Loading summary
Transcript58 lines
- [00:04]
A
Welcome to another bonus episode of the Tech Brew Ride home. I'm Brian McCullough. As always. Today we're going to talk to someone that I actually spoke to about a decade ago. He's one of my favorite authors. I've read almost everything he's written. We're talking to Sebastian Malaby and we are talking about this book, the Infinity Machine, Demis Hassabis and the DeepMind the Quest for Super Intelligence.
- [00:31]
B
Welcome. Thank you so much.
- [00:33]
A
So let me start this way. For people not ensconced deeply in the AI world, people know Sam Altman increasingly Darya Madai. Can you, in that sort of firmament of AI superstars or gurus, if you will, where does Demis fit into this?
- [00:53]
B
Sure, there's some contrast here, right? So Sam Altman dropped out of Stanford. Demis Hassabis has a PhD and a Nobel Prize. It's a different level of scientific focus. I think Sam Altman may be the greatest fundraiser in Silicon Valley history. Demis Hasabis is a breakthrough scientist. So that's one big contrast. Daria Amadeh is also a PhD scientist and I think is the closest of the other AI leaders to be compared to demisisaibis. Maybe the difference is that Demis had this conviction in the importance of artificial intelligence in the mid-1990s when he was still in his teens, and he was convinced that he was going to build an AI company, bring powerful AI into the world. 15 years before AI could even recognize the photograph of a cat. Nothing in AI was working. And he had this super early conviction, which is pretty much unique in the field.
- [01:50]
A
But he is a different generation than say the Geoffrey Hintons or like Mustafa Suleiman is maybe also of his sort of cohort. Right?
- [01:59]
B
Yeah. Geoff Hinton is older. Geoff Hinton was working as a professor on deep learning back in the 80s. In fact, I know one of his first graduate students from that period. So he has the longest roots, but he is a professor, he is not a scientist entrepreneur. And Demis Hassabis not only did science, but he also created this company out of nothing. I mean, how do you raise money? Right. When it's 2010 and you're going around saying, I've got this idea DeepMind, and people are like, well, what's the product? And he says, there is no product. Not for the next 10 years plus there will be no product, but you should still back me because this is going to be the most important invention in human history, no less. And he had the conviction and also the persuasiveness to do that. Geoffrey Hinton. None of that.
- [02:46]
A
Right. Stays in the, in the research side. So that's important framing that. He's sort of straddling the fence here in terms of the entrepreneur and the research. The book sort of functions as obviously a profile of him specifically, but of these people who have this conviction that AI is a technology that could be the most important technology ever. Right. And it's almost a profile of that mindset, I would argue. So you kind of open the book with Geoff Hinton's admission that the prospect of discovery is almost too sweet not to pursue it. Even if all of these folks seemingly have concerns, some of them deep seated fears about where this might go. They almost. The intellectual challenge is too much. Why did you start with like that sort of framing?
- [03:37]
B
Well, well, I always wanted to capture a sense of what it's like to have your hands on the 21st century version of nuclear technology. Right. This has got enormous upside. AI is very exciting, but it's also dangerous. And what does it feel like to be bringing something into the world that could actually destroy humans or many of them? I mean, why do people do that? It's risky and yet they go ahead. And I think the Oppenheimer comment, because Jeff Hinton was quoting Robert Oppenheimer, that when technology is sweet, you go ahead and invent it and you worry about the consequences afterwards. The extraordinary thing is I thought as I began the project I would have to raise this point delicately with people. Maybe they didn't want to be reminded that they were doing something existentially risky, but they bring it up themselves. Sam Altman shares a birthday with Robert Oppenheimer and instead of keeping that quiet, he advertises it to people in interviews. Demis Hassabis. I mean, I was talking to him once about the setting up of his first office, DeepMind's first office in Russell Square in London, in the heart, like near the British Museum, near University College London. I said, what was it like? And as a writer, I know that when you ask somebody to recapture the emotion of 15 years ago, they often go, yeah, it was cool. But Demis is such a storyteller. He's so vivid in his conjuring of what his journey has been that he said to me, okay, so you come down the stairs of my office, it was in the attic. Ding, ding, ding, ding. Down the stairs, there's no elevator in sight. You come out, there's the square in front of you, the green trees, and on the right, three doors down, Sebastian, there was the London Mathematical Society where Turing invented the origins of artificial intelligence. And beyond that, do you know what there was? There was that black, white, black, white level crossing across the road on the corner. And that is where the Hungarian nuclear scientist Silad in the 1930s came up with the idea of a nuclear chain reaction. And that led to the Manhattan Project. And now we're doing the new version of the Manhattan Project. So I didn't have to prompt these people. They thought in this parallel in terms of this, you know, we are the new nuclear inventors. And that's kind of at the heart of the story in a way. I wanted to capture that.
