[0:00:05] Welcome back to the final episode in this season of the DeepMind podcast. And boy, have we covered a lot of ground. From protein folding AIs to sarcastic language models, sauntering robots, synthetic voices, and much more. It has been quite the journey. But we do have one more treat in store for you. A chance to hear from DeepMind CEO and co-founder, Demis Hassabis. The outcome I've always dreamed of is...
[0:00:35] AGI has helped us solve a lot of the big challenges facing society today, be that health, creating a new energy source. So that's what I see as happening, as a sort of amazing flourishing to the next level of humanity's potential with this very powerful technology. This was my opportunity to ask Demis all the things that have popped into my head during the making of the series. Well, most things. We'll see how far I can push it.
[0:01:04] As luck would have it, the day I sat down with Demis coincided with the opening of DeepMind's sparkling new premises in London's King's Cross. There weren't many people about yet, so it felt like an exclusive preview. I feel like I'm in a high-end furniture catalogue. Let me set the scene for you. This new building is rather beautifully appointed. It's got a double helix staircase running through the middle.
[0:01:33] There are fiddle leaf trees in practically every corner and there are stylish fluted glass crittle doors between offices. And yes, those meeting rooms christened after great scientists, Galileo, Ada Lovelace, Leonardo, they are all still a feature. Sparkling, push the boat out.
[0:01:59] While sipping on my beverage of choice, some memorabilia outside Demis's office caught my eye. A nod to AlphaGo's famous victory over Lisa Dole in the game Go. There is, sitting underneath two extremely fancy black spotlights, a chessboard in a black frame. If I go over to it, there's a picture of Garry Kasparov, the legendary chess player who was beaten by...
[0:02:28] Deep Blue, the IBM computer. He signed the chessboard and it says, for the AlphaGo team, keep conquering new heights. I mean, just a chessboard designed by Kasparov on the wall. Perfectly standard. Oh, awesome. Oh, we're going in.
[0:02:54] After settling down inside Demis's office, I started by asking him about DeepMind's long-term vision of building AGI, or Artificial General Intelligence. It's an ambition that has been baked into DeepMind's DNA from the very beginning. I think it's fair to say that there's some people in the field who don't think that AGI is possible. They sort of say that it's a distraction from the actual work of building practical AI systems. What makes you so sure that this...
[0:03:23] I think it comes down to the definition of AGI. So if we define it as a system that's able to do a wide variety of cognitive tasks to a human level, that must be possible, I think, because the existence proof is the human brain. And unless you think there's something non-computable in the brain, which so far there's no evidence for, then it should be possible to mimic those functions on effectively a Turing machine, a computer.
[0:03:51] And then the second part of that, which is it's a distraction from building practical systems. Well, I mean, that may be true in the sense of what you're mostly interested in is in the practical systems. AGI itself is a big research goal and a long term one. It's not going to happen anytime soon. But our view is that if you try and shoot for the stars, so to speak, then any technologies that you sort of build on the way can be broken off in components and then applied to amazing.
[0:04:18] And so we think striving for the long term ambitious research goal is the best way to create technologies that you can apply right now. How will you recognise AGI when you see it? Will you know it when you see it?
[0:04:32] What I imagine is going to happen is some of these AI systems will start being able to use language. And I mean, they already are, but better. Maybe we'll start collaborating with them, say scientifically. And I think more and more, as you put them to use at different tasks, slowly that portfolio will grow. And then eventually we could end up, it's controlling a fusion power station.
[0:04:54] Eventually, I think one system or one set of ideas and algorithms will be able to scale across those tasks and everything in between. And then once that starts being built out, there will be, of course, philosophical arguments about, is that covering all the space of what humans can do? And I think in some respects, it will definitely be beyond.
[0:05:12] What humans are able to do, which will be exciting as long as that's done in the right way. And, you know, there'll be cognitive scientists that will look into, does it have all the cognitive capabilities we think humans have? Creativity. What about emotion, imagination, memory? And then there'll be the subjective feeling that these things are getting smarter.
[0:05:32] But I think that's partly why this is the most exciting journey, in my opinion, that humans have ever embarked on, which is, I'm sure, that trying to build AGI with a sort of neuroscience inspiration is going to tell us a lot about ourselves and the human mind. The way you're describing it there is this big goal in the future that you steadily approach. I'm wondering whether...
[0:05:54] In your mind, there's also a day where this happens. You know how children dream of lifting the World Cup? Have you thought about the day when you walk away from the office and you're like, it happened today? Yeah, I have dreamed about that for a very long time. I think it would be more romantic in some sense if that happened, where one day you're coming in and then this lump of code is just executing, and then the next day you come in and it sort of feels sentient to you. It would be quite amazing.
