Scientist Stories: Fei-Fei Li with Demis Hassabis, AI for Science

My name is Fei-Fei Li. I'm the co-director of Stanford Institute of Human-Centered AI. So welcome to Stanford High's headquarters space. It's truly an honor to be here. First of all, welcome President Marc Tessier-Lavingue for attending this. And our guest of honor doesn't really need any introduction, but nevertheless, I will get started here. We are very honored to host a talk and then later a discussion by Demis Hassabis, the founder and CEO of Alphabet DeepMind. Demis started his career in video gaming, developing new games at a few studios before launching his own. So for those of you who are taking away games from your kids, Think twice. Then Demis headed back to school for a PhD in cognitive neuroscience to better understand the architecture of brain at MIT, right? So through various places— MIT, Harvard, University College London— Demis has always been this very creative thinker and leader in thinking about the relationship between machine intelligence, human brain, and, and so on. And then in 2010, he co-founded DeepMind. You all know DeepMind. It is one of the first AI, truly AI companies of our era, and it was acquired by Google in 2014. DeepMind has made breakthroughs in data center energy consumption, generative in this case, degenerative eye conditions, and protein folding through the very impressive AlphaFold work that we all have heard of. It has made significant advances in the realm of deep learning and reinforcement learning. And true to them as gaming roots, DeepMind has also beaten a few world champions at at the very ancient game Go. Now DeepMind is looking to apply techniques from AlphaFold to nuclear fusion in hopes of halting our climate crisis. I hope we're going to hear more about that. And of course, DeepMind and Demis's teams are exploring ideas around— well, they have pioneered the term artificial general intelligence. Which Demis has called this an epoch-defining technology that will change the very fabric of human lives. And before I just turn the podium to Demis, I do want to say that when HAI was founded, Stanford HAI was founded back in 2019, Demis was part of our founding distinguished fellows. Acknowledging his profound impact in the field of AI. And I cannot think of a better person, a better name, a better leader to be here today to talk about AI, given all that is going on in the world. AI has truly had a public awakening moment, and it's no longer just a niche field that nerds like us play around. It's impacting human life, society, and our future. And HAI has been missioned to be one of the forums that will host this kind of intellectual discourse about AI and the societal impact. So especially to the audience students today, not only you're here for a treat to hear what Demis has to say, I hope you engage in the dialogue with him after this talk. So without further ado, Demis, please. Thank you.