- [05:49]
A
That's interesting because there's another framing that you do that. But this gets into Demis's specific story. Russell Square is Bloomsbury, a very literary neighborhood. Also, you say that Demis identifies with Ender from Ender's Game, which actually was a huge blind spot for me. I only read it like a couple years ago. But that was an interesting insight to me because one of the concepts of that book is, is sort of a child, a savior figure for all of humanity, but also sacrifices themselves, their personhood, for their duty to humanity. I don't want to put words in your mouth. What do you take from Demis's deep identification with Ender?
- [06:41]
B
I think it's what you said. I think it's this vision of somebody who sacrifices what works, you know, gives everything he has in order to bring about something which could save humanity. And Demis has that slightly messianic side to him. And the interesting thing, again, is he doesn't conceal it. Before the first dinner I ever had with him, right at the beginning of the project, when he was still sort of checking me out, he said, before we meet, you should read Ender's Game, and you're going to see who I really am. And so I read this science fiction story about this diminutive boy genius who literally saves humanity from the aliens. And I'm thinking, does he really want me to see him as a messiah? And the answer is, yeah, he kind of does, because that's how he sees himself. And he's not ashamed or abashed. And in some ways, with most people, you'd think, well, this is totally absurd. But this is actually Demis is somebody who really did have an early vision of AI really does work until 4 o' clock in the morning every day. Really does have this sense of mission. It's genuine. And so why not be honest about it?
- [07:45]
A
Also, the young person that you describe is a chess prodigy. Just a Prodigal young person in general, a gifted kid, as we would have called it in the 80s. Do you feel like that there was very early on a sense of destiny because of maybe identifying within himself this intellectual drive that I have these gifts and I either have to, or I am destined to do something big with my intellect.
- [08:22]
B
Right. I mean, there's a whole literature on the gifted child syndrome, and I think Demis had a lot of the inputs. One is that you're isolated from your peers. And he was, he was. And this is his word, not mine. I was an alien, he told me. And what he meant was, you know, he looked Chinese, his mother is Chinese, Singaporean. He was at the school in North London, you know, way smarter than everybody else and didn't show up at school for a sort of a year at a time because he was playing quasi professional chess on the chess circuit, even at the age of 10. So he was just this strange person and he didn't fit in. And so he taught himself, he taught himself voracious amounts because he's so clever. So he reads and reads and reads, and by the time he's in his sort of mid teens, late teens, he's reading something like Goethe, Le Scherbach, which is a book about. Which has inspired a lot of AI people about the parallels between computer code and the code in our brains. And so he has his own path, and I think that explains why he has this extraordinary ambition to develop AI. Even before he goes to college, the games developer for whom he worked wrote him this half a million pound check. So that's. I think it was more than a million dollars in today's money. And he turned it down, even though he came from a poor family, because he was desperate to study computer science at Cambridge University. And nothing was going to stop him, not even a check for more than a million bucks. And so he had this singular path. Nobody. In 1997, which was the year he left Cambridge University in Britain, set up their own company, right? There was no Silicon Valley in Britain, but he thought, why not? I'm going to do this. I want to build AI. I need to build a company to build AI. And so by a year out of college, he'd founded a company.
- [10:15]
A
Few things I want to catch here. Peter Molyneux is the man that offers him money not to go to college. People might know him from Populous, various very popular video games. You describe him as sort of a magus figure, sort of manipulative, but so interesting that he is not taken in by that, that he still has this singular vision. I want to grab. Go to Escher Bach real quick, because you and I research along similar lines. That book comes up constantly for entrepreneurs, scientists, computer scientists, developers, whatever. Can you also give. This is largely a normie audience that you're talking to here. What is it about that book that has inspired so many people in the field of technology?
