Sebastian Mallaby on Demis Hassabis, DeepMind and Europe’s AI future

23:12I think that's nonsense because it was actually a cunning way of making the Americans contribute nearly a billion dollars of research and development money into London with all the team, all the human talent, all the spillovers remaining in Britain. Fantastic. We get the American cash, but we stay in Britain. In my view, what matters, what we should look at in these deals is where is the talent staying? And they stayed in Britain. So I think it was a great deal for Britain. All right. So then you go forward a bit. The next thing that happens is a big fight between DeepMind and Google. And this was secret.

23:47Nobody knew about it until I wrote my book. But it was called Project Mario. And it was basically Demis trying to spin out of Google. He decided that Google wasn't taking safety as seriously as it had promised. And to put pressure on Google to give him the safety oversight he wanted, he went to Reid Hoffman, again, an American investor, and said, Reid, would you back, you know, for the good of humanity, Reid, would you back a spin out so that DeepMind's AI technology can have a proper oversight, a proper safety board? And Hoffman said, yes, I'll give you a billion. So they had the money to spin out.

24:25And then they were fighting with Google about whether they would spin out, whether they would get some safety oversight. And anyway, this went on for three years. It was quite, quite exhausting for Demis. He would fly back and forth between San Francisco and, you know, do these kind of grueling 11-hour flight out to San Francisco, spend 12 hours negotiating, and then come back on the next plane. So after a while of this, frankly, he just got too exhausted and gave up. He never got what he wanted. I saw these big term sheets with the spin-out terms, you know, red lines all over them from the lawyers.

25:00And there was real work that went into this, but no outcome. So then there's this slightly kind of bitter truce where Demi says, well, screw that. I'm just going to go and do some AI for science. He does this alpha fold breakthrough, which gets him the Nobel Prize. So that's not bad. But he's basically kind of like doing his thing and trying to ignore Mountain View. And then in 2022, as you say, Chatty PT is released. And Sundar Pichai says, okay, we're going to merge the two Google AI teams. One is DeepMind in London. One is Google Brain in Mountain View. We'll put these guys together to make one powerful team we're going to fight back against.

25:37We're going to fight back against OpenAI and ChattyPT. And he put Demis in charge of it. Now, my point about the business school professor is the business school professor would say, OK, you're telling me there's a merger in the middle of a intense technological fight when you're already behind because the other guys have released ChattyBT. And you're adding that there was this big fight between DeepMind and Google in the past. And oh, by the way, on the scientific team level, there were always fights between Mountain View and London about who got the compute to train the next model. And they would always be bitter that the other one got too much.

26:14They didn't get enough. And there's eight time zones of difference between Mountain View and London. So how you can actually meld the cultures. Is this going to produce a model that can rival the open AI one anytime soon? Forget it. That would have been the business school professor's projection. Instead, it took two and a half years after the merger. And by late 2025, Gemini 3.0 was better on the independent evaluations than any open AI model. And so that's why I think it's a case study for a future business school professor, because I think nobody would have predicted that the merger worked as well as it did.

26:57Andreas Munk Holm:What do you think allowed it to work? Like now you've spent all this time with Hazabiz, you've interviewed more than 100 people around him. I'm sure that you have a good view on what were the foundations of making it work. Yeah. So I think the answer is actually a point that generalizes to being an insight on how you do deep tech startups, right? Because deep tech startups where you're actually doing new science, not just trying to productize existing science. it's a different kind of organizational challenge than a product-oriented startup because you have if you're doing deep science you're trying to actually push the scientific frontier you need to hire a very smart scientist and give them a lot of freedom to explore just pursue their ideas and you hope that something will happen and this is kind of the bell labs you know at the extreme it's the bell labs mentality and that was kind of like what deep mind had at the beginning and it's also what Google Brain had in Mountain View.

28:03A lot of blue sky research, right? So a lot of freedom. But then if you're going to actually deliver on projects in a efficient way, you need to also have a second culture on top or in parallel. And the second culture Demis got from his background as a video game designer. He'd done his own startup after Cambridge, the first startup, which was purely a video game production startup. And he had learned the video game methodology from before he was at Cambridge. And what you do in those labs, apparently, is that you have these teams that have designers and they have the artists and the musicians that do the music.

