[0:00:09] So hello, lovely Cheltenham audience. Oh, you can do better than that. I've done lots of events here. I know you can do better. Hello, lovely Cheltenham audience. Fantastic. Thank you so much for coming along tonight. My name is Vivian Parry. I'm chair of the Science Festival. I've come to introduce this event particularly because DeepMind is our sponsor. And actually, it's the first time for them.
[0:00:39] Cheltenham is in the forefront of trying to bring new science, engaging with people, understanding what they think, dialogue, and we're delighted to have been able, to have been involved with DeepMind to do that. So, with no more ado, would you please welcome Jim Al-Khalili and Demis Hassibis of DeepMind. Hello, good evening everyone.
[0:01:17] Welcome to Cheltenham Festival. I don't know how many events you've... Just show of hands, how many people came along to the event with Hannah Fry on... Oh, look at that. It's not bad. He's not going to say anything that he said in that event. It's all going to be fresh material.
[0:01:36] Welcome to the festival. I just want to say a quick corporate message first. The Cheltenham Festival, as you may know, is a charity. It offers these diverse opportunities to 25,000 children, young people in the community through world-class festivals. The Science Festival is just one of four and many outreach programs.
[0:02:00] I urge you to all have a look at the festival website to find out more and hopefully how to get more involved because this is your... People of Cheltenham are lucky to have these festivals. I'm sure you know that. And it's a wonderful place. The festivals were fantastic.
[0:02:17] This is my 15th year in a row. I missed the very first festival ever since I've been coming along, and I'm now on the festival advisory group that puts together the program each year. Right. Now, DEMIS has allowed me to...
[0:02:37] Say a little bit about his life and biography and his work and just indulge him a bit, and I hope he doesn't squirm too much, because it's a fascinating story. Demis Hassabis, I think, is by any measure a quite remarkable man.
[0:02:56] He was nicknamed the superhero of artificial intelligence. I think it was by The Guardian quite relatively recently. Of course, he's a former child chess prodigy. By the age of 13, he was ranked number two in the world for his age group. That's pretty right. He retired at 14. Been there, done that. He won a place at Cambridge at the age of 16 to study computer science, but was too...
[0:03:19] Young to start and had to wait. So while waiting, he begins his career working in video games and programming. He worked with the legendary game designer Peter Molyneux, with whom he co-designed and wrote the programs for the classic game. Now, was it Theme Park or Theme Hospital? It was Theme Park, right. Very successful, released in 1994. It sold several million copies. Having done that, he then goes off to Cambridge to study computer science.
[0:03:47] And after graduating, goes back and rejoins Molyneux, who had now set up another games development company called Lionhead Studios, which is actually just down the road from my University of Surrey on Surrey University's research park in Guildford. In 1999, aged 23, so still a young man,
[0:04:13] He won the Mind Sports Olympiad, an annual international multi-discipline competition for games of mental skill. He then goes on to win it five years in a row before retiring in 2003. It's good that you're giving lesser mortals a chance at this game. After several years in the gaming industry, Demis then switched back to academia and does a PhD at UCL in cognitive neuroscience, which allows him to get back to his...
[0:04:42] real passion and interest, which is artificial intelligence. During his PhD at UCL, he published several influential papers on memory and amnesia, developing a new theory on the way the mind creates and maintains the context of remembered events. In 2011, after postdocs at Harvard, MIT, and back at UCL, he then left academia to co-found the London-based AI startup DeepMind Technologies. This was with Shane Legge and Mustafa Suleiman.
[0:05:11] Among other things, the company developed an AI system capable of understanding and playing the 1980s Atari games, but not by being taught the rules. I don't know if we're going to, we might touch on that during our discussion, but it's just absolutely fantastic what these AI programs are doing. They're not showing the rules, they're just showing these pixels on a screen.
[0:05:33] And they just play and learn and practice and get better, just essentially as a human would. In 2014, of course, DeepMind was then acquired by Google for 400 million pounds. In 2016, it...
[0:05:48] Made history when the self-learning neural net program AlphaGo beat the world champion at the notoriously difficult Chinese board game of Go. And now, DeepMind has turned its attention to applying its algorithms to solving interesting problems in areas that can...
[0:06:08] As Demis will explain, he hopes, will benefit humanity. So areas in healthcare, in tackling climate change. I think you told me that...
[0:06:20] Our monkey brains have reached as far as they can go, and maybe we need the help of something more intelligent than us. Oh, and finally, Demis, this year was made a CBE and elected Fellow of the Royal Society. So, yeah, not bad going. Beat that for next year. So, anyway, Demis, that's enough. Welcome to the festival. Since you founded DeepMind, it's...
[0:06:48] Remarkable how quickly the field of AI has developed and changed. I suspect more quickly than anyone, even people in the AI field would have anticipated. How does it compare today with what it was like a decade ago?
[0:07:02] Well, it's quite amazing, actually, if you see. We sort of founded DeepMind about 2010, and we were thinking about it already 2008, 2009. And it's only 10 years ago, but it's actually very hard to remember how AI was back then. Because now AI is the sort of new buzzword. Machine learning is everywhere. And AI is talked about in all the big companies around the world. But back 10 years ago, AI was still not really discussed. There were some techniques that were being used.
[0:07:32] In a few places in academia, things like deep learning and some machine learning going on, but it was barely around in industry. I remember with my co-founders, we were just sort of reminiscing about this this last year. Even within our academic departments, it was not really the dumb thing to say that you were thinking about working on AI. It was still a hangover from what's called the AI winter, where obviously AI as a field has been going for 60, 70 years, and it's gone through various peaks and...
[0:08:01] It went through a big down phase during the 90s when the old rules-based AI systems reached their limits. It was still not really a topic of serious research in some senses. If you said you were going to be looking at researching AI, it was regarded as a little bit of a fringe activity. I remember us secretly talking about this in our corridors.
[0:08:30] During our postdocs and not really telling our professors what we were planning because I think they probably would have been horrified that we've kind of lost our minds. But yet we felt that...