[0:06:23] From what we've seen so far, it will probably be more incremental and then a threshold will be crossed. But I suspect it will start feeling interesting and strange in this middle zone as we start approaching that. We're not there yet, I don't think. None of the systems that we interact with or have built have that feeling of sentience or awareness, any of those things. They're just kind of programs that execute, albeit they learn. But I could imagine that one day that could happen.
[0:06:49] There's a few things I look out for, like perhaps coming up with a truly original idea, creating something new, a new theory in science that ends up holding, maybe coming up with its own problem that it wants to solve. These kinds of things would be sort of activities that I'd be looking for on the way to maybe that big day. If you were a betting man then, when do you think that that will be?
[0:07:11] So I think that the progress so far has been pretty phenomenal. I think that it's coming relatively soon in the next, you know, I wouldn't be super surprised, the next decade or two. Shane said that he writes down predictions and his confidence on them and then checks back to see how well he did in the past. Do you do the same thing? I don't do that, no. I'm not as methodical as Shane. And he hasn't shown me his recent predictions. I don't know where he's secretly putting them down. I'll have to ask him. It's just a draw in his house. Yes, exactly.
[0:07:41] Like Shane Legg, DeepMind's co-founder and chief scientist who we heard from in an earlier episode, Demis believes that there are certain abilities that humans have but are missing from current AI systems. Today's learning systems are really good at learning in messy situations, so dealing with vision or intuition in Go, so pattern recognition, they're amazing for that.
[0:08:06] But we haven't yet got them satisfactorily back up to be able to use symbolic knowledge. So doing mathematics or language even. We have some, of course, language models, but they don't have a deep understanding yet still of concepts that underlie language. And so they can't generalise or write a novel or make something new. How do you test whether, say, a language model has a conceptual understanding of what it's coming out with?
[0:08:33] That's a hard question and something that we're all wrestling with still. So we have our own large language model, just like most teams in these days. And it's fascinating probing it, you know, at three in the morning. That's one of my favorite things to do is just have a little chat with the AI system. Does it tell you something interesting? Sometimes, but I'm generally trying to break it to see exactly this. Like, does it really understand what you're talking about? One of the things that suspected they don't understand properly is...
[0:09:02] Quite basic real-world situations that rely on maybe experiencing physics or acting in the world, because obviously these are passive language models, right? They just learn from reading the internet. So you can say sort of things like, Alice threw the ball to Bob, ball threw it back to Alice, Alice throws it over the wall, Bob goes and gets it, who's got the ball? And, you know, obviously in that case it's Bob, but it can get quite confused. Sometimes it'll say Alice or it'll say something random. So it's those types of, you know, almost like...
[0:09:30] A kid would understand that. And it's interesting to see, are there basic things like that that it can't get about the real world because it sort of only knows it from words? But that in itself is a fascinating philosophical question. I think what we're doing is philosophy, actually, in the greatest tradition of that, trying to understand philosophy of mind, philosophy of science. When it's 3am and you're talking to your language model, do you ever ask it if it's an AGI?
[0:09:58] I think I must have done that. Yes, with varying answers. But it has responded yes at some point. Yeah, it does sometimes respond yes and, you know, I'm an artificial system and it knows what AGI is to some level. I don't think it really knows anything, to be honest. That would be my conclusion. It knows some words. Knows some words. A clever parrot. Yes, exactly. For the moment, at least, AI systems like language models show no signs of understanding the world.
[0:10:26] But could they ever go beyond this in future? Do you think that consciousness could emerge as a sort of natural consequence of a particular architecture, or do you think that it's something that has to be intentionally created? I'm not sure. I suspect that intelligence and consciousness are what's called double dissociable. You can have one without the other, both ways. My argument for that would be that if you have a pet dog, for example, I think they quite clearly have some...
[0:10:56] Consciousness. You know, they seem to dream. They're sort of self-aware of what they want to do. But they're not, you know, dogs are smart, but they're not that smart, right? And so it's my dog isn't anyway. But on the other hand, if you look at intelligent systems, the current ones we built, okay, they're quite narrow, but they are very good at, say, games. I could easily imagine carrying on with building those types of alpha zero systems.
[0:11:20] And they get more and more general, more and more powerful, but they just feel like programs. So that's one path. And then the other path is that it turns out consciousness is integral with intelligence. So at least in biological systems, they seem to both increase together. So it suggests that maybe there's a correlation. It could be that it's causative. So it turns out if you have these general intelligence systems, they automatically have to have a model of their own conscious experience.
