# Sebastian Mallaby: The Infinity Machine

An episode of [Ground Truths](https://invisiblenetwork.ai/podcast/1728526108), 19 Apr 2026, 54 min.

Audio: https://api.substack.com/feed/podcast/192906777/deb78dcc50a0bc2d5c958f9812b49f57.mp3

This is one of my favorite books over recent years. Sebastian Mallaby is the Paul A. Cocker Senior Fellow for International Economics at the Council of Foreign Relations and author of 6 bestselling books. THE INFINITY MACHINE tells the story of AI’s progress over the past 15 years largely, but not exclusively, from Demis Hassabis as the protagonist and leader of DeepMind’, with its 2010 mission statement to achieve superintelligence by 2030. It’s a rich, informative, page turner. What We Discussed: —What is an Infinity Machine? —Influence of Claude Shannon’s Information Theory and Douglas Hofstadter’s Pulitzer Prize winning book Gödel, Escher, Bach —Origin of DeepMind in 2010. Prescient. Charter, business plan, included use of agents. How Demis Hassabis was made for the mission! —Contrasts with Sam Altman and the other AI leaders, the Oligopoly (cover of The Economist this week). For example, Nature papers vs white papers on company websites. —In March 2016, the same day when DeepMind’s AlphaGo beat Lee Sedol, Hassabis says it’s time to do protein folding (later known as AlphaFold). —Symbolic AI (historic, deductive, rule-based) vs Deep Learning (Toronto tribe) and Reinforcement Learning (Alberta tribe). —The Big Miss: DeepMind’s lack of early recognition of the importance of transformer models (leading to ChatGPT), creating a big opening for OpenAI. And why was this missed? The Comeback Story. Is this happening again with coding (not in the book)? —The AI Arms Race and Hyperscaling —How the complex relationship between Google and DeepMind evolved —The Double Cross —With the dangers anticipated (parallels to Oppenheimer, Manhattan Project, and the atomic bomb), how to promote AI safety? —Is the major build up of data centers justified? Thank you Bob Fleischman, Jeanie, Ruben Max, FelonBroke America, Seitzinator ❌👑, and more than 600 others for tuning into our live video with Sebastian Mallaby! Join me for my next live video in the app. And a big thanks to Ground Truths subscribers (> 200,000) from every US state and 212 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. Please join! If you found this interesting PLEASE share it! Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered. Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. It enabled us to accept and support 47 summer interns in 2025! We aim to accept even more of the several thousand who will apply for summer 2026. And, if you’re interested on extending healthspan, two recent podcasts/TV segments on SUPER AGERS On With Kara Swisher On PBS Get full access to Ground Truths at erictopol.substack.com/subscribe

## Transcript

[0:00] SPEAKER_01: We're going to be talking about this phenomenal book.
[0:03] SPEAKER_01: Okay.
[0:04] SPEAKER_01: This book by Sebastian Malaby called The Infinity Machine,
[0:09] SPEAKER_01: Demis Hassabis,
[0:11] SPEAKER_01: Deep Mind and the Quest for Superintelligent.
[0:14] SPEAKER_01: It's an extraordinary book.
[0:17] SPEAKER_01: And because of knowing some of the people in the
[0:21] SPEAKER_01: the book and the whole topic,
[0:24] SPEAKER_01: it really hit me with how rich and informative and just page turner,
[0:31] SPEAKER_01: and just a really riveting book.
[0:33] SPEAKER_01: So I know this is your sixth book, Sebastian.
[0:37] SPEAKER_01: Welcome.
[0:38] Sebastian is the Paul Volcker Senior Fellow at the Council for
[0:43] Foreign Relations,
[0:44] SPEAKER_01: and he is a journalist and has been
[0:48] SPEAKER_01: Very prolific throughout his career at many different places,
[0:50] SPEAKER_01: including the Financial Times,
[0:52] SPEAKER_01: the Washington Post,
[0:53] SPEAKER_01: The Economist,
[0:54] SPEAKER_01: and others.
[0:55] SPEAKER_01: And so welcome.
[0:56] SPEAKER_01: It's great to have you.
[0:57] SPEAKER_00: Great to be with you.
[0:59] So we're going to get right into this book because the first
[1:03] thing I wanted to
[1:04] SPEAKER_01: touch on was the title itself,
[1:07] SPEAKER_01: The Infinity Machine,
[1:09] SPEAKER_01: because a lot of people might not realize exactly what is meant by this.
[1:14] SPEAKER_01: And I think if I try to piece it together,
[1:18] because you really zoom in on some critical work that annotated
[1:23] and influenced
[1:24] SPEAKER_01: Dennis,
[1:26] SPEAKER_01: and Deep Mind.
[1:27] And it seemed to me there were two specific works that really
[1:32] had an impact.
[1:33] SPEAKER_01: One was Claude Shannon,
[1:35] In 1948,
[1:36] information theory that the fundamental unit of reality is
[1:40] information.
[1:42] SPEAKER_01: The reality is processing this information.
[1:44] And the other extraordinary book that won a Pulitzer Prize by
[1:48] Douglas Hofstadter,
[1:51] SPEAKER_01: Godel Escher Bach,
[1:53] SPEAKER_01: that he basically in this book,
[1:57] SPEAKER_01: of course it was many things,
[1:59] SPEAKER_01: but some of it was whatever human brains could do,
[2:01] SPEAKER_01: computers should be able to do.
[2:04] SPEAKER_01: and dismissing anything magical about human consciousness,
[2:08] SPEAKER_01: discovering patterns in a near infinity of data.
[2:12] So would you,
[2:14] is that an apt way to try to interpret the title of the book?
[2:19] SPEAKER_01: Because obviously that's a very provocative one.
[2:22] SPEAKER_00: Well, yeah, I mean, it's actually a bit more specific than that.
[2:25] And you're right that both Shannon and the book Gerd Leshebach
[2:29] were very important for Demis.
[2:32] SPEAKER_00: But I think with the Infinity Machine title,
[2:34] what I'm getting at is that traditional computer systems were
[2:39] essentially based on
[2:41] SPEAKER_00: deductive logic.
[2:43] And it's some way by the Bertram Russell project of trying to
[2:49] create a mathematical
[2:50] system that would encompass all possible truths that you could
[2:54] derive from a set of
[2:56] SPEAKER_00: basic ground premises.
[3:00] SPEAKER_00: And
[3:01] SPEAKER_00: So deduction,
[3:02] SPEAKER_00: it turns out,
[3:04] SPEAKER_00: is a very incomplete way of thinking about what intelligence does,
[3:09] SPEAKER_00: what knowledge consists of.
[3:12] SPEAKER_00: Because a lot of the time, we are inducing knowledge.
[3:17] SPEAKER_00: We are not deducting it.
[3:19] So we're looking at lots of examples,
[3:22] and we are inducing a conclusion from these examples.
[3:26] Now,
[3:26] the point about Gödel Ashabach was partly what you said,
[3:30] which is that
[3:31] SPEAKER_00: what the human brain can do, machines will be able to do one day.
[3:36] But it was also the Gödel mathematician in the title who was
[3:40] famous for his
[3:41] repudiation of Bertram Russell for showing the so-called
[3:44] incompleteness theorem,
[3:46] SPEAKER_00: which established that,
[3:47] SPEAKER_00: you know,
[3:48] SPEAKER_00: it's impossible to deduce everything.
[3:52] SPEAKER_00: If you take the example of, you know, categorizing knowledge, you know,
[3:56] SPEAKER_00: How do we understand why a carving knife,
[4:00] SPEAKER_00: big,
[4:00] SPEAKER_00: long,
[4:00] SPEAKER_00: and sharp,
[4:01] SPEAKER_00: is in the same category as a butter knife,
[4:03] SPEAKER_00: short,
[4:04] SPEAKER_00: blunt,
[4:06] SPEAKER_00: and totally different looking?
[4:07] SPEAKER_00: Well, they have the same function.
[4:09] But then there are other categories where you have a big dog
[4:13] and a small dog.
[4:14] SPEAKER_00: They look very different, but they're still both dogs.
[4:17] SPEAKER_00: Why?
[4:18] SPEAKER_00: Because they mate together and they can be fertile.
[4:22] SPEAKER_00: So how do you explain deductively
[4:25] SPEAKER_00: that you categorize dogs in one way according to reproduction,
[4:31] SPEAKER_00: you categorize nines according to function.
[4:34] SPEAKER_00: And there's just lots of things you can't simply deduce.
[4:37] SPEAKER_00: They're much more arbitrary than that.
[4:40] And when you learn through induction,
[4:44] you need a lot of examples to induce accurately.
[4:50] SPEAKER_00: And if I look at 10 New Yorkers, I will conclude that
[4:55] SPEAKER_00: All Americans drink coffee in the morning.
[4:58] SPEAKER_00: Of course, that's not true.
[4:59] SPEAKER_00: If I looked at a bigger sample, I would understand it wasn't true.
[5:02] And so good induction requires an infinity of data,
[5:05] hence the title of the infinity machine.
[5:08] SPEAKER_00: And AI is really a system that can master induction.
[5:12] SPEAKER_01: Yeah, well, that's a great summary.
[5:13] SPEAKER_01: I'm glad you clarified it.
[5:14] SPEAKER_01: You even took it to another level.
[5:17] My trying to understand why you selected that as the title and
[5:23] quite apropos.
[5:25] SPEAKER_01: hearing your side of it.
[5:26] SPEAKER_01: Now,
[5:28] SPEAKER_01: the DeepMind company is pretty extraordinary when you go back,
[5:33] SPEAKER_01: because we're talking about 2010.
[5:34] SPEAKER_01: They're talking about almost 16 years ago.
[5:39] SPEAKER_01: I mean, Dennis will be 50 in July.
[5:42] SPEAKER_01: So he's a young chap, 34, and
[5:48] SPEAKER_01: back then, or 33.