[0:08:39] All the ingredients were sort of there to make a renewed attempt at the grand goals of AI, which is human-level general intelligence. And we were going to try and do that with a whole new approach, using neuroscience-inspired ideas, using the latest neuroscience and psychology. It had advanced a lot in the decade prior, fMRI machines, animal neuroscience. So we knew a lot more about the brain than we did in the 80s and 90s.
[0:09:09] We used some of those ideas. We had a lot more computing power. So, you know, cloud computing was coming in. The use of GPUs. GPUs, interestingly, were designed for doing computer graphics. But computer graphics turns out to be just matrix multiplication. And actually, a lot of neural networks is also just matrix multiplication. So we're very lucky as a field. We could just sort of piggyback on top of this commodity hardware that was already developed for gaming. And so we could see that there was a lot more...
[0:09:38] So there was a lot of different confluences that we felt, you know what, if we really went at this, got some of the brightest minds in the world together, put a lot of resources behind them, and combined that with a lot of engineering, we might be able to make some astonishingly quick progress. This is what we felt back in 2010. And then as it turned out, the progress was sort of even better than even our most...
[0:10:02] Optimistic projections at the time. So, you know, when we look back now, we were expecting there to be some big breakthroughs. We were hoping there would be some cool technology, some big papers. But actually, what's happened in the last 10 years has surpassed even sort of our most optimistic dreams. And that's surprising because for those of people that know me, they know me as an optimist and I'm always dreaming up crazy projects. And yet somehow, the reality of this has...
[0:10:31] It's been even more sort of wonderful than what we were dreaming, which is quite something.
[0:10:38] And do you have the sense that the breakthroughs are coming in even thicker and faster? I mean, is it speeding up, or did it just undergo a big revolution which we're now reaping the rewards of? So there's a combination. You know, there were some techniques like neural networks and deep learning, which is sort of very large neural networks you can think of as, that had just sort of come about in academia, and we were just figuring out how to make use of these techniques.
[0:11:08] And then there are other techniques like reinforcement learning, which is about decision making. And a lot of what we did at DeepMind was...
[0:11:16] Combine the best from these different techniques together. But we've got this feeling that the pace of innovation is continuing to accelerate. And so we're both simultaneously as a field reaping the benefits of what has already been done and trying to apply that to the real world now in all sorts of domains. And everybody in this room interacts with AI all the time, even though sometimes we don't realize it in our everyday devices. And then still coming up with brand new ideas that we know lots more.
[0:11:45] The leaps are required to get us all the way to something like general intelligence. We need memory. We need imagination and planning in our systems. We need abstract knowledge. We need to imbue these systems with language capabilities. So there's lots of breakthroughs still required. But the pace of that is, I would say, getting faster. And the other thing that's happening is as people are realizing
[0:12:09] There's so much potential here. It turns out that a lot of the best students now, when they're thinking about what to do with their PhDs, they're going into machine learning. Actually, sometimes at the expense of physics. I hope you'll forgive me for that. A lot of the smartest mathematicians, physicists and so on are now looking at machine learning as a field where they can really help with. I think in the next few years, we're going to get this explosion of extra talent coming into the field. Just to clarify,
[0:12:39] It is a different approach to what we call classical AI, where you write code, algorithms, and you give it rules, the heuristics. Machine learning is not the same as...
[0:12:49] Giving it rules and telling it if this, then that. That's right. So the history of AI, you know, AI has been around for 50, 60, 70 years. And a lot of the majority of that history has been about the attempts to crack intelligence, have been about building what's called rule-based systems or expert systems sometimes, or what's sometimes called classic AI now. And the idea there is that, you know, you break up the problem for the machine and you express the solution.
[0:13:19] You understand how to solve the problem. You express that solution in a set of rules or heuristics that the machine can execute. That was quite successful for a while. In fact, for a lot of practical applications, it's still very much used. You can build useful systems like that. The problem is it didn't really deliver when it comes to general intelligence. There's a number of reasons for that. One is that you can only get out of these systems what you put in.
[0:13:48] So it's sort of limited by the rules effectively you put in. So effectively, if anything new or unexpected happens, it just catastrophically fails. There's nothing in its code or its database or its set of rules that will help it deal with the unexpected. Now, the problem is, as we all know, the world is very unpredictable. And when these systems interact with real human beings, things always happen that the programmers didn't cater for. And then these systems just fall over. And we see that when we interact even with our assistants on our phone.
[0:14:17] You know, if you talk to Siri, for a while it seems like it might be intelligent. And then you hit the boundaries of what these rule systems can do. You know, you ask it something in a format or in a form that it hasn't been programmed for. And then you realize it says, oh, I don't know the answer to that, right? Or comes up with some stock answer. Because you realize, actually, it's not really understanding what you're saying. It's just sort of looking up the answers from a big database and applying pre-described rules to it.
[0:14:44] So the real revolution of these new systems is instead of hand crafting the solutions to these problems by teams of programmers figuring out what those solutions are, you actually build a system that's capable of learning for itself.
[0:15:00] So now the programmers, rather than giving the solutions, they're building architectures that are capable of learning solutions for themselves directly from raw experience by acting in the world and by learning from that experience and learning directly from raw data.
[0:15:16] You can think of it as hand coding versus learning from first principles. And it turns out that if you have enough compute power, enough data, and the right architectures, the right algorithms, that these systems are capable of learning very sophisticated things that we didn't know how to express as a rules-based system. That's the really exciting thing about this new world is we're actually solving a lot of problems now that couldn't have been solved with the rule-based systems.
[0:15:44] I mean, we're talking about artificial intelligence. I love your mission statement at DeepMind is solve intelligence and then use that to solve everything else. It's just brilliant, right? Because that is, you know, you personally, that is essentially your dream, that you make use of AI to tackle the problems that we don't know how to tackle because it'll be, in those certain tasks, cleverer than us.
[0:16:09] You said DeepMind is like an Apollo mission in terms of bringing so many different people together. How do you organize scientists around one long-term goal like solving intelligence? Yeah, so you're right. We thought of DeepMind, our vision behind DeepMind was...
[0:16:26] We had that mission statement. I think it makes sense. You can imagine us trying to pitch this to venture capitalists in 2010 before AI was a thing. And we had that mission statement then. That was the initial way we looked at it. And to us, it was logical because if you could fundamentally solve intelligence and crack intelligence in a general way, then you should have some very powerful tool that could be applied to all sorts of wide range of problems, especially we're particularly interested and passionate about scientific and medical issues.
