the day of trying to do some fundraising back in 2009 and '10, and most people thinking we were completely, completely mad to be embarking on this, on this, on this journey. But we founded it with this in mind of trying to build one day an Apollo program-like effort to build AGI, artificial general intelligence. And we use this term artificial general intelligence to distinguish it from sort of normal everyday AI where we're talking about a general system that can perform well on many tasks to at least human level. And that's the sort of general aspect that we are always striving for in all the work that we do. So, we're still on this mission now, and I think we've done a pretty good job of basically staying true to this original vision that we had in 2010 when we were just a few people in a small little office in an attic in Russell Square. So, as Nigel said, our original mission statement was, um, step 1, solve intelligence; step 2, use it to solve everything else. We have updated that mission statement a little bit. Still means the same thing, but just to be a little bit more descriptive now in the last few years, just to be a bit clearer about what we mean by solving everything else, what exactly are we talking about. And so, the way we discuss our mission now is solving intelligence to advance science, and of course, for the benefit of humanity. And that's always been the cornerstone of what we think about when we think about what should we apply AI to. Now, there are two, broadly, two ways that I think AI can be attempted to be built. One is the sort of, I guess, more traditional way of building logic systems or expert systems, And these are hard-coded systems that effectively teams of programmers solve the problem. They then incorporate those solutions in sometimes very clever expert systems. But the problem with them is that they are very limited in terms of what they can generalize to. So they can't deal with the unexpected, and they're basically limited to what the programmers foresaw, the situations that the system might be in. And of course, this line of work was inspired by mathematics and logic systems. On the other hand, the big renaissance in the last decade plus is the sort of progress of learning systems. Of course, in the '80s, there was a flurry of work done on neural networks, then that died down. We now know that probably we didn't have enough computing power or data, maybe not the right algorithms as well, But basically, in essence, the ideas were correct. So, an idea of a learning system is that, you know, it learns for itself, solutions for itself, from first principles, directly from experience. And the amazing thing about these systems and their huge promise is that they can maybe generalize to tasks that it's not been programmed for explicitly, and maybe solve problems that we ourselves, as the designers or scientists behind those systems, don't know how to solve. So, of course, that's the huge potential and also the risk of these kinds of systems. And originally, these kind of learning systems took a lot of inspiration and also could be validated, some of the ideas like reinforcement learning and neural networks, by systems neuroscience, and comparing what these systems do, comparing them on a systems and algorithmic level to what we know about how the brain works. Now, everything we do at DeepMind, of course, is on the learning system side, and we've been lucky enough to be in the vanguard of this almost revolution or renaissance in the last decade of these types of approaches. So, how do we think about what's our sort of, I guess, special take on learning systems and how powerful they can be? So, there are kind of two component algorithms or approaches, one could say, that we've fused together. So, of course, there's deep learning, or deep neural networks. And, uh, the way I think about this is that the deep neural network system is there to build a model of the environment, uh, of the data and the experience. Um, and then what do you use that model for? Well, you can use reinforcement learning, um, uh, which is a sort of goal-seeking and, uh, uh, reward-maximizing system to— you can use that model, um, and, uh, use it to plan, uh, and basically plan and take actions towards a goal, a goal that may be specified by the designers of that system. So, you have the model, and then you have the kind of the action and goal-solving element of the systems. So, we, and one of our early innovations was to sort of fuse those two things together at scale. We call it deep reinforcement learning now. And the cool thing about these systems is that they can discover new knowledge from first principles through this process of trial and error using these models. So, the idea here on this diagram of the agent system is it gets observations from the environment. Those observations go towards building and updating an internal model of how the environment works and the transition matrices of the environment. There's some goal it's trying to solve in the environment, achieve, and then after its thinking time has run out, it has to select an action, from the, from the action set available to it at that moment in time that will best get it incrementally towards its goal. And then the action gets output. It may or may not make a change to the environment that drives a new observation, uh, and then the model, um, updates further. So, you can see with this type of system, the agent is actually, uh, the AI system is actually an active learner. It participates in its own learning. So, the decisions it makes, in large part governs what experiences and what data it will get next to learn more from. So, although this is a pretty simple diagram and basically describes the whole of reinforcement learning, the reinforcement learning problem, there's huge complexities, of course, theoretical and practical complexities underlying this diagram that need to be solved. But we know that in the limit, this must work, because this is how mammalian brains work, including humans. This is one of the learning mechanisms that we have in our own brains. Reinforcement learning was found to be implemented by dopamine neurons in the brain in the late '90s. So, we know if we push this hard enough, this should be one path towards general artificial intelligence. So, what do we famously use this for? AlphaGo was, you know, the program that I think— we did a lot of things before this, like Atari games and other programs, proof points, but AlphaGo was really our first attempt at doing this at huge scale, to crack a big problem that was unsolved in AI, kind of one of the holy grails of AI research, which is a program to beat the world champion at the game of Go. And I want to talk a little bit about this in hindsight now, knowing what I know now, how I've reinterpreted what we did with AlphaGo. And I think I can explain it in a much more simple and general way than perhaps how I was explaining it back 5, 6 years ago when we were in the midst of building this system. So, just for those of you who don't know, I don't know why that's not updating. There we go. So, this is the game of Go. Oops. This is the game of Go, the board game, and it's a phenomenal game, It's a much more esoteric game and artistic game, one could say, than chess. So, occupies the same intellectual echelon chess does in the West. In China and Japan and Korea and other Asian countries, they play Go. And Go has resisted sort of old-fashioned logic system and expert system approaches, whereas chess was solved by those things. Because of various factors. One is the search space is truly enormous in Go. It's roughly 10 to the power 170 possible board positions, which is way more than there are atoms in the universe. So, there's no way one could exhaustively search all of the possible board positions in order to find the right path through. Even bigger problem, actually, is that it's impossible— or thought was it was impossible to write down an evaluation function, to sort of hand-code an evaluation function, which is what most modern-day chess programs use. So, and the reason is because Go is such an esoteric game, right? It doesn't have materiality. In chess, you know, as a first approximation, one can add up the piece values on both sides, and that will tell you very crudely, like, who, which side is winning in that position. And obviously, you need to know that in order to make decisions about what to do next. So many people have attempted to, over 20 years since Deep Blue, attempted to write, to construct these evaluation functions for Go. And one of the issues is, is that Go players themselves do not know, consciously at least, what that information is. So because it's so complex a game, they actually use their intuition rather than explicit calculation in order to deal with the complexity of Go, whereas chess players, if you ask them, you know, how— why did they make a decision, a chess grandmaster will tell you, will be able to tell you explicitly the various factors involved. A Go player generally won't do that. They'll just say things like, "It felt right," "This felt like the right move," which is what I think also makes Go an incredible game. But of course, intuition is not something one would associate with computer programs, especially logic systems. And maybe in the Q&A we can discuss a little bit more about what intuition may be. But I don't think it's— it's sort of— I don't think it's— my conclusion now after doing all these games, and indeed some of the science things we've done, is that it's not some mysterious thing. It's actually information that our brain knows about and has learned through experience, of course. I mean, there's no other way one can learn information. But it's just— it's in the association cortices. So it's not actually consciously available to our high-level cortex. So, it seems mysterious to us, you know, how we ride a bike, how we swim, these sort of motor, sensory motor things we're able to do, because our conscious part of our brain cannot access those representations. So, and if we can't do that, then we definitely can't explicitly code it in some logic code, right? Which is why traditionally those tasks, including things like computer vision, have been quite hard for logic systems to solve even over the last 50 years. So, a lot about what we were doing was trying to approximate this kind of intuition in these learning systems. So, how did we work? And I'm actually going to describe not just AlphaGo here, but the whole series of AlphaX programs. So, AlphaGo, the original one that beat Lee Sedol in, in 2016, and then AlphaGo Zero that then didn't need human data to learn from, just learned for itself, and and then finally AlphaZero, which could play any two-player game. So, I'm going to sort of describe them all, roughly speaking, with this sort of demonstrative diagram. So, the way you can think of all of these systems is we're initially training a neural network through self-play, so the system plays against itself, and it learns to evaluate positions and to pick the most likely moves. That are most, you know, useful for it to look at, right? So, that's what it's got to do. Now, initially, it starts with no knowledge, right? So, you have initialized neural network, it starts with zero knowledge. So, it literally is moving randomly, right? So, that's, we can call that version 1, right? That's the neural network. And what it does is it plays roughly 100,000 games against itself, okay? And so, that then becomes a dataset. So, that 100,000 games, we take that as a dataset, And what we try to do with it is train a version 2 of that network, a new neural network, but we try and train it on this version 1 dataset to predict in the middle of a position, you know, middle of a game, from a position in the middle of a game, which side is going to win, right? So, kind of predict ahead of time, and also what sorts of moves does the V1 system choose in a particular position, right? So, that's trying to do, it's trying to be better at both those two things. And then what happens is we train that V2 system and then we have a little mini, mini tournament between V1 and V2. So it's roughly 100 games and they have a little match-off. And basically, if there's— we— the V2 system hits a particular threshold win rate, 55% in this case, then we say it's significantly better than V1, right? And if that's true, then what we do is we replace V1 with version 2 network, this new network in purple, and that, of course, plays another 100,000 games against itself, right? And now it creates a new dataset, but this dataset now in purple in the middle is slightly better quality than that first dataset, right? Because the player is slightly better, and to begin with, almost imperceptibly better. So, it's just slightly better than random now, right? But that's enough signal to then train, you know, of course, we train a version 3 system, and that plays off against version 2. Now, if the version— if you don't reach this 55% win rate, what you do instead is you take back the version 2 and you continue to generate more data with that, another 100,000 games. So then you have 200,000 to train your next version 3, right? And eventually that version 3 will be better than version 2. So after one does this around 17 or 18 times, you go from random to better than world champion. That's it, right? And you can do this with any two-player game, perfect information game, right? So the same network can do that, get to better to world champion within, you know, 20 to 30 generations of doing this. So you literally— and we got to the point where it was so fast, you literally set it off in the morning, you could play chess about it at lunchtime and maybe just beat it, And then by tea time, you know, you had no chance. Literally in the day, you could actually see the evolution in one day. It's kind of incredible to watch as a chess player. So, what is it doing then in terms of thinking about this enormous search space? So, what's happening is, and the sort of, I think, advance of AlphaGo, one of the advances was combining this neural network system or model with a kind of more classical tree search algorithm. In this case, we use Monte Carlo tree search, and you can think of the, the tree of possibilities looking a bit like this in Go, where each node here is a position in Go, obviously shown by these little mini Go boards. And you can imagine, if you're some middle game position, you know, there's just this countless 10 to the 170 possibilities in the limit. How is one supposed to find the needle in the haystack? Right, the, the good moves that could be world champion or better level decisions. So, what the neural network does is it constrains that model, constrains the search to things to make it tractable, right, to things that are reasonably likely to work, reasonably effective, and it can evaluate that at each node level with its evaluation function. And so, instead of having to do, you know, 10 to the hundreds of possibilities, one can just zoom into, you know, mere thousands, 10,000 or so searches. And so, therefore, instead of that searching that entire gray tree of all possibilities, one just looks at this far more limited, you know, search tree in blue here. And then, when you run out of thinking time, of course, you select the best path that you found so far in pink here. So, you know, we did this back in 2015, and then, in the subsequent years, we still work on this now. There's a system called MuZero which is our latest version of this that can not only do two-player perfect information board games, but can also build models of its environment. So, it can actually also do things like Atari games and video games where you actually don't have the rules of the game given to you. It has to actually figure that out for itself through observation as well. So, it's one step even more general than AlphaZero. And what we did with AlphaGo, of course, now is, as Sir Nigel mentioned, is we took it to Seoul in 2016, seen in this million-dollar challenge match with Lee Sedol. And some of you may remember this, but we won 4-1. You know, it was a huge thing, especially in Asia and in Korea. I mean, the country almost came to a standstill. There's over 200 million people watched the games, and we won 4-1. And experts in both AI and in Go proclaimed this advance to be, you know, a decade before they would have predicted. But the important thing in the end was actually not just the fact that AlphaGo won the match, but how it won was, I think, really instructive. So, I'm just going to give one example of this, but actually AlphaGo, I think, has in the end changed the way that we as human beings view the game of