← Hidden Forces
Hidden Forces
The God Machine: Demis Hassabis and the Quest for Superintelligence | Sebastian Mallaby
00:56:17
Loading summary
Transcript92 lines
- [00:01]
A
What's up everybody? My name is Demetri Kofinas and you're listening to Hidden Forces, a podcast that inspires investors, entrepreneurs and everyday citizens to challenge consensus narratives and learn how to think critically about the systems of power shaping our world. My guest in this episode of Hidden Forces is Sebastian Malaby, Senior Fellow at the Council on Foreign Relations and a contributing writer for the New Yorker, whose books have chronicled the defining figures and institutions of modern capitalism, from Alan Green,
- [00:32]
B
Span and the Federal Reserve to the
- [00:34]
A
hedge fund industry and the venture capital ecosystem of Silicon Valley, from which the defining technological innovations of the last half century have emerged. His latest work, the Infinity DeepMind and the Quest for Superintelligence, is a biography of Demis Hassabis, the co founder of DeepMind and the man regarded by many as the most important figure in the development of artificial general intelligence. Sebastian and I spend the first hour of this conversation exploring who Demis Hassabis is, where he came from and what drives him, tracing his early life as a chess prodigy in North London, his studies in computer science at Cambridge and
- [01:13]
B
neuroscience at University College London, and the
- [01:16]
A
founding of DeepMind in 2010 alongside Shane Legg and Mustafa Suleiman. We discussed the philosophical and scientific underpinnings of his quest, including the shift from symbolic rule based AI development to the inductive data driven approach of deep learning and also get into the competitive dynamics that have defined the industry. Google's acquisition of DeepMind in 2014, Hassabis, early skepticism toward language models and the transformer architecture, and the moment ChatGPT's release shattered whatever hopes remained of achieving the singleton scenario, the hope that a single safety minded lab could develop artificial general intelligence on behalf of all humanity. The second hour picks up with the launch of ChatGPT 3.5 in November of 2022 and what it revealed about the state of the AI race, including Sebastian's assessment of Sam Altman and the character of the individuals now driving this technology forward. We examined the question of whether personality and values actually matter when competitive and commercial pressures are this overwhelming, and revisit a conversation that Sebastian had with Geoffrey Hinton, the so called godfather of AI, in which he offered his honest assessment of whether humanity is going to make it through the AI transition. We also explore why the AI safety and risk conversation has receded from public discourse not because the concerns have been resolved, but because geopolitical and commercial pressures have made it nearly impossible to slow down. We also consider alternative perspectives from Meta's Yann LeCun's dismissiveness of existential risk to the technical alignment approaches being pursued inside the major labs themselves. If you want access to all of this conversation, go to HiddenForces IO, subscribe and join our Premium feed, which you can listen to on your mobile device using your favorite podcast app, just like you're listening to this episode right now. If you want to join in on the conversation and become a member of the Hidden Forces Genius community, which includes Q and A calls with guests, discounted access to third party research and analysis, and in person events like our intimate dinners and weekend retreats, you can also do that on our subscriber page. And if you still have questions, feel free to send an email to infoiddenforcesio and I or someone from our team will get right back to you. And with that, please enjoy this incredibly important and informative conversation about the most important technology of our time with my guest, Sebastian Mallory.
- [03:52]
B
Sebastian Malaby welcome back to Hidden Forces.
- [03:57]
C
So great to be here.
- [03:58]
B
It's great to see you, though not in person this time. How you been?
- [04:02]
C
Oh, I've been having a great time. We're going to discuss it, but to be embedded inside an AI lab right when ChatGPT gets released, it's pretty good.
- [04:11]
B
So this is what, your fifth book, your sixth book? I've lost track.
- [04:14]
C
This is number six, I'm a Repeat Offender.
- [04:17]
B
And you've been on the show two previous times. The first time you were on was way back in the early days of the program and you had just published a book on Alan Greenspan's career, his life and times. And then you came back on to discuss the Power Law, which was a book on the venture capital industry. Almost all your previous books, with maybe the exception of After Apartheid, seem to be either primarily biographies, the Man Hung you, which I just mentioned in this latest book, the Affinity Machine, qualify or historical accounts of institutions or industries in which key figures play a defining role, such as With More Money than God and the Power Law. But even in this book and in your biography of Alan Greenspan, the main character is used to tell a much bigger story. What is it about you? Your interests? Your curiosities? Your nature? The questions that make you feel alive and that align with your life's mission and purpose. That might explain to someone looking from the outside why you've chosen to write the books that you've written.
- [05:20]
C
Well, I'm always looking for the story, the practitioners, the real people, but also the larger issue. I want both. I'm greedy. I don't want to choose. I don't want to do an abstract argument with no actual protagonists in it. Nor do I want to do protagonists and character just for the sake of character. It's got to mean something for how the world works. So all of my books have those two tracks running in parallel. And I find that to be not only fun to write, maybe fun to read, but actually also that's where the truth is, right? Because in the way the real world, well, we have this debate about, you know, is it a great man theory of history or a great person theory of history, or are they ineluctable deep forces, as a Marxian historian would suggest? And I'm firmly of the view that the truth is in the middle. There's a mixture. Sometimes something like demographics, technological change can drive what happens. Other times, there really are individuals who make decisions and determine outcomes. And so you need to capture both in order to do a good book.
- [06:21]
B
So it seems that you've undergone a shift in focus from finance to technology in the last decade of your life. Does that track? I mean, I'm going off of obviously, the subject matters of your book, but if so, what explains that shift in focus, do you think?
