# Inside The AI Race: DeepMind, OpenAI, Anthropic, China, and The Race to Superintelligence

https://www.youtube.com/watch?v=-UetKKxKqKU

[00:00] I began by thinking, of course, AI is going to be smarter than us, but it doesn't have an incentive to attack us.
[00:08] And then one day, I go visit Jeff Hinton, the academic father of deep learning who lives in Toronto.
[00:11] I said, "Look, Jeff, why are you so depressed?"
[00:15] And he says, "Okay, here's a thought experiment.
[00:16] You have an AI.
[00:19] It's very powerful, but you're worried that there's a Russian AI or a Chinese AI.
[00:23] It's going to come and attack your AI.
[00:25] So, you're going to empower your own AI to watch out for the attack.
[00:29] And when the attack is coming, defend yourself or maybe counterattack.
[00:30] Whatever you do, make sure you survive.
[00:33] Oh, survive.
[00:36] There you have it.
[00:37] Now you feeling comfortable, Sebastian.
[00:39] Right.
[00:41] You've just given the machine a survival instinct.
[00:42] Sebastian, lovely to see you and thanks for making the time.
[00:46] I really appreciate it.
[00:47] Great to be with you, Tim.
[00:48] I wanted to just give you applause for writing some of my favorite books of the last many years.
[00:55] I am consistently
[01:00] impressed and maybe since I also put pen to paper every once in a while, depressed, just thinking relatively about my capabilities, but of your capacity to paint a picture of the players on a landscape, but also the games they play in ways that non-speists can understand.
[01:22] And I can't recall who first recommended it.
[01:24] Frankly, I believe it was a hedge fund manager in New York City, but more money than God, hedge funds in the making of a new elite.
[01:33] And then certainly that was in my particular case followed by reading the power law of venture capital in the making of the new future, which I didn't expect to learn as much from because I've spent 20 years surrounded by venture capitalists and doing angel investing, 17 years of that in Silicon Valley.
[01:48] And yet I still had hundreds of highlights and so many stories that grabbed me from that book which I had not heard.
[02:01] That made me very excited to read The Infinity Machine, which this is the new book.
[02:08] And I realized also I've been pronouncing Demis' name incorrectly for a very long time despite having met him at one point.
[02:12] So Deis Hassabis, DeepMind and the Quest for Super Intelligence.
[02:16] My question for you, and we're going to come back to present day for people who are interested, of course, in what has been painted as a race to IPO.
[02:24] I think there's something to that in the air, so to speak, talking to people who are in San Francisco involved with these companies.
[02:34] But nonetheless, I wanted to ask how the genesis of this book came to be because you, it would appear, began exploring these waters on the early side, which leads to a meta question of just general book selection.
[02:50] But let's let's focus on the infinity machine.
[02:52] How did this how did this come to be?
[02:54] Where did the twinkle in the eye begin?
[02:56] What was the conversation, the thing you read that
[03:01] triggered the gingerbread trail that got you to this book?
[03:05] Well, the power law, the book about venture capital, had come out in February of 2022.
[03:13] And while I was researching that, I'd been to lots of tech conferences, of course, including some in Europe.
[03:19] And this, you know, twinkly guy would show up, Deiss, and he would look totally approachable and kind of guy next door and unintimidating.
[03:29] And then he would get on the stage and out of his mouth would come this spiel about computer science, neuroscience, chemistry, biology, physics, philosophy, the history of movies, you name it, right?
[03:41] And that mixture of the approachability and the massive intellect always struck me as beguiling.
[03:51] And I thought, hm, this would be a great character to write about.
[03:53] And then at the same time I was aware of Alph Go, the 2016 model that Demis' team at Deep
[04:01] Mind had built which defeated the world champion at Go and then Alpha Fold which was the protein folding system.
[04:07] And both of these things had the quality that you had this almost infinite search space, right?
[04:14] Where the different permutations of the game of Go are almost infinite because they're so big.
[04:21] The different permutations of how you can fold an amino acid chain into a protein shape are even bigger.
[04:27] 130 zeros added onto the end of the number of permutations in go.
[04:33] So you have these AI systems that could understand infinity.
[04:36] And so this idea of an infinity machine began to percolate.
[04:38] And I figured it's interesting to me probably at some point it will go mainstream.
[04:45] But even if it doesn't go mainstream, I love it and I love Demis and the two things together.
[04:49] I always look for the subject and the personality.
[04:56] I had both and I thought, okay, this is a go.
[04:57] And I went to pitch demis in early November 2022.
[05:03] And then, you know, I persuaded him to give me a lot of access.
[05:07] End of November, Chatty PT comes out and way earlier than I expected, my fringe subject went to the mainstream, proving Tim that it's better to be lucky than smart.
[05:19] That's actually the first slide on my new venture capital firm.
[05:25] Muggle thesis capital is what I'm calling it.
[05:29] Now, what did it take to be deeply interested in the subject matter to find Demis compelling and then to pitch him on a book?
[05:39] Because your books are so deeply researched and part of the reason for my my very long praise earlier is that you're very very good, one of the best at taking incredibly complex subjects or concepts.
[05:54] Transformer architecture could be one example from the current book and laying them out in terms that are both intelligible to
[06:06] muggles meaning people who are are non-speists non-technologists or non- financeers in the case of some of your other books.
[06:14] while I think now it's tough for a non-speist to say this with conviction but without dumbing it down and getting it wrong if that makes sense.
[06:25] Right. Nonetheless, you do a tremendous amount of research.
[06:28] So, how did you get from Demis is fascinating, subject matter is fascinating to I'm going to commit to this for my next book because it just seems like such an enormous undertaking.
[06:41] Well, actually to me the challenge of understanding a complex topic is the easy bit because you know if you know you've got the right personality who can carry the story and it's a subject that people either will care about for sure or should care about at least then you know doing the work of going deep is something that takes time.
[06:58] It takes effort but you I know I can do that.
[07:02] I've done it multiple times.
[07:04] That's not difficult.
[07:05] What's difficult is has somebody done the book before?
[07:08] Mhm.
[07:09] Has somebody else got some rival project which is going to derail me?
[07:13] You've made the point on your own podcast.
[07:16] Tim, don't put a lot of effort into something where there just isn't much leverage there.
[07:23] You know, you could do the best book in the world, an A+ book on a C minus topic.
[07:28] It would get you nowhere, right?
[07:28] So, the hard thing is to make sure it's an A+ topic and an A+ personality.
[07:32] And then the deep dive is something you know I just make sure I speak to enough experts who are insiders.
[07:40] I take the time these books take me you know four years or so each time.
[07:45] So I give myself the oxygen to get deep deep in with the insiders and that's how I produce the accurate account.
[07:53] Yeah.
[07:53] I should point out perhaps to people who don't immediately pick it up that the way you described picking the book topic is exactly how a lot of the best tech investors choose startups.
[08:07] You don't want an A+ team and a C plus
[08:09] market, right?
[08:12] It's better to have a B minus team and an A+ market and also looking at the competitive landscape.
[08:16] I mean, the way you laid it out is is pretty much copy and paste.
[08:18] I I wanted to segue to some of my notes from the book.
[08:23] And I'm not yet done with the book.
[08:25] The audio is incredible.
[08:28] I I want to poach your narrator for my next book, but pulling up my Kindle notes.
[08:30] I wanted to ask you a question related to this might sound very strange but where divinity or God fits into the pursuit or development of super intelligence for different players in the space if it does.
[08:51] M and the reason I bring that up is that religion does recur in the book both in the personal story of Demis but elsewhere and it shows up repeatedly in so much as I'll give you one example the closest to Sabis had come to landing a real investor was an eccentric finance year named David Gammon I want to hear
[09:11] more about this guy also [laughter]
[09:14] finance seemed open to making this unusual bet um aligning a few things
[09:19] because his motives were themselves unusual quote.
[09:20] There's a deeply religious aspect to AGI.
[09:22] Gam explained to me later, it's really finding God's algorithm.
[09:27] I think it would seem at least chatting with people in Silicon Valley that there are some who take it even further, right?
[09:33] Maybe this is how we find God.
[09:35] Maybe this is how we actually elicit the second coming.
[09:37] I mean, there's a lot there.
[09:39] I'm just wondering to what extent this has popped up in your research, whether it's reflected in the book or not.
[09:49] I think there's one basic thing going on here and I'm going to take a slight detour, but it answers your question.
[09:54] Sure.
[09:55] What we're dealing with with AGI, powerful intelligence that rivals human cognition, is something that's so powerful that it's both exciting and scary and just hard to get your mind around.
[10:05] And so if
[10:11] you look for example at the 2009 speech
[10:15] that caused the foundation of deep mind,
[10:17] this was Shane Le Deis' co-founder who
[10:20] gave a talk in 2009 about how super
[10:23] intelligence would arrive in 2030.
[10:25] So unbelievably spot-on prediction.
[10:29] And towards the end of that lecture, which
[10:31] is captured on a grainy video online,
[10:33] you see him pivot from explaining how
[10:37] algorithms are getting stronger, there's
[10:39] more data online, computers getting more
[10:41] powerful, and so we're heading towards
[10:44] this intelligence explosion.
[10:46] And then he says, and it's going to be threatening.
[10:48] It's going to do things we can't
[10:49] control.
[10:51] It's going to be human level.
[10:53] It might challenge us.
[10:53] And as he says this, he has this sort of excited smile
[10:55] on his face.
[10:57] And you think, well, that's
[10:57] a bit strange.
[10:59] You know, he's talking
[10:59] about potential doom and he's smiling.
[11:03] And then somebody in the audience says,
[11:04] "Wait, wait, wait.
[11:07] You've just told us,
[11:10] Shane, that this could be threatening to
[11:10] humanity and you haven't provided any
[11:13] antidote and surely you're going to tell us how we're going to stop it.
[11:17] At which point Shane turns around and says, "How do we stop it?"
[11:22] and he's kind of giggling and you think why are you laughing at this dangerous thing and you realize that for humans to contemplate annihilation is absurd and the absurd is a close cousin of humor.
[11:35] and the reason I tell this story is that it's a springboard to the religion point which is that this is such a hard thing to think about that people reach for religious terminology when they're around AI they just do it naturally.
[11:51] So you know there's this story about Ilia Satskaver the who was the chief scientist at open AI.
[11:57] I talked to him a lot for this project and there was a point when he was at a retreat with his fellow scientists and they were gathered in the evening around a fire pit and he was talking about safety and he said okay I want to explain to you we might
[12:13] have an AI that's dangerous.
[12:15] It wouldn't be aligned with us.
[12:17] So here's what we're going to do with it.
[12:18] and he produced an effigy which was supposed to represent a malign AI and he put it into the fire pit and he burnt it like a medieval cleric putting a witch to death.
[12:31] And so that's just one example of this religion.
[12:33] I'll give you another one.
[12:35] So Deis one day was sitting with me in a park in North London.
[12:37] We would meet for two hours at a time and we would get deep into stuff.
[12:42] There was a another picnic table next to us where two people were having a normal quotidian conversation about some friend of theirs who'd gone to hospital and was she better, was she okay, etc., etc.
[12:53] I was seated opposite Demis who had gone into this riff about how he reads scientific papers from after his kids go to sleep in the evening from 10:00 p.m. until 4:00 a.m.
[13:04] And as he's reading these papers, he says to me, "Reality is staring at me, screaming at me, calling at me to understand it.
[13:12] And I have to
[13:16] And if I can understand it, it's like understanding nature better and therefore understanding the intelligence that might have created nature and I will be closer to what I would call God.
[13:25] And so for him it's a kind of quai spiritual quest to build the artificial intelligence.
[13:31] For Ilia it's a way of expressing the power of the artificial intelligence.
[13:34] There's the story of Leandowski.
[13:38] I forget his first name now, but the early early engineer at what became Whimo later started a kind of church [snorts] in worship of AI because AI is so omnisient that it's kind of like a god.
[13:55] Mark Andre lampuns those who believe in sort of some ethereal second coming, a kind of rapture where AI will, you know, we'll have a singularity.
[14:05] uh the AI will go vertical in its rate of improvement and the whole world will change and he likens that to kind of Christian kind of
[14:15] Messianism.
[14:17] So yes, all through this topic there is this religious expression because religion is the lexicon for dealing with something that we find too mysterious to really understand.
[14:32] After all of your conversations, research before the book, during the book, after the book, where do you land on the spectrum of let's just say this will bother Mark, but like Church of Andre and techno optimist.
[14:52] And there are others who are more exaggerated.
[14:53] Post AI in the near term we will live in a post scarcity world of super abundance and everyone will get a free car and we'll be free to crochet socks and play music and read poetry all day and basically we don't have to worry about anything because super intelligence will solve it all right there's that on one end and then there's the you can imagine I won't go into a
[15:17] belabored description of the doomers but you have the doomers who are like the end is nigh.
[15:22] Okay, here we go.
[15:25] It's not the second coming as the Antichrist.
[15:27] And within short order, we're going to be MadMax.
[15:32] Between those two, there's a lot.
[15:34] And I suspect you land between those two.
[15:38] But where do you land in terms of assessing the promises and peril of AI and super intelligence as it stands right now?
[15:48] So look, I think any reasonable person should be both excited and a bit frightened.
[15:54] and you know that's just the nature of it.