[5:50] SPEAKER_01: And he,
[5:52] SPEAKER_01: with,
[5:52] SPEAKER_01: of course,
[5:53] SPEAKER_01: Shane Legg and Mustafa,
[5:56] they formed this company and they have a charter to have super
[6:01] intelligence
[6:02] SPEAKER_01: achieved by 2030 at the latest,
[6:05] SPEAKER_01: I guess.
[6:06] SPEAKER_01: And what was also striking is they had agents in the business plan.
[6:12] SPEAKER_01: Now,
[6:12] only recently have people awoken to the idea that we have all
[6:17] these agents to
[6:18] SPEAKER_01: execute autonomously.
[6:20] SPEAKER_01: So they were so far ahead.
[6:22] SPEAKER_01: And what's really striking to me is that this was,
[6:25] SPEAKER_01: I mean,
[6:26] SPEAKER_01: this takes prescience to another level,
[6:28] SPEAKER_01: I have to say.
[6:29] SPEAKER_01: So tell us about this formation of DeepMind in 2010,
[6:33] SPEAKER_01: because at the time,
[6:34] SPEAKER_01: they must have been seen as kooky people.
[6:38] SPEAKER_01: But if you look back, they were so far ahead of everybody else.
[6:43] SPEAKER_00: Yeah, I mean, that's totally right.
[6:45] In a way,
[6:45] that's why I chose Demis as my central protagonist for a book
[6:49] about AI.
[6:50] SPEAKER_00: You know, one could have chosen Sam Altman or somebody else.
[6:53] SPEAKER_00: But,
[6:53] SPEAKER_00: you know,
[6:54] SPEAKER_00: Demis really was so early that through his life,
[6:57] SPEAKER_00: you can tell the story of the making of modern AI.
[7:00] And he's just perfectly,
[7:02] he's fascinating himself,
[7:03] as perhaps we'll talk about that.
[7:05] But just the span and the early nature of his conviction,
[7:07] because in 2010,
[7:09] as you say,
[7:10] SPEAKER_00: You know, it was still the AI winter, really.
[7:12] SPEAKER_00: I mean,
[7:13] the ImageNet breakthrough in terms of computer vision and that
[7:18] CATS paper,
[7:19] SPEAKER_00: this was late 2012.
[7:21] SPEAKER_00: So Demis is forming DeepMind two years before that.
[7:24] SPEAKER_00: And he's been thinking about AI, you know, another 15 years before that.
[7:30] SPEAKER_00: So it is amazing.
[7:33] SPEAKER_00: And you're right that the business plan,
[7:34] SPEAKER_00: it took me enormous amounts of,
[7:36] SPEAKER_00: you know,
[7:36] SPEAKER_00: negotiation,
[7:38] SPEAKER_00: charm,
[7:38] SPEAKER_00: and general patience to extract the business plan from Dennis.
[7:42] SPEAKER_00: But finally, he gave it to me.
[7:44] SPEAKER_00: And when you read it, it is unbelievably present.
[7:46] SPEAKER_00: I mean,
[7:46] he really did say that extrapolating the way that computer
[7:50] power was growing,
[7:52] SPEAKER_00: he expected,
[7:53] SPEAKER_00: you know,
[7:54] SPEAKER_00: superhuman intelligence to arrive around 2030.
[7:56] SPEAKER_00: Amazing, right?
[8:01] SPEAKER_00: You're right about agents too.
[8:03] SPEAKER_01: Yeah, I was stunned.
[8:05] I knew he started this company,
[8:06] but I didn't know about the depth of the plan.
[8:09] SPEAKER_01: Now let's get more into Demis because as I read this book,
[8:15] I feel like this human was made for the mission in so many
[8:21] respects because
[8:24] For one,
[8:25] he got into chess very early as a precursor to games and
[8:31] competition.
[8:33] SPEAKER_01: And,
[8:34] SPEAKER_01: of course,
[8:34] SPEAKER_01: he was already doing that,
[8:35] SPEAKER_01: sitting on books to be able to compete against older players.
[8:40] SPEAKER_01: competitors.
[8:41] But it brought out this fierce competition,
[8:44] which was later characterized by Jeff Hinton.
[8:47] SPEAKER_01: He's the only person perhaps more competitive than me.
[8:51] SPEAKER_01: And of course,
[8:52] SPEAKER_01: when he was 14,
[8:54] SPEAKER_01: or was it 12,
[8:55] SPEAKER_01: when he lost a match,
[8:57] SPEAKER_01: it changed everything,
[8:58] SPEAKER_01: oh,
[8:58] SPEAKER_01: 12,
[8:58] SPEAKER_01: I think.
[8:59] SPEAKER_01: And he said, the only way I could know if I
[9:03] pushed myself to the point just before death because his father
[9:06] said,
[9:06] you got to try your best.
[9:07] And he gave it always 100%,
[9:10] not 99% as Shane Legg,
[9:12] his co-founder,
[9:13] pointed out.
[9:13] SPEAKER_01: So he had this
[9:16] SPEAKER_01: This unsatiable drive,
[9:20] SPEAKER_01: the games,
[9:22] SPEAKER_01: the fact that he then went on to games using GPUs,
[9:26] SPEAKER_01: which really enabled deep learning,
[9:29] the fact that then he had products from the games when he was
[9:32] working at Bullfrog,
[9:34] SPEAKER_01: so he could deal with products later.
[9:36] The fact that he was a computer scientist at Cambridge and then
[9:40] later a neuroscience scientist
[9:42] SPEAKER_01: learning about the brain,
[9:43] SPEAKER_01: the hippocampus specifically,
[9:46] SPEAKER_01: that he had this ultra charisma that he could pull jetty mind tricks,
[9:52] SPEAKER_01: that he learned from failure from Elixir.
[9:55] The company started with David Silver because of the fact that
[9:59] he was far ahead of his time.
[10:02] SPEAKER_01: And by the way,
[10:03] SPEAKER_01: one other thing that wasn't in the book,
[10:04] SPEAKER_01: which I learned from a conference when I was with Dennis in London,
[10:09] SPEAKER_01: November of 2024,
[10:12] SPEAKER_01: He was on a panel with four other Nobel laureates.
[10:17] SPEAKER_01: And one of them was Jennifer Doudna, and there were a couple others.
[10:21] Anyway,
[10:22] he said he knew he wanted to be a Nobel laureate when he was a
[10:26] young child.
[10:29] SPEAKER_01: And all the others looked at him like, what?
[10:31] SPEAKER_01: No.
[10:33] SPEAKER_01: And then you have him connected with Google later,
[10:37] SPEAKER_01: which has not an infinite data set,
[10:40] but lots of data from Gmail and the internet and searches and
[10:45] unlimited TPUs and
[10:46] SPEAKER_01: computing power and its own Google brain.
[10:49] It seems like this was a whole story built around this one
[10:55] person that had every
[10:57] SPEAKER_01: piece covered to do this.
[10:59] SPEAKER_01: Am I processing this wrong?
[11:02] SPEAKER_00: No, I think that's right.
[11:03] SPEAKER_00: I mean,
[11:04] SPEAKER_00: actually,
[11:04] SPEAKER_00: Peter Thiel,
[11:06] SPEAKER_00: who wrote the first check into DeepMind in the Series A round in 2010,
[11:13] SPEAKER_00: said to me,
[11:13] SPEAKER_00: you know,
[11:13] SPEAKER_00: he often thinks of,
[11:15] SPEAKER_00: you know,
[11:15] SPEAKER_00: there's no such thing as sort of a general purpose entrepreneur.
[11:19] People generally have one company that's inside them screaming
[11:25] to get out.
[11:27] SPEAKER_00: And so Demis, the mission for which he was born,
[11:30] was building AGI because he had all those qualities you've just
[11:34] described,
[11:35] SPEAKER_00: and he was just the perfect person to do it.
[11:38] Yeah,
[11:38] so that's what's so different about the other leaders in this
[11:41] field.
[11:42] Not only was he way ahead of them in time,
[11:45] but the portfolio of experience,
[11:48] the personality,
[11:50] SPEAKER_01: the special charisma, all these features.
[11:56] SPEAKER_01: And then,
[11:56] SPEAKER_01: of course,
[11:56] SPEAKER_01: when he got forced to have to deal with products,
[11:59] SPEAKER_01: I don't think that's something he would have picked because at heart,
[12:01] SPEAKER_01: he really seems like a scientist.
[12:03] And what also strikes me is he was so proud to publish papers
[12:09] in Nature,
[12:10] have covers in Nature,
[12:13] which is the opposite of the other companies that are happy to
[12:16] put out their white
[12:17] SPEAKER_01: papers and blogs and just post them on their website.
[12:21] SPEAKER_01: They don't really care.
[12:22] And so it's the difference between the scientists and the
[12:25] people that say,
[12:26] SPEAKER_01: you know,
[12:26] SPEAKER_01: we're so big,
[12:27] SPEAKER_01: we can just publish.
[12:28] SPEAKER_01: We don't need a journal.
[12:29] SPEAKER_01: We don't need peer review.
[12:30] SPEAKER_01: We can just publish our own stuff.
[12:32] SPEAKER_01: And that's like the mind of a scientist versus a different kind of
[12:38] SPEAKER_01: leader in AI today.
[12:42] SPEAKER_00: There's also,
[12:43] SPEAKER_00: you know,
[12:43] SPEAKER_00: part of the charm of the story is that Demis Sassabis,
[12:47] SPEAKER_00: of course,
[12:47] SPEAKER_00: he was born in North London,
[12:49] SPEAKER_00: built his company in London in King's Cross,
[12:52] SPEAKER_00: not a Silicon Valley person,
[12:55] SPEAKER_00: even though he got funding from Silicon Valley,
[12:57] SPEAKER_00: both from Peter Thiel and from Elon Musk,
[12:59] SPEAKER_00: and later from Google when he sold to Google.