- [06:35]
C
It does track, And I'd say it's explained by two things. One is just my curiosity. I always want to find something new to get my teeth into intellectually. But the other thing is that the world itself has moved in this direction. If you think about the period before the 2008 financial crisis, so the sense of where the energy was in capitalism was often in finance, in the rise of hedge funds, the rise of private equity. If you thought about the phrase masters of the universe, that was a reference to Wall Street. Then if you look at the period after the 08 crash, when basically Wall street gets regulated, all of the energy dynamism, exciting new stuff is happening really in tech and specifically in Silicon Valley, or kind of imitators of Silicon Valley, it might be the Chinese tech ecosystem, which in the 2010s was really hot. So I guess my writing interests have followed the way the world has changed, and AI is a great example of that.
- [07:30]
A
So I think it was in 2018,
- [07:31]
B
when Gillian Tett was first on the podcast, and she said then that Silicon Valley was beginning to feel like Wall street pre 2008. So, like after 08, Silicon Valley felt like sort of the last bastion of freewheeling capitalism. And they could do no wrong, the titans of that industry. And that's no longer the feeling. In fact, the populist outrage that we felt post 2008 directed at the banking system now feels like it is being redirected in Silicon Valley. Does that track also with your experience?
- [08:06]
C
Yeah, I think that Silicon Valley sometimes has a slight upper hand in that the consumer products people tend to like them. I mean, they complain about them, of course, they say, I'm addicted to my phone, I spend too much time doom scrolling, et cetera, et cetera. But they actually love their phones. And although they may wish they didn't do quite so much time on Instagram, they're kind of happy that they've got messages, they're in touch with their friends, they've got a map that avoids losing your way in a different city you can call an Uber, all that stuff. So basically, people have direct contact with the products and they like them, whereas finance is sort of abstract, remote, and you don't really think about it until the moment when there's a financial crash and the government bails out the guys who created the crash. In the meantime, you lost your job. So I think the sort of animus against finance is in some ways deeper.
- [08:55]
B
How are the tech founders and pioneers different from the financiers and economists that you've written about in the past?
- [09:04]
C
Well, that's a good question. I mean, I think tech is fundamentally more powerful and is driving, in fact, a lot of the changes in finance. If you think about. I had a very interesting conversation about a month or two ago with somebody who's a veteran Wall street insider, and we were talking about why did private equity take off in the 1980s? And his story was, well, it was the arrival of the personal computer on trading floors and this first spreadsheet software which enabled you to do valuation models of private companies and then figure out, whoops, these private companies are super cheap compared to the public ones. And nobody realized that until technology came along. Well, if you think about securitization or any number of financial products, it's technology that often advances and enables that. And then of course, there's quantitative hedge fund trading and electronic market making and all this stuff. So I think at the root of most change in capitalism, there is technological change. And so in this sense, the tech sector is more powerful and also given correspondingly to more hubris.
- [10:04]
B
So how would you say the individuals are different? How are they different intellectually? How are they different constitutionally emotionally? Do you find that there are certain aggregate behaviors or archetypes that track between these groups?
- [10:20]
C
That's a great question. I mean, I'd say that the stereotype would be that the Wall street titan is some Guy in an expensive suit and shiny shoes who sort of shakes your hand a bit too firmly and as a kind of connector, networker, schmoozer poster. Whereas the archetypal tech titan might be some computer guy who prefers to write code than to talk to people. So I think that's the stereotype. I think it's way too simple. I mean, in my experience, if you look at finance, you've got both the hedge fund tribe, who can be quite introverted because it's sort of numbers on a screen, you don't really need to talk to people, and you just make up your mind individually how you're going to trade versus venture capitalists, where it's all about connections. And you got to persuade the human being who is the entrepreneur to bond with you so that you get to invest in that company, not somebody else. You got to know other venture capitalists so that you can syndicate into their deals. I mean, as I wrote in the Power Law, venture capital is the ultimate network business. And so in this sense, you've got networkers in finance. You've also got kind of analytic introverts in finance. And if you then look at tech, of course, it's a mixture, right? There are some tech people who are indeed who fit the stereotype of they'd rather look at their shoes than look you in the eye. But there's Also plenty of CEOs who make Silicon Valley work. Think of a character like Satya Nadella of Microsoft who are extremely charismatic and look you in the eye. So I think personally, I'd say there's a range in both worlds. And maybe the last thing to say is that some of the cultural divide has faded since COVID I'd say it used to be that the hoodies in Silicon Valley were the contrast with the suits on the East Coast. And now, of course, the suits, well, they don't really wear ties anymore. They might wear a suit, but they might not. Even that has been reduced.
- [12:15]
B
That's a good point, because again, I read the man who Knew, and it's not the only account of Greenspan that I've read. And of course, I grew up as a young man still, when Greenspan was chairman of the Federal Reserve, he was by all accounts an introvert. He was a data sleuth. So maybe he had more in common with Demis Hassabis, who is the subject of this book, than some of the venture capitalists that you chronicle in the Power Law.
- [12:38]
C
Good point. Yeah.
- [12:39]
B
So as I mentioned, this book is called The Infinity Machine, DeepMind and the Quest for super intelligence. How did you go about convincing Hassabis to participate in this biography? What kind of access did he give you? And were there any conditions that he demanded in order to work with you on this project?
