{"url": "https://www.youtube.com/watch?v=t03xNZ9qY1A", "page_url": "https://www.youtube.com/watch?v=t03xNZ9qY1A", "title": "Demis Hassabis: The Future of AlphaGo & Artificial Intelligence", "show_or_channel": "The Artificial Intelligence Channel", "date": "2017-05-24", "duration_s": 1920, "type": "talk", "speakers": ["Demis Hassabis"], "confidence": "medium", "notes": "Future of Go Summit, Wuzhen, May 2017 (talk at Chinese Go Association forum); third-party upload. Alt: yY8pu5qcHxQ. (YouTube upload 2017-09-28)", "people": ["demis-hassabis"], "id": "2017-05-24_the-artificial-intelligen-demis-hassabis-the-future-of-alphago", "audio_duration_s": 1919.8490625, "worker": "jc-w15", "transcribe_s": 25.7, "model": "large-v3", "segments": [{"start": 0.08, "end": 29.46, "text": "Good morning, everyone, ladies and gentlemen, and distinguished guests. A warm welcome to the AI Forum and the Future of Go Forum in Wujian. I hope today is going to be very interesting and informative, and I hope you're going to enjoy the interesting talks that we've lined up today. And I hope today will actually stimulate a lot of collaboration between Western scientists and Chinese AI scientists to progress the future of AI together."}, {"start": 30.61, "end": 60.93, "text": "So I'm going to begin today by talking a little bit about intuition and creativity and how we might give machines these capabilities. I'm also going to give you a bit of an overview of what my company, DeepMind, is about. So DeepMind was founded in 2010 in London. We have around 500 people now, including over 250 research scientists. We joined Google in 2014, about three years ago."}, {"start": 60.93, "end": 90.9, "text": "to turbocharge our mission of solving intelligence. And the way we can think about DeepMind is as a kind of an Apollo program effort for AI. So trying to bring together the world's best scientists in all sorts of different disciplines from machine learning to neuroscience together in one place with all the resources they need in terms of compute power and data and try and make as fast a progress towards solving AI as possible."}, {"start": 90.9, "end": 122.83, "text": "Another thing that we do at DeepMind is we've designed a kind of new organizational way to do science, science and scientific research. And you can think of it as a kind of hybrid combination of the best from academic labs and fusing that with the best from startup companies and creating a unique culture that has the best aspects of both. Long-term blue sky thinking from academia and the energy and focus that you get from the best startups. So what is our mission?"}, {"start": 122.83, "end": 155.6, "text": "Well, the way we describe it is a kind of two-step process. So firstly, we would like to fundamentally understand what intelligence is and recreate it artificially. So we want to solve what intelligence is. And then once we've done that, we believe you could use that kind of technology to solve almost everything else. So we think this is going to be one of the most important and powerful technologies that humanity will ever invent. So how do we propose to do that more practically?"}, {"start": 155.98, "end": 184.98, "text": "Well, at DeepMind we talk about trying to build general purpose learning systems. Let me just unpack a little bit about what I mean by that. So firstly, and most importantly, is the idea of learning. So all the algorithms that we work on at DeepMind have learning at their core. They all learn automatically how to master tasks. They learn from raw experience and from raw data. They're not pre-programmed with the solutions in any way."}, {"start": 186.48, "end": 216.3, "text": "The second thing is this notion of generality. This is the idea that a single system or single set of algorithms can operate across a wide range of tasks straight out of the box. Perhaps even working on tasks that it's never seen before. Now, of course, we have an example of such a system. It's the human mind. We do this all the time. Humans do this all the time. Use their learning from one task and translate that learning to a new task."}, {"start": 216.3, "end": 247.89, "text": "Even a new task that they've never done before. And they're able to do that task reasonably well straight away. And that's because they usefully use their prior knowledge they've built up on previous tasks. And this is something that machines are not very good at at the moment. And we would like to make them better. So some of the key techniques that we use in order to build out general purpose learning systems are deep learning, which are neural networks, hierarchical neural networks, that many of you will be familiar with. And reinforcement learning."}, {"start": 247.89, "end": 283.07, "text": "Which allows the system to learn for itself to try and maximize the rewards it's going to get in any particular task. So we call this kind of general purpose learning system artificial general intelligence. And we call it that to distinguish it from most of what people call artificial intelligence today. Which is usually still of the pre-programmed variety. So we're very interested in this idea of general learning. Now one of the best examples still are the biggest and most famous examples of narrow AI."