[13:02] But,
[13:03] you know,
[13:04] one of the differences,
[13:04] it seems to be in the style,
[13:06] is that Demis
[13:07] SPEAKER_00: cares a lot about narrative and communication.
[13:10] He does care about his reputation and he believes you have to
[13:15] communicate publicly
[13:16] SPEAKER_00: about yourself in order to build the space,
[13:19] SPEAKER_00: the resources to go after AI.
[13:21] SPEAKER_00: So he's not innocent about the importance of public relations,
[13:26] SPEAKER_00: but his style of public relations,
[13:28] SPEAKER_00: as you say,
[13:29] SPEAKER_00: published in a peer-reviewed journal,
[13:31] SPEAKER_00: that's what he thinks is good.
[13:32] SPEAKER_00: Because then you can really impress people.
[13:34] SPEAKER_00: You can say, hey, you can't question this.
[13:36] SPEAKER_00: This is peer-reviewed.
[13:38] SPEAKER_00: Whereas Sam Altman's idea is that he has a lot of followers on Twitter.
[13:43] SPEAKER_00: It's a kind of different style.
[13:46] SPEAKER_01: Yeah,
[13:46] he has even more people questioning him after the New Yorker
[13:51] piece by Ron Farrow
[13:53] SPEAKER_01: that appeared.
[13:54] SPEAKER_01: So I think what we have now is we have this setup of a unique leader
[14:01] SPEAKER_01: who has brought with him these co-founders with other talents,
[14:05] SPEAKER_01: of course,
[14:06] and then is capable of bringing many other great minds to join
[14:11] him and basically
[14:13] SPEAKER_01: follow his lead.
[14:16] SPEAKER_01: And he's on this mission.
[14:18] SPEAKER_01: One of the most extraordinary things,
[14:20] SPEAKER_01: I had no idea,
[14:21] SPEAKER_01: Sebastian,
[14:22] SPEAKER_01: was in March 2016,
[14:26] SPEAKER_01: basically the day when Lee Seedal,
[14:29] SPEAKER_01: Lee,
[14:30] Go champion of the world was overtaken by the deep mind
[14:39] algorithms,
[14:40] if you will.
[14:42] SPEAKER_01: The same day, if not within hours, he decides, well, that's not enough.
[14:45] SPEAKER_01: I'm starting Alpha full.
[14:48] SPEAKER_01: So Alpha Go,
[14:49] SPEAKER_01: Alpha full,
[14:50] SPEAKER_01: the same,
[14:51] SPEAKER_01: you know,
[14:51] SPEAKER_01: the turning point in the same day,
[14:53] SPEAKER_01: is that about right?
[14:54] SPEAKER_00: Yeah.
[14:55] SPEAKER_00: It's unbelievable.
[14:56] SPEAKER_00: You know,
[14:56] SPEAKER_00: he's out there in Seoul with David Silva,
[14:59] his friend and the reinforcement learning scientist who was
[15:02] behind AlphaGo.
[15:03] SPEAKER_00: And it's day four of this tournament.
[15:08] SPEAKER_00: And by winning,
[15:11] SPEAKER_00: I guess it was three,
[15:12] SPEAKER_00: the third three to one,
[15:15] SPEAKER_00: they knew that they were going to win the best of five.
[15:17] SPEAKER_00: So the tournament was over effectively.
[15:20] And as they walked down the street in Seoul,
[15:23] there was a camera team behind them
[15:25] that picked up on a microphone this entirely sort of,
[15:28] you know,
[15:29] this open mic moment.
[15:30] SPEAKER_00: It wasn't intended to be for quotation or whatever.
[15:32] But,
[15:33] you know,
[15:34] Demis was saying to Dave,
[15:35] his friend,
[15:36] look,
[15:37] we've done AlphaGo.
[15:39] SPEAKER_00: Now we've got to move on to the next thing.
[15:41] SPEAKER_00: And I think, you know, protein folding is what we should go after next.
[15:44] SPEAKER_00: Now we can really do protein folding.
[15:46] SPEAKER_00: So,
[15:46] SPEAKER_00: you know,
[15:47] SPEAKER_00: as Dave laconically said to me,
[15:49] SPEAKER_00: you know,
[15:50] SPEAKER_00: when Demis does an extraordinary achievement like solving Go,
[15:56] SPEAKER_00: He waits about 32 seconds before announcing his next ambition.
[16:03] SPEAKER_01: It's so extraordinary.
[16:04] SPEAKER_01: I mean,
[16:05] SPEAKER_01: you know,
[16:06] the move 37 wasn't enough to have basically the machine come up
[16:12] with something that
[16:12] SPEAKER_01: humans couldn't.
[16:14] And then,
[16:14] you know,
[16:15] right there,
[16:16] turn to what was going to be the quest for AlphaFold.
[16:19] SPEAKER_01: And as you know...
[16:21] SPEAKER_01: I think for those who are listening,
[16:23] AlphaFold went through stages where ultimately AlphaFold II was
[16:28] what was recognized
[16:29] SPEAKER_01: for predicting 200 million proteins from amino acid sequences,
[16:35] SPEAKER_01: the folding and the features of the protein,
[16:38] SPEAKER_01: which otherwise would take,
[16:40] SPEAKER_01: I mean,
[16:41] I'm in school for research where we had so many
[16:44] crystallographers on our faculty
[16:46] SPEAKER_01: and they might spend two years,
[16:48] SPEAKER_01: three years
[16:49] SPEAKER_01: to crystallize a protein to be able to understand its folding.
[16:53] SPEAKER_01: And here you could do it in seconds through alpha fold two,
[16:57] SPEAKER_01: now alpha fold three with the protein interactions.
[17:00] SPEAKER_01: I mean, this is the singular biggest contribution of AI to date.
[17:05] SPEAKER_01: And that's,
[17:06] SPEAKER_01: you know,
[17:07] SPEAKER_01: something that Demis,
[17:08] SPEAKER_01: John Jumper,
[17:09] SPEAKER_01: and the whole big team,
[17:11] SPEAKER_01: of course,
[17:11] SPEAKER_01: can be proud of.
[17:13] And what was really interesting,
[17:14] and now I know why,
[17:15] we're going to get into this,
[17:17] um,
[17:18] SPEAKER_01: When I was sitting next to John Jumper at the same conference in London,
[17:23] SPEAKER_01: and I said to him,
[17:24] SPEAKER_01: well,
[17:25] SPEAKER_01: you couldn't have done AlphaFold II without the transformer model,
[17:28] SPEAKER_01: right?
[17:29] SPEAKER_01: And he didn't like that.
[17:32] SPEAKER_01: It's an anti-transformer model thing.
[17:34] SPEAKER_01: So before we get to that,
[17:37] SPEAKER_01: and I'll qualify that because it wasn't quite an antibody response,
[17:42] SPEAKER_01: but...
[17:44] SPEAKER_01: There's,
[17:44] SPEAKER_01: as you discussed,
[17:45] SPEAKER_01: and you introduced this pretty early,
[17:46] but I want to get back to it because it's really a central part
[17:49] of the book.
[17:50] So previously there was a symbolic AI and it was rules-based
[17:56] and it was rigid deduction.
[17:58] SPEAKER_01: Okay.
[17:59] SPEAKER_01: It only could go so far.
[18:01] SPEAKER_01: And then came along the Toronto and subsequently Alberta tribes.
[18:07] SPEAKER_01: The Toronto tribe, deep learning.
[18:09] SPEAKER_01: This is Jeffrey Hinton.
[18:12] SPEAKER_01: with Ilya and Yan and Yoshio and others.
[18:15] SPEAKER_01: And this is really about perception.
[18:18] SPEAKER_01: It's about these neural networks that can classify.
[18:22] And of course,
[18:23] that's really where images took off with Fei-Fei Li's ImageNet.
[18:26] And as you said,
[18:27] that really got going in 2012,
[18:30] not when DeepMind was formed.
[18:32] SPEAKER_01: And so that deep learning era,
[18:34] SPEAKER_01: which we're still in in some respects,
[18:37] SPEAKER_01: of course,
[18:37] SPEAKER_01: that really took AI out of the winter into the hot zone.
[18:42] SPEAKER_01: Reinforcement learning,
[18:44] SPEAKER_01: which again,
[18:46] SPEAKER_01: the fact that that's the game story,
[18:48] SPEAKER_01: that the games can learn by reward.
[18:53] This is perfect to start there and then just work your way up
[18:57] to AlphaFold.
[19:00] SPEAKER_01: But this is really using agents and trial and error.
[19:05] SPEAKER_01: It's learning by acting in an environment.
[19:09] SPEAKER_01: In the book,
[19:10] SPEAKER_01: you really make a big,
[19:13] SPEAKER_01: I think,
[19:13] dichotomy between the deep learning and the reinforcement
[19:17] learning and how the best
[19:19] SPEAKER_01: of both worlds,
[19:21] SPEAKER_01: when you confuse it.
[19:22] SPEAKER_01: Can you tell us more about that?
[19:25] SPEAKER_00: Sure.
[19:25] SPEAKER_00: I mean,
[19:26] the first DeepMind breakthrough was the system that could play
[19:30] Atari games like
[19:32] SPEAKER_00: Breakout and Pong and so forth.
[19:35] SPEAKER_00: And this combined both deep learning and reinforcement learning.
[19:40] SPEAKER_00: And it was the first successful combination of the two in one system.
[19:46] SPEAKER_00: And it was revolutionary because,
[19:50] SPEAKER_00: as you indicated by using the word tribe,
[19:53] there hadn't been much communication between the reinforcement
[19:57] learning tribe based
[19:59] SPEAKER_00: mostly in Edmonton in Canada and
[20:03] SPEAKER_00: And then the deep learners who were,
[20:05] SPEAKER_00: you know,
[20:05] SPEAKER_00: Toronto was one place,
[20:06] SPEAKER_00: but I guess there was Yoshua Bengio in Montreal and so forth.