Go. But this is the most famous game of that set of 5. There are actually some amazing different games, including the one that Lee Sedol won with a genius move in game 4. And, but move 37 in game 2, I think, will go down in Go history. And this was the board position at that time. And I haven't got time to go into why this was so amazing, but suffice to say AlphaGo here was black, and Lee Sedol is the white stones. And this is very early on in the game, move 37. You know, Go games last for a few hundred moves generally. And AlphaGo played this move 37 stone, with on the right-hand side here marked in red. And the amazing thing about this was the position of the stone was on the 5th line from the edge of the board. And that, if you're an expert Go player, is unthinkable. It's like you would be told off by your Go master that you should never make a move like that, because it gives white too much space on the side of the board. But AlphaGo decided to do it. Never seen before in master play, would be recommended against. And then 100 moves or so later, it turned out this stone, this move 37 stone, was in the perfect position to decide the battle that spread out from the bottom left all the way across the board. And it was just in the right place to decide that battle, which decided the whole game. And almost as if it had presciently sort of seen that influence ahead of time. So now people play on the 5th line all the time, I'm told. So this has changed. This has changed everything, and there's, you know, multiple books now written about AlphaGo's strategies. And, you know, this is an original strategy because this is not something that AlphaGo could have learned from human play. In fact, it would have learned the opposite. It would have learned not to do this kind of move. So, if you're interested in more about AlphaGo, I recommend you, you know, this amazing award-winning documentary that was done by an independent filmmaker. It's on YouTube now. If you want to see the sort of ins and outs of it. It was, you know, very emotional as an experience for us from all sides, especially me being an ex-Games player. I could really understand it from Lisa Dahl's point of view too. So as I said, we then took this to AlphaZero a couple of years ago, 2-3 years ago now, and generalized this to all 2-player games. And these graphs show how AlphaZero did against the best machines at the time in the specialized games of chess. It beat the best version of Stockfish, which is this incredible handcrafted system, the descendant of Deep Blue, and it was able to beat Stockfish 8, which was the best Stockfish at the time, in 4 hours of training. It could beat AlphaGo. AlphaZero beat AlphaGo in 8 hours, right, at Go. And then, and we just tried it with one other game, the Japanese chess shogi, actually, which is a really interesting variation on chess, and it could beat the best handcrafted program called Elmo within 2 hours of training. The same system, all 3 games. So, so that was generalized. And then, of course, because I'm a chess player— I play a little bit of Go, but I'm not very strong— so chess is my game. And so for me, this was the most exciting part of applying AlphaZero, because I actually had a discussion with Murray Campbell, who some of you will know was one of the project leaders behind Deep Blue at IBM back in the '90s. And we just, I think we just were about to play the LisaDoll match, or maybe we just finished, and I was giving a lecture at a conference, and Murray Campbell was there as well in the audience, and he came up to me afterwards, and we were discussing, you know, I said to him, I'm thinking about maybe we should try this with chess and see what happens. And I wanted to know what his prediction would be. You know, do you think these incredibly powerful handcrafted systems like Stockfish could be beaten? Was there any more headroom in chess? You know, chess is probably the oldest application of AI, right? I mean, Turing and Shannon and people like that have all tried their hand— every AI researcher at some point has tried their hand on a on a chess program back to the '40s and '50s, even if Turing had to, you know, run the program by hand on a piece of paper and a pen. But, and then of course, in the last 25 years or so, you know, world champions have been studying with their chess programs and mapping out all of chess opening theory, all of these things. So it was a legitimate question actually to ask, is, was there any more headroom left, and what sort of chess would AlphaZero play if if we were to train it from first principles and play against these amazing, you know, hand-engineered, you know, monsters in some sense of a machine, you know, incredible calculating machines. And so, of course, we couldn't actually come to an agreement on that, and that, you know, as the scientists in the audience will know, that's a sign of a good question, I think, where either answer would be interesting, right? If we were to win and there was some new style out there, that would be incredibly interesting. And also, it'd be interesting if these handcrafted systems, at least in one domain, chess, had reached the limit. So, we got off and started doing that, and I'm pleased to say that AlphaZero not only played stronger, but it did come up with a completely new style of chess, which I think, and my chess friends tell me, is more aesthetically pleasing as well as a chess program, obviously subjectively from a human expert's point of view. And the reason it is, is because what it does— and it does many innovations— but the main one is that it favors mobility over materiality. So traditionally, handcrafted chess programs have always favored materiality. You know, the joke within the chess circles is that, you know, chess computer sees a pawn and then grabs the pawn because it loves material, because it gets +1 in its evaluation function, and then it tries to hang on for dear life in a, you know, really ugly position, but it wins because it never makes any tactical mistakes. So it's sort of very effective, but it's a little bit sort of aesthetically, you know, unsatisfying, one would say, as a style. But instead of that, actually AlphaZero does the opposite. It loves sacrificing pieces, material, to get mobility, to get more mobility for its remaining pieces. And this, this— so this is a game from— we did a, you know, 100-match between AlphaZero and Stockfish, and this is one of the very— and then we gave it to some— the British chess champion, actually, to analyze. And he picked out like the coolest positions. And this is my favorite. It's sometimes called the immortal Zugzwang game. Zugzwang is a phrase in chess, German phrase, that means any move that one makes in that position makes your position worse. So it's a special type of position where you're in Zugzwang, which means anything you do, it's going to make it worse, which is very unusual. And it's super unusual in this kind of position, for those of you who know chess, where Black, which has got more pieces, 2 rooks and a queen, So, it's got a big material advantage, very powerful pieces, the most powerful pieces remaining in chess, but they're all stuck in the corner, and AlphaZero sort of sealed them up, you know, with cement with its pieces, and basically, none of those pieces can move, right? So, this is kind of an incredible position. So, almost anything Black does, in fact, anything Black does in this position, it's Black's move, will make its position worse, even though it's got all of these very powerful pieces. So, that was one innovation. There are lots of interesting properties about AlphaZero that I won't go into, but one can think about, well, why is it that AlphaZero plays like this and traditional chess engines didn't? Nowadays, actually, interestingly, they've updated Stockfish to include some of these ideas by hand in Stockfish, and actually now it's very powerful, it's even more powerful. So it's kind of interesting hybrid system. But my feeling is that it's better evaluating positions than chess engines. So, that's one thing. So, it's got a better evaluation function. Um, and the main thing is it doesn't have to overcome these inbuilt rules. That's why it likes sacrificing pieces, because, uh, if you think about it, a hard-coded chess engine would have to calculate in its search tree that if it was gonna sacrifice a rook for a bishop, you know, that's minus 2 points, is it going to get back those 2 points of value, right, within its search tree horizon? AlphaZero doesn't have to worry about that, because there's no rules like that in there. It can evaluate things contextually based on the particular situation at hand and the patterns involved there. And also, the other big thing is Stockfish and programs like that, they have thousands of handcrafted rules. So, one problem is generating those rules, but an even bigger problem, in my opinion, is balancing those factors together, right? That's a huge sort of handcrafted juggling act. And instead of that, obviously AlphaZero learns itself how to balance out the factors that it's learned, and to do that automatically. So, one can actually see how efficient this system is based on the amount of search that traditional search engines have to do per each move they make. And a human grandmaster makes only the order of like— looks at about 100 moves per decision. So, incredibly efficient with our models. Then the state-of-the-art chess engine like Stockfish would make tens of millions of evaluations per move, and AlphaZero was, you know, sort of in the middle here in terms of orders of magnitude, tens of thousands of moves. So not as efficient as human, top human players, but far more efficient than the search one would get in, in the search engines. So again, if you're interested in the details about all your chess players and the details about what this changed, The British champion and Natasha Ragan wrote an amazing book called Game Changer, and when we gave them behind-the-scenes access to AlphaZero and what new motifs they found, at least a dozen new motifs they found in chess. And the cool thing is, that was very gratifying for me, is that people like Magnus Carlsen, who's the current world champion, incredible player, you know, he said a few years back he was one of the first people to read the book and who we sent it to. And, you know, I've been influenced by my heroes recently, one of which is AlphaZero, which is really cool to say. And he actually incorporated— he's so talented, he was able to quite quickly, quicker than all the other chess players, incorporate some of these ideas into his play. And then Garry Kasparov, he used to be a hero of mine when he was world champion when I was growing up and playing chess. You know, he wrote the foreword for the book, and he said programs usually reflect priorities and prejudices of programmers, but AlphaZero learns for itself, and I would say its style reflects the truth, which is, you know, I think a beautiful quote. So, we've been lucky enough to have several of these sort of fundamental breakthroughs in games. We started with Atari and our program called DQN, being able to play Atari games directly from pixels and maximize the score just from pixels, not being told the rules of the game. AlphaGo and AlphaZero, I just mentioned. And then we went further with programs like AlphaStar, which played the most complex video game called StarCraft II, which is a very complicated real-time strategy game with huge other challenges. It's only partially observable. It's not perfect information. There's an economy system to it. And you have generally thousands of possible actions you can take for any choice, not a few dozen. And we managed to also get to grandmaster level at that. So that was all of our games work, but really it was leading up to this moment, which in the last couple of years has been just so exciting and so gratifying for us to make progress with, which is that the games— and I love games, always will love games, playing them, designing them, and using them as testing grounds— they were the perfect testing ground for developing AI. But ultimately, the aim was not to play games to world championship level, it was to build general systems that could generalize and solve real-world problems. And the one that's particularly passionate for me is using AI for scientific discovery. And there are 3 things that I look for when currently, when we want to select a scientific problem that we believe our systems could be good at. So, number 1 is we actually search out massive combinatorial search spaces or state spaces, So, the bigger the better, actually. Why is that? Well, because we know then traditional methods and exhaustive brute force methods won't work, right? So, we're in a regime where something else is needed, and we think that we're good at that something else. Number 2 is that we want to have, you know, we like problems that have a clear objective function or metric that one can specify so that you can optimize and hill climb against it with your learning system. And then, number 3 is we look for problems that either have a lot of data, available to learn and train from, or— and ideally it's and/or— an accurate and efficient simulator that one can use to generate more data. And that simulator doesn't have to be perfect, it just has to be good enough that you can extract some signal from the data that it generates. Now, it turns out that when you look at a lot of problems with this prism, then actually a surprising number of problems can be made to fit these criteria. And of course, the number one thing we were looking at was protein folding, which I want to talk a bit about now. And, and we look for problems, not only they just fit those three criteria, but of course there's always an opportunity cost when you embark on applying AI to something major. It's going to take you many years depending on how hard that problem is. And we look for something that will have really huge impact. Perhaps, you know, we sometimes talk about root nodes that can open up whole new branches of scientific discovery if they were to be solved. And protein folding ticked all of those boxes. So for those who don't know what protein folding is, it's this classic problem of can one go from a one-dimensional amino acid sequence, you can think of it as the genetic sequence for a protein that describes a protein coded by the genome, and can you predict from that directly the 3D structure of the protein in your body, the 3D form that it takes. And the reason this is important is that proteins are basically essential for everything in life, every function in your body, and it's thought that the 3D structure of the protein, in at least in a large part, governs its function. So if one can understand the structure, then one can get closer to the function of the protein. Now, until AlphaFold came along, the way you would do this is experimentally, and it's extremely painstaking expert work that needs to be done, and using, you know, X-ray crystallography and electron microscopy. And the rule of thumb is generally that it takes one PhD student their whole PhD to do one protein, right? And that's if you get lucky. You can be unlucky. So it's hard and really painstaking and difficult. And what happened is that the Nobel Prize winner Christian Andersen, in part of his Nobel lecture in 1972, so 50 years ago exactly now, he conjectured that the 3D structure of protein should be fully determined by the amino acid sequence. I.e., this should be possible, this mapping. And it's a bit like, you know, sometimes this problem is called Fermat's Last— like Fermat's Last Theorem equivalent in biology. Because it's a bit like saying, this is possible, but the margin's too small, can't give you the answer. And so what happened instead is, obviously, it set off a 50-year quest in biology, in computational biology, to try and solve this problem. And it's been, you know, ongoing ever since the 1970s. So the big question is, is can protein structure prediction, the protein structure prediction problem, which is the specific part of protein folding that we're interested in, be solved computationally, just computationally. And Leventhal, who is another famous contemporary of van Veensten, he, in the '60s and '70s as well, he calculated, back of envelope, that there would be roughly 10 to the 300 possible conformations, shapes, of an average-sized protein that it could take, right? So 10 to the 300, so that's a good number, that's ones we like, because, you know, it's bigger than Go. And obviously, that means exhaustively sampling this is totally intractable. But, of course, the, the, the, the, the, the, the chink of light is that, in