[0:11:48] Personally, I don't see why that's necessary. So I think by building AI and deconstructing it, we might actually be able to triangulate and pin down what the essence of consciousness is, and then we would have the decision of do we want to build that in or not. My personal opinion is, at least in the first stage, is we shouldn't, if we have the choice, because I think that brings in a lot of other complex ethical issues. Tell me about some of those. Well, I mean, I think if an AI system was conscious, and you believed it was, then...
[0:12:16] Thank you. Thank you. Thank you.
[0:12:39] If it's language-based, a kind of oracle, it's like the world's best encyclopedia, right? You ask it a question and it has all research to hand, but not necessarily an opinion or a goal to do with that information. Its goal would be to give that information in the most convenient way possible to the human interactor. Wikipedia doesn't have a theory of mind. No. Maybe it's best to keep it like that. Maybe it's best to keep it like that, exactly. Okay, how about a moral compass then? Can you impart a moral compass into AI and should you?
[0:13:09] I mean, I'm not sure I would call it a moral compass, but definitely it's going to need a value system because whatever goal you give it, you're effectively incentivizing that AI system to do something. And so as that becomes more and more general, you can sort of think about that as almost a value system. What do you want it to do in its set of actions? What do you want to sort of disallow? How should it think about side effects versus its main goal? What's its top level goal if it's to keep humans happy?
[0:13:37] Which set of humans? What does happiness mean? We're going to definitely need help from philosophers and sociologists and others about defining, and psychologists probably, you know, defining what a lot of these terms mean. And, of course, a lot of them are very tricky for humans to figure out our collective goals. What do you see as the best possible outcome of having AGI? The outcome I've always dreamed of or imagined is...
[0:14:07] AGI has helped us solve a lot of the big challenges facing society today, be that health, cures for diseases like Alzheimer's. I would also imagine AGI helping with climate, creating a new energy source that is renewable. And then what would happen after those kinds of first stage things is you kind of have this, sometimes people describe it as radical abundance. If we're talking about...
[0:14:31] Radical abundance of, I don't know, water and food and energy. How does AI help to create that? So it helps to create that by unlocking key technological breakthroughs. Let's take energy, for example. We are looking for, as a species, renewable, cheap, ideally free, non-polluting energy.
[0:14:52] Thank you very much. Thank you. Thank you.
[0:15:22] So it's just not efficient enough right now. But if you had superconductors, you know, room temperature superconductor, which is obviously the holy grail in that area. If that was possible, suddenly that would make that much more viable. And I could imagine AI helping with material science. That's a big combinatorial problem. Huge search space, all the different compounds you can combine together, which one's the best. And of course, Edison sort of did that by hand when he found tungsten for light bulbs.
[0:15:50] Imagine doing that at enormous scale on much harder problems than a light bulb. That's kind of the sorts of things I'm thinking an AI could be used for. I think you probably know what I'm going to ask you next, because if that is the fully optimistic, utopian view of the future, it can't all be positive. When you're lying awake at night, what are the things that you worry about? Well, to be honest with you, I do think that is a very plausible end state, the optimistic one I painted you. And of course, that's what...
[0:16:20] The reason I work on AI is because I hoped it would be like that. On the other hand, one of the biggest worries I have is what humans are going to do with AI technologies on the way to AGI. Like most technologies, they could be used for good or bad. And I think that's down to us as a society and governments to decide which direction they're going to go in. Do you think society is ready for AGI? I don't think yet. I think that's part of what...
[0:16:48] This podcast series is about as well is to give the general public more of an understanding of what AGI is, what AI is and what's coming down the road. And then we can start grappling with as a society, not just the technologists, what we want to be doing with these systems. You said you've got this sort of 20 year prediction and then simultaneously where society is in terms of understanding and grappling with these ideas. Do you think that DeepMind...
[0:17:17] Has a responsibility to hit pause at any point. Potentially, I always imagined that as we got closer to the sort of grey zone that you were talking about earlier.
[0:17:27] The best thing to do might be to pause the pushing of the performance of these systems so that you can analyze down to minute detail exactly and maybe even prove things mathematically about the system so that you know the limits and otherwise of the systems that you're building. At that point, I think all the world's greatest minds should probably be thinking about this problem.
[0:17:47] So that was what I would be advocating to, you know, the Terence Towers of this world, the best mathematicians is that actually, I've even talked to him about this. I know you're working on the Riemann hypothesis or something, which is the best thing in mathematics, but actually this is more pressing. I have this sort of idea of like almost Avengers assembled.
