Google Zeitgeist – Demis Hassabis: Human-Level AI is Decades Away, But We Need to Start the Debate Now •

At DeepMind, we basically work on artificial intelligence. And AI is the science of making machine smart. And I believe that this is going to end up becoming one of the most important technologies that humanity will ever invent. But before I get into that, because this session is called The Theory of Everything, I thought I would start with a little bit of an explanation of my personal journey to how I got to this point. And why I’ve decided to dedicate my life and career to work on this problem. So from a young age at school, I kind of came to this realization that in some fundamental sense, there’s only two subjects really worth studying, physics and neuroscience. And for physics, of course, is all about explaining the external world. So the external world out there, including, of course, the entire universe. And neuroscience and psychology is really about conversely explaining about what’s inside here our internal world. But then when I thought about this more, I actually came to the conclusion that the mind was more important, because obviously that’s the way we actually interpret the external world out there. And in fact, the mind interprets the world is something that was an idea that was first proposed by the great philosopher, Emmanuel Kant. And it’s really the mind that creates our reality around us. So this is where AI comes in, because the ultimate expression of understanding something is being able to recreate it. And as Richard Feynman say, one of my scientific heroes, what I cannot build, I do not truly understand. And that’s one of the things that I’m excited about with artificial intelligence. I think ultimately will help us understand our own minds better. So my personal journey as Doshini mentioned was started with games. And that’s how I got into AI. I started off playing chess. I sort of taught how to play chess when I was four. I ended up playing for various England chess teams and captain in various England, junior chess teams. And by age of 12, I was a chess master. And the thing is, when you steeped teacher kid from a very young age, how to play chess. And if they have a quite reflective personality like I had when I was young, you can’t help but thinking and introspecting about what it is about your mind that is actually coming up with these moves. What are the mechanisms that allow you to make these plans in such a complex game as chess? So then when I was around eight years old, I actually took some winnings that I won from an international chess tournament. And I bought my first computer, a ZX Spectrum 48K. And I taught myself how to program. And one of the first sort of big programs I can remember creating was actually a program to play chess. I didn’t play very well, but it was able to beat my little brother, which I was very pleased about when I was small. And you know, it worked. And this was the beginning of the path to meet towards AI. Now my love of programming and chess and games sort of came together naturally in the form of video games. And my first career was in creating and designing video games. And I did this for 10 years and I wrote several best selling games. But probably my most famous game I wrote was called Theme Park, which I wrote when I was 17. And this is one of the first games to use artificial intelligence as the main gameplay component. So this game came out in the mid-90s and 94. And the idea of the game was that you designed your own Disney world and thousands of little people with their own desires and characteristics came into that Disney world and judged and decided how much fun they had in your theme park. And then if they would go and tell their friends and they would come the next day. So this actually spawned, this game spawned a whole genre of games called management simulation games. And really sort of started this whole genre of creative games where instead of shooting and killing things in games, actually you create and build stuff yourself. And the game would react to how you played the game. So no two people ended up with the same game because the AI adapted to how the player played it. And in fact all my games involved a lot of AI. And then the final piece of the jigsaw for me was after doing this for 10 years I sold my games company and I went back to university to do a PhD in neuroscience to study how the brain itself solves some of these hard problems. And I decided to focus, I chose my topics, imagination and memory and an area of the brain called the hippocampus which is responsible for imagination and memory because these are capabilities two of the capabilities that we don’t know how to do very well in AI. And I wanted to see and get some inspiration for how the brain actually solves these problems. So after a couple of postdocs, MIT and Harvard, I then decided that I had all the ingredients in the components to start deep mind and actually attack the AI problem head on. So all these experiences then culminated in 2010 in me co-founding deep mind. And the idea behind deep mind was really to create a kind of Apollo program mission for AI. And now at deep mind we have over 100 research scientists, 100 PhDs, top people in their machine learning fields and neuroscience fields working on solving AI. And the type of AI we work on is this neuroscience inspired AI, so inspired by how the brain works at a very high level, systems level. So one way we articulate our mission at deep mind, it’s very easy to articulate, actually, what was it’s hard to do is a kind of two step process. So step one, solve intelligence and then step two, use it to solve everything else. And I’ll come