- [11:03]
B
I think there's two key things. One is the idea that consciousness is not a function of sort of our biology. It's a function of the signals in our brain. It's the patterns of synapses firing on and off that really create our thinking and therefore our consciousness. And if essentially it's information that is at the root of consciousness, then information on silicon could theoretically also become conscious. So I think that vision for the parallel between human intelligence and artificial intelligence, and therefore the unbounded nature of the potential of artificial intelligence, is the first big thing. I think the other thing is that the mathematician Kurt Godel, whose name is in the title, is famous for the Incompleteness theorem. And the Incompleteness Theorem tells us that no system of logical deduction, no mathematical sort of system, such as the famous attempts by Bertram Russell to create a kind of complete system, it will never be complete. You could never have a set of baseline mathematical propositions that will account for anything that's possible in mathematics. You just can't do it. And I think what that shows us, or the kind of conclusion that Demis at least drew from this very early on when he was still at college, was that you can't just build artificial intelligence on deduction. Logical mathematical reasoning will never get you all the way to full intelligence. You. You need not just deduction, but induction. And when you induce meaning from lots of examples, you need as many examples as possible. Because otherwise, if I just look at ten New Yorkers, I'm going to conclude that all human beings drink coffee in the morning. But if I look at a million people, I will see, no, no, not all. The more examples you have when you're inducing truths, the less likely you are to get it totally wrong. In fact, you need almost an infinity of examples to be good at induction. And that's why I call my book the Infinity Machine, because it's a machine that can handle an infinity of data and therefore do good induction.
- [13:18]
A
So he does go to Cambridge, I believe. Right. But he's studying neuroscience. What we're getting at here is this idea of finding a way to do AI that it doesn't necessarily the goal is not to mimic how human consciousness works, but See if there are lessons there, ways of thinking or thinking about building that will help machines get there as well. So his neuroscience PhD is that innate of that as well. Understanding what we understand about human consciousness and thinking as a path to doing what he wants to do.
- [13:58]
B
Correct. And I think it helped him in some ways. I mean, one was just this idea that there are many facets of intelligence. The brain is made up of, you know, the neocortex, the, you know, hippocampus, all these different components, they do different things. And it's really. And this is a quote from the business plan of DeepMind, it's the interaction between these different components of the brain that really constitute intelligence. And so this led Demis to the view that it wasn't enough just to do deep learning, which was Geoff Hinton's field, where you just feed a ton of data into the system. The system recognizes patterns in the data and therefore can match the picture of a cat onto the word cat. That's all great, but you also need things like planning, you need reasoning, you need to learn through trial and error. And that's why Demis was always focused at DeepMind from the very early phases in combining deep learning with reinforcement learning, which is the art of having a computer learn through trial and error. And the Atari system. Later on the Go system, AlphaGo, which defeated the human champion in 2016. Then there was AlphaZero, a more powerful version, which could also play chess and so forth. Then there was Alpha Star, which played Starcraft ii. And then this was part of the big tradition of DeepMind successes over which Demis Hassabis presided in the 2010s. And I think the inspiration for intelligence being composed of different approaches that came from neuroscience.
- [15:33]
A
What year was DeepMind founded?
- [15:35]
B
DeepMind was founded in 2010.
- [15:37]
A
2010. So talk to me a little bit about that. As you mentioned, he conceived of this not as a lab, necessarily, but also as a company or an organization that's more than just a research lab that stays in an ivory tower. Right. So tell me a little bit about creating what becomes DeepMind and some of the folks involved like Mustafa Suleiman and others.
- [16:01]
B
Sure. So he had two co founders. One was called Shane Legg, who had a PhD in computer science and coined the term artificial general intelligence. That comes from Shane Leck. And so he'd done this PhD in Switzerland. He'd come to London to do a postdoc at the Gatsby unit, which combines both computer science and neuroscience. And Demis was there, too. Demis had done his neuroscience PhD in London and moved to the Gatsby unit. And so the two met each other. And interestingly, they met at a safety lecture, right? Shane Legg was delivering this lecture called the Halloween Scenario where he laid out a doomsday scenario with AI. So super early, right? At the origin story of DeepMind, there is this concern with safety. So that was one co founder, the other co founder, very different, is Mustafa Suleiman, and who was eight years younger than Demis, didn't have any degree from any university because he'd gotten into Oxford, done a couple of years in almost a Silicon Valley fashion. He dropped out. Very unfashionable to do that in Britain, but he did it. And he came from this interesting Muslim background. His father was a semi literate Syrian cab driver in London. His mother had converted to Islam. There were apparently no books, no music at home, you know, go to the mosque every Friday. Very traditional Islamic upbringing. And just through sheer intelligence and again, teaching himself, Mustafa kind of broke out of that. His parents actually left him by himself in London at the age of 16. They moved and, you know, left. They split up, they left the country. And Mustafa made it despite being abandoned by his parents at 16. And he made it to Oxford, did a couple of years, dropped out, and then he's kind of floating about doing different things. And he's friends with this guy called George Hassabis, who is by complete coincidence the brother of Demis Hassabis. And so Demis meets Mustafa as a family friend when he's founding DeepMind. He wants somebody else who can just help to sort of maybe write the business plan, maybe locate an office they could rent, stuff like that. So they hire Shane Legg and Demis Asabis, the two sort of PhD scientists hire this kind of younger, more junior guy to do practical stuff. But that person, Mustafa Suleiman, becomes so energetic and so smart that he parlays his way into a co founder role. And there you have it, three co founders setting up DeepMind in 2010.
