This 60-Minute Demis Hassabis Lecture Will Teach You More About AI Than 5 Years of News

Thanks Alistair for that lovely introduction and it's so great to be back at Cambridge. I always have a warm feeling when I sort of my homecoming back to Cambridge and specifically this lecture hall as Alistair reminded me I think it is the first lecture hall I was in. It's always been my favorite lecture hall. I remember telling and I see a lot of my old friends here from my Cambridge days. I think about that one day maybe I'd come back to give a lecture in here and talk about announcing AGI and maybe a robot would walk on and astound everyone. I'm not going to do that today to disappoint you but maybe in a few years time I'll come back again and I'll give that lecture. It's an amazing place. It's such an inspiring place and I'm going to talk a little bit about how Cambridge has inspired my whole career actually and hopefully is going to do the same for many of you the students in the room.

So for me it started, my journey on AI started with games and specifically chess as Alistair mentioned. So I was playing chess from the age of four years old and very seriously for the England junior teams and things like that and it got me thinking about thinking itself, you know how does our mind come up with these plans with these ideas, how do we problem solve and how can we improve obviously when you're playing chess at a young age and you're trying to play competitively you're trying to improve that process and it was fascinating to me perhaps more fascinating than even the games I was playing was the actual mental processes behind it and AI and computers. I came across computers and AI for the first time in the context of chess. Trying to use very early chess computers like the one on the right here. I think this was my first ever chess computer. There were physical boards where you had to actually had to press the squares down to make the pieces move. And of course we were supposed to be using these chess computers to train opening theory and learn more about chess. But I remember being fascinated by the fact that someone had programmed this lump of inanimate plastic to actually play chess really well against you. I was really fascinated by how that was done and how someone could program something like that. I ended up experimenting myself in my early teenage years with an Omega 500 computer, amazing home computer back in the late 80s and early 90s and building those kinds of AI programs myself to play games like Othello. Really that was my first taste of AI and I was hooked from then on, and I decided from very early on that I would spend my entire career trying to push the frontiers of this technology.

So then that led me to Cambridge, which was really my three years here were incredibly formative for me. I went to a comprehensive school in North London. No one had gone to Oxbridge in living memory. But the reason I wanted to come to Cambridge was all these inspiring stories that I'd heard about what had happened at Cambridge, all these amazing people that I used to read about their biographies and the work they'd done. Especially people like Crick and Watson in the top left there. I remember particularly a film, the race for the double helix was amazing film from the 80s if you haven't seen it with Jeff Goldblum one of his early parts, with all the enthusiasm that he plays all of his parts as he was Watson and they were having just such an amazing time discovering roaming around Cambridge working on things like DNA and I thought that's what I want a piece of that and I want to feel what it's like to be at the frontier of discovery and what could be more exhilarating and that film actually really brought it to life what that might be like. Then of course all of my heroes, my scientific heroes, a lot of them had gone through Cambridge, people like Alan Turing and Charles Babbage of course in the lecture hall that we now sit in. And even places like the Eagle Pub where if you start at Queens College, one of the tours they give you on the first day is to go and show you the putative sort of table that they were discussing the DNA structure around, and you can't help but be inspired by that and walking down King's Parade and I almost felt like the intellectual giants of the past were almost speaking to you from the stones. That's how I felt going for a late night burger at Gardi's. That was what was inspiring me around all of these amazing people that had walked those same steps over hundreds of years and that's the history that is unrivaled here at Cambridge that I think we can still draw on and take inspiration from today. And then there's a picture of me and Aaron there, one of my best friends from Queens, on the mathematical bridge.

And then finally, when Alistair mentioned the Nobel Prize and it was an honor of a lifetime to go and collect that in Stockholm in December. Amazing one week of activities but my favorite activity is when you get to sign the Nobel book in the Nobel Foundation and that's the book there and one of my pictures of and I started leafing through the book. You sign your name and then you leaf back. You wonder if Crick is in there, and of course he is. And then you go back further and then Einstein's signatures there and it's just mind-blowing really. I started to spend an hour just photoing every page of the book. So it's full circle for me of that picture and then really seeing that film in the late 80s.