back to that right at the end why that’s not so far as fanciful as it might seem. So more presently, how is it that we’re going to practically do this? Well, we’re going to go about doing this by trying to build the world’s first general purpose learning machine. And the two key words here are the words general and learning. So all the AI that we do at deep mind involves learning algorithms. So these are algorithms that learn automatically how to master tasks from raw data. They’re not pre-programmed or handcrafted in any way. So that’s unlike most AI out there that you’ve heard of. There’s also the second part. We enforce this idea of wanting the system to be general. So the same system or same set of algorithms can actually operate across a wide range of tasks. Now AI, of course, is a big buzzword at the moment. And there have been some big achievements in AI that you’ll, of course, be aware of, like, deep blue beating Gary Kasparov in the 90s. And then more recently, I began Watson besting the jeopardy contestants. But in our opinion, that’s still examples of what we would call narrow AI. So this is AI that has been specifically tailored or built to tackle one problem and one problem only. And the hallmark of sort of artificial general intelligence, the type of AI we work on, is that it’s built to be flexible and general and adaptive from the ground up. So there’s no special casing or pre-programming involved of the task involved. It has to learn everything from first principles. So we sort of think about the intelligence problem within this framework called reinforcement learning. And it’s a very simple framework to describe. I’m just going to describe it with a simple diagram. So on the left hand side here, you have the system itself, the AI system. And the AI system finds itself in some kind of environment that it’s trying to achieve a goal in. And that environment could be real world or virtual. Now the system only interacts with the environment in two ways. So firstly, it gets observations about the environment through its sensory apparatus. We normally use vision at deep mind, but you could use other modalities. And these observations are always noisy and incomplete. So unlike the game of chess, the real world is actually very noisy and messy. And you never have full information about what’s going on. And the job of the system is to build the best model of the world out there, statistical model of the world out there, based on these noisy observations. And once it has that model of the world, the second job of the system is to pick the best action that will get it closest towards its goal from the set of actions that are available to it at that moment in time. And once the system is decided which action that is, it outputs that action, that action gets executed, it may or may not make some change to the environment and that drives a new observation. And this whole system, although it’s very simply describing this diagram, if it has lots and lots of hilly complexities. And if we could solve everything behind this diagram, that would be enough for intelligence. And we know that that’s enough for intelligence because this is the way that all mammals, including humans learn. And in humans, it’s the dopamine system in our brains that implements reinforcement learning. So I’m just going to show you a few videos of the algorithm working, but before I show you that, I just want to explain clearly what it is that you’re going to see. So we used games as a testbed for testing the intelligence of our algorithms. And in order to have true thinking machines or cognition, a system has to be embedded in sensory motor data streams, sensory motor reality. And it has to figure things out for itself. So games are actually quite a perfect setting for this. And in fact, we used Atari games from the 80s, classic Atari games from the 80s, which are very challenging for design to be challenging to humans, but are not so complex that are algorithms couldn’t make progress with them. So what I’m going to show you is the AI playing these Atari games. But the only thing the system gets is the raw pixels as inputs. So it’s just like a human looking at the screen, seeing all the pixels on the screen. So there’s about 30,000 numbers per frame because the screen’s 200 by 150 pixels in size. And the goal here is to simply maximize the score. Everything is learned from scratch. And we insist that one, the same system plays all the different Atari games, the hundreds of different Atari games. So I’m just going to run this video now. This is a one minute video. This is Space Invaders, the most iconic game probably on Atari. And this first minute, this is before the first time the AI has ever seen this data stream. So they forget it doesn’t know what it’s playing, it doesn’t know what it’s controlling. And you can see it’s actually losing its three lives, it’s controlling the rocket here at the bottom of the screen. And it’s losing its three lives immediately because it doesn’t know what it’s doing. But after you leave it playing overnight on a single GPU machine, you come back the next day and now it’s superhuman at the game. It’s learnt for itself through experience how to play. So you can see now every single shot it fires hits something. It can’t be killed anymore. It’s worked out that the pink mothership that comes across the top of the screen in a second is worth the most number of points. It does these amazingly accurate shots to do that. And if those of you remember Space Invaders, as there’s less of them on the screen, they go faster. And just watch the last shot that it does, the rocket does. This is predictive shot for the hit to hit the last Space Invader. So you