An AI@GSB Special: Demis Hassabis Thinks We’re in the ‘Foothills of the Singularity’

0:09Hi, everyone. I'm Jennifer Auker, the General Atlantic professor at Stanford Graduate School of Business. And I'm Michael McDowell, a producer here at the GSB. And today we're live from the GSB. That's right. You're about to hear a conversation between my friend and colleague, John Levin, president of Stanford University, and Demis Hassabis, CEO and co-founder of Google DeepMind. Jennifer, I think it's fair to say that you made this conversation happen. So tell me why you invited Demis to the GSB. So many reasons. But honestly, the thing that drew me the most isn't AI itself. It's Demis. Before DeepMind, before AlphaGo, his dissertation explored the neurological relationship between memory and imagination.

1:00He showed that patients with hippocampal damage lost not only the ability to recall the past, but to construct any rich mental scene moving forward. To imagine, you have to have lived. This was a big breakthrough back in 2007, and it raises a pretty profound question given what Demis went on to build. It really does. If AI increasingly handles planning, remembering, scenario building, what happens to human imagination? What capacities do we most want to preserve, deepen, and expand? Yeah, it's not just what we might lose, but what becomes possible. Absolutely. I'm thinking about scientific advancement, economic empowerment, creative potential, expanded agency, and ultimately, fuller human lives.

1:54At the micro level, I think of this as flourishing, which means feeling fully alive, creative, connected, and purposeful. And if we're able to accomplish that, really exciting things start to happen at the macro level, across organizations, communities, and society. I keep thinking about something Demis said, which you're all about to hear, which is that the future is still unwritten. None of this is inevitable. I think Demis is asking many of the same questions I am, and understands maybe better than anyone, how the choices we make now may alter not only our future, but the nature of human existence.

2:34And that's why I was so excited to bring him here to Stanford, the place where big ideas meet real application. On that note, thank you so much, Jennifer. Thank you, Michael. Here's Demis Hasevos in conversation with Jonathan Levin. Demis, it's great to have you here at Stanford. It's fantastic to be here. Thanks, everyone, for coming. Really appreciate you doing this. So I'm going to ask you some questions. We'll have some questions from students, and we're looking forward to hearing your thoughts. Maybe we would – you've been chronicled a lot recently, a movie, a book. So many people have heard about your trajectory.

3:19It is quite remarkable. I'm chess prodigy, video game developer, scientist, tech entrepreneur and leader, Nobel laureate. That's just the first half of your career. So if you were going to try to draw a through line through all those different things that you've done, what would it be? Well, I think there's several through lines, actually, with what seems maybe somewhat unconnected subjects. First of all, I've always really enjoyed working at the intersection of creativity and technology and very broadly construed. So actually the games industry, the video games industry, which is my first very early part of my career in the 90s, was one of the most creative spaces in any industry that was using cutting edge technology with art and design to sort of create an entirely new entertainment medium.

4:21So that was really an amazing time. In fact, some of the most fun times I've had in my career was early in the 90s. the chess and the neuroscience I did all of those things I've tried to I had from this idea of working on AI and AGI being the most important thing one could and most interesting one thing one could spend your career working on also from a very early age so as a teenager probably I read too much science fiction reading things like Gödel, Escher, Bach these types of books and biographies of the of some of my scientific heroes you know Turing and Feynman and so on so all of these were serving to inspire me to try and understand the world around us in a really deep way.

5:03And then building AI was my expression of that mission to try and build the ultimate tool for science. And I've tried to, because life's short, I've tried to reuse and repurpose every experience I've had in service of that bigger North Star mission that I've had, you know, for more than 30 years. So, you know, my chess training is the way that I think about business and organizing things and planning and how I think I've been able to break down very ambitious plans into smaller, more manageable steps. That all comes from kind of chess thinking, I would say. And then using games, first of all, building games, learning about engineering projects at scale, running companies, startups, and then fusing this creativity with engineering.

5:48Actually, it's what we do today with AI. It's an engineering science. So you're fusing creative work, scientific work, with very hardcore cutting-edge engineering. So that all served together. And then finally on games, as everyone knows, we used games in the early days of DeepMind as the perfect proving ground for testing out algorithmic ideas, probably most famously with AlphaGo, which I think we've just had the 10-year anniversary of and was really, looking back now, maybe the start of the modern AI era. When you went into AI professionally in 2010 or so, you started DeepMind, you had this very ambitious vision.