- [18:24]
A
What is the business plan when they're raising money? What did they say to convince people to write them a check?
- [18:33]
B
Well, I think the best answer comes from David Silver, one of the great reinforcement lending scientists and a friend of Demis's from Cambridge. And David Silver had been with Demis at his earlier startup, Elixir, which just did gaming. And he had this, he told me the first time I met him, he said, you know, when Demis was younger, we called him the Jedi. I go, why do you call him the Jedi? Well, he would say, I'm going to look in your eyes and you will believe the following things. And then we believed. And that was the business plan, essentially. I mean, the business plan was, we are going to build artificial general intelligence before any AI can recognize the photograph of a cat. And you will believe that. That this is not a crazy idea. And it was very hard to get people to buy into that. But you just needed to find one anchor investor. And they found Peter Thiel, who was interested in the Singularity summits, the early AI discussions, which were kind of half science fiction, half real science. And out of this strange cauldron of Singularity community and Peter Thiel's intrigued curiosity, by that comes a $2 million check, which is just about enough to get DeepMind started.
- [19:47]
A
People might remember this at the time where you would hear These stories about AIs beating human chess players, AIs beating Go players, and eventually they're doing things as sophisticated as Starcraft and things like that. This is DeepMind doing a lot of these. Aside from the publicity that allowed all of us that weren't paying attention to even have heard of that when we weren't paying attention, what was the purpose of doing those? Was it just to get more sophisticated? Get more sophisticated in a measurable way, a way that you can tell you're making advancements?
- [20:27]
B
Yeah. They had this early debate at DeepMind, which is, obviously, we want to build systems that we benchmark and that we measure progress and we measure our mastery of the computer science necessary to build AI. But should we do this by building our own internal metrics, or should we use external metrics? And they went for the second. And this reflected Demis Hassabis, marketing genius. Right. He is not only a great scientist, not only an early visionary about AI, not only somebody who sets up a company. He has an immense flair for storytelling and for figuring out what will capture people's imagination. And so the reason why they built an Atari system was that a lot of the people they wanted to raise money from had grown up with Atari games. And so if you had a machine that could beat all these, the Atari system, that would be impressive to potential people who might write a check. And then they went on to Go, because Go was a recognized grand challenge in computer science. And notably Sergey Brin, who by this point, Google had bought DeepMind. So Sergey Brin was the major shareholder in the new parent company. And Sergey Brin said to Demis Hassabis, well, you can't solve go. It's too complicated. 19 by 19 board 361 possible moves in the first move, then 360 in the second. You start multiplying this. The factorial of these numbers quickly gets you to somewhere near infinity. That's a combinatorial complexity that you just can't solve with a computer system. And to prove him wrong, and therefore hook Sergey Brin and the parent company into funding DeepMind more generously, Demis chose that as the target, something that they would solve together.
- [22:09]
A
We did skip over the fact that, as you're mentioning, eventually Google buys DeepMind. And there has been an excerpt from this book that has gotten a lot of attention about this. But just briefly, tell me the story of selling to Google. Facebook is in play here as well, because I think that it also shows, again, Demis also having good instincts as an entrepreneur in terms of he's got two big players chasing him, and he's not intimidated by that at all. And in fact, the picture you paint is that he's almost manipulating them.
- [22:47]
B
Right. So, you know, as I said, he's a Jedi. He's a great fundraiser. He gets Peter Thiel to write a check. He gets Elon Musk to write a check, which is interesting, because later they have a fight. He gets Selena Chao, the partner of Li Ka Shing, the Hong Kong billionaire, to write an early check as well. So he has a bunch of investors, but it's just chewing up an enormous amount of his time raising money, and he's fed up with it, and so he figures, okay, if I could get a really big backer who would fund way more research, this would allow me to focus on the science. And Google pops up and as the best candidate, partly because Larry Page's father had worked on early neural networks, and Larry Page was very keen on building a stable of AI companies. And there was this birthday party held by Elon Musk at some fake castle on the East Coast. And at that party, Larry Page says to Demis, let's go for a walk. And he pops this idea that we'll buy your company, and you don't get to build your own company, but you get to build a artificial general intelligence. And Demis says, well, I'm a scientist first and foremost. I'll do that. By the way, this is proof that Demis is not trying to maximize for how rich he gets. He's maximizing for scientific breakthroughs.