}, {"start": 283.07, "end": 305.68, "text": "is IBM's Deep Blue system, which famously beat Garry Kasparov at chess in the late 90s. Now, this was a very impressive feat. Obviously, it was an incredible technical achievement. But Deep Blue was a pre-programmed system. So a group of top programmers sat down with some chess grandmasters."}, {"start": 305.68, "end": 335.41, "text": "And they figured out what the knowledge the chess grandmasters had in their heads. And then they codified that knowledge into a set of rules and heuristics, which Deep Blue then used by using brute force search to figure out what move it should make. But if you think about it, although this is an impressive technical achievement, where does the intelligence of the system reside? And I would argue it's not in the program or the machine. It's actually in the mind of the programmers."}, {"start": 335.41, "end": 367.54, "text": "It was the programmers that solved the problem of chess and solved how the machine should play chess and then codified those rules. The machine itself just dumbly executed those rules. It didn't adapt or learn or find out new knowledge for itself. So by contrast, we want to work on systems that learn for themselves. And we think of intelligence within the framework of what's called reinforcement learning. And I'll just briefly explain what reinforcement learning is in this diagram for those of you who don't know."}, {"start": 368.88, "end": 400.51, "text": "So reinforcement learning, if we start off by thinking about the AI system here on the left, which we call an agent, and the agent finds itself in some kind of environment and is trying to achieve a goal. Now the environment could be the real world, in which case the agent will be a kind of robot, or the environment could be a virtual environment like a game, in which case the agent would be a sort of avatar. Now it's trying to achieve this goal, but it only interacts with the environment in two ways."}, {"start": 400.51, "end": 429.89, "text": "First of all, the agent gets observations through its senses about the environment. So in most of our work at DeepMind we use vision and we use vision as the main sense, but you could also use sound and touch if you wanted. And the first job of the agent system is to build up a model of the environment out there based on these observations. And this is difficult because in a real world setting you never have a complete set of observations."}, {"start": 429.89, "end": 458.8, "text": "The observations are always noisy and they're incomplete. So you have to do your best to estimate and approximate what the environment is out there. And once you've built up this model of the environment, you can then use it for a second thing, which is you can run plans in your head, in the agent's mind, and decide which is the right action to take, what's the best action to take in the current moment to get you incrementally towards your goal."}, {"start": 459.38, "end": 490.51, "text": "And once you've decided what the best action is, you output that action, you execute that action, and then that may or may not make a change in the environment, and then that drives a new observation. And this sort of happens in a real-time circle. So that's the essence of reinforcement learning, and it seems quite simple on this diagram, but actually there's a huge amount of complexity and difficult technical challenges that have to be solved to get this system working really well."}, {"start": 491.02, "end": 520.98, "text": "But we know that if we can solve all of those challenges, then this is enough for general intelligence. And we know that because this is the way actually humans and animals learn from their experiences. And in fact, in humans, it's the dopamine system in the brain that implements a form of reinforcement learning. So our first big success was building agents to play Atari games. If some of you remember that from..."}, {"start": 520.98, "end": 547.6, "text": "The 1980s, classic 8-bit Atari games. And I'll just show you a video in a second of the agent learning for itself how to play an Atari game and mastering an Atari game. But before I run that video, I just want to make clear what it is you're going to see. The only input the agent gets from the environment is the raw pixels on the screen. So that's around 30,000 numbers per frame, per 50th of a second."}, {"start": 548.53, "end": 578.13, "text": "And all it's told is the goal is to maximize the score, get the highest score possible. Everything else is learned from scratch, from first principles. So it doesn't know what it's controlling, it doesn't know what gets it points. It has to learn all of those things by observation, by experiencing the environment that it's in. And finally, we insist that one system can play all the hundreds of different Atari games, the same system. So this is the idea again of generality."}, {"start": 580.59, "end": 609.71, "text": "So we built this system called DQM, and we invented this technique called deep reinforcement learning, which combines deep learning and reinforcement learning together. And we used it on this game Breakout, which I'm going to show you here. So Breakout, you control the bat and the ball at the bottom of the screen, and you're trying to break through this rainbow-colored brick wall. And you can see here, after 100 games of practice games, the agent's still not very good at the game. They're missing the ball most of the time. But after 300 games..."