nature, in our bodies, physics solves this, right? So, it's, it can, if proteins spontaneously fold in a matter of seconds, sometimes milliseconds, in the body. So, there's obviously some energy path through this. So, how do we get to this problem? Well, actually, it's quite a long winding road for me, personally. For others in the team, less so. But for me, I actually came across the protein folding problem in the '90s as an undergrad in Cambridge, because one of my friends in our sort of group of colleagues was obsessed with this problem. And he would talk about it, I remember this very clearly, every opportunity in the bar, playing pool, whatever it was. You know, if we can crack this, that will open up, you know, all sorts of things in biology. And I sort of listened to him, and I was thinking about this. I was fascinated by the problem as a problem, and I felt it was actually very well suited to, potentially, to AI, although obviously at the time I didn't know how it could be tackled. But I filed that away as an interesting thing, and then it came up again in the late 2000s when I was doing my postdoc over at MIT, and this game called Foldit came out from David Baker's lab, who works on proteins,, and it was a citizen science game. You can see it on the left here, and what they've done, really interestingly, is turn protein folding into a puzzle game, right? And they actually got, you know, a couple hundred gamers to fold proteins, a bit like, you know, playing Tetris or something, and, um, and that some of them actually became really good, uh, and, and I remember, so of course I was fascinated at this just from a game design perspective, you know, wouldn't it be amazing if we could design more games where people played them they were actually doing useful science while they were having fun. That would be amazing. And I think this is still the best example of that. And, but also, you know, again, protein folding was coming up. And in fact, it turned out that a couple of, you know, more, a few really important protein structures were found this way by gamers and published in, you know, Nature and Nature Structural Biology. And so, this actually really worked. And that, when we then got to, you know, the third piece of the puzzle was doing Go,, and trying to sort of think about what we'd done with intuition and other things, as I mentioned earlier. And I felt that actually, you know, if we'd managed to mimic in some sense the intuition of Go players, master Go players who spent their entire lives studying Go, you know, maybe one could mimic the intuition of these gamers who were only, by the way, of course, amateur biologists, right? But somehow, some of them were able to make counterintuitive folds of the backbone that were, if you just followed an energy landscape in a greedy fashion, one would not, you know, reach a local minima or local maxima, and you would not be able to find the right structure. So, it's almost the day after we got back from Korea, we then, you know, I instigated the AlphaFold project, and I thought it was the right time to basically start working on this problem. The other important piece of the puzzle was this competition called CASP, which is sometimes thought of as like the Olympics for protein folding, and it's sort of run every 2 years, and external benchmarks— an amazing thing, actually, that I think more areas of science should do. And it's been run sort of religiously for every 2 years for nearly 30 years. So, you know, huge kudos to the organizers, John Moult and his team, for for doing this and organizing it so professionally for every 2 years without fail for 30 years. And the cool thing about it is it's a blind prediction assessment, so there's no way you can accidentally sort of train on test data or any of these kinds of pitfalls, because at the time when the competition runs over summer, usually every 2 years, the experimentalists globally agree to hold back a few of their structures that they've just found, but at that point in time, they're the only ones who know what that structure looks like. What it looks like. They hold back the publication for a couple of months, and they give it to John Mould and his colleagues to put it into the competition. And then you get those. It's quite a fun tournament, because then, you know, it's quite exciting. You get the email, and then there's a new structure that amino acid sequence nobody has ever, you know, knows the structure of, and then you have a week to sort of get it back to the competition organizers before it's published. And then at the end of that 3, 4-month period, they obviously score, your predictions against the ground truth, which at that point is published, obviously, in peer-reviewed journals, that the experimental ground truth. And then you get a kind of distance measure between your predictions and the molecules in that prediction and where they really are in 3D coordinate space. So when we started getting involved in this area post-2016, you know, we looked at CASP and the history of it, and actually there'd been very little progress for for over a decade. It's sort of the field had stalled. And this graph here shows you the scores of the winning team on the hardest category of protein, where you don't have any evolutionarily similar template proteins to sort of rely on. So it's called free modeling. And this is a percentage accuracy. It's called GDT. It's a slight nuance of the measure, but you can think of it as the number of molecules, the percentage number of molecules you've got roughly in the right place to a certain tolerance. Distance tolerance. And you can see they were hovering around 40% or less, which is useless for experimentation, right? Basically, it's pretty much random. And so that was the average, and it hadn't really moved. And so, um, what we did in 2018 is that, uh, we came along with AlphaFold 1 as our first entry after a couple of years of working on this. And we sort of, um, you know, I think we revolutionized the field in a way is that we, for the first time, we brought cutting-edge machine learning techniques, the sort of techniques we developed in AlphaGo, and other new ones for this domain, and we, as the core part of the system, and we improved the winning scores by 50%, you know, got to close to 60 GDT here. And then, of course, we didn't stop there. We then re-architected based on that knowledge. We actually tried to push that system further, and it turned out it hit a brick wall, so we had to go back, back to the drawing board with the knowledge that we had, re-architect it with a brand new system, and then that finally reached, in CAS14 in 2020, atomic accuracy. So accuracy within the width of an atom, right, for all the molecules. So when we look at the scores and the results of CAS14, what you see here is that AlphaFold 2— this is the root mean squared error— is less than 1 angstrom error on average, and, you know, from the 100 or so proteins that we're supposed to predict. So, and 1 angstrom is the, you know, the width of basically a carbon atom. So that's finally, that was the magic threshold that John Moll and others of the organizers said that they always set out CASP3 to do, because that would make you competitive with experimental techniques. Which are roughly, you know, the best ones are at that kind of error rate. So if one could do that computationally, then suddenly you have a technique that could be relied on in tandem with experiments or instead of. And so AlphaFold 2 got an error of 0.96 angstroms, which was 3 times more accurate than the next best system in CASP14, even though those systems obviously incorporated the AlphaFold 1 techniques. That we'd already published by then. So this led to the CASP organizers and John Moult declaring that the structure prediction problem had essentially been solved after all of these years. And this is what the predictions look like. So the ground truth is in green, and you can see the prediction from AlphaFold 2 in blue. And you can see, firstly, proteins are exquisitely beautiful. It's one thing to note that I've learned over the many years I've been working on this now. They're like exquisite little nanomachines. And you can see how accurate the overlays are. And we were astounded, of course, when we first got these results back. And then, you know, there were many— this is the architecture for AlphaFold 2, which I don't have time to go into the details of today. But there were a huge number of innovations that were required to make this work. And the key technical advances were basically First of all, I should say there was no silver bullet. It needed actually 32 component algorithms described in 60 pages of supplemental information actually in the paper. And that was required, and every single part of that was required. So we did these ablation analyses, which sort of took out components to see if we could get away without having them. And the result of that was everything was required. And the 3 key sort of takeaways of why AlphaFold 2 was an improvement over AlphaFold 1 is we made the system fully end-to-end. So, and you can think of it as sort of going end-to-end with a recycling iterative stage. So over time, it sort of jigs the protein structure nearer and closer and closer to the final structure that it's going to predict. And our AlphaFold 1 system didn't do that. It went from the amino acid sequence to this intermediate representation called a distogram, which is a pairwise distogram of all the protein molecules and their distance to each of the other molecules, the other N molecules. And then from that, we use a different method to create the 3D structure. But with AlphaFold 2, we actually made this end-to-end. So we went straight for predicting the 3D structure. And those of you who work in machine learning will know that generally speaking, if you can make something end-to-end and optimize directly for the thing that you're after, usually your system will be— will have better performance. Performance. We use an attention-based neural network to infer this implicit graph structure of the residues of the amino acid sequences. In AlphaFold 1, we used a convolutional neural net, which was sort of borrowed from computer vision. And if you think about it, that was introducing the wrong bias into protein folding, because with computer vision, you know, pixels next to each other are obviously going to be correlated in an image, right, in some sense. So convolutions make sense. But actually, for a protein, the amino acid sequence, you know, residues that are next to each other or close to each other on the string of letters may not end up being near each other once you get the full 3D fold, or things very far away could end up folding over near each other. So in a way, we were giving it the wrong biases, so we actually had to remove that. And then finally, we built in some biological and evolutionary and physics constraints into the system, without impacting the learning. And again, usually— so you can think of it as a little bit of a hybrid system that usually, if you put in constraints, that impacts the learning. And we managed to do that without that. So this was a huge research effort over sort of 5 years, took about 20 people at its maximum, and it was a truly multidisciplinary effort. So we needed biologists and physicists and chemists as well as machine learners. And I think that's an interesting lesson maybe to learn about cross-disciplinary work in AI for sciences is you need the experts also from the domain. And then the final maybe interesting point to note on this is that normally we're always after generality. So, you can see that from the journey from AlphaGo to AlphaZero was we increasingly made things general, right? You start with performance, then you start throwing things out of that system to try and make it simpler and more elegant, and that usually makes it more general. As you understand what it is that you're doing. But that's because Go and chess and those things were, were testbeds for what we wanted to do. If you are trying to solve a real-world problem that really matters to other scientists, or, or health, or in this case, you know, biology, then actually you need to, you know, you might as well throw the kitchen sink at it, right? Because you actually are really after the, the output itself, in this case protein structures. And that's what we did here. We really threw everything we had at it. And it's, I think, the most complex system that we've ever built. Other things to note about this system is that it's also— AlphaFold 1 was relatively slow, took a few weeks of compute time to do a protein. AlphaFold 2 took 2 weeks to train the whole system on a relatively modest setup of 8 TPUs or 150 GPUs, which by modern-day machine learning standards is quite small. And then the inference, the predictions, can be done lightning fast, you know, order of minutes, sometimes seconds, on— for an average protein on a single GPU. So when we did this, AlphaFold 2, we announced the results, published the methods over Christmas. That Christmas— this is back 2020— we were thinking, okay, how should we give access to the system to biologists around the world? And normally what you do is that you set up a server, people, biologists, send you their amino acid sequences, and then you give back, you know, a few days later, you might give them back the prediction. But actually, what we realized, because AlphaFold 2 was so fast, we could actually just fold everything ourselves in one go, right? So we just fold all proteins, and we'll start with, you know, the human proteome, which is like the human genome equivalent, but in protein space. And so that's what we did over the Christmas. We folded the whole human proteome. And so it's another thing I love about AI and computing is, you know, you can have your Christmas lunch and while you're doing that, AlphaFold AI is doing something useful for the world. So the human proteome. So we published that as well in the summer of '21, last summer. So AlphaFold 2, we predicted that every protein in the human body, it's around 20,000 proteins, represented, obviously expressed by the human genome. And at the point where we did this experiment, experiments, 30 years of experiments, 30, 40 years of experiments had covered about 17% of the human proteome, right? And we more than doubled that overnight in terms of very high accuracy structures. Obviously, we folded all of them, but very high accuracy, so that's less than 1 angstrom error, you know, they're sort of up to experimental quality, we went to 36%, and 58% at high accuracy, where we call high accuracy when the backbone is mostly, you know, where you can be confident in, but the side chains may be slightly out. And then, of course, the question is, what about the rest, the other 42%? And it may be that some of those, you know, AlphaFold 2 is just bad at, but increasingly, and this is an open research question, when we look at it with biologists, and biologists often and send us in results, and it's like, oh, look at this one, folded really well, or this one didn't fold well. We often find that the ones that didn't fold well were actually what's called unstructured in isolation. So they're disordered, intrinsically disordered proteins, which means that until you know what they interact with, they're basically squiggly bits of string. And then, presumably, when they interact with something in the body, they then, another protein usually, they'll then form a shape. But we don't know what that shape is in isolation, right? We may not even know what it interacts with at this stage. So actually, people have turned this around now to use it as a, as a disordered protein predictor. So where AlphaFold doesn't do well, perhaps that's pretty good evidence that it's a disordered protein, which of course is very important in things like disease, you know, Alzheimer's, other things are thought to be to do with badly folded or disordered proteins. One of the other things we did, which was a nice innovation for AlphaFold, was have the system predict its own confidence in its own predictions. And the reason we did this is we wanted biologists to use this who maybe would not care about the machine learning techniques or not understand them, or frankly, it would be irrelevant to them. They would just be interested in the structure. And we wanted to make sure that they were easily able to evaluate the quality of that prediction and what parts of it they could rely on, right, and which other parts they maybe need to check experimentally. And so what we did is AlphaFold, basically, we produced predictions that were split into 3 thresholds. Over 90 was what we call