- [24:06]
A
Right, because they're selling early. This is, what, four years into the existence.
- [24:10]
B
Yeah, it's actually three years in. It's super, super early. But he's also strategic so, having gotten a lot of interest from Google, he knows from experience with his other fundraising exploits that sometimes they say they're interested, and then it just takes forever to get the conversation over the line. So he devises a plan B, and he flirts with Mark Zuckerberg of Facebook. And Zuckerberg wants to get into AI. He sees buying DeepMind as a shot at one stop, shopping for an AI team that would then allow him to rival Google, which has already bought a few AI boutiques at this point. And so there's this dinner at Mark Zuckerberg's house where Demis is invited, and Demis sets him this secret test. And the test is he walks in, and Zuck begins by saying, I love AI. It's amazing what amazing potential for AI. And Demis is like, yeah, sure, sure, sure. And then they have another discussion for an hour or something. And at the end, Demis deliberately raises other technologies that might be exciting to a Silicon Valley tech leader. So he says, yeah, what do you think about artificial reality? And Zach is like, oh, it's amazing. What do you think about 3D printing? Oh, incredible. And then Nemes internally is saying, this guy's not for real. He doesn't understand that AI is in a different league of importance to 3D printing. And so forget it. I'm not selling to Facebook. I'm going to sell instead to Google. And by the way, he does this even though Facebook would have made him a lot richer than Google did.
- [25:44]
A
Well, there's a calculation there that there's no real conviction behind what Zuck is doing. Right. And so even if it's more money or promises of independence, he's intuiting that Google's interests in this are more aligned with his and are more likely to allow him to continue to work. As we see with the metaverse suddenly no longer being the flavor of Mark Zuckerberg's month, even $100 billion spent on I. That. That was an interesting insight, that to intuit, even Google's still a company that wants to make a lot of money, but to intuit that their way of doing that is more aligned with what he wants to do.
- [26:31]
B
Yeah, right. And he got that judgment right. And it was largely a judgment, personally, that Larry Page, as he put it to me, in a different life, could have been a computer science professor. And that's what Demis saw in Larry Page, and that's why he liked him.
- [26:45]
A
Well, they were on their way to being computer science professors before Google happened. I'm going to skip ahead a bit so we're positing Google buys DeepMind. DeepMind exists under the umbrella of Alphabet, and Google continues their research. The research that allows this AI moment to happen is happening at Google under these auspices. But famously, it's OpenAI that does the product that allows this to break through to the mainstream. So what is Demis's. I've talked to a lot of people about this ChatGPT moment, people that are AI researchers and this. Well, we knew this was always going to be big, but we didn't. Oh, it's happening, right?
- [27:39]
B
That.
- [27:39]
A
That moment of now this is for real. Everything we've been dreaming of is starting to happen. For Demis, what was that moment like where, okay, it's Sam that has made the breakthrough in terms of normal people. It's not that the tech, the technological breakthrough had happened years before, but the fact that the AI moment is here and it's not happening with him. What was that like?
- [28:04]
B
Well, I went to see Demis to pitch him on the idea of this book a week before the ChatGPT moment happened. And then I was, you know, he said, yes, I was embedded with him. I was talking to him regularly at, you know, two hour chunks of time we would meet together and usually at a London pub. There was a secret staircase at the back of the pub, and you would go up and there's this room where nobody else would go. And Demis and I would sit there for two hours. And so I had lots of opportunities to hear what he thought about this ChatGPT moment. And I have a few memories. So one of them is that, you know, right after ChatGPT went really viral, he said to me, you know, this is war. They have parked their tanks in our front yard. I mean, he said, on the lawn, that's English, but in American, it'd be on the front yard. And so. And so, you know, it. You could see the kind of competitive fury in Denis. And he is the most competitive person I've ever met in my life. I mean, you know, he's competitive about everything. Even playing table football at Cambridge, you know, who cares about table football, right? Fusball. I mean, you know, but. But he would say, I was the best player, Sebastian. I was the best player in the whole university. I'm like, yeah, whatever. And he went, no, no, no. I really was the best. I've seen videos of the American professional league, and they don't have the snake shot where you hold the. The baton underneath and you're like, spin. Yeah, okay, fine, fine. Let's talk about something else. But anyway, the competitive spirit when ChatGPT came out was off the charts. And so that kind of presaged the comeback, which is astonishing, really, if you look at it. I mean, after ChatGPT came out, you know, OpenAI had a huge lead. Everybody thought that they had just captured the mindshare. It was like, Googling is a synonym of search, and chatgpt was a synonym of AI for most people. And a lot of venture capitalists I know in the valley thought it was game over, that the mindshare, the buzz, the sense of