Sebastian Mallaby on 'The Infinity Machine': Demis Hassabis secretly built a hedge fund inside DeepMind to beat Jim Simons

potential, but also shows I think how big physical AI is going to be.

Yeah. Yeah. Well, thank you so much for taking the time to come chat with us.

Fantastic talking to you guys. Thanks.

Goodbye.

Great to see you.

Let me tell you all about public.com investing for those that take it seriously. We got stocks, options, bonds, crypto, treasuries, and more with great customer service. And we're talking to them later today. But first, we have Sebastian Malibi. He is the author of The Infinity Machine. Sebastian, how are you doing? I'm doing great. Thank you.

Thank you so much for uh

this has been you've been on uh on John's radar for a long time dream guest list for a long time.

We met maybe four years ago at a uh at a talk you gave around the power law and uh it was very fascinating. I love that book. This book uh goes in a different direction. Um, and after that conversation, I asked you uh probably the worst question you could ask to an author. I asked you uh what what's your next book going to be about because you had selected venture capital. Venture capital had done very well and I presumed that whatever you would pick would be a great investment category because you seem to be a good picker. Uh you told me that you were thinking about biotechnology, biotech investing. You went a different direction with AI. Is there anything I should read into that?

Well, I always kick the tires on a few ideas before I I settle on one. And um I think it took me from the power law coming out in February of 2022

to somewhere around the summer uh maybe August when I really settled on AI and then it took me another I don't know three months to get the courage up to go and pitch deis on the idea of giving me a ton of access so I could write this book. And then I got lucky because I pitched him and one week later, guess what? Uh, Chachi PT came out. So what I thought was maybe a fringe subject. Yeah.

Went mainstream super fast.

Super fast. Uh, what was the response with the team? What was the process? Obviously, he's extremely busy. He also uh sleeps at very random times. We can get into that. But uh what was your actual interaction? How much time did you spend? What was the research process like? So um once he agreed to be in, he was really in. Uh it took about six meetings, two with him, four with his team to get them to agree. And you know, my pitch was, hey, if you say in every speech you give that artificial intelligence is the greatest invention in history.

Yeah.

That means you're way too important not to have a book about you and it's going to happen. So you better get used to it. Yeah.

And also, if you're going to upend our lives, um, you know, change the way we think about ourselves as humans because it's a rival form of machine intelligence, you know, you better explain your motives to people, otherwise they're not going to accept it. So, that was the pitch. He agreed.

And then once he agreed, uh, we would meet like for two hours at a time. Uh we would go to a pub near his home in North London and there was a kind of secret staircase at the back, go upstairs, kind of dusty little room with nobody else there and we would sit there for 2 hours usually

and um he would just riff, you know, uh talk about philosophy, movies, computer science, neuroscience. I mean, he's such a range of a person. By far the most fascinating person I've ever written about.

Wow. That's uh that's high praise. Um, how do you think about balancing the biographical timeline, the history, the financial impacts which you've covered in the past? When I think about the the stories that you've told in the power law about venture capitalists, there's a little bit of their philosophy, but it's a lot of how the deals came together. Uh, fly on the wall. I love that type of storytelling, but this goes a little bit of a different direction. So, how did you how did you think about balancing all the different uh perspectives that you could bring to his story?

Yeah, I mean, I always want to do the personality, the the figure, but then the landscape behind as well. So, it's always a mixture of,

you know, you need a character who drives the story, but the story is boring if it doesn't mean anything.

Yeah.

So, you have to link it to larger stuff that's going on in capitalism and how societyy's going to change and all that. Mhm.

And I mean in this case because Demis is who he is. Um and he would just riff in these extraordinary paragraphs of like storytelling and and theoretical stuff and you know it's just so fascinating that I did give him the microphone more than I have in any other book. I mean, I basically quote quote him at some length, you know, and it's broken up with me asking him questions. And so, um, I'm kind of the reader's lens through which to see Demis talking. And I'd never used the first person really before in my other books, but in this case, those 30 plus hours I had talking in a pub with this extraordinary man, that was the gold dust I had. So to really make the most of it, I did have

these passages of us talking together which kind of interperse the more analytical or narrative bits of the book. You know

how after the transformer model dropped, what was Ilas reaction and why did open AAI get ahead? You know, I cover all that stuff as well.

Yeah.

Um but I do have these passages where you see events through Demis' eyes because and I think it's worth doing because he's so unusual. What was your understanding of AI or uh view on AI in 2022 before you meet Demis for the first time? Were you aware of the the doom arguments? Were you a believer in the technology? Did you think it was 20 years away, a hundred years away, two years away? Like where did you where were you before this book? Because I feel like it probably changed you.