- [12:59]
C
Sure. Well, I went to see him proving that it's better to be lucky than smart. I went to see him in November of 2022 and I made a pitch to him, which was basically before I went to see him. I listened to every single podcast or lecture or whatever he'd done. So I knew the way he talked and I knew his belief system. And clear in his beliefs was the idea that AI was going to be the most important invention in human history. And I said, look, if this is true and you're the leading creator of it, that means you're one of the most important people in human history. And he goes, I guess. And then I say, well, if you're one of the most important people in human history, that means that somebody's going to write a book about you. There isn't a choice. It's going to happen. And he goes, I suppose. And I said, furthermore, you should welcome this because you're going to be creating a technology that changes the way we think about how we bring up our kids, how we do our jobs, what it is to be a human being relative to a machine that now can write and make poetry and make music and do all these things that we thought were intrinsically human. So all of our worldview and the way we live is going to be upended, and it's going to be upended partly by you. And people have a right to know, who are you? What is your value system? Why are you doing this? What motivates you to do something which is so disruptive? And if you don't tell that story through a credible, independent writer, you are missing out something you really should be doing. And he listened to that. I had to speak to four of his colleagues and advisors who kind of checked me out. I guess they read my other books. Then I had to have dinner with him. And finally he agreed that he would give me a lot of access. And the deal was we would speak for a long time. We didn't specify initially how long that would be, but it turned out to be more than 30 hours of one on one conversation. And I would check quotes with him. So he had the right to veto a quote, but I could use the information so I could relay to the reader the truth that he conveyed to me. And that was the agreement. And we stuck to it.
- [15:02]
B
What do you think it was about you that allowed you to overcome his reservations?
- [15:07]
C
I think maybe it was just partly timing. The technology really was getting to a point where it was breaking out. And he had already done AlphaFold in 2020, which was the system that won him the Nobel Prize. And that put him in the spotlight a lot. And I think he was feeling confident enough about what he'd achieved to be ready to tell his story. And I happened to be the person who walked in the door first.
- [15:30]
B
Did this project require you to become conversant or knowledgeable in certain domains? And if so, how did you go about acquiring enough understanding to have genuinely productive conversations with people who are, in some cases, intellectual prodigies operating at the very top of their respective fields?
- [15:48]
C
Well, I mean, that's the fun of it, right? I mean, all of my books, I try to get really deep in the subject matter and to really understand it in a way which hasn't been done before in the kind of literature which is aimed at and the general reader. So that was true of hedge funds, that was true of central banking, and that was true of how venture investing really works, how you generate alpha in that way. And so, in some sense, learning about AI wasn't that different. It was a new subject I had to get my head around. And I'm always a magpie. I talk to people. I read things that they recommend. Sometimes it's sort of online essays, not particularly a book. I read their speeches where they explain some things. But I'd say the dominant in the end, in this project, the dominant two ways I learned about how to think about AI, not at a matrix, mathematical matrix multiplication level, but at a kind of conceptual level. What is deep learning? What is reinforcement learning? How did they interact? What's the difference between the transformer model and the SEC2SEC model that came earlier, which was another way of trying to map sequences of data onto each other. So at that sort of level, in what sense does the Atari system differ from the AlphaGo system and how is memory used? All those sorts of things I did get my head around. And I'd say it's partly talking to the scientists who built the models. And they were patient with me. And that's the benefit of having two, three hours, four hours with them. But also I found as the project progressed, so to the technology itself progressed. And I increasingly used Gemini or other models to ask in advance if I was going to go see a particular scientist. I would say, hey, they've written all these pieces, these papers which have been uploaded on the scientific repository arXiv. Based on those papers on arXiv, tell me what's striking about this person's work compared to the person I interviewed last week. And so dialoguing with large language models was quite productive too.
- [17:41]
B
Did you ever feel out of your depth? So much so that it limited what you would have liked to be able to get out of an interview or out of a case study?
- [17:51]
C
Sure. I mean, you always feel, and that's again the fun of it, you always feel stretched. And so by definition, there are moments when you have to say to somebody, I didn't understand that. Can you just explain that again? And in some ways that can actually be productive because when somebody explains it the second time, they maybe come up with a better analogy to explain it, or they kind of make it more colorful by putting in some anecdote about how they struggled with this point, but then they had an insight that got them past it, and then that insight itself becomes a story in the book. So I think you just have to be honest with the people you're talking to and say, hey, I'm not technical. I have tried to educate myself, but I'm going to interrupt you now and again and say, can you explain? So if you've got the time. These books take me three, four years. And so I do have the time to go really deep with people, to speak to them for a long time, prepare for each meeting for like a week or something. So you go in really educated and then follow up with the same people, go back and see them for another two hours. I mean, you can get there. It just takes a lot of work.
- [18:51]
B
Do you have an endless amount of curiosity?
- [18:53]
C
Yeah.
- [18:54]
B
Do you feel lucky that you've built a career that essentially allows you to be a perpetual student? And not only a perpetual student, but anyone in the world who's interesting or worth talking to, you can pretty much get a sit down with that individual.
- [19:09]
C
Well, it's a huge privilege that I have had a lot of access to a lot of amazing people. And yeah, I'm curious. I love learning from them. And that also informs, I think somebody said to me, oh, your books are kind of generous. You know, when you judge human nature, it's not gotcha. And that's true. I'm trying to explain what people are doing. I'm trying to get into their thought process, figure out which bits of it. Because when people explain themselves, they often don't know why they did something. Right. But you have to listen very carefully and then maybe you've talked to somebody else who's in an analogous job and they've had a better explanation. So it's a bit like the tennis player. McEnroe was asked, why is your topspin forehand so amazing? And he couldn't explain it. But if you filmed that wrist and you then slowed it down, the slow motion clips would show you exactly when he was rolling his wrist over forwards and so on. So you want to capture the risk motion in the intellectual process of anybody you're writing about is essentially what I'm saying. And I find that explanatory task so much more satisfying than reaching a kind of slightly cheap and aggressive judgment about people. Because most of the people I write about, they're not evil people. They're certainly not dumb people. They're kind of trying to do their best. They make mistakes. But if you understand mistakes, yeah, they were in a position where there was no good choice. It's just much richer and more interesting and more true to how the world works.