momentum was, was all with Sam, Altman and Demis would not be able to catch up. But, you know, by the end of 2025, on the leaderboards, Gemini had overtaken OpenAI models. And the way they did this was they created a merger between the Mountain View Google brain research operation and the London DeepMind. Now, any business school professor would tell you mergers are super difficult. You've got two different cultures. They've been competing hitherto, they often hated each other because they were fighting over compute and who would get more compute. So they were not friendly. They had eight hours of time difference between them. How do you do a merger in the middle of a massive competitive race when OpenAI is sprinting ahead? How do you catch up? Well, they did. I mean, I think that's going to be a business school case study. You know, it's not just a technology case study. But then the other thing about this is this chatgpt moment is I would, you know, I would go to this pub and I would say, demis, why wasn't it, you know, why didn't you release the model that fired up people's imagination? And the most interesting part of the answer is a sort of philosophical point about language and how much language is a gateway to true intelligence. And it goes back to what I was saying about neuroscience and what that taught him about building AI, because his view was, look, language is a system of symbols. And just because you can manipulate symbols and then maybe match them onto pictorial symbols. So you have the word cat, you have the picture of the cat, you match the two onto, okay, fine. But none of that is grounded in the real world. And he had this view that, you know, and it comes from neuroscience. There's this field called action in perception, that if you pick up a glass of water, I've got one here for people who are watching, you can feel the weight. What is weight? Would you know that if you read all of the words on the Internet, if you downloaded all of Wikipedia. Would you understand the idea of weight? Would you understand that if you drop the glass on the floor it might break, that the water was liquid, it would fall? I mean, those ideas about the physical world Demis thought would escape a system that was trained merely on language. And so he didn't make the big bet that when the transformer model came out from Google and it allowed you to model arbitrary amounts of language, he didn't make the bet that that was going to be the future of the next phase of AI. And so he was still doing Starcraft, he was still doing reinforcement learning systems. He was still in that neuroscience driven view that look, intelligence is a mixture of different parts of your brain and it'll be the same with computers. And so it was a fundamental difference about whether language is enough to build super powerful AI world models.
- [32:43]
A
This is, we're recording this March of 2026. That's one of the big buzzwords right now in terms of what people are raising money around. World models of AI might allow robotics to be have a chatgpt moment and things like that. So what you're suggesting is, is that it wasn't him that created the AI moment breakthrough to the mainstream because he didn't think language alone was sufficient. It was too soon in his mind. He was still waiting for the world model sort of next breakthrough.
- [33:14]
B
That's right. In fact, he was building a world model at the time internally within DeepMind. It was called Gaia and it was sort of a simulation of a natural world with bushes and grass and fruit growing on the trees. And you would try and train agents to interact in that simulated world in order that they should be able to navigate the real world. So he was thinking about world models when he should have been thinking about language. But you know, it's a classic story, right? I mean, people tried to build an iPad in the early 1990s before the compute was ready. It wasn't a bad idea to have an iPad, it's just it was too early. And Demis was right that in the end you're going to need world models. It's just that in that period, from the invention of the transformer architecture in 2017 to around about 2023 I would say was the period when actually transform architectures without machine based reinforcement learning were enough to get you a huge amount of progress.
- [34:15]
A
Reasoning models come up.
- [34:16]
B
Reasoning models is what changes that reasoning models bring in a reinforcement learning approach. And this is 2024 now. And that's when in a way, demis opinion from 20172018 comes true. But there was that interim five or six years where DeepMind, you know, lost the leadership to OpenAI.
- [34:38]
A
You're probably driving, working out, or doing chores right now. Quick tip. TikTok isn't just entertainment. It's where I find fast, practical advice for real life. Download TikTok. Now, I am going to skip over, but I agree with what you said, Sebastian, that the, the business case study of how Google doesn't get left behind as roadkill from this moment. It's one of the more interesting parts of this book as well. But let's end with a couple of, again, my concept of this book as a portrait of the type of people who are doing this right now, doing this to us, to the world. Because I have to report on this every day. I've never, I've been in tech for 30 years. I've never seen an industry or even a subset of tech that is so filled with drama. What is it about the people, all of the people that we've named, from Mustafa to Sam Altman to Elon Musk. Why is it so personal? Why are they. Why is there so much drama? Why are they all so competitive with each other?