[0:16:55] And so we felt that that was a natural follow-on. And in terms of the Apollo program mission, you know, what we thought about what sort of organization would be capable of undertaking that kind of research. And I've been fortunate in my, you know, your sort of long description of my background that you gave, very generous one, and that, you know, I've had a lot of different experiences, and I've always tried to pick experiences that would teach me something new. So I've been lucky. I've spent a lot of time at some of the best institutes in the world.
[0:17:25] I've learned from some amazing professors, how they run their labs, how they run research. But I've also spent some time at some of the best startups in gaming and learned how they organize engineering projects and get behind missions. And what I was thinking is, it'd be great to create a hybrid culture, which takes the best from both of those worlds. So you can think of DeepMind as really mixing together the best of startups, so the kind of passion they have, the pace they work at, the resources they can bring, and tension
[0:17:54] They can bring to problems. And the excitement that a lot of the best startups have. They're really successful startups. So I wanted to combine that, all that energy, with what's best from academia when it works well, which is true blue sky thinking, the smartest minds, having the time and space to think creatively, and exploring knowledge, passionate about knowledge for its own sake. And bringing that together, and also interdisciplinary thinking.
[0:18:23] A lot of that actually doesn't happen at universities, but it should do. But that's the sort of idealized state of academia if it was working really efficiently. So I was thinking, well, why not? They didn't seem to me to be opposing cultures. It felt like you could take those two things and put them together. And that's what DeepMind is.
[0:18:42] You know, I've read in history about, you know, I'm passionate about lots of things, but two things I'm passionate about is, you know, obviously the AI, the research we do, but I'm also very interested in organizational design. How can we create organizations and environments that foster innovation? And, you know, Apollo program, there's lots of examples of this in history, from the Apollo program, like big government projects, to companies like Pixar, which, interestingly, you'd think, well, how could you think about Pixar when you're doing a science company and Pixar's a great company? But actually...
[0:19:11] Pixar, in its early days, is an incredible company where time after time they're producing number one hits almost by magic. And how are they fostering that? And there's a brilliant book called Creativity, Inc. that was written about Pixar. And I've tried to collect together all these different experiences and figure out how to translate them into a scientific context. So one of the secrets we have at DeepMind is that most companies are top down in terms of the way, especially startups, in terms of the way they're run. The best scientific work is off.
[0:19:40] Often bottom-up creativity. So how do you sort of marry those two things together? And the way we do that is we have a long-term roadmap inspired by neuroscience and what we know from animal psychology and child development about how intelligence develops. And that forms a roadmap we have, a research roadmap that goes many years into the future. And then shorter term, we construct goals like win at the game of Go that we think are going to push the research in the right direction for the overall roadmap, but give us some short-term...
[0:20:09] There are some metrics about how we're doing. But we specify the benchmarks and the goals that we want to crack. But we don't specify the solutions to those problems. Anybody in the company, even the most junior researcher, can come up with ideas about how to solve a particular benchmark. And then we apply the scientific method to that. So it doesn't matter if it's the CEOs come up with the ideas or the most junior researchers. All the ideas are treated equally. And then it's empirically tested. And then you go forward with the one that really works in the best editions of the scientific method.
[0:20:38] And it seems to work very well. And we managed to scale that to many hundreds of, you know, we have 700, 800 people now of the brightest researchers in the world in this topic, which is, you know, quite a large collection of researchers to really coordinate together. A lot of brainpower to solve the problem of brainpower.
[0:20:58] And one of the most wonderful things to come out of AI research in recent years has been based on this idea of deep reinforcement learning. So in 2015, your team at DeepMind published the early 2015, January 2015, paper in Nature in which your neural net mastered this whole group of 1980s Atari video games.
[0:21:22] In particular, the game of breakout. That's the one, the lovely example. There's a YouTube video. I urge people to go and have a look at it. But can you describe exactly what it did? The system that we built was called DQN, and we debuted it in 2013. And then it was our first really big breakthrough that reached the public eye, as you say, in 2015, our first big Nature paper.
[0:21:45] There was a real watershed moment for those kinds of learning techniques because we combined at scale two basic learning techniques. One was hierarchical neural networks, which is called deep learning, but we combined it with this idea of reinforcement learning, which is a decision-making technique, to train the neural network to actually figure out what it should do in a game. And the system was called DQN, and it was the first system to ever really learn anything.
[0:22:14] End-to-end, so starting with raw data, in this case the raw pixels on the screen, and nothing about the rules, nothing about how to get points. It didn't even know it was dealing with a video stream, so it didn't know that pixels next to each other are correlated in time. It had to learn all that structure for itself from first principles and what it was controlling. So we just gave it the buttons and the directions on the joystick, but it didn't know what it was controlling on the screen. It just did a stream of numbers, 30,000 numbers per frame. And it has to make sense for itself.
[0:22:44] And the only guiding light it gets is we tell it the goal is to maximize the score.
[0:22:48] So it starts off randomly, trial and error, and then over time, it builds up knowledge about how this world works and what sorts of actions are good, what sorts of plans are good, and what will happen. It learns to sort of predict the kind of data stream, if you like. So if I was to do this action, what would be the next state I would be in, and is that a good state to be in? Will I get a reward?
[0:23:15] The breakout game, right, to those who remember. So it's the paddle at the bottom of the screen that you can... That's all you have to do. Move it left and right. So this ball, which is a pickle, bounces off the wall, and you bounce it back again, and it...
[0:23:26] Try to knock down this wall at the top of the screen. So we tried it on Atari games because the Atari console games in the 1980s was the first really big console that made it to mass market. And we wanted to pick a console that had lots of games on it because we wanted to prove this notion of generality, that this system we were building wasn't just for one game. You know, Space Invaders. There's lots of famous games on the Atari that you've heard of. Pac-Man, Space Invaders. All these games were invented for the Atari system. Breakout, as you say. And we wanted to show and actually...