6:28You were going to solve intelligence and then solve everything else.

6:35How's it going? We'll expand a little bit. What has gone according to plan and sort of what has been off plan? Well, the broad arcs of it have gone, I mean, unbelievably well, perhaps. You know, when we started DeepMind in 2010, that's, can you imagine, we used to go to, try to go to VCs in the UK, in which there weren't very many, and with that as the business plan, it was literally step one, solve intelligence, step two, use it to solve everything else, and people were quite confused. but we really meant it and actually we may go back to exactly using that mission statement because by solve intelligence we meant build AGI ideally also understand the nature of intelligence on the way to building AGI and perhaps using AGI to help that help us understand our own brains and minds better you know things like the nature of consciousness what is creativity dreaming all of these deep mysteries of the mind.

7:38And one of the reasons I studied neuroscience was try to learn from what we understood about the brain as inspiration for algorithmic ideas. And sort of, so step one was to try and build AGI. And then we always had in mind what sort of happened, which is that, of course, it's a general purpose technology, maybe the general purpose technology. And if it was built in the right way, so it was a learning system that was very general, what would be the limit of what it could be applied to it's not you know it could be applied to almost anything was the dream um and i think uh that's what's borne out i had specifically in mind for that step two advancing science and medicine so that's what i meant by using it to solve everything else i meant the big questions uh in science all of them um i wanted to you know i was fascinated by all of them the nature of time the nature of reality maybe that's the most fundamental one.

8:32And I loved physics when I was at school. That was my favorite subject. And I think when you're interested in the big questions, you end up doing physics probably. But the reason I decided that there were too many interesting big questions. So how was one to, you know, try and tackle all of that in a lifetime? And that meant, in my view, building new tools and to aid us at the best scientists, the best experts to make much faster progress in the fields that they were tackling and the big questions and important questions that they were tackling. And then, of course, AI in itself is also a fascinating artifact in itself, scientific sort of object, one could call, worthy of study itself.

9:14It's almost a new field. So this is sort of, to me, it would feel like the most fascinating and most important thing to spend one's life on. And I would have been doing it, even if it hadn't worked out, I would have found some way to be doing this, you know, in academia or wherever. This is what I would always plan to spend my life working on. And all those things I did earlier were different expressions gathering the experience and I suppose the knowledge to be able to attempt something like DeepMind in 2010 when we were sort of, I felt, we felt that we were ready to make fast progress. And of course, the second part of that, use it to solve everything else, is now much broader than just science and medicine, although that's where I've tried to personally do my work in, as well as running the overall organization.

10:01But obviously, it's going to be amazing for productivity and many other things in the world outside of just science and medicine. As you've been building these different models at DeepMind, you started with games, and then you went into science. Were there particular moments where... that, I mean, you started with a lot of conviction, but I'm curious if there were sort of particular moments where you sort of saw this is actually gonna work. Yeah. The AlphaGo and playing move third. Yeah, there were many moments where I thought it wasn't gonna work, put it that way. So some of the ones I remember really well are, we started with games because they're self-contained.

10:47They obviously were designed to be, by other humans to be challenging or fun for other humans to play. They're often, I love games, they're often microcosms of a lot of real world scenarios. You know, if you think of Go or poker or chess, I often thought as, you know, one of the courses I'd have in an MBA or business school course would be a games module to, you know, study those types of games, diplomacy. They all have really interesting aspects of the best games, of real life, and you can obviously practice many times in a kind of safe scenario. That's what I think games are really useful for.

11:25And that applies to AI systems that are learning too. They're neat environments, they're challenging, and they have clear objective functions, which also was very important for our early days of reinforcement learning. Almost no one had used reinforcement learning for any kind of scaled-up problem. It was obviously an academic discipline, but it was used mostly for toy problems, like little grid worlds. It wasn't clear it could scale up to anything major. And so we started with, I would say, the most famous set of most basic games that had become, you know, world popular, which were Atari games from the 70s.