28:46And then they have the coders who write the code. And it's like different types of code, the graphics, the whatever. And you have to kind of meld all these people into one unified team. and you have quite strict top-down rules about how everybody contributes. And so Demis, who had done a PhD in neuroscience, so he brought the kind of blue sky research culture to DeepMind, but he also brought that top-down design to DeepMind. He called it strike teams. And what he would do is he would let his scientists explore blue sky research, all that stuff. And then now and again, he would have the scientific taste to see that there was one of the areas of inquiry that was getting to a mature point where if you just gave it a big push, in a year or two, you could get a really cool outcome like AlphaFold or AlphaGo, one of these mega projects that really put DeepMind on the map.

29:41And so then Demis would say, okay, strike team. We're going from exploration to exploitation of what we've discovered. And what he did with Mountain View when there was the merger was his people who had been trained in that special source of strike teams brought that to Mountain View. And Mountain View had never seen anything like it before. They were always used to this chaotic let a thousand flowers bloom kind of bottoms up development. By bringing the video game DNA, that's what turned the Google Mountain View researchers from just being good scientists who do stuff like the transformer paper to being people who actually built Gemini 3.0.

30:25That was the transformation.

30:27Andreas Munk Holm:Sebastian, one question. Where did the whole safety discussion land? Because clearly he wasn't able to get it back then. Where are they today? How does Hazabas think about this today? Yeah, so this is the big fascinating moral dilemma kind of at the center of my whole story about him, right? I mean, I kind of always thought, even before I wrote a single word, that, you know, Demis might be the Robert Oppenheimer of our century, right? So this amazing scientist who is an incredible leader as well and leads something like the Manhattan Project to build this frontier transformational technology, but then regrets it because it's not safe.

31:12And that's what happened with Oppenheimer. And I think Demis, although he thought about safety right from the beginning, right, the first meeting between himself and his scientific co-founder, Shane Legg, took place at a safety lecture in London. They bonded over safety, right? So they really cared about safety. And they did that secret negotiation with Google because they wanted safety. But when it comes to it, we have this race dynamic, multiple labs in multiple countries trying to build AI. And if Google DeepMind is safe, it won't make anything better because there'll be some other lab, maybe Meta, maybe some Chinese lab, whoever, that puts out open-weight models, doesn't control anything, and it's not safe, right?

32:00So this is now a collective action problem, and it's not solved. and I'm very worried about that and I think Demis is as well.

32:08Andreas Munk Holm:So you've spent more than 30 hours with this man. You've called him the most competitive person you've ever met. The FUSBAR anecdote alone is iconic. I'm sure you'll share that in a second but you also describe a man who refused to publish AlphaGo as the world's best Go program because the winning wasn't the point, solving Go was. Same with AlphaFold at CASP and I'd love to just have you walk through this paradox? How does someone be the most competitive person on the one hand, and on the other hand, be excessively obsessed with mission driven over money? How does this really work out in practice?

32:46Andreas Munk Holm:What does it look like in a person when you're in service with them? Sure. Well, right. I mean, and this gets to sort of the deep questions about human nature, which you also see when you look at these genius scientists and you ask kind of what motivates them, right? If you know that you might end up as Oppenheimer, why the heck do you even do this? And different scientists in this space, or not all scientists, actually, I mean, you know, Sam Altman isn't really a scientist. He doesn't have a single degree. You know, different leaders, I should say, have different motivations. I'd say that, you know, Demis is deeply motivated by scientific curiosity.

33:21He really, really wants to understand what he calls the fabric of reality. To my astonishment, he would describe his scientific quest in religious terms or spiritual terms. He would say, to understand nature more deeply, which is what I want to do with AI, is to get closer to God or some version of God. That's how he describes it. And so this is very deep for him. And he's not really motivated by money. I mean, you know, I once had this, I kind of grilled him. I said, oh, you always say that, but what do you really mean? Let's just go through the list here. Like, you know, have, you know, since you became very rich, did you move house?