[0:23:55] Indeed, we did show in our first paper that we could play more than 50 of these games at superhuman level, the best the world champion humans could play at these games, the highest scores they could get, from first principles, just from seeing the screen. And Breakout was the one that stuck in everyone's minds, the public's minds, because for us that was a further watershed moment that we've had many times since, which is that this game is, you know, you control this bat and ball, but there's this rainbow-colored brick wall you're trying to break through.
[0:24:25] And actually, you get points for knocking the bricks out of the wall, and if the ball goes past your bat, you lose a life, and you have three lives. And it's quite a hard game, because the ball comes back at you very fast. And DQN mastered this game, but then it came up with this own strategy, which was instead of knocking the bricks out of the wall in a normal sort of from the bottom up to the top, it dug a tunnel around the side and sent the ball behind the wall to the top of the screen. And yeah, so it bounced around. Obviously, it's a way more efficient way to break down the wall, and it's much less...
[0:24:54] It's risky because the ball can't go past your bat like that. And the first time we saw that, we thought, was it an error? And these amazing researchers that built that system, they're pretty good at research, but they're not so good at playing Atari. And so they had never, you know, this was like they'd learned something from their own system. They thought, wow, of course that would be more efficient if you did it like that. And so for the first time, and this has happened many times since, we saw something that the system had discovered for itself that we were then like amazed by.
[0:25:24] And then, of course, the following year, in 2016, was another big moment when AlphaGo famously defeated a human grandmaster at the Chinese game of Go, which is like a step beyond playing chess, where you do program computer heuristics, these rules. But I'm just intrigued, generally, in the fact that you use games...
[0:25:54] For your research, you know, what is it that you learn from, you know, why are you using AI to play stupid Atari 1980s games? You just do something useful. Sure. So, you know, that was one of the big competitive edges, I would say, that allowed DeepMind to be successful. So, of course, I'm personally biased in that I spent my whole life in games in one way or another, playing chess, as you said, from an early age, which got me into thinking about how the brain works. And then...
[0:26:22] Writing computer games, which sort of got me into simulations and how do simulations work. And I sort of realised that we could gain a lot of efficiency in our research if we use games as a testing platform to try out algorithmic ideas. Now, we subscribe to this idea of grounded cognition. So it's the idea that to have a true thinking machine, it needs to be embedded in a sensory motor reality of some sort, some kind of raw perceptual input. And there's quite a...
[0:26:52] A lot of people subscribe to this view. And cognitive machines, situated cognition. And generally speaking, if you subscribe to that type of view versus embodied logic systems, which don't have any connection to the real world, then you work on real robots. Because, of course, robots, by definition, are embedded in the real world. They're physical systems. So they come for free with this sort of grounded cognition component. But the problem with real robots is we love them as...
[0:27:19] as an application area. And lots of people at DeepMind love working on robots. But they're very costly and slow to work on in terms of if you're interested in the algorithms, the software, if you're interested in the intelligence of the robots rather than the motor movements and the physical capabilities of the robot. And we're interested in intelligence, the software, not the hardware. So we felt it would be a lot quicker to convert games. So we started off with board games and with computer games like Atari. And now we're up to very modern 3D games.
[0:27:48] So much more complicated games than the old Atari games. We've been going up the ladder of... We've been sort of following the progression of games development as our AI algorithms have been getting more sophisticated. And it's so much more efficient because we can run millions of experiments at the same time on the cloud. And you can run millions of experiments in one day and get back all those results and then update your algorithms. Whereas if we were to do that with real robotic hardware, that would take a lot longer.
[0:28:16] And then the other big thing about games is they're great testing ground because they usually, you know, we've invented games. Why do humans love playing games? It's because they capture, we've invented those games because they capture some aspect of the real world.
[0:28:30] Obviously, they're an abstraction of the real world, but they capture some interesting aspect of the world, some interesting challenge. And we've designed at least commercial games to be challenging and fun for humans to play. And we're after human-level performance, so it's an easy way to measure our progress against the best humans. And also, because most games come with direct metrics we can measure, which is winning and losing and scores. So it's actually quite convenient for us to test incrementally when we make a change to our...
[0:28:59] Are we actually getting closer to intelligence and are we getting a better system? Now, ultimately, as you say, we're not trying to build, we're not trying to solve games, right? I mean, that's their fun, but really what we're trying to do is build solutions to games in a general way that we think will generalize to real world problems.
[0:29:19] So it's very key the way we're trying to solve these games and crack these games, not actually just the sense that we want to master these games. As you say, you could hand-code solutions, and that wouldn't be so interesting for us because those hand-coded solutions wouldn't be transferable to real-world challenges, scientific and medical challenges, for example. So I wonder what are the remaining big research challenges?
[0:29:46] That you need to overcome before you can start using AI to tackle some of these real world problems. And I mentioned things like climate change at the start. Problems that we as humans don't seem to know how to solve yet. Right. So although there's been lots of spectacular successes in the last few years in AI, we're still at a nascent stage. There's still the early stages. These systems are pretty sophisticated for what they can do.
[0:30:13] But they're still very limited in many ways compared to what humans can do. So these systems are going to require memory. They're going to need imagination and planning capabilities. They need to learn abstract knowledge, which we would call concepts. They don't use language yet. They don't understand language. So there's a whole series of many, probably dozens, of big breakthroughs that are still going to be required, which is why it's so exciting still, before we get to something like human-level general intelligence.
[0:30:43] We think already these systems can be useful in certain situations with the capabilities they already have. And we're starting to explore that a lot. And we've already applied that in many ways. I think one of the most compelling ones was we used an AlphaGo-like system to control the cooling systems in one of Google's big data, giant data centers. So control the fans and the water pumps. And these are huge football pitch-sized data centers that we run all of our...
[0:31:13] All of our searches and things. And they use huge amounts of power. And they're amazing. The people who run these data centers are the best data center engineers in the world. I mean, Google's been building these things for 20 years. And we thought, can we try these algorithms on these cooling systems? And they were pretty skeptical, the engineers. They were saying, look, we've optimized everything we can possibly see. But you're welcome to try, but you won't find anything. And so we let these systems start learning, and they controlled a few things. And within a couple of weeks, they were saved.