[0:18:04] I know that some quite prominent scientists have spoken in quite serious terms about this path towards getting AGI. I'm thinking about Stephen Hawking here.
[0:18:22] Do you ever have debates with those kind of people about what the future looks like? Yeah, I actually talked to Stephen Hawking a couple of times. I went to see him in Cambridge. It was supposed to be a half an hour meeting, but we ended up talking for hours. He wanted to understand what was going on at the coalface of AI development. And I explained to him what we were doing, the kinds of things we've discussed today, what we're worried about. And he felt much more reassured that people were thinking about this in the correct way. And at the end, he said, I wish you the best of luck, but not too much.
[0:18:51] Then he looked at right in my eye, a twinkle in his eye. Like, it was just amazing. That was literally his last sentence to me. Best of luck, but not too much. That's lovely. I thought it was perfect. It's perfect. Along the road to AGI, there have already been some significant breakthroughs with particular AI systems, or narrow AI, as it's sometimes known. Not least the deep mind system known as AlphaFold, which we heard about in episode one.
[0:19:19] AlphaFold has been shown to accurately predict the 3D structures of proteins, with implications for everything from the discovery of new drugs to pandemic preparedness. I asked Demis how a company known for getting computers to play games to a superhuman level was able to achieve success in some of the biggest scientific challenges in the space of just a few short years.
[0:19:46] The idea was always from the beginning of DeepMind to prove our general learning ideas, reinforcement learning, deep learning, combining that on games, tackle the most complex games there are out there. So Go and StarCraft in terms of computer games and board games. And then the hope was we could then start tackling real world problems, especially in science, which is my other.
[0:20:09] And at least my personal reason for working on AI was to use AI as the ultimate tool, really, to accelerate scientific discovery in almost any field. Because if it's a general tool, then it should be applicable to many, many fields of science. And I think AlphaFold, which is our program for protein folding, is our first massive example of that. And I think it's woken up the scientific world to the possibility of what AI could do.
[0:20:37] What impact do you hope that Alpha Fold will have? I hope Alpha Fold is the beginning of a new era in biology where computational and AI methods are used to help model all aspects of biological systems and therefore accelerate our discovery process in biology. So I'm hoping that it will have a huge effect on drug discovery, but also fundamental biology, understanding what these proteins do in your body. And I think that if you look at...
[0:21:04] Machine learning is the perfect description language for biology in the same way that maths was the perfect description language for physics. And many people, obviously, in the last 50 years have tried to apply mathematics to biology with some success. But I think it's too complex for mathematicians to describe in a few equations. But I think it's the perfect regime for machine learning to spot patterns.
[0:21:27] Machine learning is really good at taking weak signals, messy signals, and making sense of them, which is, I think, the regime that we're in with biology. How could AI be used for a future pandemic? So one of the things actually we're looking for now is the top 20 pathogens that biologists are identifying could cause the next pandemic to fold all the proteins, assuming it's feasible, involved in all those viruses. So that drug discovery...
[0:21:55] Pharma can have a head start at figuring out what drugs or antidotes or antivirals would they make to combat those if those viruses ended up mutating slightly and becoming the next pandemic. I think in the next few years, we'll also have automated drug discovery processes as well.
[0:22:13] So we won't just be giving the structure of the protein, we might even be able to propose what sort of compound might be needed. So I think there's a lot of things AI can potentially do. And then on the other side of things, maybe on the analysis side, to track trends and predict how spreading might happen. Given how significant the advances are for science that are being created by these AI systems, do you think that there will ever be a day where an AI wins a Nobel Prize?
[0:22:42] I would say that just like any tool, it's the human ingenuity that's gone into it. You know, it's sort of like saying, who should we credit spotting Jupiter's moons? Is it his telescope?
[0:22:53] Mo, I think it's Galileo. And of course, he also built the telescope, right? Famously, as well as it was his eye that saw it and then he wrote it up. So I think it's a nice sort of science fiction story to say, well, the AI should win it. But at least until we get to full AGI, if it's sentient, it's picked the problem itself, it's come up with a hypothesis and then it's solved it. That's a little bit different. But for now, where it's just...
[0:23:15] A fairly automated tool, effectively. I think the credit should go probably to the humans. I don't know, I quite like the idea of giving no bells to inanimate objects, like a large hadron collider can have one. Well, exactly. Regression can have one. The Hubble telescope can have one. Exactly, I just quite like that idea. Even before AGI has been created, it's clear that AI systems like AlphaFold are already having a significant impact on real-world problems. But for all their positives,
[0:23:46] There are also some tricky ethical questions surrounding the deployment of AI, which we've been exploring throughout this series. Things like the impact of AI on the environment and the problem of biased AI systems being used to help make decisions on things like access to healthcare or eligibility for parole. What's your view on AI being used in those situations? I just think we have to be very careful that the hype doesn't get ahead of itself.