very high accuracy, so less than 1 angstrom error, experimental quality. Greater than 70 was the backbone's correct, and then less than 50 may be these red regions. So you can see in the database, that's what they look like. Is something that should not be trusted. We did a further 20 model organisms covering all of the critical model organisms used in research, and some also, some important other ones in disease, like tuberculosis, and also agriculture, like wheat and rice. And a lot of these, these genome, sorry, proteomes are much less covered than the human proteome, right? Of course, the human one has been where the most effort's been. That's at 17%. For some of these organisms, it's like less than 1%. So for the researchers in those plant scientists and other things, you know, this is a huge boon for them because they would never have the resources to spend that time to crystallize the proteins they're interested in. We then teamed up with EMBL-EBI, the European Bioinformatics Institute at Cambridge, and they're amazing as a partnership team. They host a lot of the biggest databases around the world already, and we thought, well, the best way to to host all this data, is to just give it to them and allow them to host it and plug it into the mainstream of biology tools. And so we had a great collaboration with them, and then we basically released all this data for free and unrestricted access for any use, industrial or academic, because it's so completely free. And, you know, it's amazing to see the impact of that, and we've tried to sort of maximize the scientific impact of this by releasing it in that way. The other thing we did do, and I want to touch on this at the end, is think about the safety and ethics of this. And we consulted with, you know, over 30 experts in various areas of biology, bioinformatics, biosecurity, and pharma to check that this was going to be okay to release this type of information. And they all came back with that they were not worried about this, but they were potentially worried about future things. So that's something that we bear in mind. There are now a million predictions in the database today. We— I just want to call out one thing, is we especially ourselves, we especially prioritize neglected tropical diseases because those are the ones that affect the developing world, the poorest people in the world, the most, and they're the least researched because of course there's no money in it for pharma companies. So they're often— it's NGOs and nonprofits that have to do the work there. So for them, it's amazing to get all the structures because they can go straight to drug discovery. Without having to go to the intermediate step of finding these structures. So we prioritized all these diseases, and including ones that we've been given from the WHO about potential future pathogens. And what's the community done with AlphaFold already? We've seen just in 9 months or 10 months, incredible amount of work has been done. This is really cool on the left here with some colleagues at EMBL. They used AlphaFold in an experiment to combine with their experimental data to put together what's called the nuclear pore complex, which is one of the biggest proteins in the body. It's massive for a protein. And what it is, is it's a little gateway into the nucleus of your cell, and it opens and closes to let things in. And they were able to, you know, it's beautiful if you look at it, able to put it all together and then visualize it. I talked about this disorder predictor, WHO top 30 pathogens, and actually, interestingly, it's helped experimentalists are the ones that benefited first from this. Because they can combine this with their maybe some low-resolution images they have. And if they have two sources of information, they can then make a sharp prediction from their maybe their slightly lower-resolution experimental data and then a computational prediction. So, you know, it's been really gratifying to see hundreds of papers now and applications already with being used for AlphaFold, also in industry too for drug discovery. So what has the impact been? So, you know, we already have 500,000 researchers have used the database. We think that's almost every biologist in the world has probably looked up their proteins they're interested in. 190 countries, 1.5 million structures viewed, and already over 3,000 citations, and we've had some nice accolades along the way from Science and Nature on the method. And then over the next year, we plan to fold every protein, you know, known to science, which is in UniProt, which is the massive database that has all the genetic sequences, and there's over 100 million proteins known to science, and we're steadily sort of progressing through that right now, and we'll be releasing that over time. So stepping back then, what does this mean? I think that maybe, you know, we're entering a new era of what I would like to call digital biology. So I think, the way I think about biology is that at its most fundamental level, It's an information processing system, albeit an exquisitely complex and emergent one. And I think of it as maybe the potential, the perfect sort of regime for AI to be useful in, because, you know, one thing I think of it analogous to is in physics, you know, we use mathematics to describe physical phenomena, and it's been extraordinarily successful in doing that. Of course, mathematics can also be applied to biology, and has been applied successfully in many domains, but I think a lot of these emergent and complex phenomena are just too complicated to be described with a few equations, right? I just don't really see how you can, say, come up with, you know, Kepler's laws of motion just from— of a cell, right? How would one do that? You know, just a few differential equations doesn't seem to me likely, and I think maybe a learned model is a better way to approach that, and I think And I hope that AlphaFold is a proof of concept that this may be possible, and maybe can help usher in this new dawn of digital biology. And our attempts to go further in that space is obviously we are researching further at DeepMind and the science team. We sort of doubled down on all these things within the biology team at DeepMind. And we've also spun out a new company, Isomorphic Labs, to specifically build on this work and other related work, but specifically for drug discovery to accelerate. Accelerate drug discovery, which we hope, using computational and AI methods, can maybe be an order of magnitude quicker. Currently, you know, it takes an average of 10 years to go from identifying a target to a candidate drug. So just to start closing then, I just, you know, there isn't time to go into this, but it's, for us, it's been like a renaissance year in some sense. I've been having so much fun ticking off all of the, my sort of childhood dream projects in fusion and quantum chemistry and conjectures in maths, material science, weather prediction. This has all become reality now in the last year of applying it to important problems in each of these domains and, you know, publishing nice and important work in each of these areas. And in applications, of course, there are lots of amazing industrial applications that we've been doing, and we have an applied team at DeepMind that works with Google product teams to incorporate all of our research into hundreds of products now at Google. Probably pretty much every product you use of Google's will have some DeepMind technology in it. Some of the ones I just want to call out are our data center work and energy optimization of data centers and the energy they use and the cooling systems they use. And we're looking at applying that to grid scale now. WaveNet, which is the best text-to-speech system in the world. So any, any device that you talk to that talks back to you will be using WaveNet to have super, you know, really realistic voices. Even, you know, interesting things like better video compression for YouTube. We can save 4% of the bitrate that is used to— but whilst maintaining video quality, and also things like recommendation systems. But there's just too many to mention, actually. And then, of course, very in vogue now, and we have a ton of work on this area, but it'd be a whole talk in itself, is large models. And we have our own really cool large models, AlphaCode, that can program from a text description and write code. Still amazing to me, in competitive programming level. Chinchilla, which is our large language model that uses— is compute efficient. Flamingo, that's our vision-language combined model that can describe images. And then Gato, our latest model that is super general, can do robotics, video games, all sorts of things, language, just with one model. So, this is all very exciting, but I just want to end my last couple of slides with a bit about ethics, because obviously, you know, this is hosted by the Institute of Ethics, and it's a very important topic, and not just because of that, but, and it's also what the Tanner Lectures are about too. And so, we think a lot about pioneering responsibly. This is actually two of our values at DeepMind combined, you know, pioneering and being responsible. And, you know, I hope I've convinced you, and I hope you will realize that AI has this incredible potential to help with some of humanity's greatest challenges. You know, I think disease, climate, all of these things could be in scope. But obviously AI has to be built responsibly and safely, and we have to make sure the people who are building these things, it's used for the benefit of everyone. So we've had this sort of front of mind from the beginning of DeepMind. And as with any powerful technology, and I think AI is no different, although it may be more general, more powerful than any that has gone before. Whether or not it's beneficial or harmful to us in society depends on how we deploy it and how we use it and what sorts of things we decide to use it for. And I think it's important that we have a really wide debate about that at places like this and the Institute of Ethics. I'm very excited to see that being set up and for us to interact with the new institute. And here, just one mention is that D&I has been really critical, and I think— and we've been pushing very hard on this the last few years, and I think is critical to this. This to make sure we get the broadest possible input into the design and deployment decisions of these systems, especially for the people that this affects the most, that these systems affect the most. And that's something we've been pushing very hard on. There's still a lot more work to do, but we've been making some good progress at DeepMind. And we've been also doing that with all of our sponsorship that we do. We've now done nearly $50 million worth of sponsorship of scholarships, diversity scholarships, chairs and academic institutions and projects, and also funding things like the Deep Learning in Darba, which is Africa's biggest conference on machine learning. I'm really proud to say that a lot of DeepMinders helped set that up. And so there's many, many things that we're doing across the industry that we hope also can act as a role model for the rest of industry. So then on ethics and safety, this has always been central to our mission because you saw our audacious mission at the start, and we, even back in 2010 in our little, you know, attic room, we were planning for success. And of course, what is it, you know, we had to think through as scientists, what does success mean? What will the world look like? And obviously, if one thinks that through, and it's becoming obvious now in 2022, that, but it was obvious to us then in 2010, that this, we would have to be critical, that it would be really important questions that would have be addressed. And part of that— so we've been doing this in the background all along, and we'll be talking more about this work probably in future. We were instrumental in drafting Google's AI principles, which are now publicly available, and they were partly based on our original ethics charter that we've had from the very beginning of DeepMind. And the aim of these principles— and, you know, you can look them up later if you want to look at what they say— is obviously to help realize the far-ranging benefits that clearly AI could have for everyone on whilst identifying and mitigating potential risks and harms ahead of time. And we continue to try and act as thought leadership for the AI community on many of these topics— strategy, risk, ethics, and safety. So, what should we do then? And I just want to end with this last slide here, is what I think we should not do is move fast and break things, which is, you know, sort of the Silicon Valley trope, right? And I think we've seen the consequence of that playing out, right? It can be very extraordinarily effective to get, you know, powerful systems and growth and other things. But I do not think it's the right way to address really powerful dual-use, potential dual-use technologies like AI. And the problem with it is, is that, you know, one, one of the things that falls out of moving fast and break things is actually doing live A/B testing in the world., right, with your minimum viable products and other things. And of course, the question is, if one does that, where does the, you know, and then, you know, option B turns out to be a terrible option, right? Well, where does the harm of that happen? Well, it resides in society, doesn't it? They pay that, that pays the cost of your learning because you've done it in the world. And it's probably fine if you're just, you know, doing a little gaming app or photo app or something, but we already see with social networks It's not fine when you're at billion-user scale and, you know, things really matter, right, in terms of like your A/B testing. I don't think it's responsible to do that. So, what should we do instead? Well, fortunately, we already have another method which I think would be better, the scientific method, which I do think is probably maybe humanity's greatest idea ever, and I think it can apply here. And I think we should use the scientific method when we're approaching how to deal with these very powerful, incredible potential technologies. And what does the scientific method involve here in this domain? Well, it's sort of thoughtful deliberation and thought ahead of time, and foresight ahead of time, where you— hypothesis generation on what might happen if one were to be successful with what you're trying to do, right? So how about we think about that ahead of time, not afterwards? Then there's rigorous, you know, and careful and controlled testing. I think that's one of the main things I learned from my PhD, apart from all the neuroscience, was also the value of controlled tests. I don't think you can really understand— like, in a way, I think when I started my PhD, at least, I was all about the, you know, what's the condition of interest, and that's the thing that you're gonna make, you know, your new advance with. But actually, you can't conclude anything, of course, unless you have good controls., and I think that's something I don't think engineers get first time around, actually, but, you know, scientists and researchers, of course, do get that, because that's one of the things that you learn from doing a research PhD. So, controlled testing in controlled environments, not out, you know, in the world, until you better understand what it is that you're doing. So, you know, of course, one updates on empirical data, obviously, ideally with peer review, so you get critique from the outside, and people who are independent from your work. All of these things are standard in the scientific method, right, but are not standard in engineering. And that all of this is in service of getting a better understanding of the system before one deploys it at scale, right? And then maybe you find out something. So my view is that as we approach artificial general intelligence, and it's a super exciting moment in time, as you can hopefully you know, get from my talk and my excitement over that, but we need to treat it with the respect and precaution that— and sort of humbleness, I would say— that the technology of this magnitude demands. And I think that's what we are trying to be at the forefront on, and I think I'll be talking a lot more about this in future. So I'll just end by going back to the science question. I think if we get AI right, it could potentially be the greatest and most beneficial technology humanity's ever invented. And I think of AI as this ultimate general-purpose tool to help us as scientists understand the universe better, and perhaps our place in it.