}, {"start": 609.71, "end": 639.41, "text": "You can see now the agent is a lot better and actually gets the ball back most of the time, even when the ball's coming back at very fast angles from the wall. So we thought that was pretty good, but what would happen if we just let the machine play another 200 games? And what happened was this unexpected thing was it found the optimal strategy was to dig a tunnel on the left and send the ball behind the wall, if you can see that, with like superhuman accuracy. And of course, that's the best strategy to solve this game."}, {"start": 639.41, "end": 669.46, "text": "But the interesting thing is that the programmers and the researchers who created that system, they're brilliant researchers, but they're not so good at playing Atari games, and they didn't know that strategy themselves. So they learned something from their own system, which is quite funny. So that was about two or three years ago, and then now, of course, the reason we're all here is our new system, our latest system, AlphaGo. We've been working on AlphaGo now for about three years, and with a sort of large team for about the last two years."}, {"start": 670.54, "end": 699.78, "text": "And AlphaGo, of course, is being designed to play Go, but we've also tried to design it to be general purpose, so that eventually it'll be able to do a lot of other things too, and I'll come back to that at the end. But first of all, why is Go such a great challenge for computers? And why did we decide to take this challenge on? Well, the complexity of Go, which we talked about yesterday, is so immense. It actually makes brute force search intractable."}, {"start": 699.78, "end": 729.86, "text": "There's no way, even if you had all the computers on the planet and you ran them for a million years, you still would not have enough compute power to calculate all the possible variations in Go. And the search space, in fact, is so large, it's actually, there are more possible board positions in Go than there are atoms in the universe. So that's one problem, the search space is really huge. But a second challenge, probably even bigger challenge, is that it was thought to be impossible..."}, {"start": 729.86, "end": 760.05, "text": "And in fact, it's still thought to be impossible to handwrite an evaluation function to tell the system who is winning in a particular position. So an evaluation function is a critical part of these computer game playing systems. You really need to be able to understand whether black or white is winning in a particular position to be able to determine what is the right move to play. And because Go is so complex and so intuitive, actually it's very difficult for programmers to figure out"}, {"start": 760.05, "end": 789.55, "text": "What are the right rules and heuristics to build into an evaluation function? It's much, much tougher than it is to do for chess because of the nature of the game of Go. So Go is a much more intuitive game than, say, chess, which is more about calculation. And, in fact, if you ask a great Go player, you know, why did they make a particular move, they'll often tell you that it felt right. You know, they had this feeling that it was the right type of move in this position."}, {"start": 789.94, "end": 822.02, "text": "Whereas if you ask a chess player, why did they make a particular move? They'll usually be able to tell you a very explicit plan. You know, I was planning to do A, because I thought my opponent would do B, and then I was going to do C. Now that plan may turn out not to be good in the end, but they usually have an explicit plan in mind. So it's very different, as far as I understand it, from playing chess and Go myself, the ways that the players approach playing these games. And this evaluation function..."}, {"start": 822.02, "end": 847.71, "text": "Why is it so difficult to write for Go? Well, I think there are lots of reasons why Go is harder than chess in this aspect. One of the reasons is that there's no idea of concept of materiality in Go. All the pieces are worth the same, right? They're all stones. Whereas in chess, you know, the queen is nine points, the rook is five points, the knight is three points. So, as a first approximation, you can just add up..."}, {"start": 847.71, "end": 878.54, "text": "All the points on both sides of the board. And that will tell you, as a first approximation, who's winning the game. In Go, there's no way of doing that. Secondly, Go is what I would call a constructive game. Whereas chess is a sort of destructive game. So in chess, all the pieces start on the board, and then as the game goes along, you take off pieces. So the game becomes simpler as the game moves on. Whereas in Go, the board starts empty, and you fill the board up."}, {"start": 879.5, "end": 908.69, "text": "So, when you're trying to make an evaluation in a particular position, in the middle of a game of chess, let's say, you have all the information already in front of you. The current board position contains all the information that you need. Whereas in Go, in the middle of a game, the game's only half constructed. So, in fact, if you want to understand who's winning and losing, you have to project forwards into the future. You have to predict the future and make some prediction about what's going to happen in the future."