34:02No, I've been in the same house for like 12 years. What kind of car does your family has? I think I've forgotten what it's called, but it's 10 years old. Well, you know, do you collect, you know, art or whatever? You know, do you have a yacht, a second home? No second home, no yacht, nothing. And what luxuries do you have? well, I do, he admitted, have a season ticket to Liverpool Football Club. And it cost about, you know,£5 ,000 for two tickets. So yes, that was a bit luxurious. I mean, he doesn't really care about money, right? But he is complicated. And all of these people are complicated.

34:34Maybe all humans are. Like inside each personality, there are two or three different personalities screaming to get out. And inside Demis, there is this curious, pure scientist. There is also a sort of social democrat, values-driven, patriotic European who wants to do the best for humans. And that kind of gets to the open publication of the results from AlphaFold. They just gifted it to science. They didn't make any money from the system that won the Nobel Prize. And then there's just this crazy competitor and and all of them are a real part of demis and you know you asked for the uh first ball anecdote so i i would go see him in this i would we were talking this pub in london and we normally do like two hours at a time in this dusty room upstairs that nobody knew about so it's just the two of us and so we'd be sitting there with our i want to say no beer you know like demis wants people to understand when i say pub we're actually drinking cappuccinos okay um so so we're at we're drinking our cappuccinos and our water and and he keeps on saying um i i i was the best football player the best table football player when i was at university nobody in cambridge was better than me and i wasn't particularly interested in this like he has all these other achievements why do i care about table football so i go yeah yeah whatever and he says no no no i i really was i was the best you have to you know i i was the best and i say whatever you know he says yeah but you don't understand i've watched these videos of the world champion football players you can see them on youtube and they don't know the snake shot and i knew the snake shot i maybe i invented the snake i mean i you know you know what it is you you get hold of that of the lever and you put your hand around it and you use your whole arm and you yank it and And the friction on your arm pulls on the lever and you get extra spin and talk.

36:30And then the ball just goes even faster with more spin. And then nobody can play against that. And I had that shot and the world champions don't have it. So I was the best. Like, dude, all right.

36:44And, you know, I don't think I normally get into this. But, you know, he also, there was a story about how in his first start-up, he put a table of football set in the shared office. And he said, nobody will ever beat me. And of course, this was a challenge to all the other coders. And then one day, one of them practiced so much that he did actually beat Demis. And as described to me by one of the other members of this startup, you know, Demis went off and sat in a chair and kind of held his head and looked at the fool for a long time and didn't say anything. You know, it's like this big black cloud had enveloped him because he'd lost a game of football.

37:21And then eventually, you know, he looked up at the sky and said something like, my head is on fire. Super dramatic. So, yeah, very competitive. So I think, you know, sometimes people ask me, well, how does Demis reconcile that he's both very competitive and still wants the best for the society and still wants to do science for the sake of science? And I'm like, yeah, he's complicated. He doesn't reconcile. He just has all those three personalities at once.

37:52Andreas Munk Holm:Let me ask you about this strong social democratic values that you're describing as well. He, of course, famously refused to relocate to Mountain View, partly because he sees Britain as more egalitarian. And then he tells you that he's actually quite unhappy seeing these five to ten firms that are absorbing all the wealth and power of the AI revolution. because basically he would want AI to be used for more egalitarian purposes. I want to ask you, do you believe in what he says? Because at the same time, he is also at the forefront, at the biggest of the hyperscalers, running full steam ahead, accumulating more and more power to mountain view.

38:36Yeah. So, look, I think it's important to stipulate that it's true he doesn't like the fact that five to 10 labs are building frontier AI. But it's not that he wants there to be 100 or 200. He actually is the opposite. His vision would be one lab, preferably headed by himself. And the reason he says that is that then it wouldn't be a race. And if it wasn't a race, we could really test the models before we release them to the public. We could go slow. We could be careful. We could even, you know, just hold up releases while we do alignment research. And so we only release it with better alignment technology.