Yeah. Um, you know, I had met Demis before at tech conferences because of the power law and writing about

venture capital. I would go to tech conferences in Europe and he would sometimes be there and I I I actually cheekily you know raised the issue of um hedge funds and especially uh you know the one Renaissance Technologies the the main CEO Peter Brown um had done a PhD with Jeff Hinton about AI back in the day and of course I knew that Dennis would know that and so I said you know do do you know these guys who used deep learning and applied it to markets? Um and and that got his attention. So I I I talked to him a bit. Um so I I knew that he was amazing. I knew that the technology was ripe in the sense that he'd produced this string of breakthroughs. You know, Alph Go

defeating the human Go champion,

then Alpha Zero, which was even stronger. Then there was Alpha Fold, which won him the Nobel Prize,

um for predicting all the shapes of protein in nature. So there was a series of good models. And what they all had in common was they dealt with they dealt with insane amounts of data, crazy combinatorial spaces. Like in Go, there's 361 first moves you can make. Okay, that's way more than chess.

Yeah.

And so unscrambling go and the strategies in Go was much harder than chess. And I knew that these breakthroughs were not just cool in themselves, but they represented the coming of a time when machines could make sense of an almost infinite amount of data and extract meaning. And hence the term the infinity machine, the title of the book, and hence my enthusiasm for writing about it. And you know, so I knew it was breaking out. I knew Demis was amazing. What I didn't know is it was going to break out literally the week after a met him and pitched him on the

What has what has surprised you about how the industry overall has evolved since you started meeting with Demis? Because in some ways I have to imagine you kind of maybe maybe it hasn't been that surprising at all even though a lot of the the growth is is impressive but uh do you feel like you had a view into the future from those first conversations in the pub? Yeah, I mean I think you know I was lucky that the meetings were bookended by you know going to see Demis sort of maybe the third meeting or something. Chachi Pety by this point had gone viral and him saying to me, "Okay, this is war, you know, and you could see his competitive side come out, right? This is war. These guys, he said, have parked the tanks in my front yard, you know, I'm fighting back." So, you could see that. And then after that comes the merger between Google Brain in Mountain View and the Deep Mind team in London. So the two kind of halves of Google's AI um talent base are are united and then I think you get what you know a business school professor is in the future going to write about is kind of like a textbook case in how you make a merger successful

because everyone knows that mergers are hard and when you do it with eight time zones between the two teams one in London one in California and you're doing in the middle of this knockdown dragout you know capitalist fight over building LLMs. This is this is going to be you know most people said this is going to fail and I would come to Silicon Valley while I was doing the book check in with my friends at you know different venture capital shops and they'd all say game over open one and you know the surprise is that within two and a half years um demisses you know Google deep mind model Gemini

was doing better on the rankings than the open AI model. So that was an incredibly fast comeback. Yeah. How do you think about the importance of being like a business person or a deal guy in AI? It feels like Demis has this quote he doesn't want to grow that part of his brain referring to some legal uh negotiations that took a number of years.

Solomon.

Oh. Oh. Was that Mustafa? That was

No, actually that that that was De saying I don't want part of this this part of my brain to grow. Take away these legal briefs. I completely get that and it seems incredible for him to stay in the research mode, build the research organization and yet uh because of scaling laws, we're in this weird regime in this paradigm where sometimes doing a deal to marshall an extra 10 billion of compute actually does unlock a new capability and in is almost in the research track. And I'm wondering how you perceive the relative importance or or Demis' uh perception of the relative importance of these sort of like business dealings that might be more critical path to AGI than the demis of 2021 might think. H you know I would say that the single most important business relationship in the world today is between Sundar Pichai you know CEO of Google and Deis Sabis who's running the AI brain trust because you know Sundar has Deis' back he deals with providing the resources you know supporting the notion that you're going to spend all these tens and maybe hundreds of billions on compute you know that's Sunda's that's what he delivers and he gives Demis the oxygen to then just go do the science and sure he has to build products but I think you know he said to me several times Demis

we've got to a point where building a product like Gemini is in fact advancing the progress towards AGI there's not a there's no tension between the two if you had stopped your AGI research 10 years ago and taken a sidetrack to build some widget yeah that would have been a waste of time

but Well, now that it's so mature and you're actually to build the next LLM, you got to kind of figure out, you know, a reasoning model, then it's going to be a gentic and then you're going to be scaling it even more and all this sort of stuff that we've seen. Um, and this is genuine scientific progress as well as product advance.