So then in 2010 we started DeepMind in London as really at the time it was a kind of an Apollo program effort for trying to build artificial general intelligence. AI that was truly general and could perform all the cognitive capabilities that humans are capable of. So it would be a truly general AI system. In fact, the idea for that really comes from Turing and Turing machines. So something that's able to compute anything that is computable as Turing showed with his Turing machines. And really that's been the foundation for me and one of the main things that I carried with me from the lectures here at Cambridge was all these theoretical underpinnings of computer science and computation theory that people like Turing and Shannon famously did in the 40s and the 50s.

So we started in 2010 and it's amazing like it's 15 years ago which in some ways isn't that long ago but when we started out DeepMind almost nobody was working on AI which is hard to believe today given that almost everyone seems to be working on AI today. In just over a decade things have accelerated incredibly and obviously we've been part of that very exciting journey.

So our mission at DeepMind from the beginning was building AI responsibly to benefit humanity. But the way we used to articulate it when we started out was a two-step process. Step one, solve intelligence. Step two, use it to solve everything else. This seemed very outlandish at the time in 2010. You can imagine trying to pitch venture capitalists on that basis. It seemed pretty crazy. But I still fundamentally believe in that today. I think more and more people are realizing that AI built in this general way could have this profound and transformative impact on almost any field. That involves accelerating scientific discovery itself and medicine and advancing our understanding of the universe around us.

So back when we started out there were basically two ways to build AI broadly speaking: the expert system way where you kind of pre-program an expert system directly with the solution, things like Deep Blue that beat Garry Kasparov at chess very famously in the 90s actually while I was studying here, that would be the pinnacle example probably of an expert system. But the problem with these expert systems and why they never really scaled to full general intelligence is they can't deal with the unexpected. If something unexpected happens that you didn't already cater for, there's nothing in the system that will allow it to deal with that. They were inspired by logic systems and they were quite rigid and fragile and brittle because of that. Whereas the modern day approaches are built on learning systems. These are systems that are able to learn for themselves and learn directly from experience or data from first principles and really inspired by more neuroscience ideas. Obviously the promise of these systems that we have today is that they can go beyond the knowledge potentially that we as the programmers or the system designers already know how to solve. And of course that's extremely valuable in areas like scientific discovery.

So we started in the early 2010s with games of course and I've used games many times in my life: first of all to train my own mind, then I used to build games and AI for computer games, and then finally in a third way to train up our AI systems. Games are the perfect proving ground for AI systems. You can start with very simple games like Atari games from the 70s. This system DQN was the first time anyone had built an end-to-end learning system that could learn directly from raw data. In this case the raw pixels on the screen and it's not told anything about the games or anything about what it's controlling. It's just told to maximize the score based on this video stream pixel stream input.

We were able to master all different Atari games around 2013. Then we took these systems and we scaled them up to the grand challenge of games AI, which is can you create systems that can play the game of Go at world champion level or beyond. Go is probably the most complex game that humans have ever invented. It's thousands of years old. It's also the oldest game and one of the most elegant games. There are 10 to the power 170 possible positions in Go, which is more than there are atoms in the observable universe. You cannot come up with a strategy in Go using brute force techniques. It would be impossible. So you have to do something much smarter. We famously in 2016 won a million-dollar challenge match against 10 times world champion Lee Sedol, one of the legends of the game, the South Korean grandmaster, watched by 200 million people around the world. Not only did AlphaGo win that match, importantly, it actually came up with new original Go strategies, even though we've played Go for thousands of years and professionally for hundreds of years. It was still able to find never seen before strategies, most famously this move 37 during game two, which surprised the best players in the world. It was an unthinkable move and yet it decided the game in favor of AlphaGo 100 moves later. That told me about the potential for these types of systems to invent and discover new knowledge. My dream was to generalize this to all areas of scientific discovery.

