# Talking With Azeem Azhar

An episode of [Paul Krugman Podcast](https://invisiblenetwork.ai/podcast/1877144831), 13 Jun 2026, 47 min.

Audio: https://api.substack.com/feed/podcast/201847235/3d961c704dc40c81c18d8c95992eec54.mp3

I last spoke with Azeem, the proprietor of Exponential View, 18 months ago — ancient history on this subject. So we revisited the state of AI. . TRANSCRIPT: Paul Krugman in Conversation with Azeem Azhar (recorded 6/12/26) Paul Krugman: Hi everyone. Paul Krugman back on my usual schedule of recording interviews. And today I’m talking with Azeem Azhar, who I spoke to in January 2025, basically centuries ago in AI time. And with AI on everybody’s mind, I thought it would be good to revisit. I should say Azeem is an independent researcher and founder of Exponential View, which is one of the top tech Substacks out there. So hi, welcome to another conversation. Azeem Azhar: Yeah, thank you, Paul. And it has been eighteen months, also known as one and a half centuries in AI time since we spoke. Krugman: Yeah. Let me ask sort of the dumbest question: what is this thing called AI? How does it do what it does? I mean, even skeptics have to admit that it’s really impressive how it’s sort of leapt over all of the previous barriers. How is this happening? Azhar: You know, I think we’re still figuring it out. I think of AI ultimately as a machine that does certain things, and it’s been built by passing first millions, then billions, then tens of billions, hundreds of billions of trillions of words of human output through a neural network to give it some sense of how humans have thought about the world. And because it operates at dimensions well beyond the form of space and time, it seems to be able to find relationships between quite complex concepts. And I think we’ve all had that experience, whether we’ve been using Chat GPT or Claude over the last two or three years, that it seems to be able to recognize things that are quite deeply related that don’t immediately spring to mind. And in the last year and a half or so, the labs have started to train the AI models not just on words in books, but actually on tasks, like, “what is the set of things that you do to write a piece of code that does something?” “What is a set of things you do to use a piece of software in an enterprise?” And they’ve tried to train those models on those particular tasks. Essentially it’s aping what we do, and they use various mathematical tools like reinforcement learning where the model notionally gets a reward. Of course it’s not a reward the way you and I think of it because it’s a machine. Paul Krugman: Right. Azhar: And so that’s what it is. It’s sort of reflecting back, but also I think discovering some really deep relationships in the world that we might not spot, you know, prima facie as humans. Paul Krugman: Brad Delong calls it “a vast stew of linear algebra,” which makes some sense to me because I think that Pagerank with Google was the last thing I actually understood. And that’s the eigenvector with the largest eigenvalue. Not that anybody needs to know that, but this is like a million times bigger, right? Azhar: That’s basically it. Yeah. Krugman: But it’s sort of not what artificial intelligence was supposed to be, right? Azhar: No, not at all. I mean, I sometimes go back and look at the TV series of the seventies that I grew up with as a child, and they’ll always have an AI in the spaceship. Space 1999 had an AI you could talk to. And it was very precise, it was very clipped, and it did things and got things right. And there was a sense that you could trust it. But you’d never think to say, as I sometimes do now, you know, “Find me five analogies to help make this point.” I use it as a brainstorming partner, or I give it tracts of my book, the book that I’m writing, and say, you know, “How would Paul Krugman criticize this argument?” And I get suggestions that I then work through by hand? I don’t think we really imagined it would look like that. Krugman: Yeah. In sci-fi it would talk in a monotone and would be relentlessly logical. And in fact these models are unpredictable, they’re sometimes temperamental, they’re not reliable.…

## Transcript

[0:00] SPEAKER_00: Hi, everyone.
[0:01] SPEAKER_00: Paul Krugman back on my usual schedule of recording interviews.
[0:06] SPEAKER_00: And today I'm talking with Azim Azhar,
[0:08] SPEAKER_00: who I spoke to in January 2025,
[0:12] SPEAKER_00: basically centuries ago in AI time.
[0:15] SPEAKER_00: And with AI on everybody's mind,
[0:17] SPEAKER_00: I thought it would be good to revisit,
[0:20] SPEAKER_00: talk about some of the basics,
[0:21] SPEAKER_00: then talk about,
[0:22] SPEAKER_00: you know,
[0:23] SPEAKER_00: where we are now in this whole incredible story.
[0:26] SPEAKER_00: So hi, welcome to another conversation.
[0:29] SPEAKER_01: Yeah, thank you, Paul.
[0:30] And it has been 18 months,
[0:32] also known as one and a half centuries in AI time since we
[0:35] spoke.
[0:36] SPEAKER_00: Yeah,
[0:37] SPEAKER_00: I should say Azim is an independent researcher,
[0:40] SPEAKER_00: founder of Exponential View,
[0:42] SPEAKER_00: which is both,
[0:42] SPEAKER_00: you know,
[0:43] SPEAKER_00: among other things,
[0:43] has an extremely valuable Substack and one of the top tech
[0:47] substacks out there.
[0:48] SPEAKER_00: So anyway, so hopefully can cure some of my ignorance here.
[0:54] SPEAKER_00: Let me ask what the sort of
[0:56] SPEAKER_00: Dumbest question, what is this thing called AI?
[1:00] SPEAKER_00: How does it do what it does?
[1:01] SPEAKER_00: I mean,
[1:02] even skeptics have to admit that it's really impressive how
[1:05] it's sort of leapt over
[1:06] SPEAKER_00: all of the previous barriers.
[1:08] SPEAKER_00: How is this happening?
[1:09] SPEAKER_01: You know, I think we're still figuring it out.
[1:11] SPEAKER_01: I think of AI ultimately as a machine that does things.
[1:15] SPEAKER_01: And how has it been built?
[1:16] SPEAKER_01: It's been built by passing first millions,
[1:19] SPEAKER_01: then billions,
[1:20] SPEAKER_01: then tens of billions,
[1:21] SPEAKER_01: hundreds of billions of trillions of words of human output through
[1:25] through a neural network to give it some sense of how we have
[1:30] thought about the world.