39:20And so his vision is kind of European in the sense that it's like a CERN type vision where you unite, you know, this nuclear research under one roof. So it's a sort of centralizing view. Now, that's not how things have turned out, because when you have this godlike technology, there are going to be multiple labs in multiple countries that want to try to build it, and it's not just going to be one. It just isn't how humans function, right? People are tribal and jealous and competitive. They don't just magically unite into one global effort in unity. So Demis' vision is impossible, and the fact that he even had that vision shows that he has blind spots, that he can be naive.

40:04But anyway, we are where we are. We have these multiple labs in the race. Now the question is, I think your second part of your question is, also, why is he participating in this? Because it's not what he believes is safe. And I think the answer is, yeah, but he wouldn't make it safer by quitting. Like if he left the position at Google DeepMind and went to take a professorship somewhere and somebody else just carried on running Google DeepMind, the world would not be safer. In fact, I think it would be less safe because at least with Demis in that chair, you do have somebody who genuinely has a track record of wanting to try to make it safe, as evidenced by the fact that he spent three years fighting tooth and nail with Google to get safety, as evidenced by the fact that he also had the idea which he gave to Rishi Sunak for an AI safety conference at Bletchley Park in 2023, where people from all around the world came.

41:02And it was the start of what I see as a necessary discussion to get to global agreement on non-proliferation for AI. So I think he is genuine in his safety ambitions, but he just can't do it by himself because there's a race.

41:19Andreas Munk Holm:Yeah. Sebastian, I want to ask you about a different point in time and something that I think many of us are like thinking about and not fully understanding, which is, of course, Google's reaction to ChatGPT and what happened in those pivotal months. I'd love to ask you, we have these six months where, you know, we got ChatGPT. It felt like Google was scrambling. They launched BART, which I think many of us have already forgotten about. Now it's called Gemini. what's the version of this period that you learned through talking to Demis? What did that teach you? What do you think about that? Yeah, so, I mean, when I was writing the book, I had the good fortune to pitch Demis on this idea a week before Chachipiti dropped.

42:10So, you know, then he agreed. And so I was then embedded right from the start of this whole war. And I would go sometimes to Silicon Valley while I was doing the research, because somebody would want me to speak about my power law book in Silicon Valley, whatever. And I would check in with friends of mine who are venture capitalists in Silicon Valley. And they would say to me, Sebastian, you're writing about the wrong guy. Do you realize? I mean, he's history. No one cares about demos anymore. No one cares about Google anymore. The race is over. OpenAI has won. That was the standard view at that time.

42:43And I would say, well, I'm not so sure. I think time will tell and blah, blah, blah. And they would go, yeah, yeah, yeah, you know, whatever. You're smoking something. You know, this is game over. And meanwhile, when I talked to Demis, what I saw was this steely determination. So I said to him, OK, this is what they say in Silicon Valley. What do you say? And he said, look, chat GPT, open AI, they parked the tanks on my lawn. This is war. And you could see that kind of massively competitive side coming out. And he said, anyway, people think that I'm just a scientist. I do AI for science. I do alpha fold, protein folding, but they've forgotten.

43:23What they've forgotten, Sebastian, is that I also did video games in the past and I know how to ship product. And I've studied what Steve Jobs did with all his product rollouts. I understand this game and I'm very good at it. And, you know, I can do either. I can do science and great science and Nobel level science, if that's what you want me to do. I can also do product. You wait. So that was his view. And, you know, I waited, I watched, and I think it was a bit chaotic at first. But as it turned out, as I was saying earlier, you know, this was sort of a business school case study in making the merger work.

43:59And I think, you know, it wasn't just Demis who made it work. You know, he has a great team. people like korai kevachoglu who's been his sort of chief scientist or chief maybe i guess he's called head of research for a long time at deep mind you know moved last year to go live in mountain view and and make the merger really work as best as possible so there are other people who deserve credit but all in all it has worked and i would say that demis was relationship with sundar pichai the CEO of Google. It's probably the most important buddy act in global business right now, because Sundar needs Demis because Demis is delivering the AI, and Demis needs Sundar because he's delivering the money and resources for him to build the AI.