}, {"start": 908.85, "end": 936.64, "text": "It's not enough to just assess what is currently on the board. And obviously predicting the future is much harder. Another issue in Go is that a very small local change in the position, one stone in a different place, one square different place, can totally change the entire evaluation of the position. So it's very small changes can make large changes to the evaluation. And you can see that the idea of..."}, {"start": 936.64, "end": 968.93, "text": "Go being a much more intuitive game because actually for centuries now in the history of Go, there's been this idea of a divine move or a God move. A move so profound and so beautiful, it must be as if it's inspired, divinely inspired. There's no concept of this in chess. This is only in Go. So how do we get around these two immense difficulties, the search space and the evaluation function? Well, we use two neural networks."}, {"start": 968.93, "end": 997.47, "text": "It's called a policy network to help narrow the distribution of moves that we have to consider and the value network which learns for itself how to evaluate particular positions. I won't go into the technical details of this. I'll leave it to my colleague David Silver who will be talking next to go through the technical details of how AlphaGo works and why it works so well. We published all these details in a front cover article in the international journal Nature last year."}, {"start": 997.47, "end": 1022.59, "text": "And in fact, we're pleased to see that many other companies, including some Chinese companies, took these technical details and created their own versions of AlphaGo. And in fact, all the experts in AI and in Go proclaimed that it was more than a decade before its time. And for those of you who work in technology or science, you'll know that..."}, {"start": 1022.59, "end": 1049.68, "text": "Things never arrive a decade before you expect them. It's usually a decade after you're hoping that they will appear. So this was very unusual in the history of AI research, a breakthrough to come so quickly like this on a really tough challenge like Play & Go. Of course, we won the match, but actually the most important thing was the ideas that came out of the match and the beautiful moves that were played. So for AlphaGo..."}, {"start": 1049.68, "end": 1077.49, "text": "My favorite is Move 37 from Game 2, which I think really astounded the world and announced to everyone that AlphaGo really could come up with creative ideas. So here's the position from Game 2, and this is the move that AlphaGo played as Black on Move 37. And it's a shoulder hit, but it's a shoulder hit on the fifth line. And this is kind of unthinkable in this position, I've been told, from professionals."}, {"start": 1077.55, "end": 1108.35, "text": "And the reason is, sort of just the intuitive reason is, is there are two important lines in Go. The third line, where if you place a stone on the third line, what you're really saying to your opponent is, I want to take territory to the side of the board. Instead of that, if you play on the fourth line, then what you're saying to your opponent is, I want to take influence and power into the center of the board. And the idea is that you hope that that influence will later on translate into territory elsewhere on the board."}, {"start": 1108.35, "end": 1139.62, "text": "That will compensate for the territory you're giving away locally in this area. So for thousands of years, the received wisdom has been that an exchange on the third and fourth line is about equal for the two players. The territory you get on the third line is about equal to the influence that you get on the fourth line. So if that's true, then playing on the fifth line would be very unusual. Because playing on the fifth line to get influence into the center gives away territory from the fourth line."}, {"start": 1139.62, "end": 1171.36, "text": "You're giving away a lot more territory. But it seems as though one of the things AlphaGo prefers in its style is it really likes influence, and it really values influence into the center of the board. So perhaps for thousands of years, we as human players have been slightly underestimating the value of this influence in the center. Now, the cool thing about Go is that, as we all know, Go is more than just a game. It's an art form. But what I like about it is..."}, {"start": 1171.36, "end": 1200.98, "text": "One of the many things I like about it is that it's objective art. Because unlike most art, where it's subjective and you decide aesthetically whether you like it or not, Go has an objective purpose, which is to win the game. So you can value and objectively value whether a creative move is good or not as to how it affected the ultimate outcome of the game. And in fact, we know this move was very creative and beautiful because ultimately it did..."}, {"start": 1200.98, "end": 1233.47, "text": "It turned the tide of this game. So around 50 moves later, that stone on move 37 affected the fighting that happened in the bottom left-hand corner here. And these two stones in the bottom left-hand corner, black stones, ended up running out into the center of the board and connecting up around 50 moves later with the move 37 stone. Of course, this inspired Lisa Dole to even greater heights too. And he came up with his own divine move in game four."