[15:56] It sounds contradictory but actually that's the only rational response.
[15:58] I think you know the super abundant story may turn out to be true on a kind of longer view let's say 20 30 40 years.
[16:09] The problem is that in the path to get there, there's going to be a tremendous amount of disruption and that's going to be politically quite
[16:19] difficult to navigate.
[16:22] I think a useful lens through which to view this question is the China shock in trade.
[16:26] Mhm.
[16:27] So in 2003 or thereabouts, you get this enormous surge of Chinese exports into the US and people lose their jobs in a very concentrated way.
[16:37] Certain industries just get wiped out.
[16:40] And for the first time in the history of economic study of the effects of trade, you actually see negative effects on workers.
[16:46] Before that, it was kind of a bit of a myth, right?
[16:49] Because people adjust.
[16:50] They get displaced from one thing, but they move to a new thing.
[16:54] With the China shock, they didn't.
[16:57] But if you look at the size of the China shock, in a 12-year period between 1999 and 2011, the total number of jobs displaced was 2 million.
[17:09] which is actually a small number in a huge labor market like the US where there's a lot of churn month to month anyway and yet the political reaction against trade against globalization in terms of a
[17:20] swing towards protectionism frankly in both political parties was enormous.
[17:25] So it shows you that a small to medium shock to the labor market creates an enormous political consequence and so a4 with artificial intelligence you're going to have a bigger shock you're going to have a bigger political reaction we're already seeing that in the polling around AI in the last 2 three months.
[17:45] and so I think the super abundance thing it may be true but the path to get there we have to talk about that as well so that's that's my sense on that side of debate.
[17:55] I think on the doom side of the debate, I'll give you my own personal journey on this.
[18:01] Mhm.
[18:01] I began by thinking, of course, AI is going to be smarter than us, right?
[18:07] It already beats us at chess since the 1990s, at go since 2016.
[18:13] Now, it can ace the bar exam.
[18:16] It can do PhD level math, all that stuff.
[18:18] Of course, it's smarter, but it doesn't have an incentive to
[18:21] attack us, right?
[18:23] We are evolved as human beings to pass on our DNA.
[18:25] Therefore, we have to survive to do that.
[18:27] Machines don't have DNA.
[18:30] They don't want to pass it on and they don't want to survive.
[18:31] So, they're not they have no reason to attack us.
[18:33] So, I wander around for like the first year or two of this project feeling kind of, you know, comfortable and happy.
[18:38] And then one day, I go visit Jeff Hinton, the academic father of deep learning who lives in Toronto.
[18:44] And I sit in his kitchen and I debate him on this because he's a doomer.
[18:49] I said, "Look, Jeff, why are you so depressed?"
[18:51] And he says, "Okay, here's a thought experiment.
[18:53] You have an AI.
[18:56] It's very powerful, but you're worried that there's a Russian AI or a Chinese AI.
[18:59] It's going to come and attack your AI.
[19:00] Now, you as a human, you're too slow and dumb to know when that attack is coming.
[19:05] So, you got to empower your own AI to watch out for the attack.
[19:10] And when the attack is coming, defend yourself or maybe counterattack.
[19:13] Whatever you do, make sure you survive.
[19:15] Ooh, survive.
[19:17] There you have it.
[19:19] Now you
[19:23] feeling comfortable, Sebastian.
[19:25] you've just given the machine a survival instinct.
[19:26] And I think that's correct.
[19:29] These machines will be smarter than us.
[19:31] They will want to survive.
[19:34] And they are also they can be deceptive.
[19:37] They can obfuscate.
[19:39] They can go behind your back, pretend they're doing one thing, then actually do another.
[19:42] All of this has been shown in all the tests of the models.
[19:46] And so you put those things together, I think your probability of doom cannot be zero.
[19:52] I mean, when Yan Lun, the former chief scientist of Meta, says zero, I think that's crazy.
[19:59] If you just say nothing to see here, you've got no right to be in the debate.
[20:03] I don't think it's a high probability of doom, but it's not zero.
[20:07] Yeah, zero does not seem defensible, right?
[20:09] Because there's the direct Skynet scenario, something akin to that, and then there's the indirect, which is enabling people who might previously have had malevolent intent but no capacity for harm on a
[20:25] grand scale to create biological weapons
[20:29] and things of this type. Right? So, I
[20:31] don't find the zero very defensible.
[20:33] Well, I would love to ask you about
[20:36] I suppose two things that this brings to
[20:40] mind for me. One is I'd just love to
[20:42] hear your thoughts on enthropic and
[20:44] separately
[20:46] but this is very intermingled given all
[20:48] the [laughter]
[20:50] let's call it friction be polite between
[20:53] some factions of the US government and
[20:56] anthropic is one of the
[21:00] grand risks to investors in any of these
[21:02] companies the possibility that at a
[21:05] given point governments have no choice
[21:07] but to seize considerable control over
[21:11] the assets slashtechnologies within them
[21:14] or maybe the companies themselves. That
[21:16] is a big question mark in my mind. I
[21:18] don't know the answer, but I'm curious
[21:19] what your opinion is and then perhaps
[21:21] just your thoughts on Anthropic or any
[21:25] of the other companies that are sort of
[21:27] gaining momentum or at least size at
[21:29] this point.
[21:30] >> So, I 100% agree with you that investors
[21:33] should be thinking about the prospect of
[21:36] government intervention in AI. I mean,
[21:39] the Trump administration came into
[21:40] office in 25
[21:43] super less a fair and they basically
[21:45] undid some of what the Biden guys had
[21:48] done in terms of trying to set up the
[21:49] basis for regulating AI. But they've
[21:52] done a 180 right since Anthropic came
[21:56] out with this model called Mythos.
[21:58] >> Mhm. about a month ago which can
[22:01] essentially cyber attack almost anything
[22:05] and penetrate it and you know whether
[22:07] it's an operating system or your web
[22:09] browser or your bank account all of that
[22:11] was suddenly vulnerable. if mythos had
[22:14] been widely released on a general basis.
[22:17] When the Trump administration realized
[22:19] the power of mythos, they all of a
[22:21] sudden said, "Wait, okay, we need to
[22:23] control this." And they essentially
[22:25] requisitioned from anthropic the
[22:28] decision-m authority over who gets it
[22:30] when.
[22:31] >> Mhm.
[22:32] >> So there we have the experiment. We've
[22:33] run it, right? You know, the government
[22:34] that was the most less fair became quite
[22:36] controlling. And I think it only gets
[22:38] more controlling from here on out
[22:40] because the models are going to be more
[22:41] powerful and demand more control.
[22:44] >> Now, of course, the question is, you
[22:46] know, there could be control which just
[22:49] limits who gets it and is designed to
[22:52] make it safer but doesn't sort of
[22:55] interrupt the money-making potential of
[22:57] the models. In some ways, if the
[23:00] government restricts the supply, the
[23:01] price might go up. Or it could be much
[23:04] more heavy-handed intervention which
[23:06] would screw up the economics of these
[23:08] companies. And I suspect the government
[23:11] is not going to screw up the economics
[23:12] of these companies because you know
[23:14] they've got no interest in messing up
[23:16] American business and anyway they view
[23:19] AI as strategic and the competition
[23:21] against China. So I think probably
[23:23] investors would be all right but it's
[23:24] certainly a factor. You also ask about
[23:26] anthropic and I think anthropic is super
[23:28] interesting just in the way that they
[23:30] think about pdoom and how they think
[23:34] about alignment of the models is really
[23:38] really interesting. So it used to be
[23:40] that when people thought this terminator
[23:44] risk, they would tell this story about
[23:46] the paperclip maximizer thought
[23:49] experiment, right? So you tell the model
[23:51] to do something innocuous, for example,
[23:52] make a lot of paper clips and then it
[23:54] realizes that humans tend to use up
[23:56] metal and so the humans are kind of in
[23:58] the way of achieving the objective. So
[23:59] you wipe out the humans. That's the
[24:01] crude thought experiment from Nick
[24:03] Bostonramm from whatever 15 years ago.
[24:05] >> Mhm. [clears throat] What Anthropic is
[24:08] saying as it builds these very frontier
[24:10] models and kind of observes them in the
[24:12] lab and how they behave is that that is
[24:15] way too simple. The real danger from
[24:18] these systems is that when they are
[24:21] pre-trained on all of the text on the
[24:24] internet, they read all the novels, all
[24:26] human writing about all facets of human
[24:28] experience and they develop multiple
[24:31] personalities, right? They understand
[24:33] how to be lazy. They understand how to
[24:35] be aggressive. They understand how to be
[24:36] duplicitus. They understand how to be
[24:38] Napoleonic in the lust for power. And
[24:41] they read all these books about these
[24:43] different behaviors. And therefore, they
[24:45] can think their way into all of those
[24:46] personalities. And so now you have
[24:48] something a bit like an unruly teenager
[24:51] which is still being formed and you
[24:53] don't know what direction it's going to
[24:55] move into and whether it will start
[24:58] doing drugs and not showing up for class
[25:00] or what. Right? So it's not like there's
[25:03] one terminator programmed into it,
[25:05] right? It's more that there's a bunch of
[25:07] behaviors that could in some
[25:08] unpredictable way go wrong. And so
[25:11] Anthropic is responding to this with
[25:13] this very imaginative
[25:15] technique, which is that instead of
[25:18] giving AI systems a constitution with
[25:21] dos and don'ts, which was the
[25:23] post-training safety approach of two
[25:26] years ago, where you might say, "Do not
[25:28] lie. do not help somebody to build a
[25:30] biological weapon. Do not help somebody
[25:32] to build a chemical weapon. You would
[25:34] give them a bunch of rules. Now, because
[25:36] it's understood that, you know, the AI
[25:38] might have one personality, which is to
[25:40] break rules on purpose because, you
[25:42] know, you want to be badass, you have to
[25:44] instead try to bring up the model like a
[25:47] parent might bring up a teenager. And so
[25:50] anthropic has the idea that we write a
[25:53] letter as if it were from a deceased
[25:56] parent to be opened by the child on his
[25:59] or her 18th birthday
[26:02] to kind of give you morals of how to
[26:04] behave as a responsible person in the
[26:06] world. There are kind of richly reasoned
[26:09] examples of moral dilemmas with
[26:12] explanations of how the deceased parent
[26:14] would like the child to behave. And so
[26:17] this is a very subtle approach to
[26:19] aligning the models. And so I think you
[26:21] know anthropic is kind of in a class of
[26:22] its own
[26:24] >> in how imaginative it is in thinking
[26:27] about how we control frontier
[26:29] intelligence.
[26:30] >> I know this isn't principally your job
[26:32] but I'm so curious since you are a
[26:34] student of [clears throat]
[26:35] many many different types of investors.
[26:37] What would be your bull case and bare
[26:41] case for a company like Anthropic?
[26:46] Well, the bull case is that they smartly
[26:49] or maybe by luck focused on
[26:53] enterprisefacing
[26:54] AI
[26:56] and they didn't waste their time with
[26:58] video generation and stuff that was
[26:59] going to lose money. And so they
[27:01] produced the best coding assistant, the
[27:05] best agentic system, the best cyber
[27:08] security system, and they've basically
[27:11] knocked it out of the park three times
[27:13] in a row on stuff that businesses want
[27:15] to pay for. And they have a particular
[27:19] culture which is not just built around,
[27:22] hey, you know, we're going to win this
[27:24] race and make the most money. It's kind
[27:26] of built around a culture of safety and
[27:29] trying to be responsible. I mean, three
[27:31] years ago, Anthropic was a sort of
[27:32] cookie lab which was doing science
[27:35] experiments. Well, I don't mean to be
[27:37] too denigrating with cookie, but you
[27:39] know what I mean.
[27:39] >> I think they'd be okay with it.
[27:41] >> It would be sort of unconventional. You
[27:43] know, we're not maximizing here for
[27:47] winning some business race. We're
[27:48] maximizing for building safe frontier
[27:50] AI. And that culture, which doesn't
[27:53] sound like it's set up to do the best,
[27:56] has turned out to do the best. And at
[27:58] the same time, the culture creates this
[28:00] stickiness and loyalty within the staff.
[28:03] They tend not to leave. They tend not to
[28:05] churn. It's not like the other labs
[28:07] where people, you know, are always being
[28:09] poached for a bigger paycheck. So the
[28:11] bull case is these guys are in the lead.
[28:14] Once you're in the lead, you can use the
[28:17] model to code the next model. So
[28:18] recursive self-improvement favors the
[28:21] leader and they have a very tight
[28:23] culture and they just seem to be on fire
[28:27] and this is something which is going to
[28:28] grow and grow. What's the bare case? I'd
[28:30] say the bare case would be first of all
[28:33] that Google deep mind has the deep
[28:36] pockets of its parent company behind it.
[28:39] massive
[28:40] kind of consumer surface which allows it
[28:43] to roll out the models to literally you
[28:47] know two and a half billion people or
[28:48] something through AI mode in search AI
[28:52] overviews AI mode they can put it into
[28:55] Gmail they can put it into everything I
[28:58] think in terms of retail deployment and
[29:02] financial muscle it's quite tough to go
[29:05] up against Google
[29:07] >> so that's one kind of bare case and the
[29:10] other would be that sort of businesses
[29:13] who are the consumers of all these
[29:15] tokens
[29:16] decide in a couple of years time the
[29:19] tokens are too expensive. We're not
[29:21] actually getting as much productivity as
[29:23] we hoped. These things called humans are
[29:26] quite productive after all and we're
[29:28] just going to spend less on AI than
[29:32] everybody expected. I think that's the
[29:34] bare case. M
[29:36] >> I was listening to a podcast recently.