So how do these systems work? We basically train up these neural networks through a system of self-play. This is actually AlphaGo and also subsequent systems like AlphaGo Zero and AlphaZero that generalized what we've done for Go to play any two-player game from scratch. You start off with a version one of the system that doesn't really know anything about the game, just the rules. It plays randomly. You play say a 100,000 games against itself. That creates a new database of game positions from those 100,000 games. From that, you train a second version, a slightly better version of the model, version two, trained to predict what the likely moves are to be played in any one position and also who is more likely to win. Then you use that version two to play against version one in a 100 game match off. If it wins by a significant margin, say 55% win rate, you replace version one with version two and you create a new database of games that are slightly higher quality and then you learn a version three system. If you repeat this around 17, 18 times, you go from playing randomly in the morning to 24 hours or less later being stronger than world champion level. It's quite an incredible process to see this self improvement process playing out in a very short amount of time.

So if we think about what these neural networks are doing, you're reducing down this intractable search space of 10 to the 170 possibilities down to something that's much more tractable in a few minutes of compute time. It's doing this by narrowing down using the neural network to efficiently guide the search mechanism. If you think about this tree of possibilities, each of the nodes is a Go position. Instead of having to look at every possibility, you can use the neural network to guide you just down the most interesting and most useful lines to examine. After you've run out of thinking time, you pick the best line, the most promising line that you've seen thus far.

This then leads to we then played not just Go but any two-player perfect information game and it was even able to discover new strategies and new styles of playing chess which is extraordinary given that chess computers were so strong already. Programs like Stockfish and AlphaZero was able to beat Stockfish at chess, which is almost impossible to do. Not only did it beat Stockfish, in this particular position, one of the most famous games that AlphaZero played, it's called the immortal Zugzwang game. White is winning because it favors mobility over material. Most chess computers favor material. Black has more material but actually can't move any of its pieces. They're all stuck in the corner. AlphaZero sacrificed material for this mobility. For human grandmasters and top chess players, this is not only very effective style, it's very beautiful aesthetic style to play chess in. Some of the top chess players commented about this. Garry Kasparov said "the programs usually reflect priorities and prejudices of the programmer. But because AlphaZero learns for itself, I would say that its style reflects the truth." Magnus Carlsen said "I've been influenced by one of my heroes recently, one of which is AlphaZero." He actually incorporated a lot of these ideas into his own game.

So we did all these landmark breakthroughs in games AI over the first decade of DeepMind's existence but of course these were just the training ground for what we wanted to do. It was a means to an end. It wasn't the end in itself to play these games. It was to create these algorithms that could be generally useful for tackling real world problems.

What we look for in real world problems, not only scientific problems but industrial problems as well, we look for three different criteria that make it suitable to be tackled by these types of AI systems. Number one, we look for problems that can be described as massive combinatorial search spaces. Usually far too complex, far too many combinations to brute force the solution. Maybe there's some kind of structure that we can learn about with our neural networks that can guide that search very efficiently. Secondly, we look for problems that can be described with a clear objective function or some sort of metric that you can optimize against. In games that's very easy. It's things like maximizing the score or winning the game. There's a lot of real world problems that you can boil down to a few metrics or a few objective functions. And finally, you need quite a lot of data or experience to learn from and ideally an accurate and efficient simulator so you can generate more synthetic data to augment the real data. It turns out that there are a lot of problems that can be couched in these terms, including many important problems in science. The one that I always had in mind from my days of first coming across the problem here at Cambridge as an undergrad is the protein folding problem.

Proteins are incredibly important. They're the building blocks of life. Pretty much every function in the living body depends on proteins, from your neurons firing to your muscle fibers twitching. Proteins are what makes life possible. The protein folding problem is really easy to describe. A protein is defined by its gene sequence, which then specifies an amino acid sequence, which in nature then folds up spontaneously into usually a very beautiful protein structure. You go from this genetic sequence to a protein structure. The protein structure, the 3D structure, is very important because it goes a long way to defining what function it has in the body. The protein folding problem is can you predict the protein structure directly from this one-dimensional amino acid sequence? Can you predict computationally that incredible 3D structure from that sequence?