44:46Andreas Munk Holm:We could think a ton about and talk a ton about the future of AI's head. I don't think there's any point in whether Google is going to win it or it's going to be Anthropic, because three months ago, everyone was writing off philanthropic and now they're all the rage. Yeah, exactly. I agree. It's not worth trying to predict. No, exactly. So now instead of doing that, I want to ask you, and I'm going to quote a small thing from your book here, which is that you write that Hazabas deals with the valley and takes its money, but he's remained outside of it and is furiously critical of it, often lambasting its leaders in his conversations with you.

45:24Andreas Munk Holm:I'd love to ask you, without breaking confidence, of course, what is the European founder's case against Silicon Valley as Hacibus makes it to you in private? Well, it's basically that there should be more to frontier science than shipping products. And this is a specific criticism, which I think pertains to something which has the power that AGI has, right? I mean, I think his view is, look, if you just want to produce another app to go in the app store, Silicon Valley is great. I mean, it's very good at that. It ships product fast. There's a great venture ecosystem. We want to have the same in Europe because we admire Silicon Valley.

46:11We're trying to do the same as that. It's not that we're against Silicon Valley. We just want to have our own one. But when you're doing something as consequential as AGI, which is going to change the way everybody lives and how they bring up their kids, how they do their jobs, how they think of themselves as human, because it's now a rival form of cognition, which is taking away the uniqueness of being a human intelligent being. This is so fundamental and so potentially dangerous that just to ship product as if you're doing another app is wrong. And I think what he sees in Sam Altman or what he sees in some of the venture capitalists who back the sort of upstart challengers and who basically just want to throw money at open weight projects is I think he's ambivalent about that stuff.

46:57And, you know, as you know, Google DeepMind does have a strong suite of open weight models called Gemma, and they are proud of them. I think that's a bit of a contradiction of what Demis really believes should be done, because I think that, you know, open weight is dangerous. And so therefore, I'm actually critical of Mistral, I'm critical of Meta, I'm critical of Cohere, I'm critical of all the Chinese labs. And I think ultimately, when you do an open weight model, first of all, the person who uses it can remove safeguards. And second, if they start to do a cyber attack or something other terrible thing, there's a chance that if it's from an API the model provider sees this happening and can cut it off as has happened in the case of a cyber attack in Mexico where both OpenAI and Anthropic saw that their model was being used they cut off the cyber attack.

47:56Now if you just allow people to download the weights and use the model in their own server with no connection to the provider anymore you've lost that ability so so why would you give that up i just think that we're getting to a point with mythos and so forth where ai is getting too dangerous for this to be a smart way forward so i don't like what's happening with mistral for example and i think you know the logic of demis's view having talked to him for more than 30 hours is that he also shouldn't like these open weight providers and yet they provide their own open weight but to go back to your question about what is the critique of Silicon Valley, I think it is that simply this is a very consequential technology which is going to change the world.

48:38It shouldn't just be a crude, vulgar race where people just want to win some sort of race. They want to tweet out their product before they've even released it. They want to throw money at some startup, some new startup. He also complains that there's too much noise. That phrase would come up a lot in my discussions with it. Too much noise, Sebastian, what do you mean? He said, well, you know, I want people to be really thinking, to be doing proper science, deep research, really making progress, breaking through, you know, big barriers to our progress in AI. And they're not going to do that if they're in Silicon Valley.

49:14It's hard to do that because everybody's looking over their shoulder and saying, oh, my former colleague just started his own startup and raised like, you know, half a billion dollars. Maybe I should do that too. And there's just too much distraction from actual knowledge creation in his view. And then I'll just say one more thing about this. I do think that when I go visit the DeepMind office in King's Cross in London, I come out of the tube in King's Cross. I walk through this kind of trendy area of boutiques. I go across a little bridge where there's a canal. In the summer, there are movies by the canal where kids can watch the movies for free.