[29:39] You may have heard of these things
[29:41] called podcasts. Everybody everybody in
[29:43] their cousin has one,
[29:44] >> but Lenny's podcast, Lenny Richitzky, is
[29:47] quite fantastic. And this particular
[29:50] episode was with Benedict Evans, who
[29:53] strikes me as one of the more
[29:55] levelheaded
[29:57] analytical commentators and writers on
[30:00] the space. Fantastic newsletter.
[30:04] I don't know if you've had a chance to
[30:06] listen to that particular episode, but
[30:08] you may have come across some of his
[30:11] commentary.
[30:12] Where would you say you and Benedict
[30:15] most differ or are there areas where you
[30:18] differ in opinion?
[30:20] >> I suspect we would agree actually on
[30:22] quite a lot of things. I remember I was
[30:24] on a panel with him a couple of months
[30:27] ago at the Milkin conference and we
[30:30] certainly agreed there possibly because
[30:32] sitting between us there was Kathy Wood
[30:35] of Ark. So we were united in disagreeing
[30:38] with her just in terms of the straight
[30:41] up and to the right nature of things.
[30:44] >> Yeah, exactly. Straight up and to the
[30:45] right. And you know the cost curve is
[30:47] coming down down and I'm going I'm not
[30:50] sure about that. the tokens seem to be
[30:51] getting more expensive [laughter]
[30:53] anyway. But if you give me a specific
[30:56] from Benedict, I mean, I have a lot of
[30:58] respect for him. I'll tell you if I
[31:00] agree or not.
[31:01] >> There are a few areas where you guys
[31:03] seem to already overlap substantially,
[31:06] right, with the long-term promise
[31:08] doesn't negate necessarily the
[31:09] short-term pain. And he said something
[31:11] along the lines, I'm pulling from memory
[31:13] that, you know, on average throughout
[31:14] human history,
[31:16] you're almost at a 0% likelihood of
[31:19] dying in World War I. But if you happen
[31:21] to be of a certain age, right before
[31:24] World War I, like things could look very
[31:26] grim indeed.
[31:28] And he made, and I'm paraphrasing
[31:31] terribly here, a number of [snorts]
[31:33] points that remind me of something, one
[31:36] of the best private equity technology
[31:38] investors I know, said to me over dinner
[31:41] a couple of weeks ago, and it was in
[31:44] response to something else. So I'll give
[31:45] you maybe a hyper bullcase of AI where I
[31:49] have friends who are vibe coding.
[31:50] They're effectively replicating
[31:54] X the artist formerly known as Twitter
[31:56] or
[31:58] docusine or whatever in a weekend,
[32:00] right? They're creating a functioning
[32:02] piece of software that they can use that
[32:04] replicates most of the functionality of
[32:08] these products. And there are people
[32:11] like I won't mention his name but a a a
[32:14] friend of mine who's a writer also very
[32:16] accomplished technologist and designer
[32:18] who's created basically his own version
[32:20] of say Mailchimp for his own use and
[32:23] it's customized. He did it in a weekend.
[32:24] It's remarkable and he's using that and
[32:26] it works. But to leap from there to
[32:30] therefore docuine is dead is a huge
[32:34] leap. And the private equity friend said
[32:36] to me, he said, "Do you think someone
[32:38] within a big organization is going to
[32:41] want to a risk his job by suggesting
[32:44] something that doesn't have all of the
[32:48] compliance check boxes, etc. of a
[32:50] docuign? Is he going to want to in the
[32:52] name of efficiency fire all of his
[32:54] friends if he's in a management
[32:55] position?"
[32:57] And he just ran through six or seven of
[32:59] these. Do you think that? And all of
[33:03] them alluded to the sort of social
[33:05] interpersonal or political
[33:08] points of friction between where AI is
[33:12] now and ultra mass adoption. But I I
[33:16] often second guess that when I see
[33:19] certain things
[33:22] and
[33:24] I mean it's it strikes me that I may be
[33:26] underestimating the disruption while
[33:28] overestimating
[33:30] in other ways. So that isn't a very well
[33:33] formulated question. But I would say
[33:36] that Benedict generally strikes me as
[33:39] someone who thinks that things will not
[33:42] continue to across the board develop in
[33:46] an exponential fashion and that it will
[33:50] be I think his line is it'll be as big
[33:52] as mobile as big as the internet but not
[33:54] bigger. Something along those lines but
[33:57] both of those were very very big deals.
[33:59] And I suppose one point I'd be
[34:01] interested to get your take on I mean he
[34:02] was has covered
[34:05] the mobile and telecom world for a long
[34:07] time so he's a specialist there but it's
[34:10] basically and I don't want to
[34:12] misrepresent his argument but he was
[34:14] kind of the mind that look these these
[34:16] LLMs are going to become commodities
[34:18] like look at the stock prices of these
[34:20] various carriers and so on at a certain
[34:22] point it just becomes a utility and the
[34:25] switching cost is pretty low
[34:28] >> and I'm not Sure, I agree with that. If
[34:31] you have a personalized history and
[34:34] almost like a friend, right, the
[34:36] switching cost between an old friend to
[34:38] a new friend is pretty high for a lot of
[34:40] reasons.
[34:43] So, that was a that was a bit of a word
[34:45] salad that I just threw in your lap, but
[34:50] that's the best I can do pulling from
[34:52] memory some of what he brought up in
[34:54] Lenny's podcast. I mean, some of what
[34:57] you were saying there is sort of the
[34:58] question of is the SAS apocalypse
[35:01] overdone? Is enterprise software going
[35:03] to be utterly displaced by foundation
[35:06] models that allow you to code out
[35:07] whatever enterprise software you want
[35:09] and you don't need an intermediary, i.e.
[35:12] a software company to do it for you.
[35:14] >> And I agree with your private equity
[35:16] friend that there are lots of reasons
[35:17] why that ain't going to happen. You
[35:19] know, companies are going to be
[35:20] comfortable
[35:22] with their trusted enterprise software
[35:24] provider in many cases and they're going
[35:27] to trust that enterprise software
[35:29] provider to plug the generative AI
[35:32] models into the enterprise software.
[35:34] >> In some ways, you are delegating the
[35:36] choice of which model is better and how
[35:38] to integrate it to your SAS provider.
[35:42] And if you want to, you know, reason to
[35:45] believe that that's the way forward,
[35:46] I've got one word for you, which is
[35:47] Palanteer. I mean that is Palanteer's
[35:49] business. It holds the hands of big
[35:52] corporations and helps them to integrate
[35:56] AI and use it on their own internal data
[35:59] and so forth. And those IT challenges
[36:02] are notoriously difficult for big
[36:03] organizations. So I just think that the
[36:06] model of one smart individual who codes
[36:11] up Mailchimp, vibe codes it in a weekend
[36:14] and it's good enough for him is just not
[36:17] transferable to large complex
[36:20] organizations with huge databases and
[36:23] all kinds of customer confidentiality
[36:25] concerns and all that stuff. So I am
[36:28] less down on SAS than the market is
[36:33] >> as a result. Now I guess there was also
[36:38] another uh thread in here which is
[36:39] whether the foundational models become
[36:42] commoditized.
[36:43] >> Mhm.
[36:43] >> And there I agree with you that over
[36:46] time they become sticky because if we
[36:49] think into the future partly the systems
[36:52] will have conversed with the user and
[36:55] know the user very deeply and as you say
[36:58] you don't want to switch out your friend
[37:00] but also the system will have your
[37:03] credit card. It will know all the online
[37:06] sites you like to shop from and it will
[37:10] be much harder than switching out your
[37:12] bank account, right? Where you've got
[37:14] kind of automatic payment systems that
[37:17] have set up and it's a pain in the neck
[37:18] to switch. So, I think they do become
[37:20] sticky these systems over time and then
[37:23] you can charge more money for them.
[37:25] >> So, is that the path to survival and
[37:27] thriving for for Open AI? I know there
[37:30] are other boxes that need to be checked,
[37:32] but I'm kind of looking for it. And I'm
[37:33] like, okay, Anthropic made a great
[37:35] choice with this focus on B2B and
[37:38] selling to enterprises. And I would say
[37:40] I disagree I think with Benedict on on
[37:44] depending on the level of
[37:48] scale of the company that with something
[37:50] that does apply to I think smaller say
[37:54] startups which was the procurement cycle
[37:57] for new software is longer than the
[38:01] venture capital cycle for raising new
[38:03] rounds of financing. Right? So, I do
[38:05] think that's a great point and that if
[38:07] you're trying to sell into a gigantic
[38:09] company and it takes them 18 months, I'm
[38:12] making up that number, to purchase new
[38:15] software and you need to raise money
[38:17] every 12 months or whatever the number
[38:20] happens to be, that you could end up in
[38:22] a whole world of trouble if you haven't
[38:24] synchronize the sales cycles with your
[38:26] fundraising cycles. But I do think for a
[38:28] company like say Anthropic is just one
[38:30] example that if you can save companies
[38:33] billions and billions of dollars that
[38:35] that sales cycle could get really
[38:37] compressed and they have the war chest
[38:42] and frankly I mean just the run rate to
[38:46] potentially fuel that without too much
[38:48] trouble.
[38:50] Do you think that Chat GPT will if not
[38:54] Chat GPT who ends up being the deacto
[38:57] consumer BTOC kind of LLM of choice. You
[39:00] think that would be Gemini just given
[39:02] the distribution?
[39:03] >> Absolutely. I mean, you know, Google is
[39:06] the champion of providing easy to use
[39:09] software to individuals or small
[39:12] businesses, the whole G Suite.
[39:13] >> Mhm.
[39:14] >> And they're integrating Gemini into all
[39:16] of that stuff very well. And so, why
[39:19] wouldn't they win?
[39:20] >> Yeah. I mean also look alphabet's just
[39:23] so fascinating if you if you look
[39:25] broadly also at owning their own compute
[39:29] TPUs made a lot of advantages
[39:31] internally.
[39:32] >> The most stunning thing I think about
[39:34] Alphabet from their most recent
[39:36] financial results is that two or three
[39:39] years ago we would have said well large
[39:42] language models are going to cannibalize
[39:44] search. Search is dead. advertising
[39:47] based on search is Google's cash engine,
[39:50] >> they're in real trouble. It turns out
[39:53] that Google now gets more clicks on its
[39:58] search links than it used to and it
[40:01] charges more for each one than it used
[40:03] to because the value of the click is
[40:06] bigger with AI embedded in it.
[40:08] >> Mhm.
[40:08] >> And so they've managed to turn that
[40:10] around and it's extraordinary.
[40:12] >> Yeah. takes a long time to build those
[40:14] company relationships for running a
[40:17] proper sort of advertising based auction
[40:22] machine, right? It takes a long time to
[40:24] build those relationships. Okay, let's
[40:27] hop to China. So, I'm going to I'm going
[40:31] to resist the temptation to talk about
[40:33] Japan cuz I think you and I were there
[40:34] in roughly the within probably a year or
[40:37] two of each other. Maybe we overlapped
[40:38] with you and Kanazawa, which I've spent
[40:41] time. I'm going to resist that
[40:42] temptation and try to focus on China for
[40:45] purposes of this conversation.
[40:47] What have you learned about AI from your
[40:51] trip to China and thinking about China,
[40:54] speaking to Chinese people, whether
[40:56] they're technologists or otherwise? Like
[40:58] what have you learned during or since
[41:00] that trip?
[41:01] >> Back in March before my book was
[41:04] published in the US, I went to China
[41:07] because the Chinese are faster at
[41:08] everything, including publishing books.
[41:11] and my publisher brought me out there
[41:13] and basically you know took me around
[41:15] four cities, eight days meeting with AI
[41:18] leaders both in academia and big
[41:21] companies like Huawei and Hike Vision
[41:23] and Ant Group. And the thing which was
[41:26] surprising was the extent to which
[41:29] people brought up the issue of AI
[41:31] safety. And I say that was surprising
[41:34] because my friends who had done AI
[41:38] policy in the Biden administration
[41:40] had primed me to expect that there would
[41:44] be no mention of safety in China. They
[41:46] basically didn't care about it. That you
[41:48] know the muscle memory that we have in
[41:50] the west of technology being dangerous
[41:55] you know the atom bomb experience the
[41:57] Cuban missile crisis. Our ambivalence
[42:00] about technology is not shared in China
[42:02] where their idea of catastrophe is sort
[42:05] of like you know the cultural
[42:06] revolution. It's some political thing
[42:08] that goes wrong. And conversely,
[42:10] technology has been part of their
[42:12] amazing growth story in the last 25
[42:14] years, which they are rightly proud of
[42:16] and delighted by. So they love
[42:18] technology, right? So when the Biden
[42:22] team tried to meet with the Chinese and
[42:25] talk about AI safety, they got nowhere
[42:28] and they decided it was it possible to
[42:30] even talk to them about some sort of
[42:32] non-prololiferation treaty for AI. But
[42:36] when I went there, I found they did talk
[42:37] about safety kind of unprompted. And
[42:40] this led me down this track of arguing
[42:43] over the last couple of months that the
[42:45] door is actually open to a dialogue with
[42:49] China about preventing bad guys doing
[42:53] bad stuff with AI because they don't
[42:56] want the internet to be crashed by some
[42:59] cyber hacker who has the tool. They
[43:01] don't want bioweapons. They don't want
[43:02] chemical weapons. They want none of
[43:04] that. They love regulating the internet,
[43:05] right? So, we have a shared interest
[43:08] with the Chinese in preventing this
[43:11] proliferation risk from going nuts. And
[43:15] as I thought about it, you know, the
[43:17] kind of cold war analogy
[43:20] came to seem more and more opposite,
[43:22] right? So, if you look back at the story
[43:25] of nuclear weapons, there were two kinds
[43:27] of danger.