Thanks so much, Fei-Fei, for that great, uh, very generous introduction. It's really amazing to be here, and thanks for inviting me. Um, as Fei-Fei mentioned, obviously the HAI was set up in, I think, 2019, and it's actually my first time to be here physically because of obviously COVID and all the travel restrictions. Everything got in the way. We've been meaning to do this for years to now finally I'm here and it's great to be here and I hope to be here a lot more often. I have a lot of reasons to come over to the Bay Area, so I hope to pop in here every now and again. So I'm going to cover my talk today trying to— there's sort of real passions of mine, which is to use AI to accelerate scientific discovery itself. And of course, at the end of the talk, I'll also discuss a lot about the current vogue of generative models and generative AI and large language models and the work we're doing in that part too. And maybe that can lead into some of the Q&A and discussion that we'll have afterwards. DeepMind was founded way back in 2010, almost like the middle, the medieval times now of AI. You know, this is 12, 13 years ago now, and it's been incredible to see We've been in this sort of very amazing position of being kind of at the forefront of where the whole AI field was going. And it's been sort of, we expected things to go sort of like this, but even for us, it's been amazing to actually just live through that. So back in 2010, it was very difficult to even, I remember trying to raise our seed round of, you know, a few hundred thousand dollars. It was almost impossible to do. And these days, you know, it seems like, $1 billion rounds are being done every other week. And that's just in just over a decade, a very short amount of time. And back then in 2010, if you can— some of you, most of you, not old enough to cast your minds back to that, but it was just nobody was talking about AI. Very few people were talking about AI and certainly not general AI. And it's certainly in academia where Shane Legg and I— Shane Legg's the chief scientist of DeepMind— were studying It was almost— you almost get eye rolls from professors and other things if you discussed about this sort of more general AI or the original aims of the AI field to build human-like intelligence. So it's been kind of astounding to see what's happened. And obviously things are changing incredibly fast, even month by month now. So we set up DeepMind in 2010 because we felt there was a lot of different techniques coming together that we could see deep learning had just been invented, and we've always been big proponents of reinforcement learning, and we wanted to bring those things together along with some understanding we had about the human brain. My PhD, I worked on the hippocampus and memory systems and imagination, made some interesting discoveries in that domain that I thought would also potentially carry over into ideas for AI systems and AI architectures. And we thought all of that was coming together and also the advent of a lot of compute power, and specifically GPUs, which ironically, of course, were invented for games. So everything for me, as you'll see, comes back to games one way or another. So what was our mission for DeepMind? Well, we thought of it as a kind of an Apollo program effort. If we really went for this with intense focus and ambition, We felt that a lot of progress could be made quite quickly. That's how it turned out. And our mission statement was to sort of try and solve intelligence. And by solve, I mean fundamentally understand the nature of intelligence and then try to recreate that in an artificial construct and then use it to advance science and benefit all of humanity. And we're still very much on that same mission today. So what did we start off with? Well, I think our first big result was back in around 2013, where we finally managed to scale up what then became called deep reinforcement learning systems to something that was actually significant, a significant scale, and quite challenging for humans to do. And we went to the earliest game systems that were popular at all, which was the Atari systems from the 1970s. You might recognize Space Invaders here, and the set of around 50 games that were on the emulators for Atari systems. And what we built with Atari DQN, our first sort of big successful system, was a system that learned how to play and maximize the score just by only being given the pixels on the screen, the raw pixels on the screen. So it was very much probably the first example of a kind of end-to-end learning system on something, they're working on something really challenging, perceptually challenging, like an Atari game. So that was an incredible moment for us. And I remember when, you know, back in 2012, 2011, when we were struggling to even win a single point at a game like Pong, and we were just sort of wondering, well, maybe we're just 20 years too early with these ideas of learning systems. And then suddenly it won a point, then it won a game and then it didn't lose any points. And then eventually by 2013, it was playing all the Atari games. Of course, we then took that much further and maybe still the thing we're most well known for is the program AlphaGo that Fei-Fei mentioned, where we tried to crack the game of Go, really the Everest of game AI. And we needed to do it through self-learning systems. All of you hopefully will know Go, this super complex game that's played in Asia. It's got more possible positions, 10 to the 170 by one estimate, than there are atoms in the universe. One cannot possibly brute force solutions to Go. In fact, furthermore, even people who play Go, the top masters, they don't really understand and can't explicitly explain what the rules are, what the heuristics are that they're following. So unlike chess players, you know, and that we can kind of distill chess players, grandmasters' knowledge into a set of rules and then program basically expert system chess computers to play pretty well like Stockfish, that approach is impossible, was impossible with Go. So you have to use this learning approach, you have to allow the Go system to learn for itself. Then obviously famously in 2016, we had this massive million-dollar challenge match in Seoul, 200 million people watched the match around the world, and AlphaGo famously won that match 4-1. But more important than it winning was the creative strategies that AlphaGo came up with, most notably move 37, which is shown here on this diagram, the red, circled in red in game 2 that really blew the Go world and all the Go experts sort of blew them away in terms of this novel idea that no human expert would've ever played. And because it's basically on the 5th line and it's very early on in the game, and that's sort of unthinkable to do from a space perspective. But of course, 100 moves later, that stone in move 37, stone, that black stone happened to be in the perfect position to decide the match and decide the fight that was spilling out from the bottom left. So it's almost like AlphaGo presciently put it there 100 moves later for it to be in the perfect position. So it's these creative strategies that I think, and creativity in some ways, and just want to unpack that slightly, uh, that I think is one of the promises of these learning systems. Like, the, the point is that they may be able to actually solve come up with solutions to problems that we ourselves don't know how to. And with expert systems, of course, this is not possible because you can only program directly a solution that you already know yourself how to solve. So you're sort of inherently limited with expert systems to solutions that experts, other experts, know how to do. So I think in terms of creativity, we can think of 3 different types of levels of creativity, at least 3. This may be something interesting for us to discuss. I think you can think of the lowest level or the least, the sort of most mundane creativity as being interpolation, so kind of averaging things together. So imagine you show a system a million cats, pictures of cats, and then you say come up with an average of all of those pictures. You know, that will be a novel cat. It won't be something in the that's exactly the same as any of those other million cats, but it's just a sort of simple averaging process. Then you have like extrapolation, which I think AlphaGo has shown, which is given all of human knowledge about Go, then it then plays 10 million games of Go against itself, or at least AlphaGo Zero, the successor version of AlphaGo did. And then it discovers new kind of, it can extrapolate new strategies never seen before, not in its training corpus.. But then there's still one level of creativity to go, I would say, which I'm calling invention or out-of-the-box thinking, which is, can, could these systems actually invent Go, right? So not just a new Go strategy, but actually invent Go or invent chess. And so, and our systems today still can't do that. But I don't think there's anything magical about that level of creativity. Maybe you can call that true creativity. I think it's still a systematic process that can be probably modeled by these systems. I just think one thing that's holding them back is that we don't really know how to express that demand to a system in a way that it could understand. If I was going to get a computer system to design a game like Go, I would say something like, you know, invent a game that only takes 5 minutes to learn but many lifetimes to master. Is aesthetically beautiful, can be completed in 10 hours of play so it fits into a human day. These kinds of things is what I would give as instructions, and then I'd hope it would come up with something like Go. But there's no real way for our current systems, at least, to take on board such abstract conceptual instructions and do something with that. But I think maybe that's within reach now. So how does the self-play system work? And I'm just going to combine together actually a range of systems together: AlphaGo, AlphaGo Zero, AlphaZero, and even our more recent thing, MuZero. So AlphaZero was our program that ended up, could end up playing any two-player game to better than sort of world champion level. And it's a very simple process if you break down this self-play idea, which I think is still something worth considering in the generative AI space too, actually, the analogous of this. I don't think it's been used yet. But what we did here is we started with V1 of, let's call it AlphaGo and/or AlphaZero, and it starts off playing randomly whatever game it is you've given it to play. It doesn't know anything about strategies or anything. So it's just playing purely randomly. And you play 100,000 games like that, and then you generate your V1 dataset. And you use that dataset to train a V2. And what the V2 is trying to do is predict what sorts of moves are likely in certain positions, and also who ends up winning the game from that position, and what's the probability chance of either side winning from that position. So that's the two jobs of the network, to try and kind of model what the game space is like. So then you train up V2, you then have a face-off match, a 100-game match, of V1 versus V2. And you see you have some kind of threshold. In our case, we set a 55% win rate threshold where if V2 beats V1 by above that threshold, you assume that it is significantly better. And then you replace the master system, the generator system with that new V2 system now. And you go around, of course, iterating this round. So now we'd play another 100,000 games with V2, so it's slightly stronger. So that means the generated data is slightly better quality. And then you use that to train V3, which is, of course, then a slightly better system. And then you face off against V2 and you continue. Now, obviously, if V3 doesn't beat V2, then you continue to generate more data, another 100,000 games with V2. So then now you have 200,000 games to train a new V3, right? And eventually, your V3 will beat V2. If you do this 17 times for AlphaGo Zero, you go from playing completely randomly to being stronger than the world champion in just 17 epochs, which can be done in a matter of hours. This is a very powerful system. What's really going on if you step back is that it's basically you're creating a model to model, for example, Go and the dynamics of Go and the strategies of Go and the likely positions, and that allows you to do tree search on top but do it in a tractable way. So instead of having to explore all possibilities, like in gray here, like that should be intractable, you actually use the model to narrow down your search to just the most reasonable options. And then that allows you, in a certain amount of constant thinking time, to actually find a near-optimal or very, very strong actual line that you want to play, an actual move you want to play here shown in pink. So I'm going to come back to this at the end because I think this is a very general way of thinking about AI and the idea of coming up with a solution to a problem. So we've been very fortunate over the last decade. We've been part of kind of creating many big breakthroughs in all sorts of different games, all of them kind of landmark results. The Atari one, the AlphaGo one I mentioned, AlphaZero I just talked about, generalizing that to every two-player game. And then finally AlphaStar, which was our program to beat grandmasters players at StarCraft II, which is the most complex real-time strategy game, computer game. And it has extra challenges over board games of being partially observable. It needs things like long-term planning. So it's complex in, in more challenging ways than a board game. And so this was all of our work in games. Now, of course, although I love games, always have done, playing them, designing them, using them for training for AI, I sort of use games in every way possible. But they've always, although it's been very fun to do that, it's always been a kind of means to an end, not an end in itself, right? The end was never to just win at Go or win at StarCraft. It was to build, it was to use games as a convenient proxy to test out our algorithmic ideas so that we could apply them to important real-world problems. So that's what has been so exciting for me in the last couple of years, I would say, is that that first 10 years was really about building the foundation, the building blocks of what we see today in the modern AI industry. But as of a couple, 2, 3 years ago, I felt that the time was right. Actually, we had powerful enough, sophisticated enough algorithms that we could now start really tackling some very challenging problems, both in industry, you know, through our partners at Google, and pretty much every product that you use at Google now has some form of DeepMind technology in it. But obviously excitingly, going back to the mission statement and my particular passion, of using AI for scientific discovery itself. So I'm going to go through probably our biggest and the result I find I'm most proud of, of us so far, which is our AlphaFold program. Many of you maybe will have heard about that, but for those of you who don't know what the protein folding problem is, it's this incredibly important problem in biology of basically predicting the 3D structure of a protein directly from its amino acid sequence, its one-dimensional sequence. You can think of it as a bit like a genetic sequence for the protein. And proteins are the workhorses of biology, the workhorses of life. Every function in your body, every function in life pretty much depends on proteins. And the function the protein does in your body is very dependent on the 3D structure that it ends up folding into. So of course, knowing the structure then can be really important for things like drug discovery and understanding disease. But unfortunately, it takes, you know, order of 3, 4, many years can take to even experimentally determine one structure, depending on the difficulty of the protein to crystallize and so on. And in fact, my biologist friends used to always tell me, you