- [35:53]
B
You know, I think when the technology has this almost infinite potential, a huge fight over who dominates it is just rational. I mean, it's inevitable. And one of the sort of naive mistakes that Demis Hassabis made when we look back on it, is that he thought that there would be one lap that would shepherd AI into the world on behalf of all humanity. And as Reid Hoffman pointed out to him, and he's a bit of a figure in my story as well, you know, circa 2015, it was clear to Reid Hoffman, no, that's not going to happen. You know, humans are disputatious, jealous, tribal. If there's something as exciting as artificial intelligence about to kind of be unlocked, of course you're going to have multiple labs in multiple countries trying to build it. Right? And I think that's why, you know, I mean, look, Elon Musk, one might say, well, you know, he's got Tesla, he's got SpaceX, he's got, you know, he's got his whole satellite business, you know, Starlink and so forth. Why does he need to do AI? Because the answer is because AI is the biggest thing. It is the biggest thing, probably the biggest invention literally in human history. So of course it attracts these outsized egos and of course they fight about it.
- [37:09]
A
Right. I was going to say ego it is. It's also all of These people fancying themselves the smartest people in the room because they have been the smartest people in the room for most of their lives in various context. So, right, if this is the greatest intellectual challenge, your ego won't allow you not to be a part of that, if you believe yourself to be intellectually capable of being a part of that.
- [37:34]
B
And in other words, you invoked the sweetness term, Robin Oppenheimer, Geoffrey Hinton, using this term, and the idea that this is irresistible, something that you can't resist as a scientist. Now imagine you have the biggest pot of honey in the history of science, the sweetest possible thing ever, that's just going to attract the biggest, most aggressive messianic bees who are going to buzz their way in that direction, and they're going to be interestingly different. Right? So Sam Altman is this kind of widely opportunistic entrepreneur who's great at raising money. And Elon Musk is a great engineer who sort of brute forces his way ahead. And, you know, Daria Made is sort of a, you know, slightly highly strung scientist with a surprisingly good touch for entrepreneurship. And Demis is a very relatable scientist with a great touch for entrepreneurship. They're all a bit different. And so you're right. I mean, you know, I wanted to write a book about artificial intelligence, but also about human intelligence. And what do we learn from watching these people? And you know, I think at the end of the day, you know, we as readers and as normal people, we might say, look, these guys, these Titans, are disrupting our lives. They're going to change how we bring up our children. They're going to change how we do our jobs. They're going to change how we think of ourselves as human beings when there's a rival source of intelligence. Why do they have the right to do this to us? Just because they think it's sweet and they want to go ahead. And one could be very critical of that. But as I thought about it, I have a sort of softer verdict to just float with you. And the softer verdict is, look, all human beings look at technology and they feel both excitement and fear, and they take that trade, they go ahead with it. And if they didn't, we would still be living in caves. So in some sense, the story of Demis Sosabius is an enlarged story about all of us.
- [39:33]
A
It's about seeking the fire, even though the gods will punish you. Yes, a couple more things. And again, from your perspective, having spent three years researching this, talking to all these folks, if you had to put Money on it right now. Who do you think will shape AI's future more? The demises who want to just understand reality. Maybe that's their main motivation. Or the Sam Altmans who want to deploy products. In the end, which side of that divide do you think will be more successful in pushing the technology forward?
- [40:07]
B
I think there's two separate questions. Which is the most promising business model? And I think pure foundation models are not a good business model because there's lots of them and for the moment, they're not really sticky with customers. You can switch from one to the other one. And this is why I think we see OpenAI pivoting in real time, really, really aggressively. Giving up on video generation, giving up on the idea of a shopping app. It was trying to do everything and now it's realized that it was going bust by doing that. Just, it was spending too much, the burn rate was too crazy. And here's another way in which this AI race is creating a business school case study. Because the amount of money that OpenAI raised in 2025 set a record. It was bigger than any private fundraising in history and bigger than any IPO fundraising in history. A lot bigger. And yet if you looked at the 41 billion they raised in 2025 relative to what they said they needed between now and 2030, when they're hoping to go, it's insufficient. It's way insufficient. Right? So there was just no way they were going to survive. So I think there's a set of questions which are interesting about which business model turns out to work. But I think then there's a more important to me question, which is which type of approach is really shaping humanity?
- [41:28]
A
Right?
- [41:29]
B
And I think there it is, the cutting edge of the algorithms that really determine if it's safe or is it not safe? Is it going to be used for science? Like Demis got his Nobel Prize for solving, effectively structural biology, a system that predicted all of the shapes of proteins in nature. I mean, that's a massive thing to have done for humanity and it deserved the Nobel Prize big time. I think Demis Hassabis, as probably the best positioned algorithm builder, is the single most important person in the field.