[0:31:42] They were saving 40% of the energy the cooling systems were using. They were amazed, which is 15% of the energy over all the data centers used. So that's great for climate, because a lot of the world's energy is now being used on data centers. But also, obviously, it saves a huge amount of money, hundreds of millions of dollars per year if you spread that out over all the data centers. And then the really cool thing is, after these engineers, these incredible data center engineers saw what this system was doing,
[0:32:12] That gave them more ideas about how to design new data centers so that they would be more efficient in the future. And also, potentially, so to expose more controls and parameters to future AI systems. So there's this sort of big battle between human designers and what the AI can optimize that will create increasingly efficient data centers. Now, of course, you've chosen a particularly benign example there of the use of AI. But, of course, we know that much of the positive...
[0:32:41] The social impact of AI depends on the intent and ethics of the creator. How do you and DeepMind think and plan ahead for these? Because we're already at the stage where some of these AIs, the way they're solving problems is too complex for us to appreciate. We don't know how they did that, even in games like that.
[0:33:02] How can we keep control? I think there's lots of aspects to this. So we've been thinking about that since the founding of DeepMind, even prior to that. And the reason we were is because we were planning for success. So we were hoping our approach would work, and we were thinking ahead of time, what would that mean if we created those kinds of technologies? And I think there's several aspects to it. One is to always have ethical thinking at the heart. So we've had that from the start at DeepMind. We have a whole ethical team.
[0:33:32] There's a team, a society team, that actually researches these questions and collaborates with many academics and non-profits on these questions. And we've always been at the forefront of putting out there principles that we think are going to be important in terms of deciding how society decides how to use these systems. So, for example, we've committed to not using any of the AI systems we built for military purposes. And we believe all AI researchers should commit to that. And we've signed UN letters to that effect. And I think there needs to be...
[0:34:01] It's important for us to think about how to use these tools. AI is just like any other powerful technology or tool. It's neutral in of itself. It depends on how society decides and human designers decide to use it and deploy it that will decide whether it's overall beneficial or bad for humanity. And I think AI is no different to many other technologies in that way. It may be slightly more powerful and it's different and it's more general, but it's similar.
[0:34:28] I think we need to have that kind of debate, and that's why it's great doing these kinds of things and talking to the general public and engaging them so that they understand what these technologies are capable of and also what they're not capable of at the moment. You know, that we're very far from a lot of the science fiction scenarios that a lot of the alarmists talk about. And then another aspect to this that's very important, and we have a whole team doing this, is that you mentioned at the moment these systems are quite black box, so, you know, they solve these games or these problems, but yet we don't really know what approach...
[0:34:57] What is the solution they found? We can see that they've solved it because we can look at the behaviour. But we would also like to understand the decision-making process. And actually, I'm not so worried about that because, as I stress again, we're just at the beginning of this sort of success. And AI is an engineering discipline, scientific discipline, but it's an engineering discipline. And with all engineering, it's easier to build something that works first and then reverse it.
[0:35:26] And that's where we're at the stage now. We've just built systems that can do anything interesting at all, but we've proven that that can even work. And now lots of teams, including ours, are now looking at building visualisation tools and analytic tools to kind of tease apart these internal representations, these artificial brains, if you like, and actually figure out what are the representations that it builds up, what's the knowledge, how does it encapsulate that knowledge, and what factors does it weigh up?
[0:35:56] In terms of making decisions. So I'm very confident that in the next five plus years, we're going to see really sophisticated statistical analysis tools and visualization tools, you know, equivalent to brain imaging for real brains. Imagine doing that for artificial brains, right? Building a kind of brain scanner for them. As well as the fact that these AI systems, once they learn abstract knowledge, might be able to express themselves and tell us in languages that we can understand or symbols we can understand what their decision-making process is. So I think...
[0:36:25] We're actually just in a transitionary period at the moment of building any systems at all that work at all to fully understanding them. And that's just natural in any engineering discipline. And of course, we have to acknowledge that AI is coming and it's going to revolutionize our world potentially even more than the internet has over the past 25 years.
[0:36:49] Broad societal concern is about the impact that AI, robotics, automation in general is going to have on our lives and our jobs. If it's going to do a lot of jobs better than us, and this is more than the revolution of the 1970s where robotic arms replaced people on production lines and factories. And of course, people are then worried about...
[0:37:13] If it's going to widen social economic inequalities, if the gap between the haves and the have-nots is maybe exacerbated by AI, or that it'll be in control of powerful corporations that have ulterior motives, because people already know about the problems that you might face with Google and Facebook and so on, big data. How do we...
[0:37:35] Mitigate against that. I think there's a lot of issues that we need to be discussing now that are coming down the track, like a lot of the ones you've mentioned. I think there are lots of philosophers and economists that have tried to project into the future and predict what's going to happen. And as we all know, the future is notoriously unpredictable. So I don't think it's clear yet what's going to happen. Some of those things you said could happen. And the question is, is this just another...
[0:38:02] Normal, but very large technical disruption. And we've seen that already recently with the Internet, first of all computers, then the Internet, then mobile. And what's happened with those things is that jobs have changed. So some jobs have become obsolete, but whole new jobs that we could never have imagined before become viable.
[0:38:23] And I think there'll be a lot of that when robots or other AI systems do a lot more of the mundane work or number crunching work or brute force search or finding patterns in tons of data, which really the human mind's not built for. And that may free us up, say in medicine or science, to think more creatively and to use our time more of those pursuits or hypothesis generation, those kinds of things. And I think they could be extremely complementary tools to...
[0:38:52] Enhance what we're doing and to actually allow us to make faster breakthroughs and have more meaningful jobs that have more interesting aspects to them and remove a lot more of the drudgery. So I think that could be phase one. Then there's some economists that talk about phase two, which is if it starts becoming more and more general and capable, more intellectual tasks, what happens then? And what I think about that is that...
[0:39:17] You know, along will come with that, presumably, they'll be installed, those systems, because they're bringing a lot of productivity gains. And that will result in a lot more wealth. And the key thing to me there is to make sure that those benefits accrue widely to everyone in society. Whether that's through, you know, taxation or other things, and make sure that everyone benefits from that increased productivity in the world. So that it doesn't exacerbate this inequality. And, you know, I think there are many movements towards...