- [20:36]
B
Yeah. Also, if you're adversarial from the get go, it's difficult to get your interlocutor to trust you. But also you have less credibility to push in areas where it's relevant because if you spend your whole interview pushing the individual, it's hard to differentiate between where it really matters and where it doesn't. We've spent quite a bit of time on this portion of me of the background because it's especially interesting to me and hopefully to the audience. One more question before we get to the book.
- [21:04]
A
How did you decide to go to
- [21:05]
B
press with a book about something of civilizational importance like AI that is still in the process of playing out. Did you feel the temptation to continue refining the story and not release it?
- [21:16]
C
No. I mean, partly. I thought there was a good landing place for the story, which is that I began my interviews right when ChatGPT was released and OpenAI became the leading lab in AI because of that model, overtaking Demisisabis and DeepMind for a bit. But then when I finished my writing at the end of 2025, Gemini 3, the Google DeepMind model was judged to be better than the OpenAI model. So I think there was a natural arc to that part of the story. Of course, my book also tells a whole backstory about what was going on in AI before we get to ChatGPT. But I mean, I felt there was a sort of satisfying resting place. And of course the story's going to continue, we're going to get more agentic systems, we're going to get robotic systems and all of that's going to be super exciting. But I do think I've nailed one important slice of modern AI history, and I wanted to get it out because frankly, we're just getting into the moment in 2026 when everybody is going to have their lives affected. We're just seeing the effects on the job market starting, I believe. And when your life is turned upside down, you want to understand what's going on. It's like a bereavement or something. First you're in denial for a bit, but then you need to understand what happened. And I think reading an account of why it was created, who was the individual who created it? What motivated them? What was their values? This is actually quite important in terms of how people can process the AI that is disrupting their lives.
- [22:45]
B
So this book is predicated on the argument that in order to understand the inventors of transformational technologies, we need to understand their motivations, what makes them tick. In other words, what is most important for us to understand about Demis Hassabis, the founder of DeepMind?
- [23:00]
C
Well, that in his case, the motivation for building AI, which he knows might well be dangerous, is a sort of compulsion to advance science. An absolute addiction to the idea that science has to progress.
- [23:15]
B
The sweetness of discovery.
- [23:16]
C
That's right. So the sweetness of discovery. This is a phrase that goes back to Robert Oppenheimer, the father of the Manhattan Project. And what he said was, when you see a technology that is sweet, you go ahead and invent it and you worry about the consequences afterwards. And this line was then quoted by Geoffrey Hinton, the AI professor who also got a Nobel Prize for inventing. Not quite inventing, but developing neural networks, deep learning. And so it's a phrase which is very kind of present in the AI world, that they are doing something equivalent to the atom bomb. Inventing a technology that's dangerous, but it's so sweet you can't resist it. And I think Demis is drawn to that sweetness very powerfully, so much so that he sometimes describes it in quasi spiritual terms.
- [24:04]
B
And Geoffrey Hinton famously resigned from Google in order to speak out against his concerns around AI risk. And that's something we're definitely going to have a chance to talk about. What is most important for us to understand about Hassabis life, whether that be his early experiences as a chess prodigy, his time at Cambridge and University College London, and the importance he ascribes to certain movies and novels like Ender's Game or books like Hofstadter's Godel, Escher, Bach.
- [24:32]
C
Well, Demis life is extraordinary from start to finish. I mean, it really is one of the most amazing characters I've ever met in my long career of writing about people. And I mean, clearly the smartest as well. And so I think if we just start with the early life, one of the questions we have to ponder is why did somebody living in London, of all places, not even in Silicon Valley, develop this conviction that he was going to devote his life to building super powerful AI? He has this conviction starting around 1993. This is 15 years before AI can even recognize the photograph of a cat. Like AI could do nothing in 1993. But here is this kid, 17 years old, absolutely convinced that he's going to devote his life to building AI. Where does that come from? Where does the conviction come from? And I think you can only understand it if you understand that he was the kid of immigrants. His mother was Chinese, Singaporean, living in North London. His dad was of Greek extraction, and he was just outlandishly clever. And he would drop out of school for a year at a time because he was the captain of the junior chess team in Britain. And he was focusing on chess and just sort of studying by himself in his bedroom. And so he was a total outsider. And he was weird to his parents because he was so extraordinarily precocious with his brains. He was weird to his schoolmates when he even went to school, because he's this Chinese looking kid who's like insanely clever and disappears and makes a bunch of money by winning chess tournaments around the place. So it came naturally that he developed an ambition which for anybody in 1997 was crazy for somebody in London was completely crazy. Right? I want to build powerful AI he did it by his chess led to also he was very fascinated by coding. He got hired by a games company building early video games, and he went off to stay in the house of the entrepreneur who had set up this gaming company. And this guy used to talk about AI and kind of slightly gave Demidus Bug this ambition. But it came from a very unusual place. And so I think understanding both how unusual it was his childhood, but also how driven it made you, because chess was crazy competitive. They used to have these wooden planks underneath the tables because otherwise the contestants would kick each other in the shins to distract the other guy. It was vicious. And people would break down, there would be tears all over the place. People lost games. And it was brutal competition, but also very conducive to developing A kind of autodidact individualism.
- [27:09]
B
Yeah. His early experience as a chess prodigy was informative. Not just because he came to a determination at some point that it wasn't enough and that it was a waste of his brain power, but also there's this great part of the book where he recounts his father telling him that he had to basically give it his all and that would be enough. And what he heard was basically push yourself to near death. And that's the way you know for sure that you've given it your all.