Why is this such a hard problem? Levinthal, a famous protein researcher in the 60s, described a conjecture that became known as Levinthal's paradox. He calculated roughly 10 to the 300 possible shapes that an average protein can take. Yet somehow in nature and in the body these proteins fold up spontaneously in a matter of milliseconds. That's the paradox. If there's so many possibilities, how does nature do this? How does physics achieve this? This gives you hope that this must be tractable computationally because physics does solve this problem billions of times a second in the body. Furthermore, what attracted me to this problem was that there was a biannual competition called the CASP competition, like the Olympics for protein folding, run by Professor John Moult at the University of Maryland since 1994. It's a great competition because they work with experimentalists who painstakingly find these structures using very exotic and expensive equipment like electron microscopes. They keep newly discovered structures unpublished and put it into the competition. The competition organizers know what the ground truth is but the computational teams try with their methods to predict those structures. Usually around a hundred proteins are in the competition and at the end of the summer they reveal the true structures and you can compare the predicted ones.

We entered AlphaFold one for the first time in 2018. We started the AlphaFold project in 2016, pretty much the day after we got back from the AlphaGo match in Seoul, Korea. We felt that the techniques were mature enough and ready to be applied outside of games to tackle really meaningful problems. We call them root node problems because if they could be solved, they open up whole new branches and avenues of discovery. Protein folding was a prime example. We started working in 2016. AlphaFold one was ready after a couple of years and we entered it into the CASP 13 competition. For the decade prior, these bar charts are showing the winning score in the hardest category of proteins being predicted. There was not much progress for a decade, stuck at a 60 points level. We were told by experimentalists that 90 points would be atomic accuracy, where you'd be within the width of an atom. That was the accuracy you had to reach so that experimentalists could rely on these predictions rather than having to do the laborious work. As a rule of thumb, my biologist friends would tell me it takes a PhD student their entire PhD, four or five years, to find the structure of just one protein. There are 200 million proteins known to science and 20,000 proteins in the human proteome.

With AlphaFold one, we were able to win this competition and be better by almost 50% than the next best system. AlphaFold one for the first time introduced machine learning techniques as the main component of the system. But it was not enough to reach this atomic accuracy. We had to go back to the drawing board with what we'd learned and re-architect it for AlphaFold 2 from scratch, using all the learnings from AlphaFold one to finally reach this atomic accuracy. This led the organizers to declare that the problem had been solved at the end of 2020.

This is an example of how AlphaFold works visually. On the left hand side is a very complex protein. The ground truth is in green, the predicted structure is in blue, and you can see how closely the blue overlaps the green. On the right hand side you can see how AlphaFold 2 works. It builds up that structure in an iterative process, recycling itself over 192 steps. It starts as a scrunched ball of amino acids and then builds out more and more plausible structure, refining the last parts until it has the finished prediction.

Because AlphaFold is so accurate, but also extremely fast, folding an average protein in seconds, we realized we could actually fold all 200 million proteins known to science. Over the course of a year, we used a lot of computers on the Google Cloud to fold all of them and then put them out freely on a database with our colleagues at EMBL-EBI just up the road at the Sanger Center just outside of Cambridge. We provided that for free unrestricted access to anyone in the world. If you think about how long it takes to do that experimentally, four or five years per protein, it's like a billion years of PhD time done in one year. It opened up whole new avenues of exploration because many of these structures, especially for less well-studied organisms like certain types of plants important for agriculture, would not have been found. Now all 200 million are available, and you can look at them at an aggregate level across species and see commonalities through evolution. There are really interesting new avenues of structural biology being explored. We thought about safety from the beginning. We take our responsibility very seriously at the forefront of AI. In this case, we consulted with over 30 biosecurity and bioethics experts to make sure the benefits outweighed any risks. Now over two million researchers are using it from pretty much every country in the world. It's been cited over 30,000 times and has become a standard tool in biology research.