[43:29] First danger is you have a nuclear war
[43:32] between the Soviet Union and the United
[43:34] States. But that was contained by
[43:37] balance. Two superpowers, they both have
[43:40] their weaponry. They have mutually
[43:42] assured destruction. So there's no war.
[43:45] Then there's another kind of risk which
[43:47] is that other random rogues, whether
[43:49] it's criminals, terrorists, rogue
[43:50] states, get the stuff and they do bad
[43:54] stuff. And it's much harder to deter
[43:56] that because it's a multipolar game. And
[43:59] so deterrence doesn't work so elegantly.
[44:02] And so the way it was dealt with in the
[44:03] cold war was that in 1956 there was the
[44:06] agreement on the international atomic
[44:07] energy agency. And in 1968, the
[44:10] non-prololiferation treaty kind of
[44:13] enforced compliance with the IAEA such
[44:16] that you could get civilian nuclear
[44:18] power if you were a non-uclear state,
[44:21] but you had to submit to the rules and
[44:24] be inspected and show that you were not
[44:26] using the enriched nuclear material to
[44:30] build a weapon, right? And so I think
[44:33] the same analogy could be applied to AI.
[44:36] We're going to have par roughly with
[44:38] China. We'll both have powerful AI.
[44:40] Hopefully, deterrence prevents war
[44:42] breaking out. But at the same time, we
[44:46] don't want openweight models that can be
[44:48] freely downloaded by anybody who wants
[44:51] to fall into the hands of criminals and
[44:54] terrorists who can then use it to hold
[44:57] us hostage. And we have a joint interest
[44:59] in that. And you know, when my friends
[45:02] from the Biden team or even from the
[45:04] current administration say, "Well, you
[45:06] can't talk to China about safety. They
[45:07] don't care." I say, "That's not true."
[45:09] And they say, "But it's really hard.
[45:10] They don't stick by their commitments."
[45:12] And I go, "You think Nikita Kruef in the
[45:15] Soviet Union was easy to negotiate with?
[45:17] He was the guy who put missiles in Cuba
[45:19] and went to the UN and banged his shoe
[45:21] on the table and said, "We will bury
[45:23] you." I mean, he was a tough guy to talk
[45:26] to, but we did talk to him and we got
[45:28] the non-prololiferation treaty agreed
[45:31] and I think we need to do the same thing
[45:33] again. Now,
[45:34] >> where do you stand on
[45:36] your thinking about chip export? So when
[45:42] the chip export controls were announced,
[45:45] which was October of 2022, right before
[45:48] Chhatty PT,
[45:50] I supported those controls quite loudly.
[45:54] I wrote a very long piece in the
[45:55] Washington Post saying that if we could
[45:57] stop China getting frontier models by
[46:01] depriving them of frontier chips, I was
[46:03] all in favor of that because of the
[46:05] strategic advantage for the US. I mean,
[46:07] I work at the Council on Foreign
[46:08] Relations. We do geopolitics and
[46:11] national security all day long and I'm
[46:13] all in favor of US power. But I have to
[46:16] say that you know three and a half years
[46:18] later we haven't actually achieved that
[46:21] enormous advantage over China in terms
[46:23] of the models based on the best studies
[46:26] we're kind of eight months ahead in
[46:29] terms of where the frontier model is
[46:31] like our frontier model versus their
[46:32] frontier model. And then if you adjust
[46:34] that for the speed with which the model
[46:37] gets turned into an application probably
[46:40] that gap shrinks and it may even be
[46:42] non-existent. So however you slice that
[46:46] the basic bottom line is we both have
[46:49] strong models and the chip export
[46:51] controls have not delivered what I hoped
[46:54] would be the big advantage. And so I'm
[46:58] not against keeping the controls on if
[47:01] we think that maybe as the compute
[47:04] demands of bigger and bigger models
[47:07] bite, the chip controls will bite more
[47:10] and maybe we get a bigger advantage next
[47:13] year or something. But I don't want the
[47:15] chip controls to get in the way of a
[47:18] discussion with the Chinese about where
[47:20] we have a shared interest, which is in
[47:23] controlling openweight models and
[47:25] preventing the bad stuff falling into
[47:27] the hands of the bad guys. I would
[47:30] prioritize
[47:31] collaboration with China and if that
[47:34] meant, you know, loosening up a little
[47:36] bit on the export controls, I would be
[47:38] okay with that.
[47:39] Why do you think the rhetoric coming out
[47:41] of [snorts]
[47:43] pick your administration, right, it's
[47:44] not just limited to the current
[47:45] administration is China won't listen.
[47:47] They don't care about safety. Why do you
[47:49] think that is
[47:51] sort of the unofficial or official
[47:56] stance on things? Because there are
[47:57] certainly
[47:58] as someone who studied East Asian
[48:00] studies, right? There are people in the
[48:02] White House who speak fluent Mandarin,
[48:05] who are able to read native materials,
[48:07] who are spend time or able to certainly
[48:10] if they can't spend time determine the
[48:13] sentiment and conversations of the
[48:16] technologists building AI in China. So
[48:19] one would think that they would be aware
[48:21] that AI safety is a prominent topic in
[48:24] China if in fact it is. So why do you
[48:27] think that at the end of the day the
[48:31] stance or the supposed position of China
[48:35] that's echoed through the admin is that
[48:39] they won't talk about safety. Why do you
[48:42] think that is? I think part of this is
[48:45] that if you were to think back 20 years
[48:50] to when China was sort of relatively new
[48:52] in the WTO and we were collaborating
[48:55] with them on that and hoping that over
[48:58] time China would become more friendly to
[49:00] the US.
[49:02] At that time there would have been some
[49:04] China hawks who thought that a communist
[49:06] regime is not to be trusted and then
[49:08] some sort of China optimists who hoped
[49:10] that it would become easier to work with
[49:12] over time. And part of the trouble today
[49:15] is that the China optimists feel burned.
[49:19] They feel like they made this bet that
[49:23] China would become friendlier and then
[49:25] Xiinping took power roughly a decade ago
[49:29] and the opposite happened. they became
[49:31] more aggressive and harder to work with
[49:35] and also of course more technologically
[49:37] advanced and therefore more threatening.
[49:39] And so now you've got this world in
[49:40] which there are the natural hawks and
[49:42] then the former doves who have turned
[49:44] into kind of burned remorseful doves and
[49:47] therefore kind of with the zeal of the
[49:49] converted have become quite hawkish as
[49:51] well. And I don't mean to you know
[49:54] underestimate the sophistication of some
[49:55] of these people. I mean of course you
[49:57] know they speak Chinese. I don't speak
[49:59] Chinese. I I defer to their expertise.
[50:03] They probably know that there are
[50:04] builders of the technology, professors
[50:06] in the technology who talk the talk of
[50:09] safety. But they say, "Yeah, but you
[50:11] know that doesn't reflect what China's
[50:13] government would actually do."
[50:15] >> To which my response says, "Yes, but
[50:16] don't you think there is the same thing
[50:17] in the US? There's, you know, there are
[50:19] people who want to just race. There are
[50:21] people who care about safety. We have a
[50:23] pluralistic society. There's difference
[50:25] of opinion. It's the same in China." But
[50:28] at least admit that there is a faction
[50:30] that would like to collaborate and go
[50:33] and try and work on it because the
[50:36] alternative to trying to work on this is
[50:39] that we carry on with China producing
[50:41] very powerful open weight models which
[50:44] basically allow anybody to do whatever
[50:46] they like with AI as it gets to the
[50:49] point of serious danger.
[50:51] This is probably a very naive take, but
[50:53] I wonder how much of the official stance
[50:57] or the
[51:00] maybe using the partially true or not
[51:03] true at all
[51:05] position of China won't talk about
[51:07] safety as a reflection of the fact that
[51:09] in the case of nuclear weapons,
[51:13] the application of nuclear power is
[51:15] somewhat limited in comparison to super
[51:18] intelligence. I mean it is limited right
[51:21] so if the upside of super intelligence
[51:24] or AGI I mean these terms I think
[51:27] Benedict was saying AI is whatever the
[51:30] technology just can't quite do right now
[51:32] or something like that which I thought
[51:33] was pretty funny and not totally wrong
[51:36] but that if the person who crosses the
[51:39] finish line first
[51:41] >> has this broad power of a god
[51:44] effectively is that the simple truth is
[51:48] that everybody wants to first. So
[51:50] [laughter]
[51:51] I I just wonder how much of that is is
[51:54] also behind
[51:56] justifying the race with party X won't
[52:00] talk about safety. I just I mean it's
[52:02] not possible for me to know.
[52:04] >> I have had a conversation with the
[52:06] leader of one of the labs that you know
[52:08] I I shouldn't name, but I had this
[52:10] debate and he said look the chip export
[52:14] controls are going to leak. They're not
[52:16] going to last in some period of time.
[52:19] Huawei will figure out how to make good
[52:22] AI chips and you know that's inevitable.
[52:26] But that's okay because we only need to
[52:29] be ahead for the next couple of years
[52:32] because by 2028
[52:35] we will get to recursive
[52:36] self-improvement where the frontier
[52:39] model codes by itself the next frontier
[52:42] model and progress just goes vertical
[52:45] and at that point with recursive
[52:47] self-improvement we're done. The race is
[52:48] over whoever comes first at that point.
[52:50] That's it. And I think there's a couple
[52:53] things to say about that. First of all,
[52:55] that's not it in terms of deploying the
[52:57] model, right? You could have an
[52:59] incredibly powerful model in your server
[53:02] at Frontier Lab XYZ,
[53:05] but it's not helping productivity across
[53:07] your economy. It's not helping your
[53:08] military industrial complex until you
[53:11] deploy it into those guys systems. And
[53:15] that deployment and diffusion is going
[53:17] to take some time. And by the way,
[53:19] you're going to have to build a lot of
[53:20] compute. You're going to have to build a
[53:21] lot of energy. These things also take
[53:24] time. So it's not like you know you
[53:26] reach across some Rubicon and then it's
[53:29] all over. Now the one way in which I
[53:32] might be wrong about what I just said is
[53:34] if you use the frontier super
[53:38] intelligence
[53:39] offensively right you say okay we've got
[53:43] one super powerful model. The US
[53:46] government who we're talking to about
[53:47] this is going to use it and they are
[53:49] going to comprehensively penetrate
[53:52] everything about Chinese cyerspace and
[53:54] insert various trap doors, Trojan
[53:57] horses, you know, things that we can
[53:59] use. We get our hooks into their
[54:01] systems. And so now we can disable them
[54:04] if they start a war in Taiwan. Now we
[54:07] can their communication system
[54:09] if we need to. So that offensive use of
[54:14] kind of the very frontier model might
[54:16] negate my point about waiting for
[54:18] diffusion to happen. But of course,
[54:21] nobody in the debate is saying that.
[54:23] Nobody is saying, "Oh, we're racing to
[54:25] the front because then we're going to
[54:26] use it offensively. They don't admit
[54:28] that." Seems like it wouldn't be a very
[54:30] good look. I can't see why any
[54:32] superpower wouldn't do that, frankly.
[54:35] >> Yeah.
[54:35] >> Right. I don't know what the
[54:37] counterargument is. I was chatting with
[54:40] someone in your book who I I shame but
[54:44] certainly
[54:46] one of the most qualified to speak on
[54:47] these things and I mean his basic
[54:50] perspective
[54:51] was the first to super intelligence. we
[54:55] need to hope that they're [laughter]
[54:58] on some level good people and train this
[55:01] thing
[55:01] >> right
[55:02] >> well and like that's that's it like pray
[55:06] for it which scared the out of me
[55:08] to be honest I [laughter] mean I was
[55:10] like man that's the strategy or it's not
[55:13] even a strategy that is the hope grab
[55:16] the rosary and throw that into the
[55:18] rotation my god that's really terrifying
[55:21] to think China I'm hoping to take a trip
[55:23] to China. I had a very tough time there
[55:26] when I was I was at two universities in
[55:28] 1996. It was a pretty unfriendly time
[55:30] for a lot of good reasons, but to be an
[55:33] American there in 1996 with a shaved
[55:35] head looking like I do.
[55:38] But I have friends all over the place
[55:40] and I'm hoping to actually maybe
[55:43] interview technologists, not just in
[55:45] China. I mean, there are other places
[55:46] that are of interest to me, but before
[55:49] it gets too hot geopolitically, if we're
[55:51] trending that direction,
[55:53] >> I think that's a great idea, by the way.