Thank you. Well, thank you, Demis, for that extraordinary tour de force. We do have a little time for questions, but we wanted to give you the chance to kind of give us that sense of your vision. Now, we've got an opportunity for questions from the audience. Got to wait for the microphone to be handed to them and to stand up if possible when asking questions, but I'm afraid I'm afraid there is a kind of discrimination. It's only those on the ground floor that can ask a question due to health and safety policies in the theatre. So please, if you have a question, please raise your hand and I'm happy to take questions at this point. So could I— yes, John. Perhaps I'll start with John. I'll give you the privilege. There is a roving microphone.

And, uh, just declare who you are, John, and perhaps stand up and just ask your question. Interesting to begin an ethics talk with some discrimination, Nigel. But, um, I'm John Tessoulas. I'm the director of the Institute for Ethics in AI. Thanks so much for a really fascinating and inspirational talk. I guess I want to ask two questions. One's a very general question about the nature of the project you're embarked on. So the objective is to generate powerful all-purpose tool that will help create new scientific understanding. And the nature of this tool is artificial general intelligence. So that is a tool that can replicate or outperform human beings across a wide range of cognitive tasks. The worry is, is there a tension there? If you had something that could outperform human beings across a wide range of cognitive tasks, could we still regard that as a tool, or would it become a colleague? So you talked about respecting AI at the end, but it looks like something with that level of capacity would demand a different form of respect that would preclude the original objective of now treating it as a tool. So that's one question. The second question is, you've talked about what will benefit humanity, And so I guess one question I have is along these lines: how do you make that determination? So you might say, look, some people have the view that AI applied to military applications will benefit humanity. Others don't. How do you make that determination? And I guess there's also this further dimension. There's a division of labor in making that assessment. Do you think too much has been placed on the shoulders of developers, researchers, corporations, and that really government should step in and resolve some of these issues?