[20:10] SPEAKER_00: And so they really didn't communicate much,
[20:13] but there was one graduate student who had studied at a
[20:17] graduate level in both
[20:19] SPEAKER_00: places,
[20:19] SPEAKER_00: both Edmonton and in Toronto.
[20:22] SPEAKER_00: And this was Vlad Nih.
[20:24] And Vlad Nih,
[20:25] a sort of Ukrainian-Canadian,
[20:28] was persuaded to join DeepMind early on
[20:32] And he brought with him this ambition to combine the two sides
[20:35] of his graduate
[20:36] SPEAKER_00: studies,
[20:36] SPEAKER_00: to put them into one system.
[20:38] SPEAKER_00: And he teamed up with other people who had studied in Switzerland,
[20:42] SPEAKER_00: where this divide between the two schools,
[20:44] SPEAKER_00: reinforcement learning and deep learning,
[20:47] SPEAKER_00: was less pronounced.
[20:49] SPEAKER_00: And so DeepMind resolved to make the two work together.
[20:53] SPEAKER_00: And that's what unlocked these Atari game-playing agents.
[20:57] You needed the deep learning partly just to give the system a
[21:00] perceptual tool,
[21:02] to actually see the Atari board,
[21:04] the screen,
[21:05] understand the pixels on the screen.
[21:07] But you needed the reinforcement learning to give the system
[21:11] the ability to kind of
[21:12] SPEAKER_00: understand the game,
[21:14] SPEAKER_00: to plan,
[21:14] SPEAKER_00: to develop a strategy.
[21:15] SPEAKER_00: And that's where the reinforcement learning came in.
[21:20] And through combining the two,
[21:23] DeepMind put itself in a league of its own.
[21:25] SPEAKER_01: Yeah, and that, of course, was instrumental
[21:29] SPEAKER_01: That is,
[21:30] SPEAKER_01: you could take the protein data bank,
[21:32] SPEAKER_01: which had been worked on for many years,
[21:34] SPEAKER_01: and the brilliance of using that,
[21:37] SPEAKER_01: which we don't have many data sets like that,
[21:40] SPEAKER_01: to basically bring deep learning and reinforcement learning together.
[21:46] But going back to when I challenged John Jumper about,
[21:50] well,
[21:51] for you to go into
[21:54] high gear with these 200 million proteins,
[21:56] you had to have had transformer models.
[21:58] SPEAKER_01: And he really wasn't keen on that.
[22:01] SPEAKER_01: And that gets me to the next thing,
[22:03] SPEAKER_01: which is,
[22:04] SPEAKER_01: in the book,
[22:07] SPEAKER_01: through these chapters,
[22:08] SPEAKER_01: you're really starting to appreciate
[22:10] SPEAKER_01: that Demis is an extraordinary force, a character.
[22:16] SPEAKER_01: Those are not enough words.
[22:19] SPEAKER_01: Maybe not messianic, maybe not enders, but out there.
[22:23] SPEAKER_01: He's been a very class of his own as an individual.
[22:26] SPEAKER_01: But then you see the big miss.
[22:30] The fact that after all this tremendous momentum leading up
[22:35] eventually to the Nobel
[22:36] SPEAKER_01: Prize in chemistry in 2024,
[22:40] SPEAKER_01: that in 2017,
[22:43] SPEAKER_01: the group of seven at Google,
[22:47] SPEAKER_01: the company he's part of,
[22:49] SPEAKER_01: they come up with this transformer model and attention is all you need.
[22:53] SPEAKER_01: And they basically set the stage for Ilya at OpenAI to,
[23:00] SPEAKER_01: you know,
[23:00] SPEAKER_01: he falls out of his chair or whatever.
[23:03] SPEAKER_01: He jumps out of his chair.
[23:04] SPEAKER_02: Jumps out of his chair, yeah.
[23:05] SPEAKER_02: Jumps out of, yeah.
[23:07] SPEAKER_01: And so now...
[23:08] SPEAKER_01: OpenAI is born.
[23:10] SPEAKER_01: They have now the way forward.
[23:13] SPEAKER_01: They're going to be GPT-2, 3, 4, and on and on.
[23:17] SPEAKER_01: And they're going to just get hyperscale with data.
[23:21] SPEAKER_01: And they're going to build the infinity machine.
[23:23] SPEAKER_01: I mean, they didn't call it that.
[23:24] SPEAKER_01: But all of a sudden,
[23:26] SPEAKER_01: they invented, we're going to get AGI.
[23:28] SPEAKER_01: Well,
[23:28] SPEAKER_01: they're like,
[23:30] SPEAKER_01: what,
[23:31] SPEAKER_01: 17 years late here,
[23:34] SPEAKER_01: or at least seven years late,
[23:35] SPEAKER_01: sorry,
[23:36] SPEAKER_01: after what DeepMind had initiated.
[23:39] SPEAKER_01: But
[23:40] SPEAKER_01: Then what I just cannot fathom,
[23:43] SPEAKER_01: although you bring up some possibilities in the book,
[23:45] SPEAKER_01: you bring up,
[23:46] SPEAKER_01: well,
[23:46] SPEAKER_01: maybe because there was so much angst between Google and DeepMind,
[23:51] SPEAKER_01: you know,
[23:51] SPEAKER_01: about being separated,
[23:53] SPEAKER_01: or maybe it's because there was angst between Demis and Mustafa,
[23:57] SPEAKER_01: or,
[23:57] SPEAKER_01: you know,
[23:59] SPEAKER_01: about this pride that we developed,
[24:01] SPEAKER_01: you know,
[24:02] SPEAKER_01: the whole deep learning,
[24:03] SPEAKER_01: reinforcement learning,
[24:04] SPEAKER_01: we don't need anything else.
[24:05] SPEAKER_01: But there was basically a dissing
[24:09] SPEAKER_01: of the importance of transformer architecture and generative AI.
[24:13] SPEAKER_01: Can you help explain that?
[24:15] SPEAKER_01: Because this was a catastrophic miss at the time.
[24:21] SPEAKER_00: Well,
[24:21] SPEAKER_00: you know,
[24:22] I talked to Temis a lot about this because I had the good
[24:24] fortune of pitching him
[24:26] SPEAKER_00: on the idea of this project,
[24:28] SPEAKER_00: giving me hours and hours of his time.
[24:30] SPEAKER_00: in November 2022.
[24:32] SPEAKER_00: And of course, on the last day of that month is when ChattyBT came out.
[24:37] And so I was embedded with DeepMind right when this whole topic
[24:41] of AI went from the
[24:42] fringe to the mainstream and right when generative AI had its
[24:46] breakout moment.
[24:47] SPEAKER_00: So in our discussions,
[24:49] SPEAKER_00: and we would meet up in a pub in North London near his house,
[24:53] and there'd be a dusty staircase at the back that people didn't
[24:56] know about.
[24:57] SPEAKER_00: go out the staircase,
[24:57] there was this empty room we would have to ourselves and sit
[25:00] there for two hours
[25:01] SPEAKER_00: and debate stuff.
[25:03] SPEAKER_00: And so I had plenty of chances to ask him about why he had this big miss.
[25:07] And the most interesting part of the answer is essentially his
[25:10] neuroscience PhD had
[25:13] SPEAKER_00: left him with this big conviction that,
[25:17] SPEAKER_00: you know,
[25:17] SPEAKER_00: intelligence is,
[25:19] SPEAKER_00: first of all,
[25:21] SPEAKER_00: the result of different components of the brain and different
[25:25] SPEAKER_00: ways of coming at information and processing it.
[25:29] SPEAKER_00: It's not just one thing.
[25:31] SPEAKER_00: And also that in neuroscience,
[25:33] SPEAKER_00: there's this field,
[25:34] SPEAKER_00: action and perception,
[25:36] SPEAKER_00: which posits that to really be intelligent,
[25:39] SPEAKER_00: you need to be embodied in the world.
[25:41] SPEAKER_00: You need to have a physical model of the world to experience gravity,
[25:45] to experience what it is to drop a glass and have it shatter on
[25:48] the ground.
[25:50] SPEAKER_00: Without those things, you won't really be grounded.
[25:53] SPEAKER_00: And this word grounded came a lot in my conversations with Dennis.
[25:58] And so when he saw the early efforts at OpenAI to apply
[26:04] transformer models to a
[26:05] SPEAKER_00: language model,
[26:07] SPEAKER_00: you know,
[26:09] SPEAKER_00: his idea was sort of,
[26:10] SPEAKER_00: you know,
[26:10] SPEAKER_00: good luck with that.
[26:11] You know,
[26:11] you're not going to get to full intelligence just through
[26:14] language.
[26:15] SPEAKER_00: Language is too limited.
[26:17] SPEAKER_00: Not grounded.
[26:19] SPEAKER_00: Not grounded.
[26:20] SPEAKER_00: And also he said to me at one point, you know,
[26:22] SPEAKER_00: It's a question of how big you think human experience might be.
[26:27] Could you really encompass all there is in human experience
[26:32] just by downloading all
[26:34] SPEAKER_00: the language on the internet?
[26:35] SPEAKER_00: I didn't think so.
[26:36] SPEAKER_00: Well, it turns out he was wrong.
[26:38] SPEAKER_00: You can get at least,
[26:40] SPEAKER_00: I don't know,
[26:41] SPEAKER_00: 90% of human experience by downloading everything on the internet.
[26:45] SPEAKER_00: And language is more grounded than you realize in the sense that,
[26:49] SPEAKER_00: first of all,
[26:50] SPEAKER_00: it's written by humans who are grounded.