know, the rule of thumb is that it takes— you can take one whole PhD student, their whole PhD, to basically determine the structure of one protein. Of course, let, um, Christian Anfinsen famously said in his Nobel acceptance speech in 1972 that this should be possible, like determining the 3D structure of protein should be determined, fully determined, by the amino acid sequence. So this protein folding problem should be solvable in theory. But of course he didn't say how. It's a bit like Fermat's Last Theorem— I'll just write it in the margin— and then 50 years of toil later, people are still you know, trying to figure out why this would be, you know, should be the case. So it's been literally a 50-year-old grand challenge in biology. I found out about it actually as an undergrad at Cambridge in the late '90s. A couple of my biologist friends were obsessed with this problem and really explained to me, first of all, what it was, but also how much additional research it would unlock if one could actually solve this problem. And that always stuck with me as something. And I tend to do this when I hear interesting things like that, I sort of file them away in a kind of metaphorical little book of things to do. And one day when, you know, this— I thought this would be a very suitable problem for AI. So that's the question then. Can the protein structure prediction problem be solved computationally? Why is this such a hard problem? Well, again, you know, Leventhal, one of the contemporaries of of, um, of Van't Veensten, uh, had sort of famously conjectured this paradox of, well, there's 10 to the 300 possible shapes an average protein can take, um, you know, an average size protein. That's what he calculated, back of envelope, uh, but somehow obviously nature spontaneously folds these, uh, proteins in your body in milliseconds, you know, trillions of these proteins. So how is this possible? Um, you know, is the physics is somehow solving this in a very tractable amount of time. So in a way, that should give you hope that there is some landscape to learn about, maybe it's a quantum one, but there is some landscape that your AI system should be able to learn about in order to solve this problem in a tractable timescale. The other reason we picked this problem, apart from its sort of significance in terms of enabling many other things downstream, is that there's this fantastic competition called the CASP competition, which is a little bit like the Olympics of protein folding. And it runs every 2 years and has done since 1994. So it's like 30 years in now. And it's very rigorously run, and it's to test the best computational systems every 2 years. And what the beauty of it is, is that what happens is experimentalists, when they know they're about to— they've just resolved a structure experimentally, but they haven't published it yet. They give it to the CAFS competition, uh, to sort of be a blind test, and then you, you're supposed to give your prediction in. And then by the end of the competition, they publish their, the real structure, and obviously you can compare it then to the ground truth. But it's a, it's a double-blind experiment. You don't know who the competitors are, and you also don't know what the real structures are. So it's a really clean experiment for testing computational systems. And before we came along with this, this was the— this was the basically what had been going on in the field for the last decade before we got involved with AlphaFold. We started AlphaFold in 2016, actually the day after we got back from Seoul and the AlphaGo match. This was our next big project. And you can see here, this is a bar chart of the winning scores of the top team in each of the CAS competitions going back to 2006, every 2 years. And on the y-axis, you can see the score that they rate for this called GDT. But you can sort of think of it as about roughly as a percentage accuracy of how many of the residues, the amino acids, did you get within a tolerance of the correct ground truth position, right? And that tolerance is is a couple of angstroms. So you can see that the winning teams for a decade were only getting about 30 to 40% of those correct, which is— and on the hardest category, I should say, so free modeling category where there aren't templates or other things that you can rely on that are known, so that they're truly unknown from their sequence. And essentially, this is useless for experimentalists, right? It's just it's just not nowhere near accurate enough. It's— and many error parts of the structures would be kind of completely random. And what we were told is that in order for this— by the organizers— in order for this problem to be considered to be solved, you would have to get prediction— your prediction would have to be accurate to within the width of an atom. So that's less than 1 angstrom error, and, um, and, and you'd have to do it on over 90% of the, of the residues. So you're looking for kind of 90, roughly a 90 GDT score, for that to be of use to an experimentalist. And one way to think about that is if you crystallized a protein twice, with two different labs, and you tested it, and then you compared the two structures you got experimentally, there would be some disagreement between those two experimental structures, right? So there's no perfect ground truth in a sense. So there's still, there would be experimental disagreement to that. Fidelity. So you want to get within that bound, then you're starting to become really useful potentially. So we entered in 2018 after a couple of years of work with AlphaFold 1. That won the competition by a huge margin. We improved over previous scores by almost 50%, getting close to 60 GDT. People were amazed. It was the first time, I would say, cutting-edge ML techniques had been used as the core of a protein folding system, and it had this huge improvement straight out of the box. But the problem was, is obviously we wanted to get to this 90 GDT or less than 1 angstrom error, and we managed to do that with AlphaFold 2, but actually we had to go through a whole re-architecting using the ideas and experience we'd learned from AlphaFold 1, but it needed It needed a completely different approach in the end and different architecture in order for AlphaFold 2 to work and for us to crack that. Including one of the important innovations was actually building some physics constraints and chemical constraints into the network, into the properties of the network. So it didn't have to learn things like, you know, van der Waals forces and overlapping atoms shouldn't be overlapping and things like that and bond angles. We actually solved that in a sort of expert system way, but then without harming the learning. So it's actually, I don't know there's many other examples of systems that have combined sort of prior knowledge, one could say prior domain knowledge in that way, handcrafted with a learning system. 'Cause usually that interferes with the learning system. So this was obviously amazing. We released this in 2020. The results came out in, The competition was in the summer of 2020. The results came out in the end of 2020. And then the organizers of the competition, this amazing guy called John Molt, who has run the comp— founded the competition, run it for 30 years, was almost in tears sort of declaring that the problem had been solved. And obviously, looking back at 2016, and I was talking to him about this, given the trajectory, he was despairing of the fact that it would ever be solved in his lifetime. So that was fantastic. Here's an example of really— I mean, proteins are super beautiful once you actually get into them. They're little bio nanomachines, incredible. I mean, once I started really looking into this back in 2016, I was very integrally involved in this project. Um, you know, they, they're exquisite biological nanoscale machines, and, uh, it's kind of amazing that, that this complex dance can work in our bodies and give, you know, gives birth to life, really. And so here's one really complex protein that, and you can see AlphaFold actually iteratively gets some recycling processes, and it goes through the network several times, and then eventually gets to the final shape, which you can compare is very similar. The green and blue, the green's the ground truth, and the blue is the prediction, and they're almost overlapping for something as complex as this. So what did we go on and do then? We decided that the best way to get the maximum impact from AlphaFold in the world and the maximum benefit for humanity was to publish the methods, open source the code. And then furthermore, we sort of realized, this is the Christmas of 2020 now, we had a pretty productive sort of 2020 and lockdown, was that we decided actually, how would we give this to the world? And the normal way you normally do this is you normally provide a service on a server And people give you their amino acid sequences, and then, you know, a few days later, you send back to them, you email back to them the predicted structure. And that's the normal way things were done. But because AlphaFold was so fast and also so accurate, we just decided we would just generate all proteins known to science. So actually, over that Christmas, we did the whole human proteome, you know, over the Christmas holidays. And then we did 20 model organisms. And the most important ones for research and things like plant science and so on, important crops. And then we did every protein known to science, so all 200 million. And this is really important because in humans, you know, the coverage is around, I think it was around, you know, 17, 18% of that had been found experimentally. And we more than doubled that with with high accuracy predictions. But for other organisms, let's take plants like wheat or rice, very important crop species, less than 1% was known, partly because their genomes are much bigger for evolutionary reasons, but also there's just a lot less funding and a lot less experimental work done on those types of organisms. So is it actually even differentially useful to people like plant scientists? So that's been amazing to see. And then once we've done that, obviously we wanted to put that on some kind of database. We thought about doing our own one, but actually we realized that it would be much better to just plug into the main vein of what biologists use every day. And there's all sorts of amazing databases biologists have created, and one called UniProt, Obviously there's the PDB itself that has all the crystalline structures in it that we learned from, 150,000 structures that have been accumulated over the last 40, 50 years. And we teamed up with EMBL-EBI, European Bioinformatics Institute, just up the road from us in Cambridge in the UK, and they already host some of the world's best databases. And we very quickly had an amazing collaboration with them over 2 or 3 months to create a new database and host it and put all the AlphaFold predictions onto that database and also plug it in to all the other databases, genetics databases and other things that are other ways where you might come to find the structure that you're looking for. So that's been amazing. And then on the safety and ethics side, of course, we take this very seriously. We always have done. I'm going to talk a little bit about that towards the end. We actually consulted over 30 experts in the area from obviously biologists, but also from pharma and biosecurity and even human rights to just make sure that what we were releasing wasn't dangerous in some way, and the benefits would certainly outweigh any of the risks involved. And they all came back unanimously saying that they thought that the benefits far outweighed any potential risks. And then what has AlphaFold been used for over the last, I guess, 18 months, nearly 2 years now? An incredible plethora of things. If you're interested, I encourage you to go to our unfolded.deepmind.com website, which has— we're collecting sort of use cases from fantastic scientists around the world who've used it and told us, contacted us and told us how transformative it has been for them. And if any of you are biologists in the audience, and I know some of you are, you know, and if you've been using it, we'd love to hear from you what you've been using it for. But here's just some examples of, you know, John McGeehan and team from Portsmouth are using it to design plastic-eating enzymes to deal with plastic pollution in the oceans. People using it for antibiotic resistance. It was used to help determine the nuclear pore complex, one of the biggest proteins in the body that that governs a kind of gateway to let nutrients in and out of the, of the cell nucleus. That was a combination of experimental data and AlphaFold predictions. Jennifer Dowden has been using it in her Institute for Crop Sustainability, improving crop sustainability in the face of climate change, malaria vaccines. And then most recently, I was pleased to see Feng Zhang at the Broad Institute using it to create a sort of molecular syringe. As a new drug delivery mechanism, which is fantastic, in his recent Nature paper. So it's been really amazing to see, and actually more than we could have dreamed of. And just, you know, 18 months in, we've had a million researchers use AlphaFold and the predictions AlphaFold's made. We think that's pretty much almost every biologist in the world now, from almost every country, and it's already been cited 10,000 times already. In just 18 months, the, the, the paper, the methods paper, and obviously got some nice accolades from Science and from Nature as well. So what's my kind of— all these things are happening, and every year something amazing is happening, and I'm trying to sort of make sense of this, or, you know, while being in the middle of it, and it's, if anything, you know, accelerating the pace of progress. And it's kind of interesting, there's a few things, takeaways that I maybe want to leave you with, which is, so first of all, there's this notion I'm kind of been saying and thinking about of science at digital speed. And what do I mean by that? I mean science at this sort of technology speed. So being able to do science at what we normally think of as digital technology speed. And I think AlphaFold is maybe the first example of that. And I think is, but I think it's just going to be the first of many.. And I think AlphaFold is a sort of science at digital speed in two ways. One is that obviously it can fold the proteins in milliseconds instead of taking years of experimental work. So 200 million proteins, I mean, this is obviously just a funny back of the envelope rule of thumb, but you times that by PhD time of 5 years, that's like a billion years of PhD time, right? By some measure that has been done in a year. Of course, they're not all perfectly accurate, et cetera. So we still need both experiments as well, but that has to be helpful. And then the second way it's at digital speed is the speed at which it can get disseminated. The solution can be used by actual biologists, drug discovery people, all of pharma, all of that. Now normally, as I understand it, you may invent some incredible new technique, biological technique. It's a huge breakthrough. But it still might take a decade for that to flow through into the lab. Everyone has to be trained on it. Maybe new equipment has to be built. People have to think about how to use it. Here, it's just a database. You literally do a keyword lookup and it's there, right? So for everyone in any country with, or just with a, just like a Google search. So, so because it's digital in nature, the solution, it can also be disseminated in, in, in the way we normally think about apps and other digital devices and services. So I think that's an interesting concept, and maybe other people can think of other examples, but I can't. And I feel like this is going to be more— I'm really excited about the next few years and seeing this happen more and more in many other branches of science. Another observation I have is I think we're maybe, you know, at