12:07And we started with the simplest game of all, which was Pong. You know, the bat and the ball, just two bats and a ball. And there's an inbuilt AI system, an inbuilt system that controls your opponent and uses all the information that the game has about where the ball is and so on to move the bat around. And what we wanted to do was could you play Pong just from the pixels on the screen, so the raw data, the raw visual input, and no other information, no privileged information about no access to the insides of the program about where the ball is or the speed of it and so on, which obviously the program knows.

12:45But we didn't give the DQN system, as it was called, our Atari system, any of that information. It just got the 20 ,000 pixels on the screen. And 20 ,000 pixels, I mean, seems, you know, sort of trivial now. But back in 2010, that was an enormous amount of input data. No one had ever dealt with something that complex and then multiplied by all the frames that you were doing. And for about, it felt like six months, maybe it was only two months, but we couldn't win a single point at Pop. right so it was you know jerking the bat around it was like oh is it ever gonna even be able to like control the bat and um and of course he had no notions of any of these things and it was just losing 21 nil to the inbuilt ai and i did think and we had a couple of different ways of trying to attack this and we had almost no money you know the runway you know the the the bare the couple of uh million dollars of funding that we had which wouldn't even cover an intern these days was, which is good for all of you students, was our entire funding.

13:47And, you know, we were paying, we were taking like no salary and it was running out the money. And I was like, oh, well, maybe it turns out, maybe we are still 10 years too early. Maybe we're 20 years too early. And then magically it got a point and it was like, oh, maybe it was just luck. And then it started winning a lot of points and then it started winning the games. And then it was like, okay, we have liftoff now. So now, and those of you working machine learning will know this, if you get a foothold, you can usually hill climb your way out of that. That's been the history of AI, I would say, right?

14:20Once you have something working, there's usually a way of optimizing it more. And that's what turned out with Atari. And then, so that was our first big result. And our first nature paper was the, you know, really the first deep reinforcement learning model, certainly at scale. combining deep learning to learn the domain and deal with the perceptual inputs and the complexity of the inputs and find the patterns in it and then reinforcement learning built on top of that to kind of make the decisions and and do the planning and then of course that culminated in alpha go which was the always our aim uh dave silver and i was head of that project we were undergrad friends as undergrads at cambridge and we were just had been discussing it since our undergrad about um We were there in the mid-90s.

15:04The Deep Blue-Kasparov match happened while we were at college. Of course, I was fascinated by it both from the chess and the AI point of view. But I was more impressed with Kasparov's brain than I was with Deep Blue. Because Kasparov, with his incredible mind, still one of the biggest chess geniuses there's been of all time, he was able to basically compete on an equal footing with this supercomputer brute force machine next to him. But of course, he could do all the other things with his mind, speak five languages, do his politics, drive cars, all the rest of the things humans can do. And to me, that was, you know, incredible.

15:39That's much more impressive. So there was something missing from the deep blue system. And obviously, those techniques, those expert system techniques, where you hand curate the heuristics, and then you use brute force search on top, which is still how a lot of traditional chess programs work today, that works for chess, but it's never worked for Go. Because Go's too esoteric a game. It hasn't got material. Every piece is worth the same. It's all about patterns and intuition. Even the top Go players, that's how they play it. So we realized, we sort of thought, okay, if someone could actually get to world champion level at Go, it's not just about really that was an aside getting to that level.

16:14It was more about the approach you would have taken would probably be a really interesting algorithmic approach and maybe and hopefully would generalize to other domains. And that's what turned out with AlphaGo. So, and then it went beyond our wildest dreams, really, because not only did it win the match against Lee Sedol in 2016, it also created new, famously, new strategies that had never been seen before, even though we've played Go. Go is the oldest game humanity has invented. It's 2 ,000 years old, 2 ,000 plus years old, and been played professionally for hundreds of years. And we hadn't discovered those strategies.

16:52So that was sort of double whammy for me. I was waiting for that moment that AI was able to come up with something novel. And there's more levels of creativity beyond that, but at least it was a novel idea. And then that was for me what I was waiting for to then start using AI for science. So the moment we got back from Seoul, we started the AlphaFold project. So let's talk about the science a bit, because then you went into the protein folding problem. And again, you picked a problem where there was data and where there was a clear objective function in terms of thinking about protein folding.