49:50And then there's an art school, a design school, Central St. Martins, where there's sort of a fountain that bubbles up through the cobblestones. And kids can run around, you know, three-year-olds can run around with their socks and shoes off in the fountain, you know, laughing and screaming with delight in this public space. And then you go beyond that and you get to some more trendy shops and then you get to the DeepMind office. And next to it, there is this green patch of grass. And I've seen in the summer, you know, the kids from the local government housing, you know, playing football on the green grass right by the DeepMind office.

50:23Show me the place in Sandhill Road where this kind of public space exists. It does not exist. It does not. It's all like, don't trespass. And I think that is sort of a vibe which Demis really likes. And so do I. I get it, right? And, you know, maybe some American cities have a bit of that, but in general, America is not that good at it.

50:49Andreas Munk Holm:I actually wanted to ask you just before about your own views. You got into them and you spoke about Mr. Allen, open weight, especially there with a lot of passion. I'd love to ask you about, and we spoke about sovereignty earlier, but we didn't mention Aout specifically in the sovereignty discussion. I think it would be interesting to hear your take on Europe's frailty or lack of independence when it comes to AI, because we have yet to stand up a serious data center, to be honest. We don't have the compute. We have, like everyone I ask, say it's very hard to see even that we will have a serious amount of compute in three or five years.

51:34Andreas Munk Holm:It's even further down the line. How do you think about that? We're obviously quite behind, but what worries me more is it's very hard to see a good path for the next five years. Yeah, I think there is, I'm going to maybe not remember the name right, but I believe Vertex is a European sovereign cloud provider based out of Eastern Europe, maybe Finland or Estonia. I should remember the name because full disclosure, I actually wrote a seed check, or not a seed check, I wrote a check into it, partly because I do believe in all these sovereign initiatives in Europe. And I've only ever done about three checks in my life, but that was one of them.

52:23So maybe we'll be able to make progress with building a sovereign European cloud. But I think that the way I see the AI ecosystem globally evolving in an ideal world is that I think we need to recognize that AI of the sort that is about to emerge in the hands of people who have bad intentions, whether they are terrorist groups or rogue states or what. That's just something we don't want because the cyber attack danger is serious. Bioweapon danger is serious. And so it's like nuclear material. We want civilian nuclear power to be available around the world for all kinds of governments that need power.

53:14We don't want all these governments necessarily to be making weapons because it just creates an arms race and is destabilizing. And we definitely don't want the nuclear weapons to be in the hands of Al-Qaeda or some terrorist group, right? And I think with AI, we need to come up with an AI version of the Non-Proliferation Treaty where, sure, France wants its own AI. France should be able to have its own AI that's kind of trained on French data, respects affects French culture, puts the French tradition kind of front and center in the training set. All of this I completely respect. But I think we need to have a non-proliferation agreement where governments that do this sort of sovereign AI are doing it under the safe conditions that have globally been agreed.

54:06Like what is the red teaming before you release the model? There should be no open weights. It's just too dangerous. And there should be a bunch of these rules about safety. And everybody has to sign up to these rules in order to get the sovereign AI.

54:20Andreas Munk Holm:I think you speak a lot of truth there. And I think it's incredibly important that we in Europe get ahead of this agenda. Sebastian, we're up on time. You've just authored the Infinity Machine. We've spoken a bunch about it here. I can tell you it's 15 hours packed with insights on AI and how Hazzab is, as a founder thinks and as a leader leads. I would definitely advise anyone to go and get it. I personally am listening to it right now. Unfortunately, this one is not read by you, narrated by you on Audible. Am I correct? Because I only listened to the first part and that was not. Yeah, no, I don't tend to, I've never read my own books.

54:56I think there are actors who are good at that.

54:59Andreas Munk Holm:It was actually, okay, then it was actually probably because we did, I listened to it when we did the Power Law, I listened to the Power Law book the days before we did our recording. And then we had our two-hour conversation on how much it was back in the days, which I'm sure is then why I'm tying your voice perfectly into that book. Okay, everyone out there, you will not be enjoying Sebastian's beautiful European accent, unfortunately. Instead, you'll get American who's reading and narrating for you. Sebastian, thank you so much. I hope you enjoyed it as much as I did. I hope our audience will be much wiser about how hazardous things today.

55:34It's been a pleasure. Thank you very much.