Thanks, John. You know, great question. So I think, you know, with your first question, the reason human capabilities are an interesting mapping is because the human brain is the only evidence of general intelligence we have in the universe, as far as we know. So, so I think I think, you know, there's always the question is how do you know you've got there? And so, and you can approximate it with millions of tasks potentially. So, that's one approach. The more tasks you have in your grab bag and it can do all of them and compare it against human performance, you might have done it. But you— there's always the possibility that might— one might have missed out a particular type of cognitive ability like creativity or something. So, that's why I think, and also I think AI can be applied back in neuroscience as well, by the way. That's one of our scientific areas that we apply AI to is neuroscience itself and better understanding our own minds. So I have this view that as a neuroscientist, that this journey we're embarked on with AI will be— is the most fascinating journey one can ever take scientifically, because there's not only the artificial building, it's then comparing that to the human mind and then seeing, I think, uncovering the mysteries of our own minds. You know, what's dreaming, what is creativity, what are emotions, all of these questions that we have, free will, potentially even consciousness, the big questions. I think building AI and intelligent artifacts and then seeing what is missing in them is a good way to explore that scientifically. And so then, I don't know the answer to your question. I think that's part of this journey, is at what point would these things not become just tools? And it may even be that it's a design question because whether we should build, you know, what is consciousness, we don't know. And that would be a whole, obviously, debate in itself. But should we build it to the extent of what it is? Should we build them in our systems? I would say no to begin with if we have that choice until we better understand them as tools, and then we can bring in that extra complexity of free will. And you know, who— where do they get their goals from? Right. Initially it will be designers, but if they could be self-generated, So I think we're still a long way away from those things, but I think we— that's one of the things I think we should inch towards very cautiously and with precautions, because also it will get to the heart of what it means to be human. So, and I think that should exactly be done multidisciplinary with philosophers and ethicists and theologians and the wider, you know, humanities. I think this is where the humanities comes in. As well as the science. So, yeah, so I think that's, you know, that's all to come.