[0:24:14] There are a lot of people think AI can just do anything already. And actually, if they understood AI properly, they'd know that the technology is not ready. And one big category of those things is very nuanced human judgment about human behavior. So parole board hearing would be a good example of that. There's no way AI is ready yet to kind of model the balance of factors that experience, say, parole board member is balancing up across society. How do you quantify?
[0:24:42] And then if you add in a further thing, which is how critical that decision is either way, then all those things combined mean to me that it's not something that AI should be used for, certainly not to make the decision. At the level AI is at the moment, I think it's fine to use it as an analysis tool to triage like a medical image, but the doctor needs to make the decision.
[0:25:07] In our episode on language models, we talk about some of the more concerning potential uses of them. Is there anything that deep mind can do to really prevent some of those nefarious purposes of language models like spreading misinformation? We're doing a bunch of research ourselves on, you know, the issues with language models. I think there's a long way to go, like in terms of building analysis tools to interpret.
[0:25:33] What are you going to say is the truth at that point?
[0:25:59] Then it sort of impinges on like, well, what does society think about that? And then which society are you talking about? And these are really complex questions. And because of that, this is an area I think that we should be proceeding with caution in terms of deploying these systems in products and things. How do you mitigate the impact that AI is having on the environment?
[0:26:19] Is there just a danger of building larger and larger and larger energy-hungry systems and having a negative impact? Yeah, I mean, we have to consider this. I think that AI systems are using a tiny sliver of the world's energy usage, even the big models, compared to watching videos online. All of these things are using way more computers and bandwidth. Second thing is that actually most of the big data centres now, especially things like Google, are pretty much 100% carbon neutral. But we should continue that trend to become fully green.
[0:26:48] And then, of course, you have to look at the benefits of what you're trying to build. So let's say a healthcare system or something like that relative to energy usage. Most AI models are hugely net positive. And then the final thing is, which we've proven, is that actually building the AI models.
[0:27:05] Can then be used to optimize the energy systems itself. So for example, one of the best applications we've had of our AI systems is to control the cooling in data centers and save like 30% of the energy they use. That saving is way more than we've ever used for all of our AI models put together probably. So it's an important thing to bear in mind to make sure it doesn't get out of hand. But I think right now, I think that particular worries are sort of slightly overhyped.
[0:27:32] While Demis and his colleagues at DeepMind are thinking hard about what could go wrong when AI is deployed in the real world, what really shone through during our conversation was Demis' faith in the idea that ultimately building AI and AGI will be a net positive for the whole of society. If you look at the challenges that confront humanity today, climate, sustainability, inequality, the natural world, all of these things,
[0:28:00] are, in my view, getting worse and worse. And there's going to be new ones coming soon down the line, like access to water and so on, which I think are going to be really major issues in the next 50 years. And if there wasn't something like AI coming down the road, I would be extremely worried for our ability to actually solve these problems. But I'm optimistic we are going to solve those things because I think AI is coming and I think it will be the best tool that we've ever created.
[0:28:31] In some ways, it's hard not to be drawn in by Demis' optimism, to be enthused by the tantalising picture he paints of the future. And it's becoming clearer that there are serious benefits to be had as this technology matures. But as research swells behind that single North Star of AGI, it's also evident that this progress comes with its own serious risks too.
[0:28:57] There are technical challenges that need resolving, but ethical and social challenges too that can't be ignored. And much of that can't be resolved by AI companies alone. They require a broader societal conversation, one which I hope, at least in some small way, is fuelled by this podcast. But I'm struck most of all by how far the field has come in such a short space of time.
[0:29:24] At the end of the last season, we were talking enthusiastically about AI playing Atari games and Go and chess. And now, all of a sudden, as these ideas have found their feet, we can reasonably look forward to AI making a difference in drug discovery and nuclear fusion and understanding the genome. And I do wonder what new discoveries might await when we meet again. DeepMind the podcast has been a Whistledown production.
[0:29:57] The series producer is Dan Hardoon, with production support from Jill Acheneku. The editor is David Prest. Sound design is by Emma Barnaby. And Nigel Appleton is the sound engineer. The original music for this series was specially composed by Eleni Shaw. And what wonderful music it was. I'm Professor Hannah Fry. Thank you for listening.