[1:31] And because it operates at dimensions well beyond the four of
[1:36] space and time,
[1:37] SPEAKER_01: it seems to be able to find relationships between
[1:40] SPEAKER_01: Paul Krugman
[1:58] the labs have started to train the ai models not just on words
[2:04] in books but
[2:05] actually on on tasks on the you know what is the the set of
[2:09] things that you do to
[2:11] SPEAKER_01: write a piece of code that that that
[2:14] What is a set of things you do to use a piece of software in an
[2:17] enterprise and try
[2:19] SPEAKER_01: to train those models on those particular tasks?
[2:22] SPEAKER_01: Essentially,
[2:22] it's aping what we do and they use various mathematical tools
[2:27] like reinforcement
[2:27] SPEAKER_01: learning,
[2:28] SPEAKER_01: Paul Krugman paulkrugman
[2:47] SPEAKER_00: Okay.
[2:47] SPEAKER_00: I mean,
[2:48] SPEAKER_00: it's pretty,
[2:49] SPEAKER_00: I guess,
[2:50] SPEAKER_00: Brad DeLonge calls it a vast stew of linear algebra,
[2:52] SPEAKER_00: which I guess is kind of right.
[2:54] SPEAKER_00: That's basically it, yeah.
[2:55] It makes some sense to me because I think that page rank with
[2:58] Google was the last
[2:59] SPEAKER_00: thing I actually understood.
[3:01] SPEAKER_00: That's the eigenvector with the largest eigenvalue.
[3:04] SPEAKER_00: Not that anybody needs to know that.
[3:06] And this is somehow like a thousand times or a million times
[3:10] bigger version of this,
[3:11] right?
[3:11] SPEAKER_00: So, okay.
[3:12] Yeah,
[3:12] it's pretty important,
[3:14] but it's not at all what people used to imagine you would do.
[3:18] SPEAKER_00: It's not what artificial intelligence was supposed to be, right?
[3:22] SPEAKER_01: No, not at all.
[3:23] SPEAKER_01: I mean,
[3:23] SPEAKER_01: if you go back,
[3:24] I sometimes go back and look at the TV series of the 70s that I
[3:27] grew up with as a
[3:28] SPEAKER_01: child,
[3:28] SPEAKER_01: and they'll always have an AI in the spaceship.
[3:32] SPEAKER_01: Space 1999 had an AI you could talk to, and it was very precise.
[3:37] it was very clit and it did things and got things right and
[3:41] there was a sense that
[3:42] you could trust it but you'd never think to say as I sometimes
[3:47] do find me five
[3:49] analogies to make this point as a brainstorming partner or I
[3:53] give it tract of my
[3:55] book the book that I'm writing and say how would Paul Krugman
[3:59] criticise this
[4:00] argument and I get suggestions that I then work through by hand
[4:04] I don't think we
[4:04] SPEAKER_01: really imagined it would look like that
[4:07] SPEAKER_00: Yeah, and you're right.
[4:09] In sci-fi,
[4:09] it was always,
[4:10] it was,
[4:11] we're talking a monotone and would be relentlessly logical.
[4:13] SPEAKER_00: And in fact, these things are, the models are unpredictable.
[4:17] SPEAKER_00: They're sometimes seem to be temperamental.
[4:19] SPEAKER_00: They're not reliable.
[4:20] That's one of the big problems is that,
[4:23] so it's sort of almost,
[4:24] it's wild.
[4:25] SPEAKER_00: It's not at all what we imagined.
[4:27] it's not at all and this point about reliability is so complex
[4:31] I found that a
[4:33] couple of months back one of the versions of Anthropics Claude
[4:35] came out and I found
[4:37] it so sycophantic that it became unhelpful because I like these
[4:41] things to help me
[4:43] on hard problems and to challenge me so I switched back to chat
[4:48] GPT which has
[4:49] always been a little bit more maybe a bit less friendly and
[4:52] what's going on there
[4:53] SPEAKER_01: Paul is that
[4:55] because we don't really have a good theory about how to build
[4:58] these they are
[5:01] developed almost like on a petri dish and nudged in particular
[5:06] directions so they
[5:07] take the shape that we expect them to take and to use an
[5:11] economist term they
[5:13] SPEAKER_01: improve non-monotonically with every
[5:17] SPEAKER_01: So you'll see the latest release of an anthropic model.
[5:20] SPEAKER_01: And the way we figure out how good a model is,
[5:23] is there are maybe 20 or 30 public benchmarks that they're
[5:26] measured against,
[5:27] like how well do they summarize text and how well do they write
[5:30] software code.
[5:31] And the next version of the model won't necessarily be better
[5:35] at everything than
[5:36] SPEAKER_01: the previous version because you lose something in order to get it.
[5:39] SPEAKER_01: And that's the complexity that the labs are wrestling with.
[5:43] SPEAKER_00: Wow.
[5:44] SPEAKER_00: Okay.
[5:45] SPEAKER_00: Yeah.
[5:46] SPEAKER_00: I don't think there's any historical parallel to that,
[5:48] to a technology that you really don't understand and that you
[5:51] really aren't fully
[5:52] SPEAKER_00: in control of.
[5:53] SPEAKER_00: Wow.
[5:53] SPEAKER_00: All right.
[5:54] SPEAKER_00: Second naive question.
[5:56] SPEAKER_00: So I'm actually, I don't think I'm a Luddite.
[5:59] SPEAKER_00: I'm not,
[6:00] SPEAKER_00: I've always been happy to adopt technologies,
[6:03] but I'm also not a hobbyist or I'm not maybe incurious on some
[6:09] of these things.
[6:10] SPEAKER_00: I tend to pick up techniques
[6:12] technologies but also things like mathematical techniques as
[6:16] needed because I see
[6:16] something to do and I'm feeling kind of inadequate because I
[6:20] keep on not finding
[6:21] anything to do with the more elaborate I'm using notebook LM to
[6:28] extract tables from
[6:29] PDFs that sort of thing but what should I be doing I do have
[6:35] people friends who are
[6:36] using Claude a lot but I can't quite figure out what
[6:39] particularly agentic AI should
[6:40] SPEAKER_00: be doing for me
[6:42] Yeah,
[6:43] you know,
[6:43] I'm really sympathetic to that because I have the same issue.
[6:47] These tools have been developed by software developers in a
[6:51] really particular part
[6:52] SPEAKER_01: of the world,
[6:53] SPEAKER_01: which is Silicon Valley,
[6:54] SPEAKER_01: where the culture really revolves around the art of the
[6:57] SPEAKER_01: paulkrugman paulkrugman
[7:24] I mean the way that I get around this is that I do look at
[7:28] things on on Twitter or
[7:29] X as it's called because people are sharing tips and I often
[7:32] just ask the model you
[7:33] know what could I do with you given that I'm trying to do this
[7:37] thing I'm trying to
[7:38] solve this problem and that will come back and give me a
[7:41] suggestion I have had some
[7:43] success with agents so I have an agent called R Mini Arnold so
[7:47] R is a pie into
[7:49] SPEAKER_01: Isaac Asimov's robots they're all called
[7:51] SPEAKER_01: Paul Krugman
[8:06] SPEAKER_01: paulkrugman.com
[8:35] SPEAKER_01: a lot of my life's context.
[8:37] SPEAKER_01: So it knows the music I like,
[8:38] SPEAKER_01: it knows the book I'm working on,
[8:40] SPEAKER_01: it knows the investments I'm making,
[8:41] SPEAKER_01: it knows the essays I'm doing,
[8:43] SPEAKER_01: it's got the calendar of speeches that I'm about to give.
[8:46] SPEAKER_01: And so when it goes off and does a task, it tries to figure out
[8:50] What in my world is this going to be relevant to and where can
[8:53] I draw threads from?
[8:54] And when it works,
[8:55] it is really sublime and it does feel a little bit like science
[8:59] fiction.
[8:59] SPEAKER_01: But I would say it's incredibly brittle.
[9:02] SPEAKER_01: I mean, there's breaks every four or five days.
[9:05] SPEAKER_01: So what would a task be?
[9:06] SPEAKER_01: I mean, what might you ask it to do?
[9:09] SPEAKER_01: So a good example was I was thinking about the
[9:15] Devine and Paul David research about why electrification took
[9:19] the time it took.
[9:20] And I wanted to understand what were the determinants of that
[9:27] 35-year lag from
[9:28] Pearl Street generation to productivity growth from the
[9:34] perspective of a model,
[9:36] SPEAKER_01: right?
[9:36] SPEAKER_01: What could the levers be?
[9:37] SPEAKER_01: And so
[9:39] I threw that into our mini Arnold and it set up a team of
[9:44] sub-agents which had
[9:46] personalities of key economists and was able to go off and do
[9:50] research the way the
[9:50] SPEAKER_01: AIs do,
[9:51] SPEAKER_01: but also research on all the academic positions.
[9:53] SPEAKER_01: papers that I have downloaded in the past.
[9:56] SPEAKER_01: So I have access to JSTOR.
[9:58] SPEAKER_01: I'm allowed to download 100 PDFs a month.
[10:02] It can look at all of those and start to compile an answer in a
[10:06] way that perhaps a
[10:07] SPEAKER_01: chat GPT can't.
[10:09] SPEAKER_01: Paul Krugman
[10:39] SPEAKER_00: Yeah, all right.
[10:39] SPEAKER_00: I guess I'm getting it.
[10:41] SPEAKER_00: Okay, now my question on that is still on very basic stuff.
[10:45] SPEAKER_00: So obviously, you and I are not typical, are not the market.
[10:51] SPEAKER_00: The people who are going to use this are going to be
[10:55] SPEAKER_00: middle managers, independent business people.
[10:58] SPEAKER_00: I guess,
[10:59] SPEAKER_00: you know,
[10:59] SPEAKER_00: in some way,
[10:59] SPEAKER_00: both of us are kind of independent business people,
[11:01] SPEAKER_00: but not typical ones.
[11:02] SPEAKER_00: And I find myself,
[11:04] SPEAKER_00: when I hear that lots of people are piling into that,
[11:08] I find myself thinking about what I think of as the homemade
[11:11] pasta problem.
[11:12] so you you you're probably too young for this but you know
[11:15] there was a time when I
[11:17] when you know when I was young and we were using stone axes for
[11:20] computing when
[11:21] there was a big fad of making your own pasta pasta machines
[11:25] were everywhere and
[11:26] then at a certain point it was kind of a collective what the
[11:29] hell are we doing you
[11:30] know store-bought pasta is actually better if the Italians
[11:34] don't do this and I have
[11:35] SPEAKER_00: to think that for most the range of
[11:39] tasks the range of agents in effect that that's independent
[11:44] business people can't
[11:45] be that wide there are people aren't that different and that
[11:48] that why aren't there
[11:49] sort of off the rack off the shelf what won't there just be
[11:53] agents being sold why
[11:54] would people be vibe coding why wouldn't they just take one of
[11:57] these off the shelf
[11:57] SPEAKER_00: things as I might
[11:59] SPEAKER_01: Yeah,
[12:00] SPEAKER_01: well,
[12:00] I think it's different for an independent person or a small
[12:02] business to a middle
[12:04] SPEAKER_01: manager in a big company.
[12:06] I would imagine that you will start to see people selling
[12:11] specific agents that
[12:13] solve your marketing problem if you're a barber's shop and
[12:18] you've got four chairs
[12:19] SPEAKER_01: and you have maybe 30 people a day coming through.
[12:23] Right now what you do is you go to ChatGPT and you help it
[12:25] write your collateral
[12:26] SPEAKER_01: for your website.
[12:28] That feels like it's an interim step to somebody delivering the
[12:33] actual finished product.
[12:35] SPEAKER_01: Why haven't we seen it?
[12:36] SPEAKER_01: I think we haven't seen it yet because
[12:40] SPEAKER_01: The terrain is still big enough.
[12:41] SPEAKER_01: So all the companies who are,
[12:43] SPEAKER_01: you know,
[12:43] SPEAKER_01: they're not the anthropics and the open AIs,
[12:45] but there's a lot of other companies building agents that are
[12:48] these end to end
[12:48] SPEAKER_01: workflows for businesses.
[12:50] They still believe that the prize for them is to build the
[12:54] generic platform.
[12:55] SPEAKER_01: Paul Krugman paulkrugman
[13:12] SPEAKER_01: Paul Krugman paulkrugman
[13:41] and in that instance I think it's much much harder to see how
[13:46] do you use AI to
[13:47] SPEAKER_01: really really change the way you work
[13:49] SPEAKER_00: Yeah,
[13:50] SPEAKER_00: I mean,
[13:50] SPEAKER_00: you know,
[13:51] SPEAKER_00: again,
[13:51] SPEAKER_00: we're talking about ancient history here,
[13:52] SPEAKER_00: but,
[13:53] SPEAKER_00: you know,
[13:53] SPEAKER_00: everybody still uses Excel,
[13:55] SPEAKER_00: even though it has always been horrible.
[13:58] But,
[13:59] you know,
[14:00] the constraints of corporate life mean that sort of everybody
[14:04] has to use Excel.
[14:04] SPEAKER_00: So that means,
[14:05] SPEAKER_00: again,
[14:06] SPEAKER_00: that we might,
[14:07] SPEAKER_00: I don't know,
[14:08] maybe we'll see quite a lot less coding a few years down the
[14:11] pipe because people
[14:12] SPEAKER_00: will just be able to purchase whatever it is they need.
[14:14] SPEAKER_00: I don't know.
[14:15] SPEAKER_01: I think it's a balance.
[14:18] SPEAKER_01: Often when you hear people who are proselytizing really heavily,
[14:21] and I think this technology is going to be really impressive
[14:24] and have significant
[14:26] SPEAKER_01: impact.
[14:26] SPEAKER_01: But when people pitch this,
[14:29] they actually forget that there are other actors in the market
[14:34] who might respond to
[14:35] SPEAKER_01: whatever else is going on.
[14:36] SPEAKER_01: And I do think that right now,
[14:39] SPEAKER_01: if you're a large company,
[14:40] SPEAKER_01: you do want to be building as much as you can because what you can buy is
[14:45] isn't isn't right isn't right for the market I mean it took if
[14:49] you think about
[14:50] being you know Henry Ford and you're putting together the
[14:53] Highland Park plant you
[14:55] can't go to a supply chain and buy something because nobody was
[14:57] thinking in those
[14:58] terms and I think we are slightly at that stage for large
[15:00] corporates now whether
[15:02] we'll be there in five years I I don't know and I think that
[15:06] one one thing that we
[15:07] SPEAKER_01: have to really think about is how does
[15:12] where will the value actually reside between having your own
[15:17] capabilities to design
[15:18] SPEAKER_01: software for your processes and handing that over to
[15:24] SPEAKER_01: another company who's designing software for 100 businesses like yours.
[15:28] And historically,
[15:29] it's made more sense to hand it over to another company.
[15:32] But it may just be that the cost curves have changed
[15:36] sufficiently that you'd rather
[15:38] have the nuance and the control and do whatever vibe coding
[15:42] becomes in 2030.
[15:44] SPEAKER_00: Okay,
[15:45] SPEAKER_00: yeah,
[15:45] SPEAKER_00: I mean,
[15:48] SPEAKER_00: countering my own kind of view,
[15:50] SPEAKER_00: I do know that on some things that I know about,
[15:53] SPEAKER_00: like healthcare software,
[15:55] organizations like the VA that built their own have done much
[15:59] better than the ones
[16:00] SPEAKER_00: who tried to buy it from Microsoft.
[16:01] SPEAKER_00: So, yeah, that might be a story that makes sense.
[16:07] SPEAKER_00: Actually,
[16:07] since we're talking about going for the model of models versus
[16:11] something much more
[16:12] SPEAKER_00: specific,
[16:15] How should I think about the Chinese and all of this stuff,
[16:17] Chinese versus the big USA firms?
[16:21] SPEAKER_01: I've just spent eight days in China, and I was really fortunate.
[16:26] I got to speak to developers and engineers and management from
[16:29] about a dozen of the
[16:31] SPEAKER_01: Chinese labs.
[16:32] SPEAKER_01: In many cases, they hosted us in their offices.
[16:37] The main thing the Chinese say about the US is that Claude Code
[16:44] is brilliant and
[16:45] Claude is the best model that is out there and they really
[16:49] couldn't get enough of
[16:49] SPEAKER_01: it.
[16:50] SPEAKER_01: The term is they're Claude Pilled.
[16:53] SPEAKER_01: They
[16:54] SPEAKER_01: Talk about the constraints on getting access to compute,
[16:58] SPEAKER_01: but just in a way that that's a fact of life.
[17:00] SPEAKER_01: I mean,
[17:01] there's no sort of commentary on it other than it's hard and
[17:04] they have to figure
[17:04] SPEAKER_01: out how to get around that.
[17:05] And how do you build a culture of efficiency when you don't
[17:09] have as much compute?
[17:10] SPEAKER_01: And I think they've built a culture of efficiency really, really well.
[17:13] SPEAKER_01: I think it's going to help them over the longer term.
[17:18] SPEAKER_01: They don't really talk about a competition with U.S.
[17:21] SPEAKER_01: labs the way the U.S.
[17:23] SPEAKER_01: talks about competition with China.
[17:25] SPEAKER_01: They do see themselves competing with each other.
[17:27] SPEAKER_01: And as you know, that's what the Chinese economy is.
[17:31] SPEAKER_01: It's mayors in different cities who are
[17:33] SPEAKER_01: paulkrugman paulkrugman
[18:01] that everyone was using Claude for coding as opposed to,
[18:03] you know,
[18:04] the cheaper Chinese version.
[18:06] SPEAKER_00: That's interesting.
[18:07] SPEAKER_00: But so,
[18:07] SPEAKER_00: I mean,
[18:08] you can imagine a future maybe where a lot of businesses are
[18:12] actually using these
[18:14] SPEAKER_00: less,
[18:14] SPEAKER_00: less comprehensive,
[18:15] SPEAKER_00: but much cheaper models.
[18:17] SPEAKER_00: And, but that, that would, that's, you know, I think what I'm getting
[18:21] SPEAKER_00: from you and from other people is that a lot of entrepreneurs in the U.S.
[18:25] SPEAKER_00: are still dreaming of the Uber model that solves all problems,
[18:30] SPEAKER_00: but that that probably is not going to be as God knows,
[18:34] SPEAKER_00: but probably not going to be the way it all goes.
[18:36] SPEAKER_00: In the end,
[18:37] SPEAKER_00: we're going to end up with a lot of specialized models,
[18:40] SPEAKER_00: but also the Uber models will still have a role.
[18:43] SPEAKER_01: Yeah,
[18:43] it never made sense to me that you'd have a single model that
[18:46] would do everything
[18:47] SPEAKER_01: because the single model is,
[18:50] SPEAKER_01: you know,
[18:50] SPEAKER_01: if it's going to solve the Riemann hypothesis,
[18:53] SPEAKER_01: it's going to require a lot of resources.
[18:55] SPEAKER_01: And if all you need to do is get it to root...
[18:58] SPEAKER_01: a bill to the finance department.
[19:00] SPEAKER_01: It seems a bit silly to ask Einstein to come and do that for you.
[19:03] And,
[19:03] you know,
[19:04] we've had segmentation of markets for a long time and it's like
[19:09] airlines.
[19:09] SPEAKER_01: That's why not every seat in an airline is first class.
[19:12] SPEAKER_01: Some passengers don't want it, don't need it, won't want to pay for it.
[19:15] So I do think that the ecology looks like a whole array of
[19:20] much,
[19:20] much cheaper models that are
[19:24] Serving by volume lots of corporate needs and then having more
[19:29] sophisticated
[19:30] SPEAKER_01: complex models for the more complex tasks.
[19:33] SPEAKER_01: And that you're already, I think, starting to see.
[19:37] SPEAKER_01: I don't see it, by the way, as a shock.
[19:38] SPEAKER_01: Paul Krugman
[20:02] SPEAKER_00: Let's move to kind of more macro considerations.
[20:06] SPEAKER_00: So from the beginning,
[20:09] SPEAKER_00: in all of the couple of years of history that we have here so far,
[20:12] SPEAKER_00: people have worried about bubble.
[20:15] SPEAKER_00: And a lot of us,
[20:16] SPEAKER_00: sadly,
[20:16] a lot of us still remember the 90s quite vividly and think
[20:21] about all of that.
[20:22] SPEAKER_00: But you've been doing a lot on, you've been trying to sort of do
[20:28] a set of criteria that you can use and you've been pretty I
[20:31] think not dogmatically
[20:33] but just you just aren't seeing the bubble you want to talk
[20:35] about that I do I don't
[20:38] want to get remember what it was like in the 90s I lived
[20:41] through that one and also
[20:42] the housing bubble which frankly was was far far worse and much
[20:45] more much more
[20:46] terrifying the you know I have a really simple mantra here
[20:51] which is that honest
[20:53] SPEAKER_01: customer revenues are
[20:55] SPEAKER_01: Paul Krugman paulkrugman
[21:15] SPEAKER_01: Paul Krugman paulkrugman
[21:29] SPEAKER_01: paulkrugman
[21:45] SPEAKER_01: Paul Krugman paulkrugman
[22:05] SPEAKER_01: And that's both a large number,
[22:07] SPEAKER_01: it's a much faster revenue growth rate than mobile or the internet.
[22:12] SPEAKER_01: It's also a small number because the US is a $32 trillion economy.
[22:15] SPEAKER_01: And I think the thing is that at that level of spend,
[22:21] SPEAKER_01: Paul Krugman paulkrugman
[22:40] unpicks the bubble,
[22:42] a bubble,
[22:42] is when financing starts to get a bit smelly.
[22:46] SPEAKER_01: So that was clearly the case in the global financial crisis,
[22:49] SPEAKER_01: where you had all of these synthetic collateralized obligations that were
[22:54] SPEAKER_01: magnifying the
[22:54] SPEAKER_01: paulkrugman
[23:17] SPEAKER_01: Bad,
[23:18] poor or strong or robust is the funding quality and that
[23:21] funding quality measure is
[23:22] SPEAKER_01: definitely getting worse.
[23:23] SPEAKER_01: It's worse now than it was nine months ago,
[23:26] SPEAKER_01: but it doesn't feel,
[23:28] it doesn't seem from the numbers to be at the level that it has
[23:32] been historically
[23:33] SPEAKER_01: when these things have imploded.
[23:35] SPEAKER_01: and nor does it seem to be the type of exposure that is really systemic,
[23:40] SPEAKER_01: which is what we saw in the global financial crisis.
[23:45] SPEAKER_01: You know,
[23:45] SPEAKER_01: there are companies like Oracle and CoreWeave whose,
[23:49] SPEAKER_01: you know,
[23:49] SPEAKER_01: debt looks very,
[23:51] SPEAKER_01: very risky and it's harder and harder for them perhaps to raise money,
[23:54] SPEAKER_01: although Oracle just did.
[23:56] SPEAKER_01: But it doesn't feel like it's systemic.
[23:58] SPEAKER_01: You know, when the global financial crisis popped, no one knew
[24:02] SPEAKER_01: who was in trouble,
[24:03] whereas now you'd be able to isolate it with a single company
[24:06] or a single firm.
[24:07] SPEAKER_01: So at the moment,
[24:08] SPEAKER_01: we feel that this is still a demand led boom,
[24:13] SPEAKER_01: that funding quality has definitely got worse,
[24:15] but not so bad that I would say that there is an imminent
[24:19] problem on the horizon.
[24:21] You may know this number and I don't,
[24:22] but I mean,
[24:24] at this point,
[24:24] you're saying that basically
[24:27] SPEAKER_00: sort of non-duplicated,
[24:28] SPEAKER_00: roughly speaking,
[24:29] SPEAKER_00: final demand for this stuff is about half a percent of GDP.
[24:33] SPEAKER_01: And what share are sort of AI-related stocks in market value?
[24:38] SPEAKER_00: It has to be substantially larger than that.
[24:41] SPEAKER_01: They're about 40% right now of the S&P 500.
[24:45] SPEAKER_00: So that's a kind of a huge mismatch there.
[24:48] SPEAKER_00: I mean,
[24:48] SPEAKER_00: profits,
[24:49] SPEAKER_00: revenue is not the same as profits here,
[24:51] SPEAKER_00: but you really are talking about what is still,
[24:53] SPEAKER_00: as you're saying,
[24:54] will rapidly grow and still a relatively small business
[24:57] relative to this immense
[24:58] SPEAKER_00: economy and which is kind of dominating the financial markets.
[25:02] SPEAKER_00: That would be
[25:04] at least a possible source of alarm although maybe maybe okay
[25:08] but but just let's
[25:09] just dig into that a little bit because of course you know a
[25:12] stock prices is really
[25:14] a reflection of the expected future value of that stock across
[25:17] you know aggregated
[25:18] SPEAKER_01: across the market and at that 40 percent it feels very high the
[25:26] SPEAKER_01: But if you do it by the measure of earnings,
[25:29] these companies actually have a much higher proportion of
[25:33] earnings and earnings
[25:34] SPEAKER_01: growth.
[25:34] SPEAKER_01: So that starts to feel a little bit cheaper.
[25:38] SPEAKER_01: And if you look at the US stock market in 1990,
[25:42] SPEAKER_01: So after the railway calamities of the 1850s,
[25:45] SPEAKER_01: 1870s and 1880,
[25:46] SPEAKER_01: the railroad stocks were 60% of the capitalization of the US market.
[25:53] SPEAKER_01: And we had worked our way through the busts by that point.
[25:56] I think that there's a fantastic piece of academic work by an
[26:01] American finance
[26:02] SPEAKER_01: professor called Bessenbinder,
[26:03] SPEAKER_01: and I forget his first name.
[26:07] But Bessenbinder looked at the stock returns of 23,000 US
[26:11] stocks from the 1900s
[26:14] SPEAKER_01: through to 2022.
[26:16] SPEAKER_01: And those returns are highly,
[26:19] SPEAKER_01: highly concentrated in,
[26:20] SPEAKER_01: roughly speaking,
[26:21] SPEAKER_01: if I remember the numbers correctly,
[26:23] SPEAKER_01: about two-thirds are concentrated in about 30 companies.
[26:26] SPEAKER_01: And if you look at those 30 companies, those companies are oil companies,
[26:33] SPEAKER_01: Paul Krugman paulkrugman
[26:49] a couple of the healthcare businesses like Pfizer and one
[26:52] other,
[26:53] and I believe JP Morgan.
[26:55] SPEAKER_01: So you see historically that you will get this concentration,
[27:00] SPEAKER_01: and especially when you have a new general purpose technology,
[27:03] SPEAKER_01: to a number of
[27:04] SPEAKER_01: Okay, the one last devil's advocate question on all of this.
[27:06] SPEAKER_01: Again, I have no idea whether
[27:31] this is a it's probably a totally unfair thing but I keep on
[27:33] thinking that I keep
[27:35] on thinking of the California Gold Rush and that if you had
[27:37] looked at the the
[27:40] SPEAKER_00: revenue spending on essentially Gold Rush related businesses as a whole
[27:45] SPEAKER_00: It probably looked pretty solid, pretty sound.
[27:47] It was probably quite a lot of money,
[27:48] but the trouble is it wasn't the gold.
[27:50] It was the picks and shovels and blue jeans and whiskey and
[27:53] women basically were
[27:55] SPEAKER_00: revenue streams.
[27:56] SPEAKER_00: And are we looking at it?
[27:57] SPEAKER_00: Is that a fair question to ask about AI right now?
[28:03] SPEAKER_01: I think it's a really great question to ask.
[28:06] SPEAKER_01: And the question is,
[28:08] SPEAKER_01: what is the determinant of that demand,
[28:12] SPEAKER_01: that $150 billion annualized,
[28:14] SPEAKER_01: $90 billion of the last 12 months?
[28:16] SPEAKER_01: And what we see is that more and more American companies,
[28:19] SPEAKER_01: I think it's about just under 30% of the S&P 500,
[28:22] SPEAKER_01: according to our data,
[28:23] have in their earnings calls pointed to a generative AI project
[28:29] with a quantifiable
[28:30] SPEAKER_01: measurable result.
[28:32] SPEAKER_01: Paul Krugman
[28:54] Not a single one of them could point to even a 10 basis point
[28:57] improvement in their
[28:58] SPEAKER_01: business from the investments they've made so far.
[29:01] So when we break out that $90 billion and say two thirds of
[29:05] it's in the US,
[29:06] SPEAKER_01: so $60 billion,
[29:08] SPEAKER_01: again,
[29:08] SPEAKER_01: that's a lot of money if it's a single company.
[29:10] But when it's spread across Main Street and Wall Street and
[29:13] thousands of firms,
[29:14] SPEAKER_01: it's still at the experimental stage.
[29:17] SPEAKER_01: And then what we should be thinking about is
[29:19] are these executives learning by doing are they getting better
[29:23] are they going to
[29:23] make bigger commitments and you know the messages I get vary
[29:27] from those who have
[29:27] had some success saying we're having success in the tens of
[29:30] millions we want to
[29:31] figure out how to get to the hundreds of millions to those who
[29:34] are finding it
[29:34] harder but are still persisting so I think that we're slightly
[29:39] beyond pure
[29:42] SPEAKER_01: picks and shovels right now, but still we're far from the point.
[29:48] SPEAKER_01: I think in Paul David's work,
[29:50] it was 50% of American companies getting electrified when you
[29:53] see the productivity
[29:55] SPEAKER_01: rise.
[29:55] SPEAKER_01: We're just a long way from that.
[29:57] SPEAKER_00: I can't resist.
[29:59] SPEAKER_00: Just before we had this conversation,
[30:01] SPEAKER_00: headlines flashing across,
[30:02] SPEAKER_00: but I saw something about a KPMG study that involved,
[30:05] SPEAKER_00: have you seen this,
[30:06] had case studies which turned out to have been about usefulness
[30:10] of AI that turned
[30:10] SPEAKER_00: out to have been AI hallucinations.
[30:15] SPEAKER_01: Can I throw out the one thing that I do think is quite challenging here,
[30:19] SPEAKER_01: which is we've talked a lot of,
[30:21] SPEAKER_01: the market has talked a lot about the bottlenecks because
[30:24] SPEAKER_01: We saw this in railroads, right?
[30:27] SPEAKER_01: The US couldn't make enough steel.
[30:28] We did a Bessemer's collective invention where they all
[30:31] compiled their patents
[30:32] SPEAKER_01: together to move the know-how forward.
[30:35] SPEAKER_01: But there are these bottlenecks and there's a lot of emphasis on power.
[30:39] SPEAKER_01: can you get the electricity to the system?
[30:41] SPEAKER_01: And I think it's fair to say from our data,
[30:45] we think there's more demand than there is supply capacity
[30:48] right now for AI.
[30:49] But there is starting to be a question of whether there's
[30:51] enough capital in the
[30:53] SPEAKER_01: market to do all of this.
[30:55] SPEAKER_01: And we may see
[30:59] another few trillion dollars of intention from AI companies and
[31:05] technology
[31:06] SPEAKER_01: companies to build our infrastructure to 2030.
[31:10] And it's starting to rival the new issuance of the US Treasury,
[31:14] which is $2 trillion a year.
[31:16] SPEAKER_01: And I'm trying to, you may have more insight than me.
[31:21] SPEAKER_01: I'm trying to talk to people who understand this a bit more to say,
[31:25] is this actually capital constraint going to start to be an
[31:29] issue or is this just
[31:30] something that the market knows how to how to clear but I you
[31:33] know I'm puzzling
[31:34] SPEAKER_00: about that at the moment yeah I mean one thing that
[31:37] SPEAKER_00: I mean, ordinarily, we'd expect to see that in prices.
[31:40] SPEAKER_00: And maybe we are to some extent.
[31:42] Real interest rates are certainly well off their very low point
[31:46] of,
[31:46] you know,
[31:47] pre-COVID.
[31:48] And there's lots of,
[31:49] lots of,
[31:49] and there's so much dust in the air in between.
[31:51] SPEAKER_00: But now looking at them, they're higher.
[31:52] SPEAKER_00: They're still substantially lower than they were
[31:55] at the peak of the 90s tech boom we kind of real interest rates
[31:59] were around 4% then
[32:00] and and they're more like two I guess now which is high by by
[32:05] the what we've become
[32:06] accustomed to but but not and I'm not sure that this is in a
[32:11] way it's kind of
[32:12] surprising given the AI boom and the fact that now we have
[32:15] massive budget deficits
[32:17] that that rates aren't even higher than they are so it's it's
[32:20] kind of a puzzle
[32:21] SPEAKER_00: whether this is an actual constraint I think
[32:24] One question is,
[32:25] is that this is not,
[32:26] you know,
[32:26] this is not,
[32:27] what can we say?
[32:28] SPEAKER_00: NVIDIA is not the U.S. Treasury in the end.
[32:31] SPEAKER_00: I mean, they need somewhat risk-tolerant capital, the kinds of capital.
[32:37] SPEAKER_00: Capital is not totally fungible.
[32:39] And the possibility that these guys may not be able to raise
[32:42] enough money is,
[32:43] SPEAKER_00: I guess,
[32:43] SPEAKER_00: something we need to think about.
[32:45] SPEAKER_01: Well,
[32:45] SPEAKER_01: although on that NVIDIA point,
[32:47] SPEAKER_01: I did see that CDS credit default swaps on five-year NVIDIA bonds.
[32:52] So that's the cost of insurance against the bond defaulting
[32:56] currently lower than
[32:57] SPEAKER_01: U.S.
[32:57] SPEAKER_01: Treasuries.
[32:57] SPEAKER_00: I saw that.
[32:58] SPEAKER_00: And I have to say that strikes me as being...
[33:00] SPEAKER_00: paulkrugman paulkrugman paulkrugman paulkrugman paulkrugman paulkrugman
[33:25] Actually quick in passing since this is,
[33:28] as it happens,
[33:28] we're recording this on SpaceX day.
[33:31] One thing I've been wondering about is if there are sort of
[33:34] limited pools of
[33:35] SPEAKER_00: capital for what are perceived as cutting edge investments.
[33:40] You wonder whether,
[33:40] in some sense,
[33:42] Elon Musk is diverting a lot of capital that AI might need.
[33:46] SPEAKER_00: I mean,
[33:46] SPEAKER_00: he's got what everybody tells me is a crud AI product in Grok,
[33:51] and yet a whole lot of sort of the meme money is pouring into
[33:55] SpaceX right now.
[33:56] SPEAKER_00: Is that something I should actually be thinking?
[33:59] SPEAKER_01: Well, I think what Musk showed was his willingness to adapt.
[34:04] So the AI product is now being subsidiarized against using his
[34:09] AI capacity to serve
[34:10] SPEAKER_01: customers like Anthropic.
[34:12] I think he has this incredible following,
[34:15] but he does have something else.
[34:16] I've talked to people who have worked with him in the past,
[34:19] and they say that
[34:21] His ability to relentlessly focus on one thing and demand it
[34:26] gets optimized is
[34:29] SPEAKER_01: really the piece that sets him apart.
[34:33] His first principles thinking that has him being able to bring
[34:37] down the cost of
[34:37] SPEAKER_01: space launch faster than anyone in history.
[34:41] And you have to change the rights law learning curve for an
[34:44] industry when Elon Musk
[34:45] SPEAKER_01: gets involved because he pushes the rate of learning so fast.
[34:49] you know aggressively and you know I think for all the the
[34:53] challenges and his sort
[34:54] of mercurial behavior elsewhere that's generally I think quite
[35:00] a good thing because
[35:01] you know if we the thing that technology has done is it's
[35:04] brought down the cost of
[35:06] SPEAKER_01: Paul Krugman paulkrugman
[35:21] SPEAKER_01: We're going to be much,
[35:22] much further ahead of what we can do in space than we would
[35:26] have been if SpaceX had
[35:27] SPEAKER_01: not been successful over the past decade.
[35:29] It raises all sorts of other questions about how do we now
[35:31] govern what used to be
[35:32] SPEAKER_01: this commons or this commons in a reasonable way,
[35:35] but there is a definite benefit that we've come down that
[35:38] learning curve so
[35:39] SPEAKER_01: quickly.
[35:40] Okay now that's fair enough and I have to say that the one time
[35:44] I looking at Musk's
[35:45] activities I said he's really onto something was when I
[35:48] realized that he had
[35:49] diagnosed what is the cost of space launches and it's really
[35:52] not the fuel it's
[35:53] really the it's really the rocket and and recovering the rocket
[35:56] makes all the
[35:57] difference so that that is kind of a real insight and being
[36:00] able to make it happen
[36:01] SPEAKER_00: is a real thing
[36:03] SPEAKER_00: productivity.
[36:06] SPEAKER_00: So it's a funny thing, right?
[36:08] SPEAKER_00: This is all moving so fast that we don't have time for the kinds of,
[36:12] SPEAKER_00: you know,
[36:13] where's the technology in the official productivity numbers
[36:17] issues that we had in
[36:18] SPEAKER_00: the past,
[36:19] SPEAKER_00: but we're already,
[36:20] it's feeling like there's a kind of a solo moment where people
[36:23] are saying,
[36:23] SPEAKER_00: okay,
[36:24] SPEAKER_00: we're doing all this stuff.
[36:25] SPEAKER_00: It all seems very impressive, but I don't see it in the numbers.
[36:28] SPEAKER_00: But you want to talk about that?
[36:30] SPEAKER_00: Because you were already alluding to some of that already.
[36:32] SPEAKER_01: Yeah, I mean, it comes up all the time.
[36:35] I wonder if we're all a little bit tick-tocked and we need
[36:38] things to happen much
[36:38] SPEAKER_01: more quickly than we used to in the past.
[36:43] SPEAKER_01: We aren't seeing it as far as I can tell in the numbers.
[36:45] SPEAKER_01: Eric Brynjolfsson,
[36:46] SPEAKER_01: who is probably a colleague of yours over at Stanford,
[36:49] says that he does think it's showing up in the aggregate
[36:52] productivity numbers.
[36:54] SPEAKER_01: I mean, the way that I look at this is that
[36:59] How quickly should we expect a technology like this to show up
[37:04] in any kind of number?
[37:05] SPEAKER_01: And if we're at $90 billion a year,
[37:06] SPEAKER_01: $60 billion in the US,
[37:08] SPEAKER_01: as you say,
[37:08] SPEAKER_01: that's not much of US GDP.
[37:12] SPEAKER_01: It's also those early stages.
[37:14] SPEAKER_01: Paulkrugman
[37:41] effectively what you're doing in those first phases is you're
[37:44] retrofitting all of
[37:45] your tangible and intangible capital stock and processes and
[37:50] people with the new
[37:51] SPEAKER_01: technology that is electricity but it's not until you
[37:56] you know depreciate that existing capital stock you're willing
[37:59] to change your
[38:01] processes that you get to the unit drive which is what Ford has
[38:04] in Highland Park
[38:06] that you really see the you know the productivity benefits but
[38:09] if I may I'd like to
[38:10] SPEAKER_01: just bring some kind of little a few numbers to that yeah yeah so
[38:17] SPEAKER_01: You know,
[38:17] what would we expect to see in a company that has achieved the
[38:21] Ford equivalent of a
[38:22] SPEAKER_01: Highland Park in terms of output,
[38:24] SPEAKER_01: right?
[38:24] SPEAKER_01: We'd expect to see something.
[38:27] And one thing that I thought we might see is what would happen
[38:30] in an AI native firm to
[38:32] SPEAKER_01: Paul Krugman paulkrugman
[38:51] SPEAKER_01: When you start to look at some of the AI native firms like Mercore,
[38:55] SPEAKER_01: which does data generation,
[38:57] SPEAKER_01: that number is closer to $7 million per employee.
[39:00] And if you look at Anthropic,
[39:02] it's close to $10 million per employee of revenue.
[39:06] And so you do have that thing that you would expect to see if
[39:10] someone has been able
[39:10] SPEAKER_01: to be AI native,
[39:11] which is that the enormous commercial productivity of a single
[39:15] employee can be
[39:18] SPEAKER_01: measured.
[39:19] SPEAKER_01: But we're talking about a handful of firms at this point,
[39:22] SPEAKER_01: but I do think that there is a slight shape that we're able to pick up,
[39:26] SPEAKER_01: that this isn't impossible.
[39:28] SPEAKER_01: It may be hard, it may take time, but it's possible.
[39:31] SPEAKER_00: Okay.
[39:31] Do you happen to know,
[39:32] what would those numbers look like per dollar of nested
[39:36] capital?
[39:37] SPEAKER_00: Because one question,
[39:38] one of the big worries is