It's been amazing to see what other researchers have done with all this technology. I've just called out six of my favorite examples. People at University of Portsmouth are using it to tackle plastic pollution, trying to design new enzymes that can digest plastic. We're working with the Fleming Center on antibiotic resistance. Neglected diseases like tropical diseases that affect the poorer parts of the world, we work with Drugs for Neglected Diseases Institute. For malaria, leishmaniasis, Zika virus, a lot of those structures were not known, but now they can go straight to drug discovery. There's been a lot of fundamental research on things like the structure of the nuclear pore complex, which is important for letting nutrients in and out of the cell. There's amazing work at the Broad Institute on drug delivery, designing molecular syringes. It's even being used in looking at mechanisms of fertility. Almost every area of biology and medical research is using AlphaFold.

We've continued in the last few years to develop more improvements. We released AlphaFold 3 earlier this year for academics to use. AlphaFold 3 now deals with interactions. AlphaFold 2 gave a picture of the static protein structure, but biology is a dynamic process. You need to understand how different biological elements interact with each other: proteins with other proteins, but also proteins with other molecules important to life like DNA, RNA, and ligands, so small molecules like drug compounds.

We have a separate set of work, AlphaProteo, which does the reverse of AlphaFold, but still using AlphaFold techniques. If you want to design a novel protein for a particular function, what is the amino acid sequence and the genetic sequence that will give you that structure? It's running it in reverse and trying to design new structures that will do novel things, which could be extremely useful for designing drugs, antibiotics, and antibodies.

Taking a step back, if I look at all the work we've done in the last 15 years, what are the implications for science and machine learning? Everything is about making this search tractable. You have an incredibly complex problem with many possible solutions, and you've got to find the optimal solution, a needle in the haystack of that enormous combinatorial search space. You can't do it by brute force. So you have to learn this neural network model. It learns about the topology of the problem so that you can efficiently guide the search to maximize or find the optimal solution. I think this is an incredibly general way to approach a whole myriad of problems. Thinking back to the Go example: we're trying to find the best Go move. But you could also change those nodes to be chemical compounds. Now you're trying to find the best molecule in chemical space, the best molecule that will bind specifically to the target you're interested in but nothing else, reducing side effects and toxicity. These are the same techniques we're using to design molecules now as we move into drug discovery.

I think in biology at least, we're entering a new era of what I like to call digital biology. I think of biology at its most fundamental level as an information processing system that's trying to resist entropy. That's basically what life is. It's a phenomenally complex and emergent information processing system. That's where AI comes in. Just like maths was the perfect description language for physics, AI is potentially the perfect description language for biology. It's perfect for dealing with the complexities of the emergent behaviors and interactions in a dynamic system like biology. I think AlphaFold is a proof point of that. I hope when we look back in 10 years time it won't be an isolated breakthrough but will have heralded a golden era of digital biology.

We started a new spinout company, Isomorphic Labs, to build on our AlphaFold technology and move into the chemistry space, trying to reimagine drug discovery from first principles with AI. Right now it takes an average of 10 years for a drug to be developed. It's extraordinarily expensive, costing billions of dollars. Why can't we use these techniques to reduce that from years to months, maybe even one day weeks, just like we reduced the discovery of protein structures from years to now minutes and seconds? We think of this as doing science at digital speed.

My dream one day is to be able to create a kind of virtual cell, a computational cell perhaps of something very simple like a yeast cell, so that you can actually run experiments in silico on it. The predictions you get out of the virtual cell would inform your real world experiments in the lab. You can reduce a lot of the search done in the wet lab and use the wet lab for validation steps rather than the very expensive and slow search process.