the dawn of a new era of what I like to call digital biology. So My thinking is, and why we, you know, even embarked on something like AlphaFold, is that I feel like biology at its most fundamental level can be thought of as an information processing system in some sense. Obviously, it's a phenomenally complex one and an emergent one, so that brings huge complexity. And I think that's why mathematics, you know, humans doing mathematical models of these things, whilst of course can be very useful, I think they're going to be— it's difficult for maths to fully describe, like a cell. You know, I'm not sure we're going to ever discover Newton's laws of motion for a cell. It's just too emergent, too complex. And I think that's why AI might be the perfect solution to that type of regime. So, you know, I think of maths and how well it describes physics, physics phenomena. I think AI could be the right description language for biology. And I feel, I hope AlphaFold is a proof of concept, and I hope that when we look back on it, it's actually not just AlphaFold itself being useful, but it heralds the dawn of a new era of digital biology. And I've actually— we've actually spun out a new Alphabet company called Isomorphic Labs, a sister company to DeepMind, to try and retake these technologies further in biochemistry space and improve AlphaFold further and other things to reimagine the drug discovery process from first principles using AI and computational techniques. Can you do the search mostly in silico, and then leave the wet lab work to the validation step. That would be the kind of dream. So then, you know, basically building on that and stepping even further back then, what have we done? When I look at the body of work that we've done, what is it that these things have in common? And I think the essence of what our systems are doing is we're effectively trying to find an optimal solution in some kind of enormous combinatorial space. I think that's a very general description of problems, scientific problems, but other problems too. And then what we've done with, from AlphaGo onwards to AlphaFold, is effectively learn a model of that environment, the dynamics of that environment, the manifold of that environment basically, and the context. And we've learned that from data or from simulated data. Ideally, you have both in many cases. So in games, obviously, we have this— it's effectively simulated data. Um, and then what you do is you take that model, and hopefully it doesn't have to be perfect, that model, but it has to be better than chance, of course. And then you use that model to guide a, a search, uh, process according to some objective function, right? And if it's a game, then you're trying to win or maximize the score. But if it's something like, you know, a biological problem, chemistry problem, maybe you're trying to minimize the free energy, something like that. As long as you can express that objective function, that should work. And I think this basically turns out to be a very general solution. I think this is a general way to think about a lot of problems. I'm not saying every problem can fit into that. I mean, maybe, but, but, but a lot of problems. And when you look at things in this, with this lens, you start realizing that that problem can actually be made to fit to this type of solution. And I'll give you an example. From drug discovery, which is what we're trying to do at Isomorphic, is— so this is the tree I showed you earlier, finding the best Go move, right? So each node here is a Go position, and then each, each edge is a Go move. And you're trying to, you know, intractable space, 10^170, and you're trying to find a near optimal or close to optimal Go move and Go strategy. Well, what happens if we just change those nodes to chemical compounds, right? And now we're in chemistry space, and, uh, and, you know, you start with some fragment or something at the top, and then you've got to decide what you're going to add to that in some sort of generative process. Um, and now you're basically exploring through chemistry space, but obviously you need an underlying model of biochemistry, uh, which is something we're building. Uh, and then in theory, you could search, uh, for, for optimal compounds with optimal properties that you're looking for, like no side effects, all of these solubility, all of these properties, ADMEC properties they're called in drug discovery, and do it in the same way, in the same tractable way we did with Go. And obviously, if this is true, if this works even to some degree, it would be revolutionary, I think, for drug discovery process. So I'm going to sort of come to the end now of like, well, we've had sort of golden couple of years in some sense with AI for science. We've had been lucky enough to have many Nature and Science papers published in all sorts of domains. So from quantum chemistry, better DFT functions to approximate Schrödinger's equation, pure mathematics— we've solved some important conjectures in topology with collaborating with some brilliant mathematicians. We've done work on fusion reactors with EPFL on their test fusion reactor, controlling the plasma in real time in in their fusion reactors and being able to hold the plasma safely in place for arbitrary amounts of time, being able to predict rainfall many hours ahead and more accurately than current meteorological models. So I mean, I just showed that I could have had several more slides of this. I just, but this is what should happen if you're building a general algorithmic solution. To things, right? You should be able to apply it to a kind of mind-blowing number of different, uh, spaces. Um, and in general, we usually— we have a whole approach of how to do this too. We, you know, how do we find a problem that's interesting? How do we find a problem that fits, uh, what I just mentioned? And then we usually try and find one of the world's best experts in, in this area to collaborate with, right? And then and really make sure we're actually answering a problem that they actually care about. It's very, actually quite difficult to do that, actually articulate what the right problem is unless you get really deep with a domain expert on that. And so we've been lucky to collaborate with many, many fantastic scientists and mathematicians on all these problems. And then in applications, there's a ton of those too. Fei-Fei mentioned some of those earlier. We saved— one of the early things we did at Google was saving about 30% of the cooling energy used to cool the massive data centers at Google. So it's a huge energy saving, and we're starting to explore doing that across actual whole power grids. WaveNet, our most realistic text-to-speech system that everyone uses, a variation of that, one of the first autoregressive models that we did, Aaron van der Aal, led that work at DeepMind way back, and is now the basis of all modern text-to-speech systems. So pretty much any system that you're talking to and their very realistic voices that come back are based on this work. We compress YouTube videos down by 4%, improve recommendation systems across the board, really. A huge amount of impact there on applications too. Then, of course, there's large models. All the rage right now. We've done a ton of work in this area. Just to pick a few things out is Chinchilla, one of our sort of famous results that looking at the scaling laws of large language models and actually showing that, find the important finding that they're kind of undertrained or they were undertrained back a couple of years ago. AlphaCode, our system that can program at competitive programming competitions, program at median human programmer level, and that's getting better all the time. Flamingo, our system to describe what's happening in visual images. And then Gato, probably one of our most general agents out there so far. You can think of it as an arbitrary token-to-token translator, and it doesn't care what those tokens are. They don't have to be words, they don't have to be images, they could be anything. It could also be actions. For example, like to control a robot. So that same system, Gato, can do pretty much all the things I showed you earlier on in this talk, and more besides, things like controlling robot arms and things like that, all with the same system. We pioneered a lot of stuff in large language models ourselves. We've been testing them for years now. Our biggest systems are called Gopher and Chinchilla, and there's some newer ones. But we've been using them to do a lot of research into safety of these systems. How can we make sure they do rule adherence? We pioneered areas like RLHF now that are becoming huge. We sort of published the first paper in that back in 2017. Tension retrieval, memory, all of these things are important advances that need to be included into our large models. And many people are now experimenting with these things. My view is currently that scale is incredibly important. Obviously, we have to continue on those scaling laws and adding more compute and more data is clearly helping. Maybe we're starting to see some diminishing returns now on that. It depends on how you look at it. My view is that these large pre-trained models have to be necessary, but they're probably not sufficient in themselves component of AGI. That's my current position. But, you know, I'm not highly confident on that. There is definitely plausible that they might be sufficient. But my guess, my best guess would be that they're clearly necessary but not sufficient on their own. And I think some of the areas we need more innovations on are areas like grounding and factuality. Of course, we all know about the hallucinations of these systems, and there's many ways to try and address that. Planning, missing memory, and reasoning, and quite a few other things besides, I think. And we're working on all those topics very hard now. And our best chatbot is called Sparrow internally. We published a paper on that just over a year ago. We've been improving it since. And really for us, this is a deep exploration into how we can build a dialogue agent to be helpful more of the time, correct more of the time, and harmless. So they're designed to answer questions with responses that are useful and evidence-based, and it uses search and retrieval to partly fact-check and to ground it. And it's pretty cool. The results we're getting on that, which we'll be talking about more soon, is maybe an order of magnitude more accurate and grounded than other systems out there, sort of less than 1% errors even under adversarial testing. It's pretty cool. But in the moment, there's a little bit of a trade-off in the sense that if you want things to be more accurate and safe, they tend to be a little bit less fun and engaging to interact with. Ideally, you want the best of both worlds. That's what we're working on right now. But it's actually a little bit difficult. It's a bit of like a Pareto frontier at the moment. If you make it more rule adherence, more strict and more polite and those things, and refuse to answer difficult questions and stuff like that, obviously it impacts how engaging that's going to be. That's an interesting dichotomy that we're seeing at the moment and we're trying to solve. So talking about ethics and responsibility, I think, you know, AI has incredible potential to help with humanity's greatest challenges. That's why I've worked on AI my entire career, my entire life. Obviously, I think that's going to come about in a big way through advancing science. But of course, these are dual-use technologies. We've always known that from the beginning of DeepMind. We've had an ethics charter from the very start of DeepMind. That's now part of Google's AI principles. So, you know, we helped to draft those, and those are publicly available. And so I think AI has to be deployed and used and built responsibly and safely and to ensure that it's used for the benefit of everyone. That's obviously, I know, a big part of what HAI and the people here are thinking about. And it's been central and core to us from the start. And people like Fei-Fei will know about that. We've been discussing these things for many, many years. And we continue to try and be a sort of thought leader and to provide thought leadership on these topics of AI strategy, coordinating, safety work, risks, ethics, and also engaging with the wider community to get them up to speed with what is sometimes a bewilderingly fast-paced, well, definitely is bewilderingly fast-paced of innovations and advances. So I'll just finish then by saying, you know, approach to AGI. My view is that we are approaching an absolutely critical moment in human history. That might sound a bit grand, but I really don't think that is overstating the the where we are, I think it could be an incredible moment, but it's also a risky moment in human history. And my advice would be, and I get asked about this all the time, is I think we should not move fast and break things, right? The typical Valley mantra of move fast and break things, it's obviously extremely effective. It's created the modern Silicon Valley, this type, and there are other variations on this same sentiment, right? Break things and then you ask for forgiveness later. I don't think that's appropriate for something this powerful. You can't necessarily specifically, it's not that you definitely don't move fast. I think you move at the right speed. I'm not saying we move slow, but we definitely don't want to break things that we could have foreseen ahead of time. And also depending on how powerful the technology is, it may not be possible to fix that afterwards. These unintended consequences. I think we're seeing that with social media, more broadly construed. Now we're starting to understand at the scale that it's at now, there are these unintended consequences that can be pretty bad. I think we should learn from that and do it differently this time with AI, which I think will be far more impactful than social media has been, which has already been obviously hugely impactful. So what should we do then? Well, I would advocate, well, fortunately we already have a way of doing this, the scientific method. And I think we should invoke that here on an organizational level as a community and as an industry. And that involves these typical aspects of the scientific method, thoughtful deliberation and foresight where possible, hypothesis generation with rigorous testing, if not out in the wild necessarily, carefully controlled conditions and controlled conditions, detailed analysis. I don't think there's enough analysis techniques out there. I think that's something we need to do a lot more research on as an industry and community. Update based on empirical data, ideally with external review. And the aim of all of this is to gain a better understanding of your systems before you deploy them at scale and then find out something isn't right. And of course, this is not a question about getting things perfect. It's not possible. You do have to do some empirical testing in the wild, but we should be thoughtful about that. We can do better than just dumping it out into the world and hoping for the best and then seeing what happens. I think we can If we use the scientific method, we can get a lot, maybe 90% of the way of understanding something, and then it just sort of narrows the things that could maybe go wrong or the unforeseen things. So I would say, you know, as we approach AGI, when, you know, maybe we're getting close, pretty close to it now, we need to treat it with the respect and the sort of precautionary principle that technology of this potential demands. So my view is we kind of need to be bold and responsible. So, you know, both of those things together. So I'll just end by saying, you know, I think if we get this right, and this has always been my hope, I think AGI, from a point of view as a scientist, you know, could be the ultimate general-purpose tool to help us understand the universe, much like maybe the Hubble telescope has helped us understand the cosmos.