[0:39:47] And actually, a lot of the companies that are building these systems are open sourcing all their technology, a lot of their libraries, machine learning libraries, and a lot of the algorithms, and also publishing all of their data and also their algorithmic advances. And I'd like to see that continue and the sharing of these technologies. And I think that's what's likely to happen in the future.
[0:40:10] I had a few more questions to ask you, but I'm aware of the time. We started five minutes late, so I've been told that we don't have to finish until five past seven. So we have 20 minutes. So I think maybe time to have the house lights up and hand you over to the audience, as it were. Look at this white, wonderful Panama hat being raised there to try and get a question. But let's have one there and then one across there.
[0:40:39] To start. Hi. So, thank you for your talk. So, so far, you've talked about AI solving games with outcomes that you can measure. In terms of the general intelligence, you talk about imagination. So, how can you define and measure imagination, and how do you know when you've solved it?
[0:41:02] It's a great question. So you're right. As we get more into real-world domains, then very few real-world domains have simple metrics that you can optimize. Everything's very complex and interconnected. And a lot of the things we would like to achieve in the real world have multifactorial objectives. So there's lots of different things involved. So I think our systems right now, we're training them on these relatively simple domains.
[0:41:32] We're working on systems that are more general that can cope with much more ambiguous and complex domains. And there are a number of ways you can try and bring in this measure. You could have human operators that sort of provide the reward so we could observe it and go, yeah, that's behaving sort of what we expected or that is coming up with the kinds of things we wanted. You could also use evolutionary strategies so that it evolves the kinds of the correct
[0:42:01] What types of metrics it should be optimizing for. Indeed, that's how we've evolved to have the emotions and the goals that we then create. So there are quite a few, and these are very open research questions. This is right at the cutting edge of research. So we know that that has to happen.
[0:42:18] In a lot of cases in real world applications, both commercial and scientific and medical, the metrics are not going to be so readily available. Having said that, there are quite a few things, like the power center example I gave you, where there are very simple metrics you can use to train your system. In that case, the amount of wattage that you're using in the data center. So actually, surprisingly, there's quite a few metrics like that for quite a lot of real world challenges that
[0:42:47] Thank you. Yes. You cited the availability of massive data sets as being one of the critical factors in the AI spring we're going through at the moment. Are you comfortable with the way that large technology companies such as Facebook have gone about harvesting and processing that data? And do you ever worry that...
[0:43:11] If people start to react against that by being much warier in how they hand their data over to companies, it could end up spoiling legitimate AI research for everybody else. Yeah, no, it's a great question. I think that I am uncomfortable with the sort of indiscriminate use of data and I actually think...
[0:43:30] I think that we need to do a lot more in that area, actually, as a technology industry. It's sort of nothing to do with AI specifically, but actually in terms of making sure users know what their data rights are and the new European law that's come in. I think these are all good steps. I actually don't worry about that too much from an AI perspective because a lot of people have talked about data being the key thing. I actually think compute power is the key thing. And that's the reason we joined Google, actually. We didn't need to in terms of...
[0:43:59] Funding, because we had brilliant investors that were willing to put more money in as a private company. So that wasn't the reason we sold. The reason we wanted to sell at that point was we wanted to use all the amazing compute power that Google has available to it. And that's where the power of simulations comes in and games again. Because actually, if you can simulate various environments, and we're doing this now with scientific problems, actually, you can generate your own data.
[0:44:26] By running the simulations in the cloud. And then you have the AI interacting with synthetic data. And then obviously there's no worries there from privacy or any other issue. And actually...
[0:44:41] If you get the generator and the simulation right, you can generate vastly more data that way than you can through any other means. Now, of course, the danger of that is that there are some limits to that in the sense that no matter how good your simulation is or how realistic it is, it won't be as good as real data.
[0:44:58] And so you still need an additional step at the end. Even once you've solved something in simulation, you have to test it on a real data set in the real world. You have to do a real experiment to just check what you've been doing in simulation really makes sense and grounded in the real world, where there will be inevitably some differences, however accurate your simulation is. But I think that model, which is the model we use at DeepMind mostly, we actually don't use hardly any real data at all. We actually mostly, 90% of what we use is synthetic data.
[0:45:28] It's actually a very promising direction for AI in general. Thank you. Right, so the Panama hat's disappeared. Oh, look, there it is. I'm just fascinated by the fact that... And then here, the second row at the front. And then I noticed there are some... Yes, there's a very nice... Wave your hand at the back. We'll come to you next. Yeah. Thank you. Do you find that there's a natural...
[0:45:54] There's a plateau to the learning. One of my personal mantras is never stop learning, but I guess there's only so much you can learn about breakouts. Do you have some experiments that you've been running for some time, for some years? Do you find that there is a plateau? I love that motto, too. That's how I live my life, too. There's a nice Japanese word for it called kaizen, which is the striving for continual self-improvement, and I think it's a great way to live. But I think you're right. We do actually see the limits of...
[0:46:21] Asymptotic limits of what these AI systems can learn. Partly because it depends if the domain isn't rich enough, then it will top out. It will get maximum score, and that's all you can do. But also, at some point, the exploration part becomes very difficult to find anything new in that domain, even if you haven't solved it all. And you can see that in the improvement curve. Say, like with AlphaGo, you can see it improves very quickly over the first, say, million, first few years.
[0:46:50] So I think in many domains,
[0:47:06] It gets pretty close to some kind of optimal solution. Obviously, that's different from mathematically solving something where you can prove it's optimal. We just think empirically you're definitely getting into the area of diminishing returns. And then perhaps it's better to use your compute power, which is quite expensive, for something else. I think it's a very interesting question, and we're going to explore this a bit further by perhaps looking at games that have been solved, mathematically solved. So Drafts is one of those games, probably the most complex game that we've actually solved.
[0:47:35] And then we could build a learning system and see how close it gets to the known optimal that we have mathematically and sort of measure how close these systems can get. I suspect they're getting pretty close. But, you know, we haven't actually sort of formalized that yet. Thank you. Right. So the Panama hat's disappeared. Oh, look, there it is. I'm just fascinated by the fact that... And then...