Okay, question in the front row.

Thank you so much for a great presentation. Karina Prunkel, I'm a research fellow at the institute. So you mentioned at various points the potential for dual use and the particular malicious dual use. So I'm curious to hear how you approach this topic at DeepMind. So what precautions, or how do you address the potential for dual use?

Yeah, great. Um, yes, so we have a lot of different, um, mechanisms now at DeepMind that have been built up over time. So, uh, one is the institutional review committee we have, which is formed— so it's chaired by Lila Ibrahim, our COO, and It's formed with different people from across the company, our, you know, legal. We have ethicists and philosophers as well. It's also rotating board, some senior researchers, and they get involved early with research projects and try to assess them from all aspects. And they will draw on outside experts. So, they bring in biologists, for example, for AlphaFold, bioethicists, so things we might not have in-house. And then they work with the research teams to, you know, either say no, that project should not proceed. Okay, it can with caveats, or why don't you build or do it in a different way with these safeguards? So that's our prototype, I would say, committee that does these things. And we're kind of exercising our muscle when the stakes are relatively low currently so that we can learn from what works and is effective as we get more powerful systems. And obviously over time, I think at some point There have got to be outside bodies that get involved. But the problem is, is that— is, and we've experimented with that too— is that a lot of these things are very specific to the technology itself. So one has to sort of, you know, understand the technology at a deep level, maybe even have access to it somehow, but in a controlled way, because one can't just— you know, open sourcing is not just a panacea either, because if it's a dangerous system, open sourcing it means any bad actor can can use it too for anything. So there's a lot of complicated, I think, ethical questions around this. I don't think there's an easy answer. So anyone who thinks there is one, I think, is kidding themselves. I hope everyone realizes the complexity involved. But I think it's pretty— I'm very happy with our internal system, but I appreciate more is going to be needed than that as the systems get more powerful and impact more of the world.

A question just behind you, I think, if you just pass the microphone literally behind you.

Thank you. Hi, my name is Ulrich. I'm a postdoc at the computer science department in the human-centered computing group. So DeepMind looks like it's this great example of how you can take the best from science and then sort of bring it together with a commercial company and then make very rapid progress. And you mentioned in the end here how you thought that the scientific process should sort of inspire the commercial world, as it were. I'm curious about what you think about the other way around. So what have you learned by being sort of embedded in Google that you think we as researchers should sort of learn from in order to make more rapid progress?

Yeah, you're absolutely right. That was the, that was the thinking, the original vision behind the org. So I spoke about the original vision of the, of the company, this Apollo program, but the original vision behind the organizational setup and processes was to be a hybrid, like the best of both worlds. Startups and the energy and creativity and pace that they have and nimbleness, and the best from academic research, you know, the blue sky thinking, ambitious thinking that happens there, but sometimes with a lot of bureaucracy, right? So I think that we did actually successfully combine those two things. And then when we agreed to get acquired, we combined it with a third thing, which is scale and resources of a, you know, large, very successful company like Google. And I think that's the main lesson, is to make sure you do things at huge impact and have, um, the ambition and realize that, you know, you can scale things to that and the consequences that come with that, but also the potential of that. So I think we've done that now, um, and very well, like marry all three of those aspects together. Uh, it's a daily challenge because as we get bigger one tends to get slower as an organization. So I have to, you know, we have to fight against that all the time. But it's pretty unique, I would say, the, you know, the organizational and cultural feel of DeepMind. But it could be a blueprint for other, I would say, grand projects could be organized in a similar way.