Yeah. Uh, do the products also help sort of shift the culture in Google? I'm interested in understanding uh this concept of like AGI pilling becoming uh becoming a believer in the demis worldview of the impact of AGI what artificial intelligence will do that mindset has to diffuse through Google the chatbt viral moment clearly had an impact I'm sure agentic coding has had has has had a similar impact but what has demis' role been in being the culture carrier of that belief belief in AI progress internally.

Well, I think he's used all his sort of visionary communication skills to unite the Mountain View team and and the and the London team.

And I think the the the one sort of organizational contribution he's made which is sort of super powerful is that from a long time ago, Deep Mind had always two different cultures going on at once. There was a kind of blue sky research for scientists. You get a lot of freedom. You could publish papers and all that. You know, go go go find what you want to find. And then there were moments when Demis decided that if you pushed really hard on a particular product or a particular project, you could get a, you know, breakthrough

achievement that would really shock the world. And so he did this repeatedly with, you know, Alph Go and Alpha Fold. And this was sort of like his judgment, scientific taste being applied to knowing when, you know, the moment was ripe to to really go for it. And once he decided that there was, you know, the blue sky research kind of bottoms up stuff, you know, that went out the window and it became a top- down strike team they called it.

And in a strike team, there was a lot of top down direction and kind of everybody had to work on the same code base. You couldn't just like go off and code your own experiment on the side. You had to be contributing to the main one. and that drives towards uh you know a team is driving in a united way towards an outcome and I don't think um Google brain had that Google brain had you know much more of the bottomup stuff

and there was no strike team component and but and and so Deis brought this strike team idea and it came from video game design you know in earlier in his life he'd been a a builder of video games and he in fact started a company doing that

and so this was like how you ship product and and that's been a key insight for Gemini.

Yeah,

Google and Gemini have every advantage just due to the massive cash flows that they have from their other businesses. How do you think that has impacted the culture of Deep Mind given that they have something very real to lose? Right. It's not just about the, you know, maybe Demis' personal desire to be at be, you know, be at the forefront of this research, but if you're not successful, then you you lose one of the greatest business. You have the potential to lose or or have your kind of core business threatened. Yeah. Google search in such a big way.

Yeah.

Yeah. Um and I think you know Google search stands for a more general point that you know Google's whole reputation stands on providing reliable information. And so the penalty for screwing up is very high. You've got this very valuable company. You don't want to support its reputation. And so I think they were more worried about you know releasing a chatbot um fast. And so they had they had prototypes of a chatbot in the fall of 2022 and they didn't want to release and then OpenAI went ahead and did it. Um and so that kind of forced their hand but their first instinct was this is going to hallucinate. Um this is going to do bad stuff. We can't afford that hit to our reputation. Um and you know Demis was quite honest with me in saying well you know the surprise was actually the public's quite happy to play with the tool that hallucinates. you know, they still it still went viral, you know, so we should be less inhibited, but but that was an example of how

being at Google could be a kind of inhibition in moving ahead.

Uh how do you how does Demison uh he's always told a very optimistic story about AI? I I love him as a science communicator, as a as a as an optimist. How has he interfaced with the effective altruist commu community more of the uh the AI doom crowd? Uh does he because he doesn't talk about it that often but does he think about it often?

He does think about it. In fact, you know, he met his co-founder Shane Le at an AI safety lecture, right? They bonded in a safety lecture.

Yeah.

And so, right from the beginning, safety has been a big part of the agenda. And when Demis sold his company to Google in early 2014, part of the deal was you're not going to use this technology for weapons ever. And you're going to have a special independent oversight kind of committee, you know, which will decide on AGI deployment because we don't want that to be just up to the corporate board of of Google.

Now, you know, he's he's slipped on some of these things, particularly the military stuff. Um but he has been thinking about safety and um you know the question is what has he got to show for it? It's all very well to think about something but what can you deliver? And I think you know this is why towards the end of my book he's talking to me quite honestly about how it's a sort of paradoxical moment. He's doing great as an AI inventor. He's doing terribly badly as a sort of AI steward as as making it safe. It's just it turns out that there's a race dynamic. the race dynamic includes China. Yeah.

You know, how do you control this technology when everybody is racing to jam it out the door?