- [27:35]
A
How important was that period as a
- [27:37]
B
kid, both the competitive aspects of chess,
- [27:40]
A
but also what he learned about himself
- [27:42]
B
and what mattered to him. What was important?
- [27:45]
C
Well, I mean, I thought this story that his father said, hey, what matters when you play a game is just try your best. When I say that to my son, I kind of mean it doesn't matter if you lose. What Demis heard from his dad was, I've got to try so hard, as you said, I'm going to exhaust myself. I'm going to kill myself. I'm going to just everything I have to give everything. And I heard this story initially from one of Demis co founders, Shane Legg, and I said, is that really right? And he said, oh, yeah, there's no 50% mode in DEMIS. There's no 99% mode in DEMIS. It's always more than 100% all of the time. And I said, what, you mean he has no hobbies? And he goes, well, maybe he might go to a game of the British soccer team Liverpool once in a while, but that's it, that's it. Otherwise, it's just the mission, build AI. And so, yeah, I think chess created this incredibly competitive Persona maybe also. And I kind of play with this idea a little bit. There were other reasons too. Some people, if they can make a lot of friends, they don't need to prove themselves so much by winning competitions. And because he was weird in his school, too clever, too Asian looking, just different, very small physique. So he wasn't going to impress people, overpower them through strength. He needed to assert who he was in a different way, and that was by winning games. And he won not only actually chess, he started to compete in other games and was the five times winner of something called the Mind Sports Olympiad. And he took it incredibly seriously. I'd never heard of it before I met him, but anyway, it was a big part of his life for a while, just beating everybody at all these different games. Backgammon, Shogi, Go. Chess. There's a game called, I guess, Risk. Just lots of games. Diplomacy.
- [29:35]
B
So I mentioned Ender's Game and Hofstadter's girdle, Escher, Bach. Because. Well, for two reasons. One is that in the book you write about how Hassabis was very drawn to stories. He was a great storyteller, and a lot of his motivations and the way that he tells the story of what drives him are told in some ways indexed to certain stories or certain novels. And also, his objectives in AI are not what you would call traditional entrepreneurial objectives. So, like philosophy and eschatology and epistemology play a very big role in his motivations. What can you tell us about what motivates him on a deeper level and how this almost feels kind of like a spiritual or ontological pursuit for him?
- [30:21]
C
Yeah. Well, I mean, he told me that he would stay up late or always go to bed around 4am and when he was sitting by himself, reading, studying scientific papers, thinking about things at 2am he would feel this sort of reality staring him in the face and saying, I am. I need to be understood. He had this sort of sense that to understand nature as a scientist was to draw closer to some maybe divine force that had created reality. He would say to me, why is it that things are set up in the world in a way that we can study? Isn't it a coincidence? And then he would sort of slam the table with his fist. Isn't it a coincidence that this table is made of atoms jumping around electrons and yet it's solid and you can put your laptop on it, and that laptop is made of bits of silicon, and yet with electricity passing through, it can think? I mean, what's going on here, Sebastian? Why is the world set up like this for us to work, for us to study? I need to understand why. Yeah, I need. Before I die, I've got to understand this. And so that is a really intense feeling for him that he has to understand. And I think that's why he fundamentally is building AI. And by the way, initially I thought, wow, this is crazy. I've never met somebody so intense before. But then, as I thought about it further, over the next three years of talking to him, I realized that we all have an element where we look at technology. We love it, we want it, but at the same time, we're frightened about what it's going to do to our society. And yet we move ahead. If we didn't take that trade of embracing a technology that's both exciting and scary, we would all still be living in caves. So in some sense, Demis is just like an enlargement of the human condition, that desire to understand. It's not just, I think therefore I am, I invent therefore I am. And I think we all have some of that. It's just that Demis has it in very large scale.
- [32:22]
B
Well, there are a number of threads that I extracted from the book that I wrote down as notes as I was reading through it. And one of those threads had to do with the competitive dynamics of the industry. So you go to great lengths to show how these aren't decisions that are taken in isolation. It's not like Hassabis is sitting in a room and saying, I have a choice to hit a button here, and if I don't hit the button, nothing will happen. We won't get any progress in artificial intelligence. And if I do, we might get a 5050 existential extinction event. It's that everyone is competing with everyone else. And so the competitive dynamics almost drive inexorably toward innovation and whatever this outcome is going to be, whether it is cataclysmic or beneficial for humanity. Was that a dynamic that you feel like you were aware of early on and that maybe you gained more appreciation of for as you sort of worked on the book, or did it really emerge throughout the course of writing it?
- [33:18]
C
I understood that there was a race dynamic the moment ChatGPT came out. And indeed, I remember going to see Demis right after it came out, and he said, oh, they've parked the tanks on our front garden on the front lawn. This is war.
- [33:32]
B
There was a sense of betrayal, too, wasn't there? A sense of sort of communal betrayal against what was on officially agreed upon consensus?