[26:52] SPEAKER_00: Secondly,
[26:53] SPEAKER_00: these systems get reinforcement learning from human feedback,
[26:56] so they get some grounding through the feedback after they've
[26:59] been trained.
[27:01] And so you get a very good approximation of a grounded
[27:05] intelligence just by
[27:07] SPEAKER_00: training on language data.
[27:09] SPEAKER_00: And that was the key mistake.
[27:12] SPEAKER_00: Neuroscience had led him in the right direction early on with games.
[27:17] SPEAKER_00: It misled him when it came to language models.
[27:20] SPEAKER_01: Yeah, well, of course, he figured that out.
[27:23] SPEAKER_01: And as I think everyone who's used AI now,
[27:26] SPEAKER_01: that generative AI was going to be another very,
[27:29] SPEAKER_01: very big factor in the future as AI moves forward.
[27:33] SPEAKER_01: And then,
[27:35] SPEAKER_01: of course,
[27:35] it isn't just the large language models and large reasoning
[27:38] models and the agentic
[27:40] SPEAKER_01: and on and on.
[27:41] And we'll see other models undoubtedly in the years ahead that
[27:46] move beyond what we have today.
[27:48] SPEAKER_01: But
[27:49] SPEAKER_01: What really was interesting is this set up an extraordinary competition.
[27:55] SPEAKER_01: You had Sam Altman and Greg Brockman on the one side,
[28:00] SPEAKER_01: and you had Antropic because Dario and his sister had exited,
[28:07] SPEAKER_01: and of course,
[28:08] SPEAKER_01: many others.
[28:09] SPEAKER_01: But principally, you had Google...
[28:13] SPEAKER_01: And DeepMind and Google really had all these assets.
[28:16] And then you had these two startups that are now trying to have
[28:22] initial public
[28:22] SPEAKER_01: offerings and hundreds of billions of crazy stuff,
[28:26] SPEAKER_01: right?
[28:27] And here Google's kind of caught with their pants down that
[28:30] they're way behind,
[28:33] SPEAKER_01: especially with this chat GPT moment,
[28:36] SPEAKER_01: as you said,
[28:37] SPEAKER_01: end of November 22.
[28:39] SPEAKER_01: But amazingly enough,
[28:41] It wasn't that long before Gemini 3 eventually evolved and
[28:46] caught up and even transcended.
[28:48] SPEAKER_01: And now we have this basically arms race with hyperscaling.
[28:51] SPEAKER_01: And these three entities are basically,
[28:55] SPEAKER_01: you know,
[28:55] SPEAKER_01: kind of in a way going after each other,
[28:57] SPEAKER_01: but they have different strengths.
[28:59] SPEAKER_01: And one of the questions I have for you that's not in the book,
[29:02] SPEAKER_01: because it kind of evolves subsequently,
[29:04] SPEAKER_01: is the coding,
[29:05] SPEAKER_01: vibe coding story.
[29:07] SPEAKER_01: So, come on.
[29:09] SPEAKER_01: Anthropic differentiated itself with, uh, Claude code.
[29:14] SPEAKER_01: And of course, open AI had codecs, but, um,
[29:18] SPEAKER_01: That's another thing I wonder.
[29:20] SPEAKER_01: Is there another miss here?
[29:22] SPEAKER_01: Because the coding, you can't do that.
[29:25] SPEAKER_01: I use Gemini 3 for a lot of things, right?
[29:28] SPEAKER_01: My deep research, my notebook LM, Nana Banana.
[29:32] SPEAKER_01: But it wouldn't be the place for coding.
[29:35] SPEAKER_01: Is that another big miss?
[29:36] SPEAKER_01: Or is that something that,
[29:38] SPEAKER_01: you know,
[29:39] SPEAKER_01: I think eventually they're going to put a lot of weight into?
[29:43] SPEAKER_00: Look, I'll come back to that in a second.
[29:44] SPEAKER_00: I just want to double click on one thing you said about...
[29:47] SPEAKER_00: the comeback story.
[29:48] Because I think if you'd asked any business school professor to
[29:52] predict the success
[29:53] SPEAKER_00: of a merger,
[29:55] SPEAKER_00: where you take Google Brain in Mountain View,
[29:59] SPEAKER_00: and you smush it together with DeepMind in London,
[30:02] and you do this in the middle of a really intense race with
[30:06] your rival open AI,
[30:08] SPEAKER_00: and you factor in that there's eight hours of time difference between
[30:12] SPEAKER_00: Mountain View and London,
[30:14] and that the two labs didn't like each other because they'd
[30:16] been competing over
[30:18] SPEAKER_00: computing resources for a long time.
[30:20] I think the business school professor would have said,
[30:23] hey,
[30:24] mergers are always difficult.
[30:25] SPEAKER_00: This one looks impossible,
[30:28] and certainly not in the time that you need to make it work to
[30:31] catch up with OpenAI
[30:32] SPEAKER_00: in this race.
[30:33] SPEAKER_00: And yet, as you just said, they did catch up.
[30:36] SPEAKER_00: It took two and a half years from ChatGPT coming out, basically, for
[30:41] SPEAKER_00: for Google to have pretty much drawn even or even gone slightly ahead.
[30:46] SPEAKER_00: So I think not only is this a frontier technology,
[30:49] but there's been some frontier business management stories
[30:53] coming out as well.
[30:56] SPEAKER_00: On the coding,
[30:57] SPEAKER_00: I mean,
[30:57] I think it's too early to say that's really a big miss because
[31:01] I suspect that it
[31:04] SPEAKER_00: won't be long before Gemini is offering coding capability
[31:09] SPEAKER_00: on a par.
[31:10] My feeling is that right now,
[31:13] all of these labs are racing,
[31:14] they overtake each other.
[31:15] SPEAKER_00: If you'd asked the question,
[31:18] SPEAKER_00: which is the top model in November 2025 last year,
[31:23] SPEAKER_00: it would have been clearly Gemini.
[31:24] SPEAKER_00: Now, if you ask that question, it's pretty clearly explored.
[31:28] SPEAKER_00: In three or six months from now, I assume it will be somebody else.
[31:32] I don't see the coding breakthrough as that much of a key miss
[31:37] for Google.
[31:38] SPEAKER_01: Right.
[31:39] SPEAKER_01: I mean,
[31:39] SPEAKER_01: I think it's more for the people that are,
[31:42] SPEAKER_01: they are kind of open claw world that are going to have,
[31:45] SPEAKER_01: you know,
[31:45] SPEAKER_01: the agents that are doing efficiently,
[31:48] SPEAKER_01: autonomously getting so much done.
[31:50] SPEAKER_01: And it isn't for the,
[31:52] SPEAKER_01: you know,
[31:52] SPEAKER_01: everyday person that's,
[31:54] SPEAKER_01: you know,
[31:54] SPEAKER_01: more using it as an information resource,
[31:56] SPEAKER_01: but we'll see how it evolves.
[31:58] SPEAKER_01: It's really fun.
[31:59] SPEAKER_01: And the competition is really fierce.
[32:01] SPEAKER_01: And that's another question I had for you because knowing Dennis as I do,
[32:06] SPEAKER_01: and,
[32:06] SPEAKER_01: you know,
[32:06] SPEAKER_01: he,
[32:07] SPEAKER_01: he,
[32:07] SPEAKER_01: I,
[32:08] SPEAKER_01: of course, have the highest regard for him.
[32:11] SPEAKER_01: He's actually a hero to me,
[32:13] SPEAKER_01: which is another reason why I love this book,
[32:15] SPEAKER_01: because I identify with him.
[32:17] SPEAKER_01: I mean,
[32:17] SPEAKER_01: he's a scientist and he wants to do things the way science is done,
[32:21] SPEAKER_01: not flash in the pan.
[32:24] SPEAKER_01: But he's been put, because he was made in charge of this whole effort,
[32:29] SPEAKER_01: as you said,
[32:30] SPEAKER_01: the coalescence of brain and Mountain View resources of talent,
[32:36] SPEAKER_01: human talent,
[32:36] SPEAKER_01: and computing resources with what is in London,
[32:40] SPEAKER_01: he now is put in a position that takes him away from the research.
[32:44] SPEAKER_01: He's now a product.
[32:46] SPEAKER_01: He now has to compete for not just benchmarks,
[32:49] SPEAKER_01: but,
[32:50] SPEAKER_01: you know,
[32:51] SPEAKER_01: gotta be better than this model or that model.
[32:53] SPEAKER_01: Like every week, it's completely nuts.
[32:57] SPEAKER_01: I wonder,
[32:58] SPEAKER_01: when I asked him if he's okay with that,
[33:00] SPEAKER_01: I mean,
[33:00] SPEAKER_01: he kind of said he was,
[33:01] SPEAKER_01: but what do you think about that?
[33:04] SPEAKER_00: Yeah, I mean, you're right.
[33:06] SPEAKER_00: He has these two different personas in one person.
[33:10] SPEAKER_00: The first is the scientist.
[33:12] SPEAKER_00: And I think he really is driven by that intellectual curiosity,
[33:17] SPEAKER_00: the burning desire to understand how the world works.
[33:20] And that's sort of his primary driver,
[33:24] but also the part of himself that he
[33:27] SPEAKER_00: likes to think of as the dominant one.
[33:30] SPEAKER_00: But then there's another side of him,
[33:32] SPEAKER_00: which is just super competitive,
[33:34] SPEAKER_00: going back to his youth chess playing and,
[33:39] SPEAKER_00: you know,
[33:39] SPEAKER_00: just kind of silly things he's competitive about,
[33:41] SPEAKER_00: like table football or fussball,
[33:43] SPEAKER_00: you know,
[33:44] SPEAKER_00: that table game.