- [42:07]
A
I guess the way I meant that was if the goal here is AGI is machines that can think equivalent to or eventually obviously superior to human beings. You've spoken to so many of these folks right now. How soon do they think this? I talk to some people and they're like, there are people in this industry and in this discipline that believe that it's less than a year or okay, a decade or at least by my lifetime. What is your sense of. From all these smart people. I'm sure some people have different views than others. If you had to put your money on, do they believe AGI is nigh?
- [42:52]
B
They do believe it's nigh. I think pushing on exactly when there are two discussions in this AI field which I think go around in circles. One is, can machines be conscious? Well, it depends on your definition of consciousness. And so that becomes a dead end quite quickly. And then are we close to artificial general intelligence? Well, it depends on your definition of AGI and they don't share a definition. And you could say, look, I look at the Gemini model. It's artificial, it's general, and it's very intelligent. Maybe we have AGI. I mean that would be one view. I'm just saying it's a little arbitrary where you draw the line where you're crossing that Rubicon. I would say though, that one, I think important metric here is the extent to which the machines are building the next generation of machines. It's that recursive self improvement which then leads to faster and faster cycles of improvement and ultimately an intelligence explosion. And this goes back to the idea of the singularity and all that carries well on. And I think we're kind of there. You ask Demis about how much of the next generation of Gemini is being coded by Gemini, Gemini itself. It's quite a lot. Right. So we're getting there. It's not that humans are out of the loop yet, but we're definitely getting there.
- [44:10]
A
Okay, final question. And this comes back to the psychology around these people doing this. I'm going to quote Demis told you this is a paradoxical moment. It should feel amazing, but it doesn't feel like how I imagined it would feel. What do you think that meant?
- [44:30]
B
Well, you know, when he was founding DeepMind in 2010 with Shane Leg, he had this view that, you know, this was a scientific, high minded, controlled endeavor and he didn't foresee the crazy race that has emerged, although he should have foreseen it, but, you know, he didn't. And so, you know, in some sense, not surprisingly, the closer we are to super, super intelligence, whatever we, however we define that, the more people crowd in, the more people are desperate to get involved and the more noise there is in the field. And if you came to AI as Demis Hassabis did, as a scientific explorer and you thought that deep research, deep thought this was sort of a beautiful intellectual endeavor which is kind of the aesthetic with which he approaches it, and then it turns out to be this shouting match on Twitter. That's what he's talking about. It's noisy, it's cacophonous, it's rivalrous, it's unsettling. It's not safe. It can't be safe when there's a race dynamic, because if one lad does the right thing and the others don't, then the world isn't safe. So this is what leads him to fantasize to me about retiring to Princeton to the Institute of Advanced Studies, where both Oppenheimer and Einstein spent somewhere there.
- [45:53]
A
Or go to an island in the North Sea.
- [45:55]
B
Yeah, exactly. Exactly.
- [45:57]
A
Thinking, yes.
- [45:58]
B
You know, he has that. And this is what's attractive about him. You know, ultimately what he loves is science. And so he has that side to him that would like to just go do science. But he's also a very competitive person who can't vacate his seat, you know, in the middle of this crazy capitalist contest. And so he's both, which is what makes him so fascinating.
- [46:19]
A
Final, final one. And this is psychological from you. The book ends with Demis saying, maybe we being the world will muddle through somehow. I'm optimistic still. So after three years of reporting on this, what's your level of optimism about what AI is going to do?
- [46:36]
B
You know, as an analyst, sort of intellectually, I'm very worried, because I think we are. You know, this technology is coming to fruition at a time when US China relations are bad, which makes collaboration with China over AI safety very difficult. It's also a time when we have a US Administration that doesn't really care about regulation or safety, and I think that's very, very troubling. So I'm worried at the same time, these judgments about how you feel about things, they are a rawsack test for your own temperament. And I just get up every morning. I'm happy about the world. I'm excited to go around and meet people. And, you know, I'm curious. I just can't be depressed. And so I don't feel it at an emotional level, but analytically, I am worried again.
- [47:22]
A
The book is called the Infinity Demis DeepMind and the Quest for Super Intelligence by Sebastian Malaby. I would recommend this to anyone listening as people have taken stabs at telling the history of the recent history of how AI got here. Doing it through the lens of this one man, I think is this is one of the best general audience ways if people want to find out how we got here. This is a good book. That'll tell you.
- [47:50]
B
Well, thank you. It's great to be with you.