[0:48:04] Thank you. Do you find that there's a natural plateau to the learning? One of my personal mantras is never stop learning, but I guess there's only so much you can learn about breakout. Do you have some experiments that you've been running for some time, for some years? Do you find that there is a plateau?
[0:48:32] I love that motto, too. That's how I live my life, too. There's a nice Japanese word for it called Kaizen, which is the striving for continual self-improvement, and I think it's a great way to live. But I think you're right. We do actually see the limits of, asymptotic limits of what these AI systems can learn, partly because it depends if the domain isn't rich enough, then it will top out. It will get maximum score, and that's all you can do. But also, at some point, the exploration part becomes very difficult.
[0:49:02] It's very difficult to find anything new in that domain, even if you haven't solved it all. And you can see that in the improvement curve. Say, like with AlphaGo, you can see it sort of improves very quickly over the first few million games, and then it tops out. And then once you get to about five, six million games, it doesn't seem to improve much anymore. And then at that point, we usually leave it for a little bit longer, and then we stop the training. So I think in many domains, it gets pretty close to
[0:49:31] I think it's a very interesting question, and we're going to explore this a bit further by perhaps looking at games that have been solved, mathematically solved. So Drafts is one of those games, probably the most complex game that we've actually solved as researchers.
[0:50:00] It was solved about 10 years ago. And then we could build a learning system and see how close it gets to the known optimal that we have mathematically and sort of measure how close these systems can get. I suspect they're getting pretty close, but we haven't actually sort of formalized that yet. Thank you. Right, so we had someone there, and then the next question will be right towards the far corner.
[0:50:24] I've got a question about consciousness. With your background in neuroscience, do you personally believe that consciousness will one day emerge in artificial intelligence? And if so, what do you anticipate will be the consequences?
[0:50:37] I think this is something quite a lot of us at DeepMind think quite a lot about, especially, as you say, the ones that have neuroscience and psychology backgrounds. I think, first of all, we need to define consciousness and agree on the definition. So that's still argued about even in philosophy, let alone amongst the neuroscientists. But let's just take our intuitive understanding of consciousness as a sort of straw man. And my view is that it's likely...
[0:51:04] That consciousness and intelligence are, what I would say, double dissociable. So I think you could have an intelligent system that would not feel conscious in any way to you, in the same way you do to me and I do to you. And certainly none of the systems we've built at the moment, they feel like machines. They don't feel in any way some kind of semblance of consciousness. And I think the other way is possible too. Any of you have pets, some animals exhibit sort of clear...
[0:51:31] I would say demonstrations of some form of consciousness, even dreaming. Self-awareness. Self-awareness of some sort. And yet they don't have the same level of intelligence that humans have. So it feels like they might be too continuum, although they're obviously connected. And the fascinating thing for me as a neuroscientist is that...
[0:51:51] I believe that trying to build AI in this neuroscience-inspired way, using neuroscience ideas, and then you can think of it as distilling intelligence into an algorithmic construct. I think one day in many years in the future, we might be able to compare a system like that with the human brain, and then we'll be able to see really what the differences are.
[0:52:12] And maybe that will shed some light on some of the unique aspects of our minds. Questions of consciousness being the biggest one, but also purpose of dreams, creativity, emotions, all the big questions that we have in neuroscience. I think that this journey we're on might be one of the best ways to answer that by trying to sort of like a controlled experiment, isolating intelligence from these other aspects of our minds. Thank you.
[0:52:39] So could you talk about the future of adversarial machine learning? So we're starting to see hacking and attacks of machine learning in the lab, but as we roll out machine learning and AI across society, how are we going to secure it?
[0:53:00] That's a great question. And I think this is a massively sort of active research topic at the moment. And again, I would say we're at the beginning of this. We've just built systems that are able to generate images, these kinds of things. So obviously, the next step was to interrogate.
[0:53:14] Investigate how robust were those systems and what allowed them to classify images and so on. And it turns out that these systems are not as robust as we would like them to be. So the classic example that the gentleman's referring to there is classifying a cat from a dog or maybe even more different, a bus from a bird. And then it turns out that you could change one pixel in the bus picture and now it thinks it's a bird.
[0:53:44] And even though clearly to any human, this would still be a bus. And so somehow what it's showing is that...
[0:53:51] These systems aren't as general as we thought they were. And I think what that means is we need to do more research on the generation system to make sure that it really is capturing something fundamental about the structure of what it's seeing, not overfitting to particular pictures, which is probably what that's showing that's going on there. And I think another defense against that is you can ask, why does the human visual system not get fooled by these tricks? And that's because we also have conceptual knowledge.
[0:54:21] So we know that when we change one pixel on a picture, it doesn't change the semantics of what's in that picture. Whereas the machine, certainly the ones we're building at the moment, that just do image recognition from the pixels and has no conception about what the picture is representing, the semantics of the picture, then it has no way to error correct when it's now one pixel's flipped from the thousands of pixels and now it's totally changed its mind about what the object is. So I think we'll start seeing more robust systems as they become more sophisticated.
[0:54:50] But this is going to be a whole area of its own sub-area of research in AI, which is called rigorous AI or robust AI sometimes. And actually, one of the ways we'll do that is automatically create adversarial examples of AIs to test our generated models. And you can actually think of the way we train AlphaGo as a sort of adversarial setup. It's trying to improve by beating
[0:55:16] Older versions of itself. So in a way, it's being tested by older versions of itself. So I think this is actually going to be this idea of multi-agent or AI testing each other to the limit. It's going to be one good way of us stress testing these systems automatically. Okay, so come to the front and then we'll go to the back in the middle. Demis, you've touched on what AI might discover in neuroscience, but are there other areas of science?
[0:55:45] Where AI and its use may make extraordinary discoveries. What fields of science do you think AI...
[0:55:54] Could really be used to tackle. And maybe Jim has a few too. So that's a great question. And the neuroscience part actually is not so much, although I think this will happen, AI being applied to neuroscience, but actually AI telling us something about our brain directly. So I think those two subjects are very interconnected in some fundamental way, which is why I studied both of them. We have both of those types of people at DeepMind. I'm really excited about AI applied to many areas of science. And in fact, even our...