Yeah, it was interesting last week we were talking about how uh you know, Bernie Sanders has come out with a with a push uh to to pause data center development and he quotes Demis and some other lab leads talking about how they would agree to a pause if other countries were were kind of in agreement on it. Uh, and Bernie obviously left out the fact that, uh, I don't think any lab lead could see China pausing on development. So,

um, uh, how has your personal definition of AGI evolved over the last few years because everyone has their own definition. Half the people that come on the show says it was

it was six weeks ago. Exactly. Uh, and then and then and then yet we're still in so many of these kind of more highlevel conversations. lab leads talk about race. You know, we're AGI is six months away. You know,

you're you're you're completely right. Everybody has their own definition. And so my solution is not to have that discussion. I mean, who cares what you know this like you could say right now it's Gemini is artificial? It's general. It's intelligent. Game over. But

you know just a definition thing. I think the other one which is sort of usually fruitless is is AI conscious? Could it become conscious? What is consciousness? Nobody has a good idea. So

let's just uh sidestep those.

I heard one good idea which was something around uh training a model specifically that would hold back all data and all and all training data related to consciousness and the idea of consciousness. And so it cannot just pull from the archive and and and reference or simulate consciousness. And if it develops consciousness from that and can talk cogently about consciousness without having any training data, ever seen the idea of consciousness, then maybe that's conscious. I don't know. It was just an interesting thought experiment. I don't know that anyone's actually run the uh run the test and I don't know that I would even accept it if they did, but uh it is something to think about.

Did did you feel an acceleration in your personal writing process due to AI? I felt an acceleration in my learning process which is sort of what I do before I go interview people. Um so because all of the

computer science papers are basically on archive and you know recently there's been less publication but you know certainly up to about 2022 you could go see a scientist either at DeepMind or one of the rival labs and and just have a conversation with the model about okay this person has done four papers. What's the connecting thread? Why is this person different to the one I interviewed last week? That was a super efficient way of getting up to speed and I didn't worry that it might be wrong because I was going to speak to the human um and cross-check it. But uh that was helpful.

Do you think uh Demis' uh having a having a home base in the UK is in what ways do you think it would have helped or or hurt the company so far?

Uh like has it been beneficial from a talent war standpoint? I'm sure I'm sure lab us some of the other lab leads have taken a trip out to the UK to

Mark Zuckerberg's hoping hosting nightly dinners for AI leads at his house which is just a couple blocks from all the other labs. Can't do that if you're in if the researchers are in the UK,

right? Yeah. I mean, there is movement across the Atlantic, but I think the deal is if you're in London, it's a little harder to recruit people, but once you've got them, they're probably stickier than they would be if if you're in Mountain View or somewhere.

Um, you know, I think Demis has stayed in London because actually he is weirdly patriotic. You know, he comes from this

mixed up sort of, you know, Greek heritage father, a Singaporean Chinese mother.

Um, in a way that makes him a typical Londoner.

Yeah. because London is a really a melting pot. Um but you know he stays there because he feels he wants he believes in sort of British social democratic values. It's maybe ironic to an American audience but you know actually he thinks it's more egalitarian in Britain than it is in the US. you know, the US, if you go to the Deep Mind Office in London, you know, you go past the sort of public spaces like this kind of fountain with toddlers playing in it and there's a kind of free movies in the summer by the canal and then you get to this green space where the kids from the local housing project are playing soccer and then you get to the deep office and it's hard to imagine you'd find that out on the way to the, you know, the the Apple headquarters or something, you know, it's just a different vibe. Could you ever see Demis as CEO of Google or would he have to do too much paperwork?

You know, this gets to the heart of the dichotomy about Demis. It's so difficult because he's so many different things at once. I mean, you know, he is this Nobel Prize winning scientist who would love to just do pure research. And he often would sort of fantasize to me about, hey, I want to retire to the Princeton Advanced uh Institute of Advanced Studies uh and and do what Einstein and Oppenheimer did before me and go there. And that, you know, I think he really means it when he says that.

At the same time, he also wants to be in command of, you know, an AI lab. and he he likes the capitalist competition.

Mhm.

So if he had the opportunity to be chief executive of uh the whole of Google, I suspect he's too competitive to say no. Um but but but who knows? I mean there is that science side. So it's genuinely unpredictable.

He probably doesn't know himself.

Were you left personally optimistic after this process about just our AI future broadly? net positive impact from AI.

I mean clearly there's a lot of upside especially in medicine um uh and pharmaceutical discovery. Um I think you have to be honest and say there's also downside significant downside. Um, and as I continued to do the research for this project, you know, I I I became more worried about the downside because, you know, people like Jeffrey Hinton, I went up and spent two hours in his kitchen in Toronto and sat there debating not necessarily whether machines would be more intelligent than humans, obviously that they will be.