}, {"start": 1233.47, "end": 1261.46, "text": "And Moose78, the wedge move that confused AlphaGo and won him the game. It's another very beautiful move from the incredible creative genius of Lee Sedol. So the impact of the AlphaGo match in Seoul was huge. We had 280 million viewers of the match online and 35,000 press articles. And the thing I'm most pleased about is it really popularized Go in the West."}, {"start": 1261.46, "end": 1292.06, "text": "So the online board sales of Go boards in the West was up by 10 times in the couple of months following the AlphaGo match. So hopefully many more Westerners now are learning to play this beautiful and elegant game. So what was even more inspiring was the reaction from the players. So Lisa Dole told us that he found new inspiration from playing against AlphaGo. It was the greatest experience of his life."}, {"start": 1292.06, "end": 1321.86, "text": "And he said, I think this will bring a new paradigm to Go. I feel like I found the reason to play Go, which is really an amazing thing to hear and still sends shivers down my spine when I hear his voice on that. Other people commented on this too, and I will talk about that in a second. So let's just go back to the idea of intuition and creativity. So what is intuition? Well, one..."}, {"start": 1321.86, "end": 1350.22, "text": "Our operational definition of intuition is that it's implicit knowledge that's acquired through experience, but you can't consciously access it or express it to someone else. So it's knowledge you have, but you can't really access it consciously. So you might ask, if it's this kind of unconscious information, how do we know it's even there? Well, you can test it behaviorally. So you can test the existence and quality of that knowledge behaviorally."}, {"start": 1350.48, "end": 1376.56, "text": "So in Go, this is very easy. You can give someone a Go position and ask them to tell you what is the move they would play in this position and then evaluate that move. And then you can tell from that how good their intuition is on the game of Go. So we think we've mimicked this human behavior, at least in the domain of Go, with AlphaGo. So what about creativity?"}, {"start": 1376.56, "end": 1403.92, "text": "Well, again, one operational definition of creativity, and I'm not claiming this is all of creativity, I'm just saying this is one pragmatic definition of it, would be the ability to synthesize knowledge you already have to produce a novel or original idea. So you synthesize knowledge you've already collected through experience, and you output a new idea or a new piece of knowledge."}, {"start": 1404.94, "end": 1422.58, "text": "And I think AlphaGo clearly demonstrated these capabilities. So we continued with AlphaGo over the last year, and there were many reasons why we did this. There were many other scientific questions we wanted to answer. But one thing we wanted to do was to try and reach perfection and to see if we could fix..."}, {"start": 1422.58, "end": 1450.51, "text": "The knowledge gaps in AlphaGo that Lisa Doyle demonstrated it still had in the match in Seoul and see if we could build an even better system that was even more general and had fewer gaps in its knowledge. We called this new version AlphaGo Master, Master version, and we tested it for the first time online on the Go servers in January of this year."}, {"start": 1450.8, "end": 1479.86, "text": "And AlphaGo went on a 60-match winning streak, won all its games, and also created some new beautiful ideas on the way, which I know many professional Go players have been studying these 60 games and actually using some of the ideas that AlphaGo created. And here's two of them here. The idea of playing in the 3-3 point into a bare corner like this, and the idea of crawling on the second line."}, {"start": 1479.86, "end": 1510.0, "text": "Both of these things, I think, would have been pretty unthinkable before AlphaGo did them in these games and successfully won these games. So KJ himself said, after he saw these AlphaGo games, humanity has played Go for thousands of years and as yet, AI has shown us we have not yet even scratched the surface. And I particularly like this next statement, the union of human and computer players will usher in a new era."}, {"start": 1510.0, "end": 1536.82, "text": "Together, man and AI can find the truth of Go. I think that's really an amazing journey that we're on. And obviously, that's partly why we're here this week. And in fact, I had the honor of meeting Go Saigen's daughter last night at the gala dinner. And it was a real honor and pleasure for me because I've read a lot about Go Saigen, obviously an incredible player who revolutionized Go in the 30s and 40s."}, {"start": 1537.1, "end": 1564.94, "text": "And some professional players have told me that maybe AlphaGo will be the advent of a new revolution and a new era in Go playing, like Go Saigen did in the 1930s and 40s. And what is interesting to me is that chess programs, I would say, the top chess programs are better tactically than the best human players, but it's really the tactics that they're better at. They're better at calculation."}, {"start": 1565.01, "end": 1597.97, "text": "Whereas I think, interestingly, AlphaGo may come up with more interesting strategic ideas. And I think what's cool about that is that we as human players can look at those strategies and incorporate some of those ideas into our own play. And actually use that, I hope, as inspiration to come up with even better strategies themselves. I just reviewed a book by Cary Kasparov, who I know well. And he was writing a book about his experience of playing against the..."