Thank you. Thank you, Demis. We get to see each other quite often, but still love, love listening to your talks, both online or offline. So I think especially with lots of students here, I think we have a lot of questions. So we do have microphones, I think, on the side, but I'll start with a couple of questions and we'll open it to the audience. So You know, you wear many hats, right? You're founder and CEO of DeepMind. You're really a world thought leader in AI. You're a gamer all your life. I think in your heart's heart, you're a scientist. I say this partially because I am too. So I'm going to ask you a question to just begin with, as scientist to scientist, really, Now thinking about all your slides, what excites you the most in terms of a scientific question today?

I think you're right about my heart is a scientist, and then these other things I do to enable that science, right? I'm an entrepreneur because I thought that was the fastest way to get the scientific mission going. Even teaming up with Google was to get the compute power so I could push the mission faster. I think for scientifically, so the thing I'm most excited about in applications is things like AlphaFold, right? And more things like that, the drug discovery. I wanna make that happen. I think we can revolutionize human health and disease and things like sustainability. All of the projects I've picked are in these key areas. So that's one thing. But maybe in a kind of even more scientific way, what I'm excited about is I've always thought this journey that we're currently on and making a lot of progress now with AI would reveal a lot about the human condition. Because I've always been fascinated, and one reason I did a neuroscience PhD is I'm fascinated by the brain, our brains, right? How are we coming up? And I've been fascinated about that through games because of playing chess when I was a kid and for the England junior teams and so on. And I was trying to improve my own thought processes. So that's what got me started thinking about thinking, how are we coming up with these ideas? And Why is my brain making this mistake? Why can't I stop it from doing that? And then of course, you know, you read philosophy and maybe too many science fiction books I read when I was a kid, and you start thinking about like, what is all the big questions, you know, really? Consciousness, creativity, dreaming, um, what, what are all of these things? And neuroscience can go some way to explain that, and I tried that very far. And then probably there are many neuroscientists in the room, but I think it'd probably take us another 50, 100 years to get to the bottom of all of that. And in some cases, maybe not, because what do we have as a comparator? You know, like, how can you dissociate intelligence from consciousness and things like that? But actually, I, I always thought attempting to build AI in this way, uh, a general learning system that learned and was general, um, and then seeing— of course we can, you know, that's— it's an engineering science, AI, right? You build the artifact and then you can use analysis techniques to pull it apart like we do in natural sciences, and then comparing that to the human mind. I think that will actually ultimately give us these answers to the deepest questions we've wondered since the classical Greeks or before. I think it's a tool in many ways as a comparator, but also it could help us with neuroscience too. Of course, neuroscience is one of the sciences I'm thinking that AI could help with. A lot of people are using AI for analysis, but we could also do it for models and other things. So that maybe is the thing I'm most excited about, is finding out these deep questions, ultimately the nature of reality with our AI system.