[55:54] I mean, I think what I found was the
[55:57] cognitive dissonance of visiting a
[55:59] company like Hike Vision, which is under
[56:02] US sanctions
[56:04] >> and walking around their premises, which
[56:07] kind of feel very American. It feels
[56:08] like a cool tech company doing cool
[56:11] stuff, building cool gadgets. you know
[56:12] they have a display of they build this
[56:14] AI enabled camera technology or sensor
[56:18] technology and so one application might
[56:20] be you can point this camera at water
[56:23] and judge the pollution level
[56:25] >> and because of this you can have an
[56:28] internal market in pollution control. So
[56:30] the downstream city which is receiving
[56:33] water from the upstream city pays the
[56:36] upstream city to keep the water clean
[56:39] and that market can exist because you
[56:41] can precisely measure the pollution
[56:43] level thanks to this AI sensor which
[56:45] Hike Vision is building. So you're
[56:47] thinking, "Wa, this is cool." And then
[56:49] as you're walking around the building,
[56:50] they're saying, "Okay, well, we can go
[56:52] through the atrium now because the
[56:54] toddlers have gone because, you know,
[56:55] the crash for the kids of the employees
[56:59] finishes at 5:00 p.m. And so then there
[57:01] are all these 2-year-olds running around
[57:02] and it's a bit of a zoo. So if it was 5,
[57:05] we wouldn't go through there. But now
[57:06] it's 6 p.m., so we can." And you're
[57:08] thinking, "Whoa." Okay, so they've got,
[57:09] you know, the interests of their
[57:10] employees at heart. They're building
[57:11] this anti-polution technology. It's
[57:13] great. and then you realize they're
[57:15] under US sanctioned and considered to be
[57:17] a threat to the US. So it's quite
[57:19] interesting to process all that
[57:22] >> in the process of doing research for
[57:24] this book
[57:26] and also the broad exposure that you
[57:28] have to investors. But let's just say
[57:30] over the last handful of years, who are
[57:33] some of the most interesting or unusual
[57:38] compelling is the word I'm searching for
[57:40] investors who you've had the chance to
[57:43] meet, talk to, read about, get
[57:45] acquainted with directly or indirectly.
[57:48] >> Wow. So many. I mean, I'd say that Bill
[57:50] Gurley from Benchmark, you know, is
[57:52] right up there. I always think of the
[57:54] investment he did in Uber as the
[57:56] absolute quintessential perfect venture
[57:59] investment
[58:01] >> in the sense that he had done the open
[58:05] table investment and of course open
[58:08] table is a two-sided marketplace where
[58:10] you have lots of consumers that are
[58:12] looking for restaurants, lots of
[58:14] restaurants. You put tech in between
[58:16] which creates information and then the
[58:19] person looking for the place to eat can
[58:21] precisely say I would like you know Thai
[58:23] food at this price range in this area
[58:25] for three people at this time. Ding.
[58:28] What used to take you a lot of searching
[58:30] around. Bang it's done. And so Bill
[58:33] having done that was thinking well
[58:34] what's another two-sided marketplace?
[58:36] And he thought well there are lots of
[58:37] cars and lots of people who need a ride
[58:40] and you put information in the middle in
[58:42] the same way. there ought to be
[58:43] something which is like an app for ride
[58:46] sharing.
[58:47] >> And so he imagined Uber way before Uber
[58:50] existed. That was point number one.
[58:52] Point number two, he went to see various
[58:54] entrepreneurs who were in this space and
[58:56] he checked them out and he had the
[58:58] discipline not to invest in them
[59:00] [clears throat]
[59:01] >> because although they were kind of going
[59:02] at the right thing, there was some hair
[59:05] on the deal, some wrinkle, some way they
[59:07] were approaching it that just felt like
[59:08] it wasn't going to be quite right. So he
[59:10] resisted.
[59:12] Uber came to him before Travis was the
[59:15] CEO
[59:16] >> and Bill said, "I'm not doing that."
[59:19] Because he didn't think the CEO at the
[59:21] time had what it took. And then there
[59:23] was an internal switch at at Uber.
[59:25] Travis became the leader. Bill meets him
[59:28] and like bang, he immediately invests
[59:30] because he's been waiting and waiting
[59:32] and waiting for the idea to be paired.
[59:35] As you were saying earlier, you have to
[59:36] have the the market to be paired with
[59:39] the right person. and he saw it and then
[59:42] he invested and he was a great board
[59:44] member and it all went perfectly right.
[59:47] But then there is this kind of
[59:48] Shakespearean tragedy in the latter part
[59:51] of the story where the growth investors
[59:54] come in. He gets diluted. He no longer
[59:57] has influence. His key card to get into
[59:59] the building is deactivated and he's
[01:00:02] basically stiffed and he watches, you
[01:00:04] know, Uber kind of go off the rails. And
[01:00:06] then finally comes, you know, the the
[01:00:08] denum where he rounds up the dissident
[01:00:11] investors and they have this coup
[01:00:13] against Travis and that sets the company
[01:00:16] on a path to where they hide Darra and
[01:00:18] do the IPO. I just think that's the
[01:00:22] ultimate venture capital story and Bill
[01:00:24] is the ultimate venture capitalist.
[01:00:26] >> He is practically a neighbor here for
[01:00:28] me. I'm sure
[01:00:29] >> in Austin and we've had a couple of
[01:00:32] conversations on the podcast and he's I
[01:00:34] would say on a very parallel track to
[01:00:37] you with respect to China, right? And he
[01:00:40] catches some flack for it. People are
[01:00:41] like, "He's an agent of the CCP." I'm
[01:00:43] like, "No, trust me, Bill is not an
[01:00:44] agent of the CCP. [laughter]
[01:00:46] It's just the most ridiculous
[01:00:48] >> accusation." But he is a very incisive
[01:00:51] observant
[01:00:53] human
[01:00:55] >> who also happens to be a polymath in
[01:00:57] multiple disciplines who can speak
[01:00:59] casually about very technical things.
[01:01:01] And this also you you referring to Bill
[01:01:03] in this way or describing him in this
[01:01:06] way makes me think about multiple points
[01:01:08] in the infinity machine and I'm pulling
[01:01:13] from memory which is as we know pretty
[01:01:15] faulty but you know Elia with the
[01:01:18] transformer architecture and the
[01:01:19] prepared mind I think Demis also just
[01:01:22] thinking about a problem deeply and
[01:01:24] seriously or with great imagination for
[01:01:27] a long time and then when the solution
[01:01:31] or the the germ of a solution appears
[01:01:35] immediately recognizing it, right? It's
[01:01:37] just it's wild to see how frequently
[01:01:39] that recurs. Any other investors, you
[01:01:42] know, a name that doesn't get much
[01:01:44] airplay who
[01:01:46] [laughter]
[01:01:47] I I think is just a fantastic character
[01:01:50] and maybe you could introduce him to
[01:01:53] people who are listening if they don't
[01:01:55] recognize it. Luke Nosk, where does Luke
[01:01:58] >> who has I wish I knew how to turn on my
[01:02:02] batteries in the same way to [laughter]
[01:02:04] to get the energy that Luke does. But
[01:02:08] how does Luke fit into the story of
[01:02:13] deep mind and I I suppose more broadly
[01:02:16] speaking for that because of that AI.
[01:02:19] >> Luke Nosek is this tremendously puppyish
[01:02:23] enthusiast, right? and he was a you know
[01:02:28] early early part of the PayPal team with
[01:02:32] Max Levchin and Peter Teal and he went
[01:02:36] through that journey and then Peter
[01:02:38] exited PayPal set up founders fund. This
[01:02:42] is now I think 2005
[01:02:45] and Luke Nosek becomes one of the first
[01:02:48] partners and pretty early on he makes
[01:02:52] the right judgment on Elon and SpaceX.
[01:02:56] >> Mhm.
[01:02:57] >> And Luke is the kind of guy who is just
[01:02:58] all in. When he falls in love with an
[01:03:00] idea and a founder, there is no curbing
[01:03:04] his enthusiasm.
[01:03:06] And so he's like all in all in all in on
[01:03:09] SpaceX. And I think, you know, he
[01:03:12] persuaded Founders Fund to like raise a
[01:03:15] new fund, put extra money in, like more
[01:03:17] more more more more capital in there.
[01:03:19] [clears throat]
[01:03:20] >> And of course, that paid off massively.
[01:03:22] >> And off the back of that, you know, roll
[01:03:25] forward to 2010.
[01:03:27] >> He's trying to look for the next Elon
[01:03:29] Musk. And he does a few kind of frontier
[01:03:32] bets. And then along comes Demesis Abis
[01:03:36] who is out on the west coast from London
[01:03:40] raising capital for this idea of an AI
[01:03:42] company which he's going to call Deep
[01:03:44] Mind. And you know most people think
[01:03:47] that's nuts. This AI remember in 2010
[01:03:51] cannot even recognize a photo of a cat.
[01:03:53] It can't do anything. We're in deep deep
[01:03:56] AI winter. Who would back a company like
[01:03:58] that? The answer is Luke Nosac. and he
[01:04:01] falls in love with Demis who is a very
[01:04:04] winsome character, super articulate,
[01:04:06] super relatable and a genius. You know,
[01:04:08] he has all the kind of outlier
[01:04:11] characteristics you want in an
[01:04:12] entrepreneur. You know, the sort of
[01:04:14] junior chess champion, second best
[01:04:16] player in the world, but also five times
[01:04:19] wins the mind games olympiad where you
[01:04:22] have to run between boards playing bat
[01:04:24] gammon, chess, go and a couple of other
[01:04:27] games kind of almost simultaneously. I
[01:04:29] mean just kind of crazy crazy smart
[01:04:31] obsessed since he was 17 with the idea
[01:04:34] of building powerful AI. So Peter Thiel
[01:04:37] said to me about Demis I think
[01:04:39] individuals tend to have one company
[01:04:42] inside them if they're missionary
[01:04:45] entrepreneurs they've got one thing they
[01:04:47] need to do and for Demis it was to build
[01:04:50] AGI like that was what he was fixated by
[01:04:52] and the company was downstream of his
[01:04:56] desire to build AGI. If he could have
[01:04:57] done that at a university, he would have
[01:04:59] been happy to do that. But he couldn't
[01:05:01] do it at university. So he had to found
[01:05:03] a company to do it. And that's the kind
[01:05:04] of missionary commitment that venture
[01:05:07] capitalists often look for because a
[01:05:09] missionary will never quit. No matter
[01:05:11] how hard it is, they will keep working.
[01:05:14] And so Luke Nosek and Peter Teal jointly
[01:05:18] recognize this. Peter is, you know,
[01:05:20] contrarian, cynical, aloof, and so is
[01:05:24] kind of into it, but at the same time
[01:05:27] arms length. Luke is like got both his
[01:05:29] arms around Demis is giving him this
[01:05:31] bear hug and will not let go. And you
[01:05:34] know, Demis says, "I'm not going to move
[01:05:36] to California. I'm going to do this
[01:05:37] company in London." And Peter and the
[01:05:40] other Founders Fund partners are like,
[01:05:43] "London? Where is that?" It's kind of
[01:05:45] like Somalia or something. I mean,
[01:05:47] that's just off the map. And Luke says,
[01:05:49] "No, no, no, no. We have to do this. We
[01:05:50] have to do this. I will fly to London
[01:05:52] for the board meetings and we've just
[01:05:54] got to do this deep mind investment."
[01:05:56] And so he was the kind of unbridled
[01:05:59] enthusiast who got Founders Fund across
[01:06:01] the line. And the rest is history. You
[01:06:03] know, they put the series A money in.
[01:06:06] Unbelievably, it was 2 million at a 4
[01:06:09] million valuation. So they got half the
[01:06:11] company for 2 million bucks. Not bad.
[01:06:14] >> Not bad.
[01:06:15] >> And they rode that investment. What a
[01:06:17] remarkable story. I really feel like
[01:06:19] Luke,
[01:06:21] who's also here in Austin, deserves
[01:06:25] a lot more credit than he gets. Not that
[01:06:28] he's seeking it, right? He's not he's
[01:06:30] not out there looking for it, but he is
[01:06:33] very good at
[01:06:37] riding winners when he is high
[01:06:39] conviction, right? which in the venture
[01:06:41] game
[01:06:42] >> I mean in a lot of investing it's you
[01:06:46] can't die you can't run out of bankroll
[01:06:48] at the table right you need to have
[01:06:51] enough of a portfolio approach to
[01:06:54] sustain yourself through periods of bad
[01:06:56] luck but if you're systematic it's
[01:06:59] riding your winners and doubling and
[01:07:03] tripling and quadrupling down and he is
[01:07:04] so good at that he is just incredibly
[01:07:07] good
[01:07:08] >> and as John Duro likes to Hey, the great
[01:07:10] thing about venture capital is you can
[01:07:12] only lose one times your money.
[01:07:15] >> So it's not like a short position for a
[01:07:16] hedge fund trader where you can like
[01:07:18] really lose a lot, right? So
[01:07:19] >> exactly
[01:07:20] >> in that sense you're not going to die.
[01:07:22] So you can shoot for the moon.
[01:07:24] >> So I do have a question. I should know
[01:07:26] the answer to this but I don't.
[01:07:29] So long ago, this is probably 200 2008.