Of course, we've been using AI not just in biology but for science, mathematics, medicine more generally. We've had a whole range of breakthroughs: in health, identifying eye disease from retinal scans; discovering new materials; helping with plasma confinement in fusion reactors; faster algorithms, so AI discovering better algorithms for itself like faster matrix multiplication; doing weather prediction; and even helping with quantum computers and error correction in quantum computing. AI will be applicable to pretty much every field. I always encourage universities to start thinking very seriously about multi-disciplinary work where you apply AI to the right questions in a particular specialist field. I think there are many advances to be made over the next five to 10 years by doing that.

I'll just end with a more general view about not just AI for science but the path to AGI and how close we are to that. We've been making a lot of advances in general understanding of the world. We sometimes call them world models. We're particularly proud of our new video model called Veo 2, which was just released at the end of last year. It's state-of-the-art video generation. It's able to generate these videos just from a text description or from a single static image. Although some of these videos may not seem that impressive, if you think about the chopping the tomato one, this is like the Turing test for video models because usually you get the tomato coming magically back together or you're chopping through the fingers. It's actually understanding the physics of the world. The system had to really understand the physics of the world, or the bubbles around this blueberry, generating from text. Blueberries dropping into a glass of water, doing all the physics correctly. It's mind-blowing. Even if you told me 5 years ago that this would be possible without building in some special understanding of physics, I would have told you that seems unlikely. Yet somehow these learning systems are able to learn about real world physics just from watching many YouTube videos.

We've gone a step further with Genie 2, bringing my games hat back. This takes those video models a step further. Now with a text instruction, you can generate a whole game. At the bottom here, we said "generate a playable world as a robot in a futuristic city" and it just comes up with this and you can control it with keys. At the moment it's only consistent for a few seconds, but we're working to extend that so that the consistency of the game world lasts for many minutes. Then you've really got what I would call a world model, a real understanding of the real world and how interactions and physics work.

We've been working very hard on the safety aspects from the very beginning in 2010. We were planning for success even though almost nobody was working on AI back then. We imagined it would be a 20-year mission and amazingly we're on track 15 years in. We were planning for success because if we were to build transformative systems and technologies, it would come with a lot of responsibility to make sure they get deployed in a safe and responsible way. One of the technologies we built is called SynthID, which invisibly watermarks using an AI system, adversarially adjusting the pixels or the text or the audio imperceptibly to the human ear or eye, but it can be detected by a detection system that these were synthetically generated. It's going to become increasingly important as these technologies become widely deployed that we're able to easily distinguish between synthetically generated and real images.

AI has incredible potential to help with our greatest challenges from climate to health. But this is going to affect everyone. I think it's really important that we engage a wide range of stakeholders from society. I've been really pleased in the last couple of years that one of the consequences of AI becoming mainstream is that many governments have got interested in it. It's been great to see international summits. The UK hosted the first one at Bletchley Park a couple of years ago, bringing together heads of government with academia and civil society to discuss these technologies, how to put the right guardrails on it, how to make sure we embrace the opportunities but mitigate the risks. I think that's going to become increasingly important given the exponential improvement we're seeing. My shorthand for this is to say the mantra in Silicon Valley is "move fast and break things." That has created a lot of advances and technologies we use every day, but I think it's not appropriate for this type of transformative technology. I think instead we should be trying to use the scientific method and approach it with the kind of humility and respect that this technology deserves. There are a lot of unknowns about how this technology is going to develop. With exceptional care and foresight, we can get all the benefits and minimize the downsides, but only if we start the research and debate about that now.

We're now building our own big multimodal models that try and take the best of all of these different models and put it into one system. We call it the Gemini series. Our latest one is Gemini 2.0, which is state-of-the-art across many leading benchmarks. We're using it to further the next generation of assistants. I call it universal assistance, we call it Project Astra, where you have it on your phone or some other devices like glasses, and it becomes an assistant you can take around with you in the real world, helping you in everyday life to enrich your life or make you more productive.

The next step in AI is combining what I've shown you with AlphaGo, these kinds of agent-based models that can efficiently search through and find a good solution to a problem in a limited domain. In this case, in games, but we want to build those types of search and planning systems on top of much more general models like Gemini, these world models that understand how the real world works and can then plan and achieve things in the real world. That's key to robotics, which I think in the next two or three years is going to be a huge area that's going to have massive advances.