So after all these years from doing from energy to drugs to games, it is still the science, the brain, the intelligence that's still—

The brain intelligence and maybe physics, you know, like the nature of reality, like what's going on here. I just find it so fascinating. Always have. And I think for me, I think most people— I mean, physics was my favorite subject at school, right? And I read Stephen Weinberg's book, you know, Dreams of Final Theory. Feynman was my favorite scientist of recent scientists. Roger Penrose— I've had many chats with him about, about consciousness, and we disagree on most things actually, but it's, it's, uh, it's always fun talking to him. And, um, but then I just thought, you know, actually be better to build a tool that could help us maybe understand all of these big questions, physics questions, the nature of reality ultimately. And so, um, and on the way have this incredible engineering, uh, uh, project as well. So it just— to me, it was the most fascinating project one could ever do.

You know, we almost become lab mates as a physics undergrad. I got into Poggio's lab, but I went to his student's lab at Caltech. Nearly overlap there. We almost become labmates. So physics, neuroscience, we share. Actually, do you think what you're building will eventually be able to generate the set of Newtonian laws of intelligence for us?

I think to the extent that there might be Newtonian laws. I don't see why not, but I actually think we'll be able to use it to understand a new set of Newtonian laws for physics, you know, like what is going on with quantum mechanics and quantum behavior and all these things. I don't see why that would be not tractable at some point with our systems.

I want to follow up on that because I was listening to your talk and this incredible engineering feat as well as, you know, just mathematical feat in terms of AI, in terms of scientific discovery. And you are sitting here in the headquarters of Human Centered AI.

So I have to ask you, what is humans' role in scientific discovery going forward? Yeah, well, there's this next period I think is very exciting. Next period meaning, you know, next decade or two, let's say. Hard to say exactly where. I think the AI systems are going to do a lot of the drudgery, pattern matching, searching literature, a lot of this sort of stuff. You could imagine a science assistant language model. We're working on these kind of things where they have to be a lot more accurate than they are now. Right now, I'm sure you've all tried it, they just hallucinate likely paper titles and papers that don't really exist. They sound very convincing. But, you know, once you fix all those problems, it could be pretty amazing.

And I think that— Are you saying that the human roles are like prompt engineers?

No, I think coming up with— it's always the classical thing of like actually defining the right question to ask. That's the hardest thing in science, right? These systems can't do that. That's the invention bit of AlphaGo I was talking about, right? Inventing Go. Okay, how are you going to do— you can come up with a Go move, We've not even got a line of sight of how it's going to invent Go or something as amazing as Go or as amazing as chess. Now, I don't think it's impossible. I'm not in the camp of like, we need some quantum thing to do it or whatever. Maybe someone like Penrose would argue, but it's given we're talking about Penrose. I think a Turing machine can do it. I often see my role in this way. You've got Penrose in my head now because I've had many interesting debates with him. I see myself as Turing's champion, right? Because I think that Turing machines and classical computers can do a lot more than we thought they could. And I think that is what AlphaFold is showing and AlphaGo is. These are massive combinatorial spaces. I mean, proteins are quantum systems right down at that level. Some people claim that it's NP-hard, you know, all of these things, right? So, and somehow we admit we have um, with AlphaFold mimicked effectively a quantum system. We've modeled it, and it's tractable. It's in P time. You get the answer back in milliseconds. So that's something pretty important there. I think I'm thinking a lot about and talking to my, you know, theoretical, you know, computer science friends about like, what does this mean for P equals NP and these kinds of things? And so I think, yeah, there's just a lot of, you know, complex questions here that we've got to, you know, trying to tackle.

Well, let's get Penrose out of your head. At least Einstein said what you just said. I think in the Nobel Museum, there's a quote by Einstein about much of science is asking the right questions.

Yeah, so this is, I think this, but asking the right questions isn't just the top-level thing. It's sort of deciding what these systems should investigate, how we should investigate them, what kinds of experiments we should do. Hypothesis generation, none of that is possible at the moment. But maybe in a number of years' time, we will have more advanced systems that can do that. Then we're in a different stage.

But right now, I think that this should be incredibly useful for human scientists, I would think. So a little bit of a switch of topic, because you talk about DeepMind's effort in LLMs at Stanford. We call them foundation models. Doesn't matter. It's all good. We have a Center for Research of Foundation Models, you're aware of. Our NLP faculty like Percy Liang and so on are leading that. One of the missions at Stanford CRFM is pushing for norms. And you're one of— I remember back in like, feels like ages ago, you were talking about all this safety. The norms. What do you think the norms should be right now in, as you call it, dangerous moment? And LLMs and foundation models are leading in terms of that.

Yeah, I mean, um, you know, I would say it's risky moment, and we have to— it's, it's, it's important what we do from here, I think, as a community, as an industry, and as a society. Um, I feel like we need, I mean, there's several things. There's much more rigorous red teaming of these systems. We do a lot of that ourselves, but it'd be great to have maybe a nonprofit body that does that or governmental body that does that. HAI or nonprofit? Could be, yeah, could well be, exactly. I think we need to understand the systems better. So better analysis tools, interpretability needs to be advanced. Some of the things I mentioned for Sparrow, I think they're all safety features like rules adherence. So one question is, can you get it to obey the rules that you want? The next question is, even if you can do that, what rules should you put in? But there are a lot of obvious things that should go in that we have, but there can be unintended consequences of that as well. So yeah, I think there's a lot of, I mean, there needs to be analysis of of data creation and the type of data you're putting in. I don't think we've done enough on that. Like, we try and fix it at the output with our RLHF, but I think a lot of it would be good also if we filtered incorrect things and inappropriate things from the inputs, from the training data. So I think there's so many things to be explored and done, and I think the problem is the capabilities are moving very fast, and some of this other work there needs to be more focus on some of this other work, safety and analysis and those types of things. Do partnerships play a role? Yeah, I think so. I think we should be discussing that. I mean, we've offered our latest models to people like the Turing Institute in the UK. I mean, I think, and we should talk about, I think we've also reached out to people here. We've got to figure out the right way to interact like that where maybe access to models can be given to external partners or collaborators, and they can independently sort of investigate these things before they're put out into the wild. It's quite hard to do that because the problem is, at the moment, a lot of this work is very manual, you know, in terms of testing out all the edge cases of effectively an emerging system. It's quite difficult to get that unless you have millions of people using it, but then it's already out.