- [33:42]
C
Well, I mean, I'd say at least it's true that both DeepMind and also anthropic had been working on a chatbot, a bit like ChatGPT, and they had been reluctant to release it because they weren't sure it wouldn't be toxic. They worried about hallucination and all that stuff. And ChatGPT came out because OpenAI was less inhibited about those safety considerations. And so there was a sense that this wasn't a fair race. It was more like one side had just been opportunistic and unscrupulous in releasing ChatGPT. So there was that. And then the other thing is that Microsoft, which of course was the major backer of OpenAI, didn't make any effort to conceal its glee that ChatGPT had gotten ahead of its rival Google. DeepMind and Satya Nadella famously came out and said, we've now overtaken them and they're gonna have to show us if they can dance, because can they dance? Was the challenge, which was, I know from my friends at Google, just super insulting and infuriating. And so this was a take the gloves off, bare knuckle fight kind of moment. And so that's when the race became totally clear. And that was at the beginning of my writing. So that was not a surprise to me. I think what was a surprise was understanding the backstory of how much Demis had tried to avoid this dynamic, this race dynamic, this race to the bottom. Because he could only justify this quest for superhuman intelligence by telling himself that he would make it safe. It would be fantastic for advancing science, but it would also be safe. And so he had a number of theories about how to make it safe. And one of them was there would be a singleton scenario. In other words, only one lab would develop AI on behalf of all humanity, and everybody would club together and make sure it was safe before it was released. And the moment when OpenAI was founded in 2015, which was five years after DeepMind, it was evident that singleton scenario wasn't going to happen. And so that was just one out of several moments of disillusionment where a theory about how to make AI safe got blown up. And so part of the story here is like a sort of an intelligent, good person trying to make AI safe, but theory after theory about how you would do that doesn't work.
- [36:06]
B
So let's use this opportunity to dig in a bit more into the foundations of the company as you suggest it was. DeepMind was founded in 2010. Help me trace how Hassabis took this kernel of an idea that you mentioned in 1993 to build an AI company and turned it into a reality through the eventual founding of DeepMind along with his co founders Shane Legg and Mustafa Suleiman in 2010.
- [36:33]
C
Well, right, so he had this idea. He wanted to build AI before he got to college. Then he read computer science for three years at Cambridge. In the meantime, he was still consulting on video game design with the guys he'd worked with before college. In fact, he persuaded the boss of the company, it was called Bullfrog Games. And the Bullfrog boss gave him a really fast car. It was a Porsche 911. And he would drive at like crazy speeds from his campus to the gaming company in the middle of the night when there was no traffic, and put in a couple of days of work with the gaming company and then come back and do classes again. So he was always juggling both computer science and kind of game design. And when he got out of college, he worked a bit again with the same Bullfrog studio. And then he started his own game design company called Elixir Studios, which did okay, he made a bit of money, but he kind of eventually had to close it because it wasn't doing well enough. And then he was kind of burnt out at that point from startups. Failing can be a pretty traumatic experience. And so both he and his co founder, David Silver, who is a big figure in my story because he was the scientist behind AlphaGo and various other systems, they both wanted to recuperate from entrepreneurship. So to relax, they did PhDs and David Silver went off and did a PhD in reinforcement learning in Canada. And Demis took a sort of more eccentric route. He did neuroscience and the theory was to build machine intelligence. I need to understand how human intelligence works. If I understand the kind of main existence proof that you can even have intelligence is the human mind. So if I do neuroscience PhD, then I will understand the endpoint of where AI is trying to get to. So he did this. Neuroscience PhD was crazy good at it. He published a paper when he hadn't even finished his PhD, which became one of the most cited neuroscience papers of that period. And then after that is when he founded DeepMind in 2010.
- [38:37]
B
So it's actually important to note now that we're talking about neuroscience, that this was still in the period where symbolic approaches to AI were the norm. Correct?
- [38:47]
C
That's right. Yeah. Much programming had been kind of what we think of as logical deductive programming. If A is bigger than B and B is bigger than C, then A is bigger than C also these kind of hard and fast logical rules. And the big jump in order to build AI was to go from deduction to induction. So with induction, there's no logic, there's just tons of examples. And you absorb all those examples and you kind of intuit the patterns from the examples. And the thing about induction is that if I was to study some examples of humanity, let's say I went, looked at 10 different New Yorkers and I saw what they did in the morning, I would conclude that all humans drink coffee in the morning. But if I studied a million human beings and they were dispersed geographically, I would understand, no, they don't all drink coffee in the morning. Any inductive conclusion is subject to revision when you have more examples. And so you need lots of examples Maybe even infinity of examples in order to do good induction. And that's why I call the book the Infinity Machine, because this shift from deduction to induction only worked when you had a machine that could make sense of an infinity of data.
- [40:04]
B
Do you think it was also an implicit concession that we couldn't build artificial general intelligence, and certainly not intelligence that surpasses human intelligence by simply manipulating formal representations, and that the move to inductive reasoning was an implicit, if not explicit, acknowledgement of the limitations of formal systems of knowledge?
- [40:26]
C
Yeah, and just that as humans, and this is where understanding how the human mind works was very important to understanding how AI was going to work. As humans, we don't think in terms of logic all the time. Most of what we do, some face appears at the door and you recognize who the person is. There's no logic there. You just look and you recognize. Right. And so when I talk, I'm not always grammatical. So even if you taught a machine the rules of human grammar, actually we as humans don't observe those rules very often. So just the whole idea that knowledge and communication could be decomposed into logical constructs is wrong. A lot of what we do and how we think is not actually logical. It's inductive, not deductive. And you mentioned Godel, Escher, Bach, the book that Demis read when he was a teenager. I mean, the most important thing about that book to Demis was really that Godel, who is the mathematician in the title, Kurt Godel, was most famous for what's called the Incompleteness Theorem, which shows that it's impossible to create a system of mathematical logic that completely explains everything that might be true in mathematics. So if it's not even possible for math, it's certainly not true for much Messier emergent systems, like how our biology works, or look at the irregularities you find in nature. Bushes of all kinds of shapes and irregular floor. I mean, the ground is not even. And it's just everything is super irregular and it doesn't follow clear rules. If I say, why is a butter knife in the same category as a carving knife? One is small and blunt, the other is long and sharp, you would say, well, they're both knives, meaning they both have the same function. But then if I say, why is a dachshund in the same category as a Labrador? They have the same function, they're both dogs. But what does that mean? Well, they're both the same species because they can mate together and have fertile offspring. But how do you know? How do you teach An AI that some categories of objects we classify according to their reproductive potential, others by their function, others by their color. I mean, there is no rule that describes that. And so just the point being that logic is super limited. And so AI was a way of developing software systems that could think inductively, not deductively.