can see how perfectly it sort of modelled the game world and that data stream. So accurately, it can predict ahead of time what is going to happen. It’s going to come from the pixels on the screen. So this is a second video. It’s my favorite video actually. So this is the game of breakout. And there’s more gradations here of the agent getting better, the system getting better. So this is after 100 games. So just 100 games. And you can see, again, here, the system is pretty terrible. But you can probably convince yourself that maybe it’s starting to get the hang of the fact that it should move the back towards the ball. Now this is after 300 games. So it’s now hitting the ball back pretty consistently and it’s almost never missing. So it’s as bad as good as the best humans can be at this game. And then we thought, that’s pretty cool. But what would happen if we just left the machine playing the game for a couple more 100 games? And this amazing thing happened. What happened was it discovered the optimal strategy was to figure tunnel around the left hand side here and then send the ball with this unbelievable accuracy around the back. So that’s really cool because actually the brilliant programmers and researchers are on this program. They’re brilliant at programming and coming up with algorithms, but they’re not so good at playing Atari. So they didn’t actually know that strategy for themselves. So this is something that their own creation taught them. So all this work was then actually published on the front cover of Nature a couple of months ago. She’s the biggest science journal in the world. And so if you’re interested to read more about these details, you can check it there. And then now we’re moving on to adding things capabilities in like concepts and learning abstract concepts and long-term memory. And these are things that are inspired by my work and other people’s work in neuroscience and around mimicking the workings of this part of the brain called the hippocampus. And of course, we’re not just building these algorithms just to play Atari games. We’re moving now towards 3D games, go, simulations, and then ultimately real robots at some point. And more near term in terms of applications, we’re using it to improve recommendation systems like on YouTube. And also moving into health care, predictive health care applications. Now I just want to end by talking a little bit of the coming back to the sort of this theme of theory of everything. So I think, you know, two of the biggest challenges facing society today, facing us as a society, are information overload, just the sense that there’s so much data around. You know, everyone talks about big data, but the problem is is what to do with it once you have it all. And I think loads of areas like genomics and entertainment, you know, all sort of suffering from this deluge of data. How do we sift through this data to find the insights in that data? And of course, personalization technologies are one of the technologies that is trying to help us with that, but they don’t work very well at the moment, partly because they’re not very personalised and they work by sort of averaging the crowd. Secondly, there’s the problem of system complexity. You know, many of the systems we would like to master as a society like climate, disease, energy, economics, even physics are getting so complex now, you know, it’s difficult for even the best and the smartest humans to master it in their lifetimes, there’s still even enough time for them to innovate. So one of the reasons I work on AI and why I think it’s going to be one of the most important technologies out there is that solving AI is potentially a kind of meta solution to all these other problems. We can use it to help us solve all of these other problems. And the dream for me, the thing I get most excited about working on AI is in the future being able to make and create AI scientists or AI assisted science, making that possible, working in tandem with human experts and human scientists. Now of course, there’s been a lot of news at the moment about ethics of A round AI and like with all new powerful technologies, you know, it must be used ethically and responsibly. And we’re actively researching and doing these things and we have an ethics board that governs the use of this technology. But the technology itself is neutral. It’s always how humans use this technology that is the end up deciding whether it’s ethical or not. And of course, human level AI is several decades away but we should start the debate now. And I just want to end by just talking about we sort of joined forces with Google early last year. And Google’s mission statement of course is to organize the world’s information and make it universally accessible and useful. And one reason we decided to join forces with Google was that we felt our mission fitted very well with this mission. And one other way of sort of describing that mission is to think about empowering people through knowledge. And another way to think about this kind of AI or artificial general intelligence is that it’s a process that automatically converts unstructured information into actionable knowledge. So just to finish with a slide about the theory of everything, you know, of course I’m a neuroscientist as well as an AI researcher. And I think that by trying to distill intelligence into an algorithmic construct and comparing it to the human mind, that might help us to unlock some of the deepest mysteries of the mind like consciousness, creativity, and even dreams. And finally I would say that, you know, in order to find the theory of everything, it might turn out that we have to solve intelligence first. Thanks for listening.