So maybe we need to also come up with better automated testing these things too. Right. I started my question asking you to wear your scientist hat. I'm going to end my part of the question asking you to wear the hat of a world-leading AI thought leader. You really talked about this risky moment a couple of times, more than a couple of times. Going forward, the next just say 10 years-ish. It's hard to imagine the next week. Yeah. Pace of AI right now. Uh, what is a couple of most important things you think as a society in this AI revolution we collectively should be working on that even the mighty DeepMind cannot work on yourself?

Well, look, I think there are, I mean, there's many complicated things I think society needs to debate and discuss. For example, you know, the values these systems should have, the rules you want, as I discussed earlier, what rules do you want them to follow? What things should they be allowed to be used for? What do we want to deploy them for? These are not, some of them are technical questions, but they're also societal questions. And I think there are quite reasonable but different answers to those things. Different cultures, different countries would give out different answers to those things. I think there are issues to do with, I mean, it's mostly in the social media realm at the moment, but I think it's going to impact AI too of like, you know, if you ask a question about something and there's a political answer, left or right, what is the right answer that you should give there? Should you give the answer that the user wants to hear, but maybe it's not accurate, that answer, right? But then what is accuracy? So these are debates that go beyond AI, but I think will be maybe amplified by AI. So I think it's even more urgent that we try and find the right answers to that as a society, global society. And that seems quite difficult because it's intermixed with the geopolitical situation, political situations, and that's hard.

Maybe it is part of the reason we have all this situation, right? So There's a lot we can talk about, but I do want to give the Q&A time for especially students. How about we start with a student and then, you know, there are microphones on the side. Just go up to, walk up to the microphone. I would really love to start with a student question. Honors? Oh, you look young.

You're like a student. Hi, thank you for the talk and the insightful discussions. I'm a PhD student who works on the intersection of robotics and AI. I'm aware that DeepMind also has some efforts that push the frontier of robotics. So I'm curious about your view of the future of robotics and what people at DeepMind are doing to push us towards that future. Thanks. Yeah, great question. I mean, my view, this isn't the view of everyone at DeepMind. There's different views on robotics. Some people think embodiment is going to be critical to AGI itself. So one has to do that. I'm of the view that it's going to be an unbelievably important application of the general AI systems. I'm not sure it's the fastest path to get to AGI, to do it, to what, you know, to insist that it's embodied. So for us, we see it as very important for two reasons. One is it's, I think, going to be a really important industrial application area. With huge implications and impact. It's also used in our fundamental research as a grounding problem. So, you know, we're usually often away in some simulation games or science or whatever that is, and sometimes it's good to just ground yourself on a real robot arm or something like that to see if your simulation work really holds, right? And it also pushes on really interesting things like low data regimes transfer learning, things like that, that I think is the best setup for that. So we use it as a challenge task quite a lot. And the kinds of works we're doing, we're doing a lot of different types of work from locomotion, robot arms, to soccer teams trying to learn coordination. So it's quite a wide range of work that we do.

Thank you. Let's do this. We alternate on both sides. So you and then you. So we'll go that way. You've got a lot of questions.

Yes, this is your time. Okay, I'll try and answer faster. Yeah, thank you, Demis, for your talk. Um, as a student, a PhD student here, and a scientist, I'm concerned about a technical monoculture where increasingly large resources are devoted to an increasingly small number of future directions. At least that's the perception. Um, What are your thoughts on sort of our roles as scientists and your role as a scientist and a funder of scientists in this problem? Yeah, great question as well. Look, I kind of agree with that in the sense that that is the direction things are going in and that's where the money— because it's sort of like, you know, it's been so successful, it's kind of monopolizing a lot of the, I guess, scientists and money and talent. DeepMind's always been an incredibly broad church of research. I think we've historically been I mean, we're quite large, but we've also been the broadest in terms of, you know, from neuroscience ideas to obviously deep learning, reinforcement learning, all of these things. We have many and we have theorists as well. And then on top of that, we have a whole science team. So we've always had multidisciplinary, very multidisciplinary what we do. But even there, we've had to slightly change our emphasis, you know, to put more emphasis on the large models and the pre-trained models. Especially in terms of compute. If you look at the compute resources, it's just the way it is. And as I said, I'm of the opinion that they are necessary. So I think one has to do that work. I think that's clear now after several generations. They're not just going to simply asymptote. I think if you saw GPT-1 or 2 or one of our early language models, maybe you think, oh, maybe it's still just memorizing or something. But I think we've definitely gone past that point. I think the only scientific thing to do is to explore that to its limit. But while still, at least at DeepMind, the resources I'm in control of exploring, maybe more than half the organization is still exploring other innovations that might be needed, whether that's memory, planning, some of the other things I mentioned. And I'm kind of 50/50 whether that's going to be needed and whether the large models will be enough on their own. By 50/50, that's not a precise number, it just means I don't know, right? Really don't know. When you're in that much uncertainty, my view is you've got to press hard on both those sides, exploiting what you know and exploring the new things. But I think, as an industry, we're probably over in— probably that isn't going on everywhere else. I think mostly they're like 90% or maybe even 100% all in on the current techniques.

But yeah, we're more like 50/50. Do you know what you're going to do in PhD now after his answer?

Just kidding.

Hi there.