- [42:50]
B
Do you think that the inherent contradictions within such systems stem from the limitations of language that show up, for example, in image recognition? Like how do you define that something is a cat, for example, you can't just say that it's a small furry animal with whiskers because it could be raccoon or it could be a baby lion. How do you formalize that into a set of rules that computers can understand from first principles? And is that an indication of the limitations of language?
- [43:18]
C
You can't formalize it into a language that a computer will understand from first principles. Which is why both facial recognition and language recognition have required AI systems that don't think in terms of first order logic. They are way more kind of subtle and complex than that. And the way that they can recognize a cat is they use these neural networks which have thousands of decision centers that are picking up on very kind of subtle relationships between, okay, these two bits of black in the picture are a continuous line that delineates an eye and then this pupil in the eye. I mean, it's just a much more subtle process which is learned through trial and error. So you show the system a labeled picture of a cat, and then you mix up the cat's pictures with non cats, it could be tigers or fairy blankets or something that you might kind of mix up with a cat. And the system tries to guess what's a cat. And if it gets the answer wrong, then it will tweak its own internal code, its parameters, so that it gets it the next time it answers slightly better. And through trial and error, it lands on a set of parameters, basically on a code that unlocks the trick of recognizing a cat. And the programmer, the human programmer, doesn't know how that worked. It's the machine that has discovered through trial and error how it works. And so this is an illustration of the limits of logic.
- [44:48]
B
So how does this relate to Hassabis early skepticism around language models and their limitations and why he felt that was not the path to general artificial intelligence?
- [45:00]
C
Yeah, this is very interesting and important. I mean, there was this moment when the Transformer paper came out which was this scientific breakthrough in how you find patterns in long sequences of data. And the one clear example of this is Long text, sentences and paragraphs that go on and on. And that was a kind of unsolved problem in computer science. How do you have a good translation machine or a good language recognition machine? In 2017, this paper comes out from Google Research and it's a huge breakthrough. And immediately, on the very day that paper comes out, the chief scientist, OpenAI Ilya Satskeyeva, tells me he jumps out of his chair, runs down the corridor, finds his collaborator Alec Radford, and says, drop what you're doing. We're going to do this, we're going to apply this transformer thing, we're going to apply it to language. Meanwhile, over at the rival lab, DeepMind, Demis is like, nothing to see here. I'm not interested in pursuing language. And he had a few people vaguely working on language and DeepMind. And Demis says to me, you're being unfair. We did have a language team. But frankly, he didn't focus on this language team. The language team they did have was not really doing the right kind of research to make use of the transformer. And so it was sort of just like a missing piece of DeepMind's agenda. At the time, they were more interested in these trial and error systems that were playing games. They'd done AlphaGo, they'd beaten the human GO champion, so they were interested in agentic systems that played games, not in language. And the reason was that partly that Demis felt that language was not enough to get to intelligence. Language is a system of symbols and mapping symbols onto some other symbols, for example, pictures. So you have the word cat and you have a picture of a cat, you map one single one to the other. How is that connected to reality? Is what Denmist was thinking. You need to be grounded in the real world to be properly intelligent. And that was his thought. And just language ain't going to get you there. And it took about three years between 2017 and 2020 for Demis to change his mind on that. And the tipping point came in the summer of 2020 when OpenAI released GPT3, the third iteration of its language model. And that was just way too good to ignore, really. It did seem that a language trained system could conduct highly intelligent, cogent conversation. And so that was the beginning of DeepMind waking up and racing to catch up.
- [47:35]
B
So I'd like to at least cover the Google acquisition before we move it to the second hour. Sebastian. That acquisition by Google happened in 2014 and it gave DeepMind essentially unlimited resources, but it embedded it inside of a global corporation with its own priorities, its own culture, and an incentive structure that could tend to view the kind of technological disruption that DeepMind was pursuing as inherently threatening to the company's core business. What informed Kasabis decision and the decision of the team, and what do we know about how aligned the co founders were, as well as some of the early investors in proceeding with the acquisition?
- [48:12]
C
Well, one of the things I discovered is that Demis was such a dominant founder that the other two couldn't really challenge what he wanted to do. So it was all on Demis. He had, at the founding, 90% of the shares, and the rest of the shares were with the other two guys. And so he was dominant. And for him, the decision was really dealing with raising money from venture capitalists, was taking up a huge amount of his time and taking him away from the scientific research he wanted to focus on. And at the same time, the amount of money he needed was starting to go up fast because it was just the beginning of deep neural networks getting deeper. And so you needed more chips, more hardware, and you needed to compete to hire the best scientists, because other labs were now trying to hire them as well. And so he needed money. And the only question was, would he sell himself to Google or would he sell himself to Meta or some other company? And he actually went to see Mark Zuckerberg, I guess it was Facebook then, not Meta, and talked to Zuckerberg and he had this amusing test that he administered on, and he said, I'm working on AI and I think it's super important. And Mark Zuckerberg said, yes, very important. And so then Demis carried on having dinner, and towards the end of the dinner, he mentions how great augmented reality is a technology. And Mark Zuckerberg says, yeah, that's fascinating, and we're really going to invest big in that one. And then Demis a bit later mentioned something else. I don't know what it is, artificial reality or some other kind of technology. And again, Zuckerberg is like, yeah, that's great, we're going to do that. And Demis thought, okay, I understand this guy. He says everything is great. He doesn't really mean it about AI. He doesn't get that AI is just way more important than augmented reality. And if he doesn't get that, I'm not selling to him. So then he pivots back to Google and he does the deal with Google. He has to fend off Elon Musk, who is very keen to buy the company, hates Larry Page of Google at that moment, and wants to kind of compete with Larry page by putting DeepMind inside SpaceX or Tesla. And Demis says, no, he doesn't want to do that. So he goes with Google. And look, it worked out incredibly well because Google was a good parent company. It did give Demis tons of resources to do compute and to hire great scientists. And so he made the right call.