Um, but whether they whether machines have a motivation to harm humans, and he's very persuasive in arguing that they will. I mean, my point was, look, humans evolved over centuries to survive, to pass on their DNA. We're hardwired to survive. That's why we fight each other. Um, machines aren't like that. So, why would they attack us? Why are you so worried, Jeff? And Jeff is like, well, you know, imagine you've got this super powerful AI and it's going to be attacked by the enemy AI and you have to defend it. So, you tell the AI, if you see a cyber attack coming, you've got to fight back. You got to defend yourself. And now all of a sudden you've given your AI a survival instinct. And so don't tell me that evolution has to happen as it happened to humans. The evolution can happen in a machine way.

Mhm.

But these systems are going to want to survive and they're going to be more intelligent than us.

So we're in trouble. And I think you can't dismiss that. So

I'm kind of both worried and excited at the same time. And I I I think that's how humans generally respond to all technology. And if we didn't take that trade and move forward with both the excitement and the scariness of technology, you know, if we didn't take that, we'd be still living in caves.

Yeah.

So, in some sense, the story of Demis is like, you know, the story of all of us, but magnify 10x. There's a documentary that was just released or might be releasing right now that features an interview with them uh the AI doc and it it it spends more time talking to all the different lab leads and voices in the industry. Uh more focused on the doom question uh can we be apocalyptic? Should we be optimistic? Uh but what I found most interesting was that uh the creator of the documentary uh you know summed up his full takeaway and said that AI is a Ponzi scheme. And I'm wondering if you got any any uh any vibes that uh everything will collapse that this is uh just not financially viable.

No, my view is that there's no AI bubble but there's only just an open AI bubble. um you know I in the sense that um the technology clearly uh is getting better and better pretty fast right you know we had the first chatbot in 2022 and it hallucinated then they killed the hallucination then they had longer context windows then they had you know multimodal systems that could handle video and pictures and then they went to reasoning models and now we're getting agentic models and now next there's going to be you know world models that will be built into these things. This is a lot of progress in just three and a half years. Um, so I think it's it's accelerating in progress and therefore it's not a bubble. Mhm.

But what is true is that it's super expensive to develop and if you're not attached to a really deep pocket like you know Demis is attached to Sunda um you're in trouble. Uh because I don't think OpenAI can raise enough money to bridge from today when they have a a very popular chatbot but almost none of those customers pay for it to some future where the product is stickier somehow and they can charge money. And so I think you know open AI has been running two simultaneous experiments. One is with a new frontier technology and the second is how deep are global capital markets and they already raised you know 41 billion last year which was a record for any private fundraising uh bigger than any IPO by the way as well. Um, and so, you know, kudos to Sam Wolman for raising that much money, but can you pull that trick like on a bigger scale every year until 2030 when they hope to break even? No. Uh, and that's why they're cutting products like except for right now

cuz they just raised the 20.

But if you look at the 100, it's kind of smok and mirrors. That's that's, you know, a lot of that is contingent on you you get this money if you go public. Uh, you get this money in the future. you get this money in kind in terms of you know compute or something. The 100 wasn't really 100.

Isn't there a little bit of a dynamic where you could wind up with like an anti- Googlele alliance? Maybe there's, you know, there's like a tension between the the industry and Google. This is like the foundational like myth of the AI industry and the AI labs broadly, although they of course have fractured many many times at this point.

Yeah. I mean, you know, of course there's always rivalries. Um, you know, one might say people will gang up on Nvidia. Um, you know, that's the occupational hazard of being the leader, right?

I think they'll deal with it. I mean, I don't think it's a winner takes all market, by the way. I I think that, um, you know,

there'll be space for others.

Yeah, that makes sense.

Last question from my side, then we'll we'll let you go. Uh, did you have any takeaways or kind of ideas around uh the diffusion of physical AI and robotics? Did anything stand out? Do you have a strong opinion there? Right now it feels like we've entered the sort of software singularity. Um but uh we we just had Kaser from Applied Intuition on and we're talking with him about how autonomy and AI is diffusing through the physical world, but I'm curious if you had any takeaways. Yeah, I mean look, I think that, you know, one thing sometimes people don't quite understand is that large language models and the transformer architecture that underpins them is super consequential for lots of applications, not just chat bots. And so robotics is being improved by this technology and you know I fully expect to see a huge breakthrough um you know over over the next two or three years with the capability of robotics and so I think that's that's going to be the the big story uh I kind of agree with um the guy from Kesar you had before uh that you know the movement of atoms is going to be affected as much as as anything else. Um, so that's part of why I don't think this is a bubble. I think, you know, the potential in, you know, this super powerful AI is is enormous.

Yeah. Well, thank you so much for coming and joining the show. It was a pleasure talking to you. The book is The Infinity Machine. It's available everywhere books