[0:56:24] Our AI today, just deep learning on its own, has been applied to things as different as exoplanet discovery, to controlling plasma in fusion reactors, to analyzing data from the Large Hadron Collider. We've used it in medicine with Moorfields Eye Hospital to diagnose macular degeneration from retinal scans, 3D retinal scans. Other people have used it to...
[0:56:53] They detect skin cancer just from a 2D picture of the mole. And that's just the beginning. I think all branches of science are going to be touched by this. I think of it a little bit like statistics. It's this general tool that could be applied to all sorts of different branches of science. And I feel we're kind of just at the beginning of that. And what I say to things like...
[0:57:20] You know, what can you apply alpha zero type techniques already today? And I think it's any area where there's sort of three criteria, really. One is that it's, you know, you're dealing with a huge combinatorial search base.
[0:57:34] That's a good thing for these types of systems. Secondly, you can specify a clear objective or metric. In chemistry, that might be minimizing the energy in the chemical system. And then thirdly, you need either a lot of data or really a fast and accurate simulator. Now, if you have any domain that satisfies those three properties, you could potentially apply the techniques we've just talked about.
[0:58:03] We've talked about or ready to, let alone later in the next few years when we get these more sophisticated capabilities. So that's what really excites me, and that's why I spent my whole life working in AI, is because I really wanted to, what I wanted to do was advance our scientific understanding of the world around us. And I just felt a lot of the, and usually when that's what you want to do when you're a kid, you go into physics, right? Because traditionally physics is for people that want to, you know, it's the subject you go into.
[0:58:32] When you want to answer the big questions, whether that's the micro quantum mechanics questions or cosmology and the big questions of the universe. And I love physics still as a hobby. But I was thinking some of the things we...
[0:58:44] I have to say, the first time I met Demis, we were going to be talking about AI. He hits me with, tell me about the delayed choice experiment and the two-slip quantum mechanics. He's thinking about quantum mechanics, and that's his hobby. Right. And one day I'd like to go back to physics, but only after we solved AI, because it would be great to have that amazing tool and assistance to help us understand the even bigger questions about the universe.
[0:59:14] Yes, somewhere at the back. There was a hand and the microphone was almost... I think you said there was someone there. No? Maybe not. Yeah. Hi. I read an article recently about a technology that DeepMind had developed that was a neural network that had been trained to navigate a maze. And what struck me as quite profound is that...
[0:59:50] There was a structure in the network that they'd seen that resembled the formation of a structure of the actual mind, something called grid cells used to navigate. And I just wondered whether you thought that deep neural networks were the technology that had the potential to actually create a human-level intelligence, or whether new technology will need to be developed to get there.
[1:00:20] Yeah, that's really cool, actually, because the study you're referring to, we just published it in Nature a couple of weeks ago. People are interested to see it. And it was one of my favorite pieces of work we've done because what it was was we're now using quite rich 3D worlds, graphics worlds, not 2D Atari games anymore. And one of the big...
[1:00:43] One of the challenges we set our AI systems is to navigate through a maze and the maze changes all the time and it has shortcuts and doors open and close and you're trying to exit in a certain time. And of course this is exactly like the test that you give to rats.
[1:00:59] in laboratory experiments. And as you're referring to, the Nobel Prize was won a couple of years ago by some colleagues of mine who discovered these cells in the brain called grid cells that are kind of like, you can think of them as mapping out space like graph paper. Actually, when they first discovered, which was 2005, when I first saw them, I thought, wow, this can't be a real, this can't be a natural system. It looks like they're hexagonal patterns, really like a hexagonal graph paper.
[1:01:28] And what was amazing in our experiment was we were just trying to teach these AI systems how to navigate, so they were just learning how to do that. But when we analyzed their virtual brains or artificial neurons, what we found is that grid cell-like
[1:01:51] Neurons had developed of their own accord. We didn't program it in there with the exact same properties that the grid cells seem to have in real life. And it was amazing. And the Nobel Prize winners, we know them quite well, Edvard Moser and May-Britt Moser, they were amazed. We sent them an early copy of the paper, and they loved the study because for them, it meant that it really helps them in their neuroscience because they know now what it seems to suggest is that this grid cell encoding of space
[1:02:21] May not just be something the brain had to do as a quirk of biology, but maybe in some sense some kind of optimal way of coding space. So it really helps their search for, well, how is our brain coming up with grid cells? Maybe it's not pre-programmed. Maybe it's something, you know, our brains are these general systems, and they learn to...
[1:02:39] Break up space in that way, the same way our AI systems do, because that's the most efficient way sort of mathematically to do it, which would be really cool. And so that was a kind of amazing, for me, way of AI and neuroscience really interacting in a very positive feedback loop, kind of inspiring each other. And I think we're just at the beginning of that, and we're continuing that work in collaboration with the neuroscientists. Actually, second part of the question, are neural nets enough for intelligence?
[1:03:06] I think they are a critical component. I think they will be a critical component of solving intelligence. I don't think my personal betting is they won't be enough on their own. I think that we're going to need other types of algorithms. I mean, we use reinforcement learning to do the decision-making part, but I think we're going to need at least half a dozen probably other modules or other types of algorithms working in tandem or embedded within these neural networks. And I think those are yes.
[1:03:35] But other people think...
[1:03:38] You know, diehard neural network people think that it's just a matter of time and making these neural networks even bigger, and one day we'll get abstract knowledge and language and other things out of it. I'm a bit skeptical of that, that I think we're going to need other things. And I think the brain suggests that there are other types of systems that work together with the way these neural networks work that are required for intelligence. I mean, I think the easiest way to think about it at the moment is the analogy to the type of neural networks we have at the moment is...
[1:04:07] It's equivalent to sensory cortex, so like visual cortex or auditory cortex. And there's a lot more to the brain and our cognition than just our sensory cortices. Thank you. Well, we could carry on talking for a very long time, and I know we have various questioners still waiting, but I'm afraid we've run over the time, and I know maybe a lot of you are going on to other events, so we will have to draw this to a close, unfortunately. Ladies and gentlemen, please join me in thanking my guest, Demis Hassabis.