[33:46] SPEAKER_00: You know,
[33:46] SPEAKER_00: he would say to me,
[33:47] SPEAKER_00: well,
[33:47] SPEAKER_00: you know,
[33:47] SPEAKER_00: you understand,
[33:48] SPEAKER_00: Sebastian,
[33:49] SPEAKER_00: when I was at Cambridge,
[33:50] SPEAKER_00: I was the best player on the campus.
[33:53] SPEAKER_00: I mean, nobody else at university was better than me at that game.
[33:57] SPEAKER_00: And I sort of didn't really care one way or the other,
[33:59] SPEAKER_00: so I would kind of laugh and say,
[34:00] SPEAKER_00: no,
[34:00] SPEAKER_00: no,
[34:00] SPEAKER_00: no,
[34:00] SPEAKER_00: no,
[34:00] SPEAKER_00: no.
[34:01] SPEAKER_00: You have to understand, I was the best.
[34:02] SPEAKER_00: And I tell you,
[34:03] I watched the YouTube clips of the professional players in the
[34:07] American League,
[34:08] SPEAKER_00: and they didn't know the snake shop,
[34:10] where you loop your arm under the lever and you kind of get
[34:13] extra oomph on it.
[34:14] SPEAKER_00: And I'm like, whatever.
[34:16] SPEAKER_00: But he's so competitive.
[34:19] SPEAKER_00: So I think, you know, on this question of is he okay about being...
[34:23] the chief executive of Google DeepMind and being basically a
[34:27] corporate leader,
[34:28] SPEAKER_00: you know,
[34:29] SPEAKER_00: part of him loves it because he is competitive.
[34:32] SPEAKER_00: And part of him is a scientist and,
[34:34] SPEAKER_00: you know,
[34:34] SPEAKER_00: would openly talk to me about,
[34:36] SPEAKER_00: well,
[34:37] SPEAKER_00: you know,
[34:37] SPEAKER_00: maybe I should take up a professorship and just think.
[34:43] Probably he'll do more corporate leadership first and then the
[34:46] other thing afterwards.
[34:47] SPEAKER_01: Yeah,
[34:47] in the book you mentioned his affinity to go to a place like
[34:51] Princeton and just,
[34:52] SPEAKER_01: you know,
[34:53] SPEAKER_01: get into deep thinking mode.
[34:55] SPEAKER_01: Yeah, absolutely.
[34:56] SPEAKER_01: I have to say,
[34:57] SPEAKER_01: Sebastian,
[34:57] SPEAKER_01: I'm a little concerned because taking him out of his science mode,
[35:02] it could be this next big model that is the next chapter in the
[35:07] evolution of AI
[35:08] SPEAKER_01: that we could be missing.
[35:11] SPEAKER_01: So we'll see how that plays out.
[35:12] SPEAKER_01: But just extraordinary.
[35:14] SPEAKER_01: Now,
[35:15] SPEAKER_01: before I go into my last topic,
[35:18] SPEAKER_01: I want to just get back to this oligopoly of companies and players.
[35:24] SPEAKER_01: and the kind of killer instinct between them.
[35:29] SPEAKER_01: So one thing I didn't know about was the ultimate double cross.
[35:34] SPEAKER_01: When you had Sam Altman, Elon Musk, and Reid Hoffman,
[35:39] SPEAKER_01: when they were together to try to work on AI safety,
[35:42] SPEAKER_01: and then they just basically just went,
[35:45] SPEAKER_01: I mean,
[35:46] SPEAKER_01: that whole story,
[35:47] SPEAKER_01: I had no idea.
[35:48] I don't know if it had been publicized previously,
[35:51] but that just kind of typifies this.
[35:56] SPEAKER_01: It's not just fierce competition.
[35:57] SPEAKER_01: It's that whatever it takes to take each other on, it seems like.
[36:02] SPEAKER_01: Is that right?
[36:04] SPEAKER_00: Yeah.
[36:04] I mean,
[36:05] actually,
[36:05] Sam Oldman was not part of the meeting you're referring to,
[36:08] but it's true that
[36:09] SPEAKER_00: Elon Musk and Reid Hoffman were,
[36:10] and this was a safety and ethics oversight board which was
[36:14] supposed to advise
[36:15] SPEAKER_00: Google on the safe deployment of AI.
[36:18] SPEAKER_00: And they had their inaugural meeting,
[36:20] SPEAKER_00: actually Elon Musk was asked to host it,
[36:21] and that was a way of trying to bind him into the kind of AI
[36:26] safety aspect of
[36:27] SPEAKER_00: DeepMind's work.
[36:29] SPEAKER_00: And Elon had expressed a lot of concern about existential risk from AI,
[36:32] SPEAKER_00: so that seemed fitting that he should be involved.
[36:37] SPEAKER_00: And then he sat through the meeting,
[36:40] SPEAKER_00: sucked up all the presentations from the DeepMind leadership,
[36:44] and a few months later announced that he was founding OpenAI
[36:48] with Sam Altman and
[36:49] SPEAKER_00: Reid Hoffman would be one of the funders.
[36:52] SPEAKER_00: So yeah, it was pretty ruthless.
[36:55] SPEAKER_00: And Google's reaction to this was to say,
[36:58] SPEAKER_00: you know what,
[36:58] SPEAKER_00: we've had enough of these oversight committees.
[37:00] We're just going to brief people who will use the information
[37:04] against us as competitors.
[37:07] SPEAKER_00: But,
[37:07] SPEAKER_00: you know,
[37:07] SPEAKER_00: interestingly,
[37:08] SPEAKER_00: and to Demis' credit,
[37:09] SPEAKER_00: and also Mustafa Suleyman,
[37:11] SPEAKER_00: his co-founder,
[37:12] SPEAKER_00: you know,
[37:12] SPEAKER_00: they really were serious about wanting safety oversight.
[37:16] So then they embarked on this three-year negotiation with
[37:19] Google to try to force
[37:20] SPEAKER_00: Google to create a new safety oversight board.
[37:24] And they threatened actually to spin out from Google and go
[37:28] independent again if
[37:29] SPEAKER_00: they didn't get what they wanted.
[37:30] SPEAKER_00: But
[37:31] SPEAKER_00: That's a whole other story.
[37:32] Demis wasn't so happy that I wrote about that because it was
[37:35] based on all kinds of
[37:36] SPEAKER_00: internal leaks from DeepMind.
[37:40] SPEAKER_00: But it is true and a remarkable story.
[37:44] SPEAKER_01: And paradoxically,
[37:46] Reid Hoffman was one of the people that they were looking to to
[37:48] help fund it and
[37:49] SPEAKER_01: all kinds of,
[37:50] SPEAKER_01: the cross talk of these big players,
[37:53] SPEAKER_01: you know,
[37:54] SPEAKER_01: just
[37:55] SPEAKER_01: The ruthlessness,
[37:56] SPEAKER_01: competition,
[37:57] SPEAKER_01: you know,
[37:59] SPEAKER_01: there's not that many examples that we can see like this.
[38:02] And obviously,
[38:02] it's going to continue because everything's at stake here,
[38:06] the dominance of AI.
[38:08] Now,
[38:09] that gets me to the last topic I wanted to get in with you,
[38:12] which is the danger of
[38:15] SPEAKER_01: The fact that we're threatening human sovereignty,
[38:20] SPEAKER_01: we have the world,
[38:23] SPEAKER_01: as we saw,
[38:23] the clash between anthropic and open AI about using AI in war
[38:30] for surveillance,
[38:32] SPEAKER_01: no less all sorts of doomsday or even just short of that,
[38:38] SPEAKER_01: horrendous scenarios that we could be seeing.
[38:40] SPEAKER_01: And reasonable people would fear this.
[38:44] And obviously you get into this various times in the book about
[38:48] the safety,
[38:49] SPEAKER_01: the ethics,
[38:50] SPEAKER_01: the need for readiness and whatever,
[38:55] SPEAKER_01: you know,
[38:55] SPEAKER_01: not to have the loss of control.
[38:58] SPEAKER_01: Now, as you got to in the book, there were times, do we have a pause?
[39:02] Well,
[39:02] that doesn't seem very smart because you have a pause for six
[39:06] months and you're back.
[39:08] SPEAKER_01: How do we aim high?
[39:10] SPEAKER_01: How do we go that route?
[39:12] SPEAKER_01: And especially when you have,
[39:13] SPEAKER_01: you know,
[39:13] SPEAKER_01: it's not just the oligopoly based in the U.S.,
[39:16] SPEAKER_01: you've got China,
[39:17] SPEAKER_01: you've got all sorts of other players here.
[39:20] SPEAKER_01: How are we going to get this right?
[39:23] SPEAKER_00: Yeah.
[39:23] SPEAKER_00: I mean,
[39:24] SPEAKER_00: this is the very deep question I wrestled with all throughout the book,
[39:27] SPEAKER_00: because Demis,
[39:29] SPEAKER_00: on the one hand,
[39:29] SPEAKER_00: excuse me,
[39:32] SPEAKER_00: he's a good person,
[39:33] SPEAKER_00: as you've been saying.
[39:35] SPEAKER_00: He sincerely does want the systems to be safe.
[39:38] And he had various theories from the time he founded DeepMind
[39:40] as to how he would
[39:41] SPEAKER_00: try to make them safe.
[39:42] SPEAKER_00: One was to have that safety oversight committee.
[39:45] SPEAKER_00: If that didn't work,
[39:46] he would negotiate with Google about having a new safety
[39:48] oversight committee.
[39:49] SPEAKER_00: He had a set of principles on AI usage at one point.
[39:53] SPEAKER_00: And none of these things really worked because of the race dynamic.
[39:59] SPEAKER_00: You could make one lab safe,
[40:02] SPEAKER_00: but if there are four other labs that are not safe,
[40:04] you haven't really advanced the ball in terms of how safe
[40:08] society might be.
[40:10] SPEAKER_00: And so, you know,
[40:11] This is the dilemma,
[40:12] and this is something that Demis would talk to me a lot.