I'll just finish then by having a slight conjecture about what this all means if we think back to Turing and all the work he did to lay down the foundations of computer science. I see myself as a kind of Turing's champion in a way, seeing how far Turing machines and the idea of classical computing can go. One of my favorite things to think about is the P vs NP problem, which is a famous problem in computer science of what sorts of problems are tractable on classical systems. There's a lot of great work going on in quantum computing, many here in Cambridge and also at Google. A lot of things are thought to require quantum computing to solve. My conjecture is that classical Turing machines, basically classical machines that these types of AI systems are built on, can do a lot more than we previously gave them credit for. If you think about AlphaFold and protein folding, proteins are quantum systems. They operate at the atomic scale. One might think you need quantum simulations to find the structures of proteins, yet we were able to approximate those solutions with our neural networks. Any pattern that can be generated or found in nature, i.e. that has some real physical structure, can be efficiently discovered and modeled by one of these classical learning algorithms like AlphaFold. If that turns out to be true, I think that has all sorts of implications for quantum mechanics and fundamental physics, which is something I hope to explore with these classical systems to help us uncover the true nature of reality. That leads me back to the whole reason why I started my path on AI many years ago. I always believed that AGI built in this way could be the ultimate general purpose tool to understand the universe around us and our place in it. Thank you.

Yeah, there's many. That's what I studied for my PhD actually, was memory but also imagination, future thinking, planning. I really wanted to understand how the brain did that and it turns out the hippocampus is involved in both, so that we could maybe mimic that with some of these algorithms. I think there are many key things in that. Of course, there are all the big questions around creativity, dreaming, consciousness, all these big questions that I think building AI and then comparing it to the human mind is one of the best ways we'll make progress with those sorts of root node problems, like what is the nature of consciousness and is there something special about the instantiation of the substrate of the brain versus algorithmically mimicking that in silicon.

They're both great questions. For the first one, look, that's why we called it DeepMind, partly because the deep refers to deep learning. Deep learning had the early parts of it, it wasn't called deep learning then but it was just becoming common with Boltzmann machines and things that Geoffrey Hinton had invented just a few years before in 2006, 2005, these hierarchical neural networks. It seemed like a super promising idea even back then to us who came across it in academia. The other thing we bet on was reinforcement learning and the combination of that, which again is coming back into vogue but it was also important for us to solve something like AlphaGo. You need both parts: deep learning to model the environment and the world, and reinforcement learning to make the plans and solutions and take action in the world. There were two reasons we bet on that even when it was just the beginning. We knew that the classical methods, these expert systems, would not scale. That's something I learned here and also at my postdoc at MIT, that they were the kind of churches if you like of the classical methods. You learn here not just what to do but also what not to do and why it may not work. I felt they would never scale to the kinds of problems I wanted to solve with AI. Whereas the learning system seemed to have unlimited potential, although they were a lot harder to get anything significant to do at the beginning. They weren't scaled up enough. The other reason we started DeepMind in 2010 was we could also see the computing paradigm was shifting on the hardware side with GPUs and other things, which of course were also invented for gaming and turns out everything's a matrix multiplication, intelligence and gaming and computer graphics. All of those different influences came together. The understanding of neuroscience and fMRI machines had advanced a lot in that previous 10 years too. I felt it was the perfect time to sort of bring all of that together back in 2010. We were betting on it not necessarily because we knew it would work but because we were pretty confident the other methods would not work, and that's what the AI winters were about basically. For the second question, it is surprising that some of these things converge and actually we weren't sure. For the first couple of years, nothing worked. We couldn't even get one point on Pong, the simplest game you could imagine. We were wondering whether we were 10 or 20 years too early, like Babbage turned out to be with his difference engine. He had amazing ideas that worked but he was 50 or 100 years too early. I always say you want to be like five years ahead of your time, not 50 years ahead, otherwise you'll be in for a lot of pain like Babbage was. We were worried about that but then it did converge and that gave us confidence to tackle the harder problems. For the last part of your question about things in nature: they're not suboptimal. They're probably pretty optimal because they've gone through an evolutionary process, not just life with biology, but geologically and physically. Asteroids and physical phenomena combined together survive some amount of time because they're stable over time. If they're stable over time, then there's probably some structure that's learnable. That would be my conjecture.