Hi. First of all, thank you so much for coming to give the talk. Definitely the highlight of my day so far. And second of all, I wanted to pose a more meta-level question. And so you mentioned during the talk that a lot of research is learning how to ask the right question. And so what I wanted to do is kind of reflect that back to you and ask you, considering that DeepMind is one of the leading places where research is done on AI, how do you approach asking the right question? What is your general approach to research to get to effective conclusions, to do the groundbreaking work that you've done? Yeah, look, I spent many, many years, I guess, intuitively trying to hone that ability. And then more recently, I think I sort of hinted at it in the talk, which is now we've done this several times, we can, you know, I've really boiled down as to what it is that we're good at, what is it that we've actually done. And then I've, you know, trained myself to be quite a broad scientist, to be able to kind of dive deep, sort of dip into something and then try and understand maybe from an expert in that domain what their problem is and then remap that. To problems that I know it may not be, you know, it won't be exact mapping, but it's close enough. And then the differences can be, you can make an assessment about how likely that is to work. And that's served us, you know, that's served us very well. Sort of picking a problem in the sweet spot, I often call it in the S-curve problem. And games are great for this because you wanna pick a problem that's not too easy, 'cause then it's gonna be trivial and not stretch your general algorithms. But if it's too hard, you won't be able to perceive if you're making any progress at all. So, you wanna always be in that sort of sweet spot of the S-curve. And you can do that with something like games, but also science, by keep on picking a harder and harder thing as your understanding of what your capabilities of your systems can do. So that's kind of one of the big things I've used. Another thing, I mean, I have many tricks, but maybe I should write a, you know, an op-ed or something about it at some point. But another thing I've used in the past that's been good is if you surround yourself with lots of smart people and multidisciplinary people, and then you present a problem to that group, um, often the speed and quality of the ideas that they brainstorm, that I found quite a good measure of the the suitability of the right time to tackle that problem. So if it's sort of flowing easily, and it doesn't matter if the ideas are actually good or not in the end, but if they're kind of plausibly good, plausibly interesting ones, and they're pretty easy to come up with, and there's lots from lots of people, that usually suggests to me that there's a lot of progress potential there. I found that as quite a good measure. Maybe other people already know that, I don't know, but That's what I found that's quite useful. Thank you so much. Hello. Yeah, thank you so much for your talk. I'm also a PhD student here. I think when I see DeepMind's work, it's really fascinating to see how you're really pioneering the whole AI for science field. And you spoke of many different applications. It also seems like when you spoke about the common motif across all of scientific research, like this search problem, so it feels like it's really pointing at a space for a base foundational model for scientific research in general. And it seems like your work at DeepMind really sort of motivates that. So is this something that your group has thought about? Yeah. What would that look like? And I think you've mentioned an AI research assistant, but I think maybe it's a bit more than just research, right? I agree. I agree. It would be more than just literature search to have the ultimate one. I think that is a way, definitely a very promising route. And I think our work, I think you correctly identified our work does point in that direction and we are exploring that. We call our models base models. Basically their foundation models. And, um, and I think there is a potential for building a kind of general model that would be good across scientific space, not just do literature search, but maybe do some experimental things, probably, possibly, you know, suggest experiments, this kind of thing. Um, I think that is very likely to be possible to work. So yeah, I would, I would encourage that direction. Hey, I'm also a PhD student. This question is sort of related to my own experiences, but how do you know when it's time to double down or give up? So you put out AlphaFold 1, you get 60%, and the next competition is in 2 years from now. Yeah. You have a decision to make. Yeah. And you guys really doubled down on sort of that decision. You rewrote everything. It worked out great. Yeah. But are there ever times when it doesn't work out as well? And sort of what is your mental framework around assessing that decision? Yes, that's a great question, and we do have a lot of heuristics for that. So, um, with AlphaFold, you know, it is case by case, obviously. Like, you have to see why you've hit a brick wall. And, and one of the things I use for that is what I mentioned earlier. Like, you get a different set of people in the room to look at that problem and look at what you've just done, how many new ideas is it, and how— what, what's the the flow of that, right? It's not just how many, it's the quality and how hard was it to do that. That is a good— so there was a lot of ideas around an AlphaFold 2 in that particular case from a different set of people. We actually had to change the team quite a lot too to get a different approach and a different technique. But there are other things too. We've been many times through this dark valley of— I mean, most of research is like that. You do something super hard, it's going well, and then, then it's like some big problem. And this is pretty normal. And once you've been through a few times, you first of all not to panic. Secondly, um, to look for orthogonal— this is where I think multidisciplinary can help you— look for orthogonal signs that you might be on the right track. So one thing I did with AlphaFold is— one reason we picked AlphaFold, and I started that after AlphaGo, was that— do you remember there was this Fold It game? Um, so I don't know if some of you know, in protein folding, there's like a citizen science game where they I think it's the Baker Lab. They, they turned protein folding into a puzzle game. There was literally a 3D puzzle game and you bent a protein's backbone and things like that. I came across it when I was studying at MIT in about 2009, 2010. I think it just come out then. And amazingly, they got— it wasn't that fun a game, but there's quite a few gamers who wanted to do science with their gaming skills. And maybe like 10,000 people played it and some of them got really good. And they actually discovered 2 or 3 real protein structures, and they were published in Nature, I think. And I saw that, and that made me— that, that, you know, I was thinking, well, that's incredible, because someone who's not an expert biologist has used their intuition to make some— sometimes counterintuitive moves that look energetically wrong but that actually would end up being right. And, uh, and what I thought was— what gave me confidence with AlphaFold 2 specifically is we'd already mimicked the intuition in AlphaGo of, in my view, some of the world's top pattern matchers. Lee Sedol, he's played Go since he was 2. He didn't go to normal school, he went to Go school. His whole mind is Go. He's an incredible guy. It's his whole universe. And yet we were able to mimic his level of intuition, let's say, somehow with our AlphaGo system. I thought if we could do that, and then amateur gamers were able to solve using this tool with their presumably pattern matching, presumably what was going on, proteins, then, um, that must be tractable too. So although that's not a definitive reason to continue, but that was one reason. And then the other reason was that somehow physics does solve this seemingly intractable problem, right? We live, we, we exist, um, the proteins get folded in our bodies all the time, um, sometimes they get misfolded in disease, but but they're folded. So, um, so, you know, in, in theory, that should then be tractable if you believe a bunch of things about classical systems and quantum systems, but that's out of scope. So I, I used some of these heuristics to try and give you confidence to double down.

Thank you very much. Great. Um, we have entered the fast questioning, fast answer phase. Yes, sorry, we did a fast answer Great question. Hello, I'm a graduate student and I'm also involved in AI research, and my question is, um, how should technologists, scientists, those investigating and researching these challenging problems bring in the perspectives and ideas of people from marginalized communities, um, particularly those who do not come from a scientific background? What infrastructure do you imagine being created to accomplish this?

Yeah, I think it's a really important question, and we think about this all the time. And when I say multidisciplinary, I don't just mean scientists. We have actually ethicists, social scientists, philosophers at DeepMind, and also we work externally with those types of people. So we try to bring many perspectives into our work, and when we, you know, design things and deploy things, and we pay a lot of attention to that and I think the community itself could do better and we could also still do improve, but it's very much at the forefront of our mind that to get as many inputs as possible into the designs of these systems and how they get deployed and especially the types of people that it might affect. I think that's a very key thing that we should be doing as part of the thoughtful deliberation that we, I was mentioning in the scientific method and the understanding part is to understand the consequences of what you're doing, butterfly effects, if you like, not just the maybe the intent that you had, but also the unintended consequences better too. I think that's all part of making sure it's good for everyone in society.

I do also want to add that here in America, part of the effort is democratizing resource for AI research. So Stanford and HAI have been leading the leading lobbyist of this bill called National AI Research Resource. It's an AI compute and data platform that we're asking the federal government to set up. And one of the key clauses or asks in that is reaching to the traditionally underrepresented communities. So it's major policy effort that Stanford is leading. Again, we're having increasing number of people asking questions, so you do need to keep your answers short.

Hi, Demis. I'm an undergrad here. A big takeaway I learned from your talk is the long-term perspective that DeepMind takes to its research. And along that line, what were some of the core founding philosophies that you helped cultivate at DeepMind in its early ages to enable it to continuously innovate at the frontier for as long as it has? Yeah, I mean, there's so many, it would be— it would take a whole talk to sort of go through that. But I think there was some techniques we bet on early that when we were very rigorous about that. So we're— so learning systems, not expert systems, right, that we made that decision early. That wasn't obvious in 2010, and that led us to reinforcement learning and deep learning as being core. To what we do, using inspiration from neuroscience and things like memory replay, episodic memory, these kind of things. That was also core. Understanding the compute would become a huge factor. And we knew about GPUs and all those sorts of things from gaming. So this was all different threads that we backed early. And then the idea of using simulations and games as a test platform rather than— we could have chosen robotics or something like that, which I think would have made our progress much more slow. And I think a lot of people who were doing that kind of work back then, even at MIT, were working on robotics platforms. And I spent some time with them. What I realized was that they were spending all their time fixing the servo motors and the hardware and almost had no time left over to work on the software, which I thought was the problem, right, with AI. So that was another reason to choose games. And obviously, also, I knew lots of games engineers and brought them on board early. So that was a—, and I knew which games to choose and all of those things. So that was easy for me, but it also made strategic sense, I would say. And so, so those are all things that we stuck true to. Also, to be honest with you, starting in London and being in the UK, that's where I'm based, and where I did most of my studying. And I just knew there was a lot of untapped talent in Europe, in the UK. And I also thought, and I think this goes to geographical diversity, Although, you know, we're part of Alphabet, but we're still very much a UK-based company and European in our outlook. And I think it's important that in the global debate about AI, you know, of course there's China and there's US, but there's also, I think we sort of represent Europe at the forefront of this technology and at the top table of what's going on. So I feel that responsibility too. That was something that was recommended to me not to do. Like, you've got to move to Silicon Valley. Like, all of everyone was saying that, and, you know, all our early funders and other things, but I resisted and, and, uh, for, you know, for various reasons. But I think that was one of the reasons, and I'm pleased that we bring that perspective as well.

You really, really, really inspiring and great to see what you're contributing to the world of science and world of AI.

And, uh, students, great questions today. Yeah, thank you. Really good questions. Thanks, everyone.