17:30And it worked. I mean, you actually managed to solve this longstanding problem of predicting protein structure. You did something very interesting when you came up with LFOLD, which was it was obviously a huge Nobel-worthy scientific breakthrough. probably also of commercial value. And you just gave it away for free. I'm curious, how did you come to that decision? Was it something, did you think about other ways of going about it? Why give it away? Yeah, so we picked the protein folding problem. I had my kind of eye on that since also, you know, my undergrad days at Cambridge. That's when I first came across it.

18:15I had a few biologist friends who were obsessed with the protein folding problem. and actually they ended up becoming structural biologists, of course, in their career. And one specifically, I remember he was, you know, every time we were in the pub playing table football or something, he would be talking about obsessively how this was the most important problem in biology. And more importantly, I think of it as a root node problem. Like if you could unlock that and find the structures of proteins, that would unlock whole new avenues of research. Things like drug discovery, obviously we're trying to push that, but also fundamental biology and disease understanding.

18:52So this was a problem worth really spending a lot of attention and time on because of the downstream effects that it would have. It also felt to me, it was a fascinating problem. It felt to me like the ultimate puzzle, you know, 3D puzzle of how does this sort of, you know, amino acid sequence, think of it as genetic sequence, fold up into this 3D structure. It's amazingly interesting, intricate thing. And The more I looked into proteins, the more incredible my respect and wonder is for biology. These unbelievable little bio nanomachines, everything on life obviously depends on proteins. And as you start looking at their structure, you start understanding their function.

19:32So this is fascinating to me as a science question. And then, yes, there was the clear objective. It's sort of like minimizing the free energy in the system. Presumably, this is how physics is doing. It's why the body, you know, these proteins fold in milliseconds in your body, you know, billions of times a second. So somehow physics has solved this. And it can't be, there must be some topology, let's say, that you could learn, maybe with a deep learning system, that would guide the search. Just like we've done with AlphaGo to find a great move in Go, a great strategy out of the, you know, more possibilities than there are atoms in the universe in Go.

20:09and protein folds are even larger search space than that. But there's some way to narrow that down in a sensible way. You learn a kind of heuristic using the deep learning models to then guide your search for that to become tractable. And it felt like a really analogous problem in science to what we'd solved in Go, sort of applying some of those same approaches, those same theories to this domain. And then the other thing was there was obviously 50 years worth of painstaking crystallography, structural biology work by many great labs and people. And after all of that effort, there was about 150 ,000 structures in the PDB, the main database, which isn't a lot, actually.

20:53Obviously, it's a huge amount of effort that's gone into that, but there's 200 million proteins. And 150 ,000 also for machine learning systems is a very small amount of data. So most people thought it was at least 10, 20 years away before we would have enough data and the right types of algorithms to tackle that. But we felt that using every technique we knew in the end that we could make progress with that. And it turned out to be the case. And then when we decided to, well, how would we make the maximum impact with this? It was obvious to me that we should fold all the proteins so that, because not only was alpha fold accurate, it was extremely fast, it could fold a protein in a matter of seconds, and then collaborate with, in the end, European Bioinformatics Institute in Cambridge, which hosts many of the biggest databases, biology databases scientists use, and just host the entire 200 million protein structures on their database, and just allow it to be as simple as a kind of Google search to just find your protein structure, along with the confidence intervals the machine learning system had about which parts of the protein structure it was confident on, which is very important for biologists to know.

22:04So we put that all together, and it was, of course, could have been very valuable. I don't know how many billions of dollars or whatever. I mean, it depends how you calculate it. To do that experimentally would be incalculable cost, but it would have been hugely valuable to keep proprietary. But for us, it felt like we would only be able to scratch the surface of the downstream impact that putting all those structures out in the world could have on our own. Because, you know, there's three million researchers around the world that use AlphaFold pretty much every day. Almost every biologist medical researcher in the world.

22:38There's no way one organization could have done that. So it was obviously the right thing to do. We also had depended on public data to train the first versions of AlphaFold. So it only felt right to give back to that community, the structural biology community, this amazing resource that was amplifying the resource they had paced stakily built. And so it was just, it wasn't even a question for me. And, you know, it was great that also the executives at Google also, you know, love science and totally got that. I don't think all companies would have made that decision. So I give them a lot of kudos on that too.