[40:14] SPEAKER_00: You know,
[40:15] SPEAKER_00: the race dynamic,
[40:16] SPEAKER_00: that language came up almost in our first conversation,
[40:20] SPEAKER_00: right after ChatGPG went viral.
[40:23] SPEAKER_00: And, you know, he looked worried.
[40:24] SPEAKER_00: He would say, you know, oh, they know this race dynamic's got worse.
[40:28] SPEAKER_00: I don't know how we're going to control this.
[40:30] And it didn't change throughout the three years in which we
[40:32] would have those
[40:33] SPEAKER_00: meetings upstairs in the pub.
[40:35] And I think the way you get out of this race dynamic,
[40:38] there's only really one thing.
[40:40] And that is that governments have to enforce restraint on all
[40:45] the labs at once.
[40:47] SPEAKER_00: Because as I say,
[40:48] SPEAKER_00: you have to have everybody observing the restraint,
[40:50] SPEAKER_00: otherwise there's no point.
[40:53] And this means essentially you need both the US and China to do
[40:56] it together because
[40:57] SPEAKER_00: China has a very big,
[40:59] SPEAKER_00: excellent and independent ecosystem from the US.
[41:04] It's able to distill models from American models,
[41:07] meaning that it can basically reverse engineer
[41:10] SPEAKER_00: any new model that comes out of the US within a few months.
[41:14] SPEAKER_00: And so whether we like it or not, China is a player.
[41:19] SPEAKER_00: Now,
[41:19] I think if China and the US were able to agree on some sort of
[41:24] safety standards,
[41:26] SPEAKER_00: then these could be enforced on all the other labs.
[41:28] SPEAKER_00: Because basically,
[41:29] SPEAKER_00: if you're Mistral in France or Coherent Canada,
[41:33] SPEAKER_00: or even if you're one of the labs in the Middle East or somewhere,
[41:37] SPEAKER_00: you probably want to train on American clouds,
[41:40] SPEAKER_00: with American chips,
[41:41] SPEAKER_00: with funding maybe from the US,
[41:45] SPEAKER_00: maybe selling to US users of your AI.
[41:49] SPEAKER_00: So you have these touch points with the US,
[41:52] SPEAKER_00: so you're going to be forced to respect US regulatory restraints.
[41:57] SPEAKER_00: And then China would take care of the other half of the ecosystem.
[42:00] SPEAKER_00: And I think that's the vision.
[42:03] SPEAKER_00: Getting there,
[42:04] SPEAKER_00: with the current state of Sino-American relations is very tough.
[42:08] SPEAKER_01: Right.
[42:09] SPEAKER_00: But we just have to remember that in 1962,
[42:11] SPEAKER_00: we had the deepest,
[42:12] SPEAKER_00: darkest,
[42:13] SPEAKER_00: scariest point of the Cold War with the Cuban Missile Crisis.
[42:17] SPEAKER_01: Right.
[42:18] SPEAKER_00: And then in 1968, we had the Nuclear Non-Proliferation Treaty.
[42:21] SPEAKER_00: So Cold Wars go in phases, and you do get detente at some points.
[42:27] SPEAKER_00: And I hope that happens with AI.
[42:28] SPEAKER_00: Right.
[42:29] SPEAKER_01: I sure hope so,
[42:30] SPEAKER_01: because this is,
[42:32] SPEAKER_01: you know,
[42:32] SPEAKER_01: when you start thinking about what could go wrong and that we,
[42:36] SPEAKER_01: the U.S.,
[42:37] SPEAKER_01: there's basically no government kicking in.
[42:42] SPEAKER_01: And as you say, we've got the China arc to supremacy.
[42:48] SPEAKER_01: So whatever wars are going on between these companies here is,
[42:53] SPEAKER_01: of course,
[42:53] SPEAKER_01: separate from the U.S.
[42:55] SPEAKER_01: versus China side.
[42:57] SPEAKER_01: competition, which also is equally fierce.
[43:02] SPEAKER_01: Now,
[43:02] SPEAKER_01: we have over 500 folks on board still,
[43:05] SPEAKER_01: and if anybody has questions,
[43:06] SPEAKER_01: you can put them in the message and we'll try to answer them.
[43:10] I can't do justice to this book in a 40-some minute discussion
[43:13] because I view it as
[43:16] SPEAKER_01: extraordinary.
[43:17] SPEAKER_01: I mean,
[43:18] SPEAKER_01: if there's a masterpiece about AI that's been written,
[43:20] SPEAKER_01: this is it because it's not just about demos,
[43:23] SPEAKER_01: it's not about deep mind.
[43:25] It's about the whole 15-year-plus,
[43:28] including the background,
[43:30] and what is superintelligence?
[43:32] SPEAKER_01: What are we aspiring for?
[43:34] SPEAKER_01: What are the risks?
[43:36] SPEAKER_01: It is all-encompassing, and it uses a protagonist quite
[43:40] an extraordinary one to get there and a company which was
[43:44] hyper-prescient.
[43:45] SPEAKER_01: But it isn't just about these people.
[43:47] SPEAKER_01: It's about the whole...
[43:48] SPEAKER_01: Now,
[43:49] SPEAKER_01: some people say,
[43:49] SPEAKER_01: oh,
[43:49] SPEAKER_01: well,
[43:49] they didn't give enough justice to Sam Altman or Ilya or Jan
[43:55] Lacoon or whoever.
[43:56] But the point is that you have to start somewhere and you can't
[43:59] really have symmetry of all.
[44:02] The reason I think this book is so exceptional is because if
[44:07] you want to get educated about AI,
[44:09] and how it grew up,
[44:12] and how it was in the winter,
[44:14] how I got to deep learning.
[44:15] SPEAKER_01: We get forthright learning, generative AI, and all the ins and outs.
[44:19] SPEAKER_01: I mean, this is it.
[44:20] So I think you've got a tough act to follow for others who will
[44:24] write a book about
[44:25] SPEAKER_01: AI,
[44:26] SPEAKER_01: Sebastian.
[44:28] SPEAKER_01: I haven't seen any messages.
[44:30] SPEAKER_01: Is there anything that I've missed that you wanted to add on?
[44:34] You know,
[44:34] one thing which I found interesting,
[44:36] and it gets to your bringing up of the safety,
[44:40] SPEAKER_00: question,
[44:41] SPEAKER_00: is the way that,
[44:42] SPEAKER_00: you know,
[44:42] SPEAKER_00: going in,
[44:42] SPEAKER_00: I thought,
[44:43] SPEAKER_00: okay,
[44:43] so I'd like to draw the comparison between the Manhattan
[44:48] Project and some of these
[44:50] SPEAKER_00: AI labs.
[44:51] SPEAKER_00: You know,
[44:51] SPEAKER_00: there's this,
[44:52] Robert Oppenheim obviously did an amazing scientific feat with
[44:56] the Manhattan
[44:56] SPEAKER_00: Project,
[44:56] SPEAKER_00: but he also regretted the consequences of his own work.
[45:01] And that sense of regret struck me as,
[45:04] you know,
[45:05] the scientists in society,
[45:06] the dilemma of
[45:08] SPEAKER_00: creating new technologies which could be dangerous.
[45:11] SPEAKER_00: Right.
[45:12] SPEAKER_00: Just struck me as super interesting.
[45:13] But I was a bit nervous about bringing it up with Demis or with
[45:17] other people I was
[45:17] SPEAKER_00: talking to because I thought,
[45:18] SPEAKER_00: well,
[45:19] SPEAKER_00: maybe they don't want to be reminded that,
[45:20] SPEAKER_00: you know,
[45:20] SPEAKER_00: their hand is on the modern version of the nuclear material.
[45:22] SPEAKER_00: You know, that's a scary thing to have in your power.
[45:26] But it turns out you don't have to raise it with people in this
[45:30] field because they
[45:31] SPEAKER_00: raise it for you.
[45:32] SPEAKER_00: They're all thinking about it.
[45:34] SPEAKER_00: Starting with, you know, Jeff Hinton, who,
[45:37] SPEAKER_00: essentially quotes Robert Oppenheimer when he says, why am I doing this?
[45:41] SPEAKER_00: It's for the sweetness of discovery.
[45:44] SPEAKER_00: Scientists can't resist the sweetness of discovery.
[45:47] SPEAKER_00: Moving on to Sam Altman,
[45:48] SPEAKER_00: who will volunteer to interview us,
[45:51] SPEAKER_00: that he's got the same birthday as Robert Oppenheimer.
[45:53] SPEAKER_00: Isn't that interesting?
[45:55] SPEAKER_00: And then with Demis,
[45:57] SPEAKER_00: one day I said to him,
[45:59] SPEAKER_00: so Demis,
[45:59] let's go back to the moment in 2010 when you'd finally raised
[46:02] the money for
[46:03] SPEAKER_00: DeepMind.
[46:03] SPEAKER_00: You had your first office in London.
[46:05] SPEAKER_00: It was in Russell Square.
[46:06] SPEAKER_00: not far from University College London where you did your PhD.
[46:09] SPEAKER_00: Now, how did that feel?
[46:11] And I know from experience as a writer that normally when you
[46:15] say this to people,
[46:15] SPEAKER_00: you're asking them to reconstruct the emotion of 15 years ago.
[46:21] SPEAKER_00: The answer you get is disappointing.
[46:23] SPEAKER_00: They go, yeah, it was cool or something like that.
[46:26] SPEAKER_00: They've got no emotional recall of what it was like.
[46:30] SPEAKER_00: But with Demis, he's so articulate and fizzy
[46:33] SPEAKER_00: But he said, yes, it was amazing.
[46:34] SPEAKER_00: So I was,
[46:34] SPEAKER_00: you know,
[46:35] SPEAKER_00: we had this attic office,
[46:36] SPEAKER_00: you see,
[46:37] SPEAKER_00: it was low,
[46:38] SPEAKER_00: low ceilings,
[46:38] SPEAKER_00: a little bit stuffy.