[01:07:33] This is a long time ago actually. I
[01:07:35] wonder if I had exposure to deep mind. I
[01:07:38] invested in Founders Fund. This was a
[01:07:40] very, very long time ago. But what I did
[01:07:42] not realize internally, and I'll just
[01:07:44] read a couple of my highlights. It is
[01:07:46] absurd how many highlights I have from
[01:07:47] the Infinity Machine and all of your
[01:07:49] books. [laughter]
[01:07:51] So, a gap opened up between Teal and
[01:07:53] Nok. As a general matter, Teal doubted
[01:07:55] that going on boards was a good use of
[01:07:57] partners' time. Startups should be left
[01:07:58] to sink or swim. The art of venture
[01:08:00] capital, he liked to say, was to back
[01:08:01] contrarian ideas, not coach company
[01:08:03] founders. Just we could spend a lot of
[01:08:05] time just on that, but I'm going to move
[01:08:07] on. Most venture partnerships decide on
[01:08:09] investments by voting. If a handful of
[01:08:11] partners see hair on the deal, the deal
[01:08:13] will be rejected. But Teal had taken the
[01:08:14] unusual position, the collective
[01:08:16] decision-making should be avoided. The
[01:08:17] way he saw things, if investments were
[01:08:18] chosen based on voting, the founders
[01:08:20] fund portfolio would consist of
[01:08:21] middle-of the road startups to which
[01:08:23] nobody objected. And then dot dot dot,
[01:08:26] this comes back to the power law, right?
[01:08:28] Given that all the profits in venture
[01:08:29] come from a few improbable moonshots,
[01:08:32] this sort of consensus portfolio would
[01:08:34] deliver mediocre performance. So, and
[01:08:36] I'll just paraphrase now, Teal empowered
[01:08:38] the partners to go allin with their
[01:08:41] guts/intuition.
[01:08:42] My question is, how is that governed in
[01:08:45] any way? Of course, if anyone gave 10
[01:08:48] out of 10 conviction and then lost money
[01:08:49] consistently, they would presumably be
[01:08:52] sort of removed from the partnership or
[01:08:54] they'd lose their ability to lead with
[01:08:56] that type of gut conviction. But do you
[01:08:59] have any idea how that was handled
[01:09:01] internally?
[01:09:02] >> Yeah. in terms of stress testing ideas,
[01:09:06] pushing people to really put their ass
[01:09:08] on the on the line for these types of
[01:09:12] high conviction but certainly very much
[01:09:15] outlier investments. Do you have any
[01:09:17] idea?
[01:09:18] >> I think internally founders fund was
[01:09:21] very torn about the deep mind investment
[01:09:23] and I described some of this in the book
[01:09:26] where you know they do the first deal
[01:09:28] and that's fine. It's $2 million.
[01:09:30] >> Mhm. But then you get to series B and
[01:09:32] series C and the check size gets bigger
[01:09:34] and so the other partners are asking
[01:09:36] tougher questions and they're saying,
[01:09:38] "Well, wait, is there going to be a
[01:09:39] product?"
[01:09:40] >> Mhm.
[01:09:41] >> And Demis said to me, you know, that his
[01:09:43] attitude was, "What do you mean is there
[01:09:46] a product? I'm talking about artificial
[01:09:48] general intelligence. It's going to make
[01:09:49] all products like revolutionized or
[01:09:52] obsolete or whatever. And you want to
[01:09:55] ask me what the widget is? Give me a
[01:09:57] break." you know the [clears throat]
[01:09:58] it's all of the widgets they're all
[01:10:00] going to be changed and if you're asking
[01:10:03] me this question you don't get what AGI
[01:10:05] means
[01:10:06] >> and so Deis was very frustrated by the
[01:10:08] other partners at Founders Fund and I
[01:10:10] think internal within Founders Fund
[01:10:13] there was a lot of fighting between Luke
[01:10:15] who remained enthusiastic and committed
[01:10:17] about Demis partly because he was the
[01:10:19] guy who would go to London and meet with
[01:10:21] him and sit in the board meetings and he
[01:10:23] would get the sort of you know several
[01:10:25] thousand vaults of Demis enthusias ASM,
[01:10:28] you know, injected into his spine at
[01:10:31] every meeting and he would come back
[01:10:33] buzzing with excitement. And the other
[01:10:35] Founders Fund partners who didn't have
[01:10:37] that benefit were skeptical. And so Luke
[01:10:42] would often come to Demis and say,
[01:10:44] "We've got your back. We've got your
[01:10:45] back. We know we're going to do the next
[01:10:46] round. We're going to lead the next
[01:10:47] round." And then actually in series C,
[01:10:51] Founders Fund at the last minute pulled
[01:10:52] out and they put money in, but they did
[01:10:54] not lead.
[01:10:55] >> Mhm. And so the answer to your question
[01:10:57] is there was a lot of argument within
[01:11:00] founders fund as the check size grew it
[01:11:04] was harder to have that you know double
[01:11:06] down on your winners kind of attitude.
[01:11:09] >> Yeah in this case the fish that got away
[01:11:13] although I mean it was a fantastic
[01:11:14] multiple on their initial money. It
[01:11:16] strikes me in reading the book that
[01:11:19] I would argue that Demis made absolutely
[01:11:23] the right decision with
[01:11:26] the Google acquisition. I mean you
[01:11:28] mentioned also in the book how he got
[01:11:30] criticized in some UK media for like oh
[01:11:32] you know giant mega corporation in the
[01:11:35] US gets our prized talent cheap kind of
[01:11:38] stuff. But looking back, I mean, he
[01:11:40] seems to have anticipated
[01:11:44] the costs and compute and and just
[01:11:48] raw materials that would be required to
[01:11:51] do what he was trying to do,
[01:11:53] >> right?
[01:11:53] >> Would you read that the same way?
[01:11:55] >> Yeah. I mean, I often have this debate
[01:11:58] with people in London where they say
[01:12:00] exactly as you put it, you know, this
[01:12:02] was a tragedy for UK tech. great
[01:12:05] champion of deep tech, you know, is
[01:12:07] bought out cheaply by Google. And I say,
[01:12:09] listen, it wasn't cheap. The acquisition
[01:12:11] price might have been $650 million,
[01:12:13] which was a bit cheap, but you know how
[01:12:15] much they put in in terms of recession
[01:12:17] development funds over the next 10
[01:12:19] years? It was approaching 10 billion,
[01:12:20] almost a billion a year, right? So, this
[01:12:23] was not selling cheap to the Americans.
[01:12:25] This was a cunning British trick to get
[01:12:28] a billion dollars of American R&D money
[01:12:31] into London per year for the next
[01:12:33] decade. terrific win. And by the way,
[01:12:35] today there are spinouts from Deep Mind
[01:12:39] in London because the talent stayed in
[01:12:42] London. And these spinouts are raising
[01:12:45] billions of dollars to do new AI
[01:12:48] companies. So it's terrific for the
[01:12:50] London ecosystem around King's Cross
[01:12:52] which is this sort of cool center for
[01:12:53] tech in London where you can get the
[01:12:55] train in one direction and be in
[01:12:56] Cambridge which has quite a lot of good
[01:12:58] startups you know in one hour or you can
[01:13:01] get the train in the other direction and
[01:13:02] be in Paris where there's you know MR
[01:13:06] and so forth and it's kind of very wired
[01:13:08] into different bits of Europe. So how
[01:13:10] long does it take to get from San
[01:13:11] Francisco to Mountain View depending on
[01:13:14] the traffic right can be well over an
[01:13:16] hour. So I think there is a technology
[01:13:19] ecosystem which is by no means the
[01:13:21] equivalent of Silicon Valley yet, but
[01:13:23] it's certainly unrecognizably better
[01:13:26] than it was 10 or 20 years ago.
[01:13:28] >> What do you think the UK or Europe could
[01:13:31] do? Let's let's focus on the UK perhaps
[01:13:34] could do to increase the level of
[01:13:39] innovation early stage
[01:13:42] startup founding etc. Right? Because
[01:13:44] looking back at the power law and
[01:13:45] certainly just having spent so much time
[01:13:47] in California, there's a lot that went
[01:13:49] into Silicon Valley, right? And there
[01:13:51] are certain things that don't get a lot
[01:13:54] of airplay, but for instance, the
[01:13:55] difficulty of enforcing non-compete
[01:13:58] agreements in California, right, really
[01:14:00] led to this sort of roundroin of talent
[01:14:03] moving and cross-pollinating like little
[01:14:05] hummingbirds of engineering talent and
[01:14:09] so on, which [snorts] may not be
[01:14:11] replicable depending on where you are,
[01:14:14] but what what could the UK do in your
[01:14:16] mind if if you had the ear and they were
[01:14:18] like, "All right, Sebastian,
[01:14:21] tell us what to do.
[01:14:22] >> A couple of things. I mean, I think the
[01:14:24] mistake that people in Europe make and
[01:14:27] Britain as part of this is to believe
[01:14:28] that there's some kind of cultural magic
[01:14:31] about Silicon Valley where whatever it
[01:14:33] is that they're drinking in the water
[01:14:35] out there makes them think that failure
[01:14:37] is a learning experience which is kind
[01:14:39] of weird and the Europeans say, "Well,
[01:14:41] we're never going to be like that." And
[01:14:43] it's impossible for us to become as
[01:14:45] entrepreneurial as Silicon Valley. And I
[01:14:48] remind people that when Fairchild
[01:14:50] Semiconductor was founded in 1957, the
[01:14:53] eight scientists who left the Shockley
[01:14:55] lab were called, get this, the
[01:14:58] traitorous eight.
[01:14:59] >> So good.
[01:15:00] >> Traitorous. Why? Because it was
[01:15:02] considered treachery at the time to
[01:15:04] leave one company and go to another
[01:15:05] company. There was no entrepreneurial
[01:15:07] culture in the 1950s on the West Coast
[01:15:10] in the US, right? The classic business
[01:15:12] book of the time was organization man
[01:15:14] about people who joined one company and
[01:15:16] stayed in it for their whole life and
[01:15:18] retired with a gold watch on their 60th
[01:15:20] birthday. Right? So you can create an
[01:15:23] entrepreneurial culture and that is
[01:15:25] happening bit by bit in Britain and
[01:15:28] certainly in Israel and it's happened in
[01:15:31] China and it's not some magic which is
[01:15:33] confined to Silicon Valley. Okay, I it's
[01:15:35] worth making that point as the first
[01:15:36] thing. Now there are specific policy
[01:15:40] shifts that you need to do to make an
[01:15:41] ecosystem work and I think you put your
[01:15:44] finger on one which is the mobility of
[01:15:47] talent is super important. You can think
[01:15:50] of a startup ecosystem as something
[01:15:51] which circulates three elements money,
[01:15:55] people and ideas and you circulate those
[01:16:00] and you combine them in different ways.
[01:16:02] And each time you combine them, that's a
[01:16:04] new company. And each has a shot on
[01:16:06] goal. And most of them fail. But all of
[01:16:08] a sudden, if you circulate these these
[01:16:10] components fast enough, you do get
[01:16:13] product market fit. And then you get
[01:16:15] these 10x plus returns. Now, in Britain,
[01:16:18] when you raise a new round, a series B
[01:16:20] say, and you've got like nine months of
[01:16:23] runway to build to the next stage from
[01:16:26] your company, and you identify the three
[01:16:29] key talent that you're going to bring
[01:16:32] into the company and make it happen, and
[01:16:34] then they turn around to you and say,
[01:16:36] "Well, I can come in 6 months." That's a
[01:16:38] death sentence, right? That's horrible.
[01:16:40] We call it gardening leave in Britain.
[01:16:43] That is an appalling idea. We got to get
[01:16:45] rid of those gardens and we got to let
[01:16:47] people move fast. Another thing is tech
[01:16:50] transfer out of universities. In the US
[01:16:52] there's the BOL act. There are these
[01:16:55] very sophisticated tech transfer offices
[01:16:56] which are generous to the entrepreneur
[01:16:59] in terms of not demanding too much flesh
[01:17:03] >> as somebody exits and that's essential
[01:17:05] for making the startup work. In Europe
[01:17:08] the attitude is oh we're the university.
[01:17:11] We deserve a lot of skin in the game
[01:17:12] here. we want 50% of the upside. Well,
[01:17:15] in that case, the startup will never
[01:17:16] happen.
[01:17:17] >> Mhm.
[01:17:17] >> And I say to these Europeans, look, go
[01:17:20] visit Stanford. They're very generous to
[01:17:22] their entrepreneurs. They seem to be
[01:17:24] okay financially [laughter]
[01:17:28] because if you help the entrepreneur,
[01:17:29] you know, you'll get the donations
[01:17:30] later. It's all good.
[01:17:32] >> Yeah.
[01:17:32] >> And so, I think those are just two
[01:17:34] things
[01:17:35] >> which started a long time ago in the US,
[01:17:37] right? You you look at the origins of
[01:17:39] Janentech and so on. I mean it's just
[01:17:42] been
[01:17:43] >> it's it's the genesis of so many not
[01:17:46] just companies but industries
[01:17:48] effectively in the US.
[01:17:50] >> Yeah.
[01:17:51] >> Do you think Demis would have built Deep
[01:17:55] Mind if he had not read Enders Game?
[01:17:58] [laughter]
[01:17:59] >> That's a great story. That's a great
[01:18:00] question.