I love that area and I've carefully followed it and helped people building like EEG caps and things. Of course the problem is resolution of these things to get the readouts from the brain and also ideally you'd want read and write. But I'm very fascinated by projects like Neuralink, chips in the brain obviously right now that's for veterans and people to get function back in their body. I think it's going to be amazing things where people will be able to walk again if they've broken their back and things like that. There's going to be some credible advances in medical sciences that will be amazing. But then beyond that, if things become routine and surgery is safe, I could imagine that would be one way for us to keep up with the technology. In some senses, it's no different to what we already have today with our technologies all around us. We all have our phones 24/7 with us and computers. We're already symbiotic almost with our technology. It would be one step further to have that, but I'm not sure. Maybe that's for the philosophers in the room to answer what the difference is if the technology is attached to you versus it's just something you carry around with you all the time.

It's a very complicated one because as you say things are changing at lightning speed. We were just discussing it with Alistair earlier, even designing three-year computer science courses is quite difficult given that the underlying material changes in less than 3 years. I think the only thing we can say for sure is there's going to be a lot of change but that brings disruption and opportunity. I'll give you an example on coding. I still would recommend that you get good at coding and math because I think you'll be able to use these new tools in a much deeper way if you have an understanding of how they're built. But on the other hand, coding is going to be more available to many more types of people because AI will allow you to program in natural language rather than a complicated computer language. That will open up fields for creative people to build games, make films, make applications, and the balance might shift more towards the creative side than the engineering side. But it will also enhance engineers to be able to do 10x of what they can do today. It's difficult to know. What I would say is focus on embracing those tools in your spare time and train yourself to be really good at picking up new information extremely quickly because that's what's going to happen in the next 10 years.

There's certainly lots of processes that can't be modeled today. But this goes back to what I said at the end of the talk. I'm not sure in the limit that there are. In the end, if physics can solve it and there's some structure there to be learned, probably with enough examples one could reverse engineer a model of that. I don't see any theoretical reason why a classical system could not make predictions or simulations of that biological system. So I don't really see what the limit could be. There are lots of abstract things like factorizing large numbers, cryptography, where there may not be any structure. If there is structure in the natural numbers, that will also be learnable. If there isn't and it's a uniform distribution, then you would need a quantum computer to crack cryptography. Those are open conjectures. But I think most things in nature have evolved over geological or biological time, so that suggests to me there is some structure to learn. That makes the search or prediction potentially tractable.

There are many ways AI is going to come in. One is the tools to build the assets you need for games, like 3D models and animations. I think that's all going to come in the next couple of years. You can think of AI for game balancing. You design a game and overnight it could play a million playthroughs and in the morning you could get a report as a game designer of things that are unbalanced. I think also bug testing for open world games. I used to make simulation games, open world games. They're a nightmare to bug test because the whole point is that the player can do almost anything and the game will react. How do you test 10 million people having their own unique journey? Having AI players play it before you release it could help solve a lot of those bugs. Excitingly as well, AI characters that are much more lifelike that move the storyline on. I used to dream about massive multiplayer worlds where the AI characters were intelligent and actually updated their beliefs and storylines based on what the players were doing, so it felt like a much more living, realistic world. I think we're on the cusp of building those types of games. Finally, the world model we're building is more about general AI and can you actually model the world. Does your model understand the world? If it can generate it for some amount of time, then it must be understanding something about the underlying physics. That's more for general intelligence. Maybe one day we'll have this holodeck thing where you can just imagine and it's all around you. We'll probably be able to have that once we have AGI, but that's still a ways off.