[46:40] SPEAKER_00: So I would go out sometimes to get some,
[46:41] SPEAKER_00: some air and,
[46:42] SPEAKER_00: and,
[46:43] SPEAKER_00: you know,
[46:43] I'd come down the stairs because of course there's no elevators
[46:45] and old buildings.
[46:46] SPEAKER_00: I go down the stairs, ding, ding, ding, ding, ding.
[46:48] And then outside in front of me,
[46:49] there would be the trees of Russell Square.
[46:51] SPEAKER_00: And just to the right, Sebastian, you know what there is?
[46:53] SPEAKER_00: There is the London Mathematical Society.
[46:55] SPEAKER_00: Of course,
[46:55] that's where Turin used to deliver his lectures about the
[46:58] origins of computer
[47:00] SPEAKER_00: science.
[47:00] SPEAKER_00: So now we're completing that.
[47:01] SPEAKER_00: So that was fantastic.
[47:02] But then beyond the London Mathematical Society,
[47:05] Sebastian,
[47:05] you know what there is there?
[47:06] SPEAKER_00: There on the corner,
[47:07] SPEAKER_00: there is a pedestrian crossing where it goes black,
[47:09] SPEAKER_00: white,
[47:10] SPEAKER_00: black,
[47:10] SPEAKER_00: white,
[47:10] SPEAKER_00: black,
[47:10] SPEAKER_00: white,
[47:10] SPEAKER_00: black,
[47:11] SPEAKER_00: white as you cross the road.
[47:12] SPEAKER_00: And do you know who crossed the road in the 1930s?
[47:15] SPEAKER_00: It was the Hungarian nuclear physicist,
[47:17] SPEAKER_00: Szilard,
[47:17] SPEAKER_00: who had the idea for the nuclear chain reaction as he crossed that road,
[47:21] SPEAKER_00: which then gave rise to the Manhattan Project,
[47:23] SPEAKER_00: Sebastian.
[47:24] SPEAKER_00: And now we are the modern version of the Manhattan Project.
[47:27] SPEAKER_00: Isn't that amazing?
[47:30] So that just gives you a taste of what it's like to speak to
[47:32] Demi Sosabis for more
[47:33] SPEAKER_00: than 30 hours.
[47:34] SPEAKER_00: I mean, it's just wild.
[47:36] SPEAKER_00: And he brought up that parallel approach
[47:41] SPEAKER_00: with, with nuclear technology, which I found so, not amusing, amazing.
[47:48] SPEAKER_00: It is.
[47:48] SPEAKER_00: And there are a lot of parallels and, um,
[47:51] SPEAKER_01: Your quote,
[47:52] SPEAKER_01: it's in the book,
[47:52] SPEAKER_01: of course,
[47:53] from Jeff Hinton and the sweetness of discovery and what you
[47:56] just went through with
[47:57] SPEAKER_01: Demis.
[47:58] SPEAKER_01: So we got a couple of questions.
[48:00] SPEAKER_01: One asked about Mustafa Silliman's book, The Coming Wave.
[48:04] SPEAKER_01: Is that a good book to anchor in?
[48:07] SPEAKER_01: And I reviewed that book.
[48:09] SPEAKER_01: I thought Mustafa did a nice job of outlining the pluses and minuses.
[48:13] SPEAKER_01: It was published not quite a couple of years ago.
[48:16] SPEAKER_01: I don't know if you have read it, Sebastian.
[48:20] SPEAKER_00: Yeah, I have.
[48:20] SPEAKER_00: It might even have been two or three years ago.
[48:23] SPEAKER_00: Yes, I did read it.
[48:25] SPEAKER_00: And,
[48:26] SPEAKER_00: you know,
[48:26] I thought it was a very good attempt to lay out some thinking
[48:30] about both the upside
[48:32] SPEAKER_00: and the downside of AI.
[48:35] It's a tough one to do in the sense that you're trying to
[48:38] project forwards and
[48:40] develop policy ideas about a technology that is still changing
[48:45] all the time.
[48:46] SPEAKER_00: We don't know where it's landing.
[48:48] SPEAKER_00: And so...
[48:49] I thought perhaps almost the essay that Mustafa also wrote in
[48:54] Foreign Affairs was
[48:57] SPEAKER_00: better because it's more succinct.
[49:03] It is tough to write a policy book about policies relating to
[49:06] technologies that
[49:07] SPEAKER_00: haven't been quite built yet.
[49:09] SPEAKER_01: I'm completely concordant with your points there.
[49:13] SPEAKER_01: The other question that's come in was
[49:16] are the large language models of today you can also add the
[49:20] large reasoning models
[49:22] um is that the path to artificial super intelligence or is
[49:26] something else going to
[49:28] have to happen and i guess that also depends on how you define
[49:32] agi because some
[49:34] people will say we're there now some people say we still have
[49:37] years to go um
[49:39] SPEAKER_01: Do you see a gap?
[49:41] Do you see something else that's missing here before we go to a
[49:45] level that will
[49:46] SPEAKER_01: transcend where we are today?
[49:49] My sense from hanging around all the scientists at DeepMind and
[49:55] in the rival labs
[49:56] and talking to them for three years is that there may well be
[50:00] new innovations like
[50:02] SPEAKER_00: the Transformer paper which need to happen before we get to superhuman
[50:07] SPEAKER_00: intelligence.
[50:09] But these are going to build on top of the large language model
[50:13] architecture,
[50:15] SPEAKER_00: the transformer architecture.
[50:16] I mean,
[50:17] that is just such a useful,
[50:18] flexible,
[50:20] you know,
[50:21] scalable architecture.
[50:23] SPEAKER_00: But I,
[50:24] SPEAKER_00: you know,
[50:24] SPEAKER_00: I think it's,
[50:26] SPEAKER_00: you know,
[50:26] SPEAKER_00: there's a sort of minority critique of the industry,
[50:32] SPEAKER_00: which says,
[50:33] SPEAKER_00: oh,
[50:33] SPEAKER_00: you know,
[50:34] SPEAKER_00: industry is spending insane amounts on computing power.
[50:37] and building all this electricity generation and inflicting
[50:41] brownouts on nearby cities.
[50:43] SPEAKER_00: And it's disgusting.
[50:44] SPEAKER_00: And really, you know, we don't need all that massive scaling.
[50:48] SPEAKER_00: I think that's wrong for a couple of reasons.
[50:52] SPEAKER_00: First is the point about what an infinity machine is.
[50:55] An infinity machine is something that has nearly an infinity of
[50:58] data because you
[50:59] SPEAKER_00: can't do good induction without lots of examples.
[51:02] SPEAKER_00: So just fundamentally, you need a lot of data.
[51:05] SPEAKER_00: And the more, the better.
[51:06] And then secondly,
[51:08] obviously the private labs don't love spending $100 billion on
[51:12] their R&D.
[51:14] SPEAKER_00: If they could spend $10 billion, they'd be very happy.
[51:16] So they have every incentive to be as efficient as possible in
[51:19] how they build this
[51:21] SPEAKER_00: technology out.
[51:21] They're spending all this money because they really think
[51:24] there's going to be a
[51:25] SPEAKER_00: return because they think it's necessary.
[51:27] SPEAKER_00: So I think that there isn't an alternative path,
[51:32] SPEAKER_00: but there might be an extension of the current path
[51:35] SPEAKER_00: in order to get to some human intelligence.
[51:38] SPEAKER_01: Yeah, no, I think that's...
[51:40] That's great because I think we have some features of it now
[51:43] where we've seen
[51:44] SPEAKER_01: superhuman capabilities of certain tasks.
[51:47] SPEAKER_01: And obviously for some to conclude we're there,
[51:49] SPEAKER_01: it would have to be much broader than where we are today.
[51:53] SPEAKER_01: Well, this has been a fascinating discussion.
[51:57] SPEAKER_01: Anybody interested in AI, this is the book you should be reading.
[52:01] SPEAKER_01: I think it's going to hold up.
[52:03] SPEAKER_01: I don't think it's one that is going to just
[52:06] SPEAKER_01: because of new things happening in the field,
[52:09] because you get at it in such an insightful way that it's not
[52:14] time-limited,
[52:16] SPEAKER_01: which is very difficult,
[52:18] SPEAKER_01: because you had to get this to the printer months ago,
[52:22] SPEAKER_01: and the field keeps coming up with new models,
[52:24] SPEAKER_01: but that didn't change anything.
[52:26] SPEAKER_01: That's what was, I think, particularly
[52:31] SPEAKER_01: surprising.
[52:32] SPEAKER_01: I did an interview with Al Gore yesterday at an AI conference here,
[52:36] SPEAKER_01: actually in San Francisco,
[52:38] SPEAKER_01: and he had known about your book,
[52:39] SPEAKER_01: but he didn't know it without.
[52:41] And I said to him,
[52:42] that's one of the things that's really remarkable is here it is
[52:47] SPEAKER_01: in April 2026.
[52:50] And it doesn't matter that there's been some new models since
[52:53] we went into print
[52:54] SPEAKER_01: because it's all there.
[52:56] SPEAKER_01: So congratulations, Sebastian.
[52:58] SPEAKER_01: It's really a joy to have this discussion with you.
[53:01] SPEAKER_00: It's been wonderful.
[53:02] SPEAKER_00: Thank you so much, Eric.
[53:03] SPEAKER_00: I've really enjoyed it.
[53:05] SPEAKER_01: All right.
[53:06] SPEAKER_01: Take care.
[53:06] SPEAKER_00: Take care.
[53:07] SPEAKER_00: All right.
[53:08] SPEAKER_02: Bye-bye.
[53:10] SPEAKER_02: Bye-bye.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.