23:15That was an easy discussion. And then we've tried to ourselves push that downstream with Isomorphic Labs, an alphabet spin out that is building, you can think, several more type of alpha fold level breakthroughs, putting them together into a way that will accelerate, hopefully, drug discovery, you know, take it down from years to months, maybe even one day weeks, just like we did with protein structures, which used to take, you know, years for a single one, and then we could do it in seconds. We'll be back with more of Demis and John after this.

24:16Thank you. create accountability for all people. These things have real consequences for the welfare of countries and the people that live in them. Join me, Kevin Kuhl, for If Then, a podcast from Stanford GSB, where we explore the forces transforming business and society today. Find If Then on Apple, Spotify, or wherever you get your podcasts.

24:49I want to turn for a minute to something you said earlier this week. You were in the news this week because at a big Google event, you said that we're in the foothills of the singularity. Yes, it got quite a lot of pickup, that line. It got a lot of pickup. And I understand that maybe the Google press team might not have been so thrilled about it. But since you're out there saying that, what did you mean by that? Yeah, so the full thing I said to sort of close the conference with was when we look back at this time, I think that, you know, maybe I'm thinking sort of 10 years from now, I think we will realize that we were standing in the foothills of the singularity now.

25:39What I mean by that, and the reason I chose that word, is that, so there's the technology, which is AGI. We've been calling it AGI, this next version of really general artificial intelligence. I believe that we're only a few years away from that, maybe like 20, 30, plus or minus a year, which is astounding to think, really. And then the era, I think that will be such an enormous transformative technology, it's going to effectively be a new human era. and that's what I'm meaning by the singularity is that and what many science fiction writers have written about that is that it's sort of the it's describing the era that we will be in in and around when AGI the advent of AGI happens so and I think we can feel this year I would say even though I've been working towards this for 30 years I think this year with the way that agents are working and tool use it started to become really useful for you know still early days of but genuinely useful in people's workflows.

26:37And we can sort of see what extra things that need to be done. And all of us, the leading labs are working on that. I think this is the beginnings of that, but the foothills, I still think there's a lot more work and it's just the beginnings, but, and it's not any one thing. It's several different technologies, several use cases that I see, several things that I thought were maybe a bit further out, turned out to be now, that are coming together that make me feel that in aggregate. and to the extent that I wanted to say that. Because I think society needs to hear that because we don't have long to prepare for what that means.

27:14And it's going to be enormously profound, I think. And the future, in my view, is still to be written, but these next few years are going to be very critical as to which way that will go and how we collectively want that to look like. If you look at surveys of how people perceive AI in this country in particular, it's very negative right now. And maybe more negative here than even in other countries. And there's probably lots of things driving that, concerns about privacy or state control or the size of the tech companies or jobs. How do you, I mean, you're running one of the leading labs, and how do you think about that public concern about the technology?

28:08I think the public's right to be concerned. I think that there are things that, and I'm concerned about several aspects of what the technology is. It's a dual-purpose technology. It's something this profound. You know, I sometimes describe it, quantify it as 10 times the impact the Industrial Revolution was, 10 times faster. and so you know it's taking a place over a decade instead of a century so that's like 100x of the industrial revolution and it's probably an underestimate to be honest but that's probably enough for us to try and comprehend right and deal with and so of course uh i think there are uh it's super exciting there's going to be amazing things that are going to happen like we're trying to do that with solving all disease i think you know a lot of the other challenges facing society today from climate to energy to disease will be helped by AI.

29:01I'm sure of it. And in fact, I'd be much more worried about those challenges if I didn't think something like AI was coming down the line. But it's going to cause a lot of change and disruptions and actually both on the technical side, economic and philosophical. And I think we've got to think through very thoughtfully and bring together all parts of society to discuss this not just the technologists the technology and the safety of the technology is just one piece of this it needs economists social scientists human and humanity experts to kind of chart out what is going to happen next and I think one of the reasons it's negative here is that specifically because it's different in other countries for example when I go to we came back from the summit in India it's hugely popular with the youth of India because they see the opportunities that it's going to democratize for them, having access to basically the same tools that you would have needed to go to Silicon Valley for.