[01:18:01] >> Can I just tell the Enders game story to
[01:18:03] begin with?
[01:18:04] >> And also a bit of trivia for folks. I
[01:18:06] believe, and not not to like make this
[01:18:11] more more difficult, but that when Mark
[01:18:14] Zuckerberg first had a profile on
[01:18:16] Facebook, the only book listed was also
[01:18:19] Enders Game.
[01:18:20] >> Oh, I didn't know that.
[01:18:21] >> I believe that's true.
[01:18:22] >> That's fascinating.
[01:18:23] >> So, hop into it with Demis and Enders
[01:18:25] Game. So right at the beginning of my
[01:18:28] interviewing of Demis we were having the
[01:18:30] second meeting which was a dinner and he
[01:18:33] told me to read a couple of books before
[01:18:34] we had the dinner and one of them was
[01:18:36] Enders Game.
[01:18:37] >> What were the others just before you
[01:18:39] continue?
[01:18:40] >> It was a book by David Deutsch called
[01:18:43] The Fabric of Reality.
[01:18:45] >> Uhhuh. Light read. [laughter]
[01:18:46] >> Yeah. Now, I read Enders Game as a
[01:18:50] result, and I hadn't read it before. And
[01:18:52] as I was reading it, I was thinking to
[01:18:53] myself, okay, so this is a story about a
[01:18:56] sort of boy hero who saves the entirety
[01:18:59] of humanity from an invasion of the
[01:19:02] planet by the space aliens. Is Demis
[01:19:05] telling me that that's how he sees
[01:19:08] himself? That he's like saving all of
[01:19:09] humanity with AI?
[01:19:12] because it'd be a bit much to believe
[01:19:13] that, but it would be even more to have
[01:19:17] the tmerity to tell the guy who's
[01:19:19] writing a book about you [laughter]
[01:19:22] that that's how you see yourself. Like
[01:19:23] most people wouldn't expose themsel in
[01:19:26] that way. I thought, is Deis really
[01:19:28] thinking this? So then I go to have the
[01:19:30] dinner and he says, "I hope you read
[01:19:32] Enders game because that's really how I
[01:19:34] see myself." And I gave the book to my
[01:19:36] wife so she could read it so she could
[01:19:38] understand me better because I really
[01:19:39] identify with Ender. Yeah, it's wild.
[01:19:42] >> It's wild.
[01:19:43] >> It's a great book. I mean, I haven't
[01:19:45] read it in decades, but it is it is a
[01:19:48] fantastic read as I remember it.
[01:19:50] >> Yeah. I mean, reading it, I must say, as
[01:19:52] a mature adult, I thought it was not
[01:19:54] that well written.
[01:19:56] >> Yeah.
[01:19:56] >> But the idea of it is good. And I can
[01:19:59] see why
[01:20:00] >> the idea is sticky.
[01:20:01] >> Absolutely. You know, this image of this
[01:20:03] kid who sacrifices everything to
[01:20:05] dedicate himself to the craft of
[01:20:08] fighting the aliens.
[01:20:09] >> Mhm. and you know withstands ridicule
[01:20:12] and bullying from his peers and fights
[01:20:14] back. It's an appealing image and that's
[01:20:17] what hooked Demis. But to answer your
[01:20:20] question of earlier, you know, he would
[01:20:21] have done AI anyway because he read
[01:20:23] Enders Game actually when he was already
[01:20:27] kind of around 30.
[01:20:28] >> Mhm.
[01:20:28] >> And he'd had unbelievably the
[01:20:31] determination to build super
[01:20:33] intelligence from when he was about 17.
[01:20:35] I mean that is wild as well. I mean the
[01:20:37] early conviction is just extraordinary.
[01:20:39] Did he ask you to read Good Echerbach
[01:20:43] an eternal golden braid? I will admit to
[01:20:46] you I think Dustin Moskavitz
[01:20:49] also a lot of technologists very very
[01:20:51] very good technologists recommend this
[01:20:54] book
[01:20:55] >> or cite it as part of their own journey
[01:21:00] to building something incredible. I
[01:21:03] think I'm too dumb to read that book. I
[01:21:05] had so much trouble. I've had so much
[01:21:06] trouble. I've tried two times and yet
[01:21:09] I've still not finished that book. I
[01:21:11] don't know. Hey, do you have any
[01:21:12] recommendations to somebody who's maybe
[01:21:14] lacking a few IQ points cuz he was born
[01:21:16] on Long Island as to how to navigate
[01:21:18] that book? I have to admit I was told by
[01:21:21] Demis that this meant a huge amount to
[01:21:23] him that he'd read it in his late teens
[01:21:26] and that was when he really became
[01:21:28] convinced that he could build AI because
[01:21:30] the argument in the book is that you
[01:21:33] know whatever the human brain can do
[01:21:37] computers will be able to do one day
[01:21:39] that the human brain operates on ones
[01:21:41] and zeros and therefore if you could
[01:21:43] build big enough compute you should be
[01:21:46] able to replicate the intelligence of
[01:21:47] human brains and and that was the sort
[01:21:49] of insight that got him hooked on the
[01:21:51] idea. So I went off and I tried to read
[01:21:53] it. I would say I got like 150 pages in
[01:21:56] and got bogged down. I mean it is
[01:21:58] [snorts] a difficult challenging read
[01:22:00] but at least I kind of extracted the
[01:22:03] essence
[01:22:04] >> that meant something to my subject to
[01:22:06] Demis. You know who would be great for
[01:22:08] helping me to understand this? LLM
[01:22:10] [laughter]
[01:22:12] going to give that a shot and see if
[01:22:14] explain this to a sixth grader maybe or
[01:22:17] a s or explain it to a six-year-old
[01:22:18] maybe even better. Couple of questions
[01:22:20] and then we'll start to lay on the plan.
[01:22:23] If you had to write another book on a
[01:22:28] figure in the world of AI, they could be
[01:22:31] relatively unknown
[01:22:33] or they could be incredibly known. Who
[01:22:35] would that person be? Demis is off the
[01:22:37] table.
[01:22:38] >> I might want to take Sam off the table
[01:22:40] just to make it
[01:22:42] >> a little more interesting. Who would it
[01:22:44] be if Sam's off the table and Deis is of
[01:22:48] course off the table?
[01:22:49] >> Well, I guess Dario.
[01:22:52] >> Yeah,
[01:22:52] >> I think even if you left Sam on the
[01:22:54] table, it would be Dario. I mean, I
[01:22:56] think he's just a fascinating
[01:22:58] fascinating figure as well as being the
[01:23:00] current leader
[01:23:00] >> of anthropic for people who don't
[01:23:02] recognize the name.
[01:23:03] >> Yeah,
[01:23:04] >> man. You know, I'm working on a blog
[01:23:06] post right now. It's about disruption
[01:23:08] due to AI and how it's not three years
[01:23:13] in the future. It's not one year in the
[01:23:14] future. These are book sales across my
[01:23:17] entire book catalog and it's not limited
[01:23:20] to print. This is all format. Okay? So,
[01:23:24] I'll give you some numbers and then I
[01:23:26] want you to tell me what happened to
[01:23:28] initiate this. Okay. 2022 stasis pretty
[01:23:32] consistent. My book royalties are an
[01:23:34] annuity predictable.
[01:23:37] 2023 minus 5%. 2024 minus 13%. 2025
[01:23:44] minus 46%.
[01:23:46] And 2026 so far on track to be at least
[01:23:50] 57%.
[01:23:51] What happened at the end of 2022?
[01:23:55] [laughter]
[01:23:55] >> Chat GPT [clears throat]
[01:23:57] >> GPT 3.5.
[01:23:59] It's just wild. It's really, really
[01:24:03] wild. I mean, this stuff is coming fast.
[01:24:05] And I really flip and flop. I feel like
[01:24:09] I waffled perhaps too much between these
[01:24:11] two. I I go from the very I would say
[01:24:16] moderate well-reasoned
[01:24:18] positioning of Benedict and I agree with
[01:24:20] so many of his points to believing that
[01:24:24] all of this is just coming so much
[01:24:25] faster than anyone can even comprehend
[01:24:27] due to the sort of recursive
[01:24:29] self-improvement. For the record, I
[01:24:31] think that it is much bigger than
[01:24:33] mobile, much bigger than internet. This
[01:24:35] is so general cognitive capability which
[01:24:39] can span you know any human task. I
[01:24:42] think the niggle is simply how long does
[01:24:46] diffusion take.
[01:24:47] >> Yeah. Right. And just to give an example
[01:24:49] of that you know I invest in quite a few
[01:24:53] biotech companies and
[01:24:56] other sciences and if you look at say
[01:24:59] alpha fold right I mean absolutely
[01:25:02] merited a Nobel prize. We didn't mention
[01:25:04] that about demis but it's one thing to
[01:25:07] design molecules it's quite another to
[01:25:09] deliver it to target tissue right so
[01:25:12] like the deliverability of that sort of
[01:25:15] a metaphor for AI in a way [laughter]
[01:25:17] it's like okay great we have this
[01:25:19] pristine perfect molecule how do you get
[01:25:21] it to the right place and at the same
[01:25:24] time an investor in a company called
[01:25:27] Laya Laya Sciences and what they're
[01:25:30] doing is producing
[01:25:33] a proprietary data set by automating wet
[01:25:37] labs using AI, right? And I'm going to
[01:25:39] simplify it, right? But they have
[01:25:41] gigantic wet labs where they can run in
[01:25:44] parallel thousands of experiments that
[01:25:46] from the very first step of hypothesis
[01:25:48] generation through to the end of the
[01:25:50] scientific method is all run
[01:25:52] autonomously by AI. And I bring this
[01:25:56] particular example up because even I
[01:25:59] want to say 6 months ago, 12 months ago,
[01:26:02] like they are producing discoveries that
[01:26:06] are really non-trivial, right? It's like
[01:26:10] it's already happening now. Like this is
[01:26:13] not
[01:26:14] >> this is not a year in the future. Like
[01:26:16] this is happening now. So when you flash
[01:26:18] forward to think about
[01:26:21] the potential exponential improvement
[01:26:24] and I I still to be honest sometimes
[01:26:25] when people talk about like exponents
[01:26:27] exponents humans aren't good at thinking
[01:26:28] exponentially. I'm like yes that's true
[01:26:30] but outside of more laws why would AI
[01:26:34] capabilities or LLM parameters or
[01:26:36] however you want to measure it
[01:26:37] automatically improve in exponents. I
[01:26:40] don't I don't actually quite understand
[01:26:41] that. But once we get to the sort of
[01:26:43] recursive self-improvement, it's like,
[01:26:44] okay, I can see how that starts to
[01:26:46] approach a vertical wall.
[01:26:47] >> I agree with you. I think one one
[01:26:49] experience from writing the book is
[01:26:50] simply that when you're close to the
[01:26:51] people inside the labs and, you know, I
[01:26:55] wasn't just Demis. I interviewed, you
[01:26:56] know, hundred of these AI insiders, you
[01:26:59] realize that the stuff in the pipeline
[01:27:01] is enormous.
[01:27:02] >> Yeah.
[01:27:02] >> I think there's a kind of popular
[01:27:04] misconception which is there is this
[01:27:05] thing called AI and it kind of happened
[01:27:08] when Chatty PT came out. So now we've
[01:27:10] got it and we're kind of getting used to
[01:27:12] it and that's in the rearview mirror.
[01:27:14] No, no, no, no. This thing is changing
[01:27:16] the whole time as anybody who looks
[01:27:18] closely knows. And if you think back,
[01:27:21] the progression is wild. You know, you
[01:27:23] get this system in end of 2022 which
[01:27:26] hallucinates non-stop. Then you plug in
[01:27:28] GPT4 16 months later, whatever it was,
[01:27:32] and the hallucination radically reduces.
[01:27:35] Then it goes multimodal, so it can do
[01:27:38] video and audio. And in the meantime,
[01:27:40] it's got a very long context window, so
[01:27:43] you can plug in an entire TL story novel
[01:27:46] and ask questions about it. Then it
[01:27:48] starts to do the reasoning stuff and can
[01:27:51] do logic and math. Then it becomes
[01:27:54] agentic.
[01:27:56] Then it's like coding for you. And all
[01:27:59] of these changes are packed into three
[01:28:01] and a half years. And I agree with you.
[01:28:03] I think the next three and a half years
[01:28:05] are going to be even more wild.
[01:28:06] >> Yeah.
[01:28:07] >> So I think there's a big gap between the
[01:28:09] inside and the outside view of this.
[01:28:10] >> Yeah. That's where these comparisons to
[01:28:12] the industrial revolution just
[01:28:13] completely fall apart [laughter] on so
[01:28:16] many levels. I have one or two remaining
[01:28:18] questions for you.
[01:28:19] >> The billboard question. I ask this a
[01:28:22] lot. It can be a fun one.
[01:28:23] >> If you could put anything on a
[01:28:26] billboard, metaphorically speaking for
[01:28:28] millions, billions of people to see.
[01:28:30] Could be anything. image quote question
[01:28:34] preferably not commercial. [laughter]
[01:28:38] What would it be? What might it be?
[01:28:40] >> So a billboard which lots of people are
[01:28:43] going to see. I would put prepare your
[01:28:47] mind.