- [50:29]
B
You know, I love this story because it highlights how perceptive he is, how unusually perceptive he is as people.
- [50:36]
A
And yet in the book, you also
- [50:37]
B
show how he could also be incredibly naive. What explains that? This dual character in Hassabis.
- [50:43]
C
I think he was naive about how Silicon Valley worked because he wasn't based there, he was based in London. He's quite patriotically British. He believes that the social democratic values of Britain mean a lot to him. If you go to Silicon Valley, it's quite hard to find any public space. If you go to the DeepMind office in King's Cross in London, there are barefooted kids playing soccer in the summer who come from the local government housing and they're playing on the government soccer pitch. And so it's just a different vibe. And Demis cares about that. So I think he was naive in terms of how to play the venture capital raising game. He believed too much in Peter Thiel, who was the series a financier, and he thought that Peter Thiel would give him more capital. And in fact, he was wrong about that. Peter was quite reluctant. He's such a contrarian that he invested at the beginning when nothing was working in AI, and then when it starts to work, Peter Thiel's like, I think it's probably overheating now it's time to sell. He's always against the consensus, which I find hilarious.
- [51:42]
B
But anyway, that's Ian's streak. That's Rene Girard in him.
- [51:47]
C
He's certainly an interesting thinker. In any event, I think clearly the venture capital game is something that Demis was naive at. And also, I'd say the inevitability of competition amongst multiple different AI labs is something he underestimated. I mean, the notion that would be a singleton lab, like one lab building AI on behalf of all humanity, obviously underestimates the truth that if you've got something as exciting as AI, multiple labs with multiple ambitious entrepreneurs in multiple countries are going to come after this goal. It's just too rich, too sweet not to chase.
- [52:23]
B
In retrospect, do you feel like Demis thinks it was the right decision to sell to Google, or do you think there's some regret there?
- [52:30]
C
Well, now and again he would say to me, oh, if only I'd been in that position trying to raise money and it had been not 2013, I could have just raised my money independently. I could have remained independent. But I think that's the minority share of his mind. Right? The majority of the time he says, it worked out great. I got all the money I needed. I had a choice in my career. Did I want to build an independent company and build it into a Google, or did I want to use the resources of Google to build towards artificial general intelligence and do great science? And he has no doubt that he cares much more about the science than about company building. He's quite good at company building, but his primary objective is science. And so when he got the Nobel Prize, that was way bigger for him than it would be to be a 10 times billionaire or something.
- [53:22]
B
So, Sebastian, I'm going to move us to the second hour. I think where I'd like to pick up there is ChatGPT35, the launch in late November of 2022. And this also coincides with when you first went to him to pitch him on this project. So I think that's an important date to look at. That'll also offer us an opportunity to talk about Sam Altman, because he has emerged to be one of the most seemingly despicable, dishonest, and untrustworthy characters in Silicon Valley. And I'd love to sort of explore some of the commentary in the book around him. And this also raises the larger question, which we can come back to, which is how important are the individuals at the center of this movie? Does it really matter at the end of the day what their characters are like?
- [54:04]
A
You also have this great conversation in
- [54:05]
B
the book with Geoffrey Hinton in his kitchen, where he gives you his assessment on whether we're going to make it or not.
- [54:12]
A
And this conversation about AI risk is
- [54:15]
B
something that you mentioned, Nick Bostrom, early on in this episode. Nick had been on the podcast before, and his book Superintelligence was, for me, very formative. It was a book I read early on that exposed me to many of the philosophical ideas and thought experiments around AI risk.
- [54:32]
A
And it feels like this conversation has
- [54:34]
B
receded into the background, and not for good reasons, but rather because of the competitive pressures that you talk about in the book, both commercial and geopolitical, that are now driving this industry forward. And that I want to talk about in more detail in the second hour.
- [54:48]
A
And I'd also like to explore some
- [54:50]
B
of the alternative viewpoints, like Yann Lecun's perspective, his dismissiveness of many of the AI risk concerns, as well as Jeffrey Irving and his technical alignment approach.
- [55:00]
A
But we're going to do all of that in the second hour Sebastian for anyone new to the program, Hidden Forces is listener supported. We don't accept advertisers or commercial sponsors. The entire show is funded from top
- [55:12]
B
to bottom by listeners like you.
- [55:14]
A
If you want Access to the second hour of today's conversation with Sebastian, head over to HiddenForces IO, subscribe and sign up to one of our three content tiers. All subscribers gain access to our Premium Feedback, which you can use to listen to the rest of today's conversation on your mobile device using your favorite podcast app. Just like you're listening to this episode right now. Sebastian Stick around. We're going to move the second hour
- [55:38]
B
of our conversation onto the Premium feed.
- [55:42]
A
If you want to listen in on the rest of today's conversation, head over to HiddenForces IO, subscribe and join our Premium feed. If you want to join any in on the conversation and become a member of the Hidden Forces Genius community, you can also do that through our subscriber page. Today's episode was produced by me and edited by Stylianos Nicolaou. For more episodes, you can check out our website at hiddenforces IO, you can follow me on Twitter cofinas, and you can email me at infoiddenforcesio. As always, thanks for listening. We'll see you next time.