Okay, I'm going to just switch to this side, then I've got a question there and a question about that. So Tim, yeah. So, uh, move fast and break things was a quote from some people who built a social network. If DeepMind was to build a social, uh, social network using the DeepMind way of doing things, not, uh, then what metrics would you use, would you optimize for your— to judge the quality of your social network? And the second question that comes with it Do you in fact have a moral obligation to build that social network?

Wow, okay, so two— thanks, Tim— I mean, two, two complicated questions there. I haven't— it's, it's actually just generally— so let's see, I have to be careful what I say, but I think, you know, social networks have never really been my thing, first of all, so I haven't, I haven't really thought a lot about it relative to scientific advances and the sorts of things that are my personal passion. I would question actually the premise of your question, which is that do we— how much value does weak ties like that give, like a sort of superficial connections like that, versus deeper ties that you get in real life with your real family and friends? I think is an interesting thing to understand. Like, are we sacrificing deeper, more meaningful moments for hundreds of more superficial moments. It's not entirely clear to me that the metric of, you know, it sounds seductive, connect the world, right? Like, why would that be bad? But this is the thing I'm talking about with the scientific method, is to try and think through the full consequences of what that would mean—echo chambers, you know, manipulation, all the rest of it—that we all know very well. I don't need to go into. So I think if I was to do something like that, I would, you know, use the scientific method again to try and really think through think through ahead of time, you know, what do you want as the outcomes and metrics. In fact, often trying to find the right metrics that actually drive the right behavior that you think is good in the world is half the challenge. It's like asking the right question in science. Everybody who does science knows that asking the question is the hardest thing. What is the right question? And it's especially hard— oh, you want an answer? Well, I wouldn't want to give you an answer on the spot, but we can, you know, talk about it over dinner. But I, I, I, at least one should attempt to start with serious thinking about the question first, right? That's the first part. I don't know what the answer is because I've not given it enough thought, but one should at least understand the meta-level of like, that's how one should start, including whether one should do that thing at all.

Potentially could be the answer of, of that hypothesis generation. Okay, I'm going to try and get 3 more questions in. We're right up against the clock, got about 7 and a half minutes.

There was a question here from Helen. Thank you, I'm Helen Landemore from Yale University and visiting fellow at the Research Center for Ethics in AI. Thank you for a brilliant talk. So you showed us how AI can help us figure out the truth of the universe, pretty much. How about the moral world? How about the political universe? Philosophy starts with Plato's Republic, which is an attempt to figure out the best constitution Surely, unless one is a complete moral relativist, there are some invariants we're trying to figure out about the moral world. Could AI help us map that out? Could it figure out the best social organization, you know, borrowing from, I don't know, all the things we've tried— capitalism, socialism, libertarianism, egalitarianism? Would it help expand our imagination and perhaps, assuming you have an objective function like satisfying majoritarian preferences subject to constraints to protect minority rights or something like that. What do you see in the future? We took 2,000 years and we haven't made much progress.

Good question. I mean, look, I think morality and political science, I think, is one of the hardest things that AI, you know, I think it can contribute in some way, but I would say it's far harder than the physical sciences, right, or the life sciences, because the most complex things in the, in the world are humans, human beings, to understand and to, to model and to, and to understand people's motivations, especially in aggregate. I think one way it could help is, um, there's also the question of even if an AI, theoretical AI, could come up with a better political construct, would humans, beings, and society accept that? Or even care or understand it? So there's all those questions to try and— and would it be implemented correctly? Obviously there's obviously implementation problems. I think more interesting maybe would be in— and I've talked to economists about this— is, and we did quite a lot of research on multi-agent systems. So again, having a little sandbox or simulation of millions of agents with interacting with each other with motivations and some goal-seeking things. And I think we're missing taking that experimental testbed actually from political science and economics quite a lot, because again economics is one of those things where, and political science, where you sort of have to test it live, A/B test it in the world. It's like, are we going to go for this political system or not? Should we raise inflation or not? Well, you've got models, but then you actually just have to do it and then see, oh, it's, you know, caused a recession or something where maybe we shouldn't do that next time. And so it would be better if, I think, if we had a simulation or a sandbox perhaps populated with AI systems that are approximates to, you know, idealized forms of humans. And then we can maybe make some interesting— we can do some interesting experimental work in that, much lower stakes. So I think that could be a really fascinating exploration area for things like market dynamics and setting the environmental settings to create, you know, more cooperation or something. I would be If I was an economist, I would be trying to use all those things. I used to be fascinated when I was a kid with Santa Fe Institute, and they used to do lots of really cool models of agent-based systems and little grid worlds. And I loved, you know, artificial— Growing Artificial Societies, I think, by Axelrod. I love those kind of work. Actually, I used to dream about going to Santa Fe to work on something like that. I still think that would be pretty cool to to have some sort of system like that.

Let's see if we can squeeze just a few more. There's a question, uh, chap in there who caught my eye there. Yes, you just— very— and try and squeeze them in because there's two more questions over here and I'm not going to get everybody's questions in.

You're going to have to come— super. Yeah, I just have a quick question, to be honest. So I think you at the end you mentioned to kind of creating AI in the image of scientific method and, and The title of your lecture is Advancement of Science Through AI, but in what sense do you think that neural networks, or the limited understanding I have of AI, is— in what sense do they follow the notion of scientific method we have? Do they also— is there any sense of talking about hypothesis and then testing? Because it doesn't seem that neural networks work in that way. They're opaque for most practical purposes. Sure. And if they do outperform us, should we just get rid of the scientific method?

Okay, that's a good question. Thank you. So, by the way, it's not in the image of scientific method, just to be clear, right? It's using the approach of the scientific method. I'm not sure what image in scientific method means. But, and yes, today that is true that a lot of the systems we have are kind of black box-like, but I think that's exactly what we should be doing more work on, is making them less opaque. There's no reason why they should be. The way I say it to my neuroscience team is, look, we understand quite a lot about the brain now, the ultimate black boxes. We have MRI machines and amazing tools and single-cell recording. So it's amazing, and that's why I got into neuroscience in the mid-2000s. So we can actually look into— we don't have to do philosophy of mind necessarily, although we should know about that, but we can actually, you know, empirically look at this, not just do introspection. And so as a minimum, we in the field of artificial minds, we should know as much about them as we do with the real brain. And we don't know everything about the real brain, obviously there's tons still we don't know, but still there's a lot more that we do know than we do about these artificial systems. And it should be the other way around, that should be the minimum we understand, because we have access to every neuron, you know, neuron, artificial neuron in the artificial brain, and we can completely control the experimental conditions conditions. So, as a minimum, what, you know, so I sometimes say this is a challenge to the team. What's the equivalent of fMRI for a neural network, right? What's the equivalent of single-cell recording? We do ablation studies. So, we have a whole neuroscience team that's thinking about this and bringing neuroscience techniques, analysis techniques, over to AI. Now, in the defense of the engineers, one of the reasons that this has happened is because the brain's obviously a static system we're all fascinated by, of course, right? But artificial systems change over time. Like AlphaGo is now in ancient history of AI, right? Although it was very meaningful at the time. So, and it takes years to study a system, right? Takes years to build it and then years to study it. So should you use that researcher time on studying a system that itself will be out of date by the time you come to any conclusions about it? So I think only now are we reaching the point where we have systems that are interesting enough, do enough interesting things in the world, like, you know, large models and AlphaFold-type things, that probably it's worth spending the researcher time on that. And so I think over the next decade we're going to see a lot more understanding of what these systems do.

I don't think there's some weird reason why that can't happen. Okay, there are so many more questions. I am literally in the red now. I'm going to have to call this to a close. I I do apologize. There is so much pent-up, I think, interest and questions for you, Demis. All I can say at this point is absolutely wonderful lecture. We're 5 minutes later than we should have been. The Sheldonian runs to a strict regime when it comes to timekeeping. You gave us the most fascinating insights, and you have given, I think, to the world with your company and your own talents, a quite wonderful vision of a future in which AI can help us flourish, empower us, and not oppress us.

So thank you very much.