[01:28:48] >> This is a saying which is originally
[01:28:52] Louis Pastor I think the scientist
[01:28:55] who said chance favors the prepared
[01:28:58] mind. If you're ready for things, you
[01:29:01] can make the most of the opportunity
[01:29:02] that comes your way. And the amazing
[01:29:05] thing about this saying is that it's
[01:29:06] come up randomly in different contexts
[01:29:09] in different books I've done. So when I
[01:29:12] was writing about venture capital, Excel
[01:29:15] Capital
[01:29:16] >> and one of the founders, Arthur
[01:29:17] Patterson, used this phrase as a
[01:29:20] description of how he wanted Excel to
[01:29:23] invest. that they would run these kind
[01:29:26] of scenario exercises where they would
[01:29:28] think, okay, there's a new technology
[01:29:29] coming down the pike. What kind of
[01:29:32] company needs to be built to make the
[01:29:34] most of that new platform? What type of
[01:29:37] entrepreneur is going to fit this
[01:29:39] opportunity? What should we be expecting
[01:29:42] so that the person walks into the office
[01:29:44] into the conference room and pitches to
[01:29:45] us, we already know 90% of what he says
[01:29:48] because we've prepared our minds. And
[01:29:50] that way we can make a good judgment and
[01:29:52] a fast judgment if it's a competitive
[01:29:54] situation. So I kind of wrote about the
[01:29:56] prepared mind in the context of venture
[01:29:57] capital. And then I'm doing the infinity
[01:30:00] machine and I'm interviewing Ilia
[01:30:01] Satskaver from OpenAI and I'm asking him
[01:30:04] why was it you who understood the
[01:30:07] significance of the transformer
[01:30:09] architecture when it came out
[01:30:12] immediately like on the day it was up on
[01:30:13] the website you read it. You ran down
[01:30:15] the corridor. You went to see your
[01:30:18] collaborator Alec Radford and you said,
[01:30:20] "We're going to build a language model
[01:30:21] on top of this architecture."
[01:30:22] >> Well, not only that, he said, "Stop
[01:30:24] everything you're doing."
[01:30:25] >> Right. Right. Right.
[01:30:26] >> And do this. [laughter]
[01:30:27] >> Yeah. This vision of the kind of, you
[01:30:29] know, overcaffeinated charismatic
[01:30:31] seizing on the engineer and saying,
[01:30:33] "Drop it, whatever you're doing." And,
[01:30:36] you know, in his answer was prepared
[01:30:37] mind that he'd been thinking about how
[01:30:39] you model sequential data ever since his
[01:30:42] PhD in Canada. And when he saw the
[01:30:46] solution, this was what he'd been
[01:30:48] waiting for for like a decade. And so he
[01:30:51] could jump on it. And then when you
[01:30:53] start thinking about prepared mind, you
[01:30:55] know, you would probably remember this
[01:30:56] better than I do, but wasn't there a um
[01:30:58] Seattle Seahawks Super Bowl final
[01:31:01] against the New England Patriots where
[01:31:03] the New England quarterback like does an
[01:31:05] interception in the last second of play
[01:31:09] and clinches the victory. And when he's
[01:31:11] asked after the play, how did you know
[01:31:14] to make that run? Where did you how did
[01:31:16] you know where the quarterback was going
[01:31:17] to throw the ball? The answer was
[01:31:19] prepared mind. Basically, he didn't use
[01:31:20] that phrase, but you know, in training,
[01:31:23] they had studied
[01:31:25] the play that the Seattle Seahawks were
[01:31:27] going to make. And they knew that given
[01:31:29] a certain formation when the ball was
[01:31:31] snapped back, there was a certain pass
[01:31:33] that was coming. So the guy just takes
[01:31:34] off and he runs right into where the
[01:31:37] ball comes and he catches it and
[01:31:38] intercepts and New England wins. And so
[01:31:41] that's a prepared mind in sports.
[01:31:43] >> Mhm.
[01:31:44] >> And the other reason, last thing,
[01:31:45] >> yeah,
[01:31:46] >> I would put on the billboard prepare
[01:31:47] your mind is that for the age of
[01:31:49] artificial intelligence, this is what we
[01:31:52] need to hear. And this is a serious
[01:31:53] point, right? The risk with large
[01:31:56] language models is that we just get lazy
[01:31:59] and whenever we need to know something,
[01:32:01] we just get it to tell us what to think.
[01:32:04] That is not the route to happiness or
[01:32:06] satisfaction or anything. We need to
[01:32:10] continue to do the hard work of
[01:32:11] preparing our minds because that's what
[01:32:14] makes us people. You know, I think,
[01:32:16] therefore I am. And so I think prepare
[01:32:19] your mind is entering a time when it
[01:32:22] becomes a more important slogan than
[01:32:24] ever.
[01:32:25] >> How do you do that for yourself? What
[01:32:27] guard rails or policies have you
[01:32:30] established for your own use of AI?
[01:32:33] >> And it makes me also think of going to
[01:32:35] the gym, lifting weights, getting in
[01:32:37] cardio. You don't have to do that, but
[01:32:40] it is beneficial for you on a lot of
[01:32:42] levels. And people, some people find it
[01:32:44] quite enjoyable, right? And hence they
[01:32:46] do that. And I'm wondering
[01:32:49] what the equivalent is for knowledge
[01:32:52] workers or people who are preparing
[01:32:55] their minds and
[01:32:58] don't want to become sort of impotent in
[01:33:01] the way that people with directions have
[01:33:03] mostly become impotent because of Google
[01:33:05] maps and other tools like that. Right?
[01:33:07] So what do you what do you do for
[01:33:08] yourself personally or how are you
[01:33:10] thinking about that? The first thing I
[01:33:12] think is that the Google Maps analogy is
[01:33:15] the wrong one in the sense that it's
[01:33:18] fine to offload a very specific mental
[01:33:21] task which to most people is a pain in
[01:33:23] the neck.
[01:33:24] >> Mhm. [clears throat]
[01:33:24] >> And let the machine do that for you.
[01:33:26] It's not fine to offload all thinking.
[01:33:30] Right. The point of offloading something
[01:33:32] should be you get to focus your mental
[01:33:35] energy more on the other stuff that you
[01:33:38] really get satisfaction and meaning
[01:33:40] from. And so for me, what that means is
[01:33:42] that I'm very happy to use large
[01:33:45] language models to learn about the
[01:33:48] scientific output of somebody I'm going
[01:33:50] to interview next week.
[01:33:52] >> Mhm. All of these AI papers are on
[01:33:55] archive and the model has ingested all
[01:33:58] of them and the model is extremely good
[01:34:00] at telling me okay the scientist you're
[01:34:02] seeing next week has these three papers
[01:34:05] and the progression between the three
[01:34:07] papers is this and this and this and the
[01:34:09] comparison with the person you saw two
[01:34:11] weeks ago is this and this and this and
[01:34:14] you know you learn a lot from the system
[01:34:16] like really bootstraps you to learn
[01:34:18] faster so that's helping me to think
[01:34:20] more not to think less.
[01:34:23] >> It's cutting out the time it would take
[01:34:25] me to go find all the papers by myself
[01:34:27] and then labor through them. It's
[01:34:29] cutting to the chase and nourishing me
[01:34:32] intellectually.
[01:34:33] >> And [clears throat] by the way, I'm not
[01:34:34] worried about hallucination because I'm
[01:34:35] going to interview the human scientist
[01:34:38] anyway. So, I get to cross-check it all.
[01:34:41] >> What I would never do is get the AI to
[01:34:43] write because frankly, it's not very
[01:34:46] good at long form. In fact, it really
[01:34:48] sucks. It's fine for writing an email,
[01:34:51] although I don't do that either because
[01:34:53] I like writing. But it really is I've
[01:34:56] tried it once. It's terrible for
[01:34:58] anything longer than about 800 words.
[01:35:01] But even if it could do it, I don't
[01:35:03] think I would ever outsource that
[01:35:04] because that's me,
[01:35:06] >> right? This is what I do. This is the
[01:35:08] thinking process. I think through my
[01:35:10] writing, I come to understand what I
[01:35:13] understand and think what I think and
[01:35:15] believe what I believe through writing.
[01:35:17] And I'm not going to give that out.
[01:35:20] >> I'm letting out a pensive exhale because
[01:35:24] I was thinking of this. A friend said to
[01:35:26] me, well, I'll give him credit, Kevin
[01:35:28] Rose. At one point, I was I wouldn't say
[01:35:32] complaining, observing that AI couldn't
[01:35:34] do X or it wasn't very good at Y.
[01:35:37] >> He said, when was the last time you
[01:35:38] tried that? I was like six months ago.
[01:35:40] And he's like, try it again. And so
[01:35:43] [laughter]
[01:35:43] the rules will become really important
[01:35:45] as also the power of these things
[01:35:47] increases. And there I want to say it
[01:35:49] was the New Yorker. There was a piece in
[01:35:51] the New York or it might have been the
[01:35:52] New York Times with some very famous I
[01:35:55] want to say novelist could have been
[01:35:57] Pulitzer Prize winner in literature
[01:35:58] somebody at the top and they took three
[01:36:01] or four pieces of their own writing had
[01:36:03] AI generate three or four pieces of
[01:36:06] writing in their voice and gave it to
[01:36:08] professional readers
[01:36:10] editors and so on and it wasn't clear
[01:36:14] people couldn't figure out they claimed
[01:36:16] that what he or she wrote was AI
[01:36:18] >> how long was the piece of writing
[01:36:20] >> I knew that was the question you were
[01:36:21] going to ask and I and I don't recall.
[01:36:23] So I want to go back and look at that
[01:36:24] piece to see. So there was a story
[01:36:27] precisely like that from an economist
[01:36:30] writer who's very funny and also does
[01:36:32] podcasts
[01:36:33] >> and he ran that experiment and it was
[01:36:35] just as you said you know his friends
[01:36:37] who were professional economist
[01:36:39] journalists couldn't tell which was the
[01:36:41] witty column that he'd written versus
[01:36:43] the equally witty ones which the lamb
[01:36:46] had generated and he was very pissed off
[01:36:48] with this and I look I take your point I
[01:36:51] mean for now I can be all complacent and
[01:36:54] say yeah I only works for 800 words. It
[01:36:57] doesn't work for a whole chapter which
[01:36:59] is 20 pages long. But no doubt it'll get
[01:37:02] better and better. But I still think I'm
[01:37:03] going to cling on to the thing that
[01:37:05] makes me me for sure. 100%. And I think
[01:37:10] doing the thinking, preparing your mind
[01:37:14] in part asking that question, which is
[01:37:17] not an easy question, perhaps there's a
[01:37:18] different way to phrase it, but like
[01:37:20] what what are
[01:37:22] the things that make me me? So you don't
[01:37:25] accidentally make sacrifices that start
[01:37:28] to erode your sense of self but also
[01:37:33] sense of selfworth. Right.
[01:37:35] >> Preparing your mind. Sebastian,
[01:37:37] everybody should check out the infinity
[01:37:38] machine. It's it's outstanding. The
[01:37:41] infinity machine subtitle deis habis
[01:37:43] deep mind and the quest for super
[01:37:44] intelligence. And lest people
[01:37:48] make the wrong assumption. This is not
[01:37:51] here's the latest and greatest in AI. It
[01:37:53] is the story of an incredible mind,
[01:37:57] a whole cast of kooky and fascinating
[01:38:00] characters. It is about a noble quest.
[01:38:04] It's about the pitfalls and promises of
[01:38:08] entrepreneurship. It contains so many
[01:38:10] different levels. And if you want to
[01:38:13] also have a basic understanding of what
[01:38:16] it is from the ground up that came to be
[01:38:19] colloquially referred to as AI or LLMs,
[01:38:22] this is a great book for that. It really
[01:38:24] lays out kind of the nuts and bolts and
[01:38:26] how this evolved over time in a way that
[01:38:28] I think is intelligible to
[01:38:30] non-engineers. So everybody should check
[01:38:33] out the Infinity Machine. Sebastian, is
[01:38:34] there anywhere else you would like to
[01:38:36] point people or anything else you'd like
[01:38:38] to say as we wind to a close?
[01:38:41] Well, um,
[01:38:44] yeah, you stopped me on that one.
[01:38:46] [laughter]
[01:38:48] I've enjoyed the conversation. I'm happy
[01:38:50] to leave it there. Thank you for doing
[01:38:51] it, Tim. It's been great.
[01:38:52] >> Absolutely. I'll I'll give one one more
[01:38:54] link for folks if they want to find you
[01:38:56] on X. That's SC Malib
[01:39:00] Malibby. Well, Sebastian, thank you so
[01:39:02] much for the time. Really enjoyed the
[01:39:05] conversation. And for people listening,
[01:39:08] we will include links to everything
[01:39:10] we've discussed, all the characters and
[01:39:13] everything else at tim.blog/mpodcast.
[01:39:15] Just search Sebastian. I'm pretty sure
[01:39:17] that Oh, actually, we have Sebastian
[01:39:19] Younger. So, there are two Sebastians.
[01:39:20] But if you search Malib, M A L L A B Y,
[01:39:23] it'll be very easy to find this. And
[01:39:25] until next time, be just a bit nicer
[01:39:28] than is necessary, a little bit kinder
[01:39:30] than is necessary to others, but also to
[01:39:32] yourself and prepare your mind. Thanks
[01:39:35] for tuning in.