}, {"start": 1597.97, "end": 1628.61, "text": "Deep Blue in the 90s. And he had a great quote in there that I was pleased to see, which he said that Deep Blue was the end. AlphaGo is a beginning. I really believe that is true because of the way AlphaGo is built. It's built to be this general purpose system that can do many things and hopefully will end up doing many things. Whereas Deep Blue was built specifically to play chess and only chess. So I think in the future, that's what we're going to see more and more."}, {"start": 1628.61, "end": 1657.7, "text": "The power of the union of humans and machines working collaboratively together with different strengths and weaknesses. And I think together, human ingenuity augmented by AI will unlock our true potential of what we can truly do. So I see AI and programs like AlphaGo as tools. They're a little bit like a Hubble telescope or a microscope. So we use tools all the time to advance our civilization."}, {"start": 1657.7, "end": 1688.53, "text": "Physical tools. And we use things like telescopes to explore our universe around us and try to understand the world around us better. And that's how I see AI being as a tool for us as humans to use to explore the scientific world around us and to make faster breakthroughs in areas of science and medicine. And that's what we really wanted to do with AlphaGo is we wanted to try and push it further and see how far off optimal we were."}, {"start": 1688.59, "end": 1713.01, "text": "What would perfect play be? It turns out that 3,000 years of professional play is not sufficient yet to find optimal play. I think AlphaGo has helped us explore some of these mysteries. I think there are many other domains in science and technology and engineering that also suffer from this combinatorial explosion of possibilities."}, {"start": 1713.01, "end": 1744.29, "text": "Like the game of Go has. So many possibilities that it's impossible to explore all of them. So I think AGI could be this ultimate tool for us to investigate and explore the universe. And some areas of science that I'm thinking about and that we're going to be working on are areas like material design, designing new types of materials with the right properties, or perhaps drug discovery, discovering and designing new types of drug structures that target particular diseases. These are both kind of big..."}, {"start": 1744.29, "end": 1775.33, "text": "Search Processors that have lots of combinations. And we think that processors like AlphaGo might be able to help. And we also want to apply this to the real world. We love games as a training ground for our algorithms. We think it's the most efficient way to train our algorithms. But ultimately, what we're doing this for is not to play games, but to actually apply it to the real world. So we're looking at areas of healthcare, smartphone assistance, or education. And we already had our first success."}, {"start": 1775.33, "end": 1802.67, "text": "Applying AlphaGo techniques to optimizing the data centers at Google and the energy they use. And by using these AI techniques, we managed to save 15% of the power that the data centers used and the energy they used by saving 40% of the cooling systems, the energy the cooling systems used. So here's a graph of what happens when we switched on the power."}, {"start": 1802.67, "end": 1834.37, "text": "And the amount of saving that happened. And then when we switched on and off the AI system. And so now we're thinking, and my colleague Mustafa Suleiman will talk more about this later this afternoon, of applying this to a grid scale to an entire energy grid. I'm just going to finish by talking about going back to the initial mission of DeepMind. So we think of AGI as maybe a kind of meta solution. So if we think today, what are some of the biggest challenges we face today?"}, {"start": 1834.37, "end": 1862.58, "text": "Well, it's information overload. There's just so much information and data, and the systems that we're trying to understand are so complex, it's very difficult for even the top human experts to understand these systems and master these systems unaided. You think of things like climate change, the economy, or diseases. Things that we desperately need solutions to, but are very, very difficult and complex systems."}, {"start": 1863.25, "end": 1893.42, "text": "So we think solving AI in this general way is like a meta-solution to all these other problems. If we solve AI, then we can use it to help us solve these other problems. And for me, personally, my dream is to make AI scientists or AI-assisted science possible. And of course, one word of caution here, as with all new powerful technologies, and artificial intelligence is no different, we must make sure that it's used ethically and responsibly, and for the benefit of everyone."}, {"start": 1894.64, "end": 1919.85, "text": "And finally, as a neuroscientist, I'm very interested in our own minds and how our own minds work, as well as building artificial systems. And I think that this journey to build general intelligence might end up being one of the best ways to understand our own mysteries of our own minds. Things like consciousness, dreaming, and creativity. And us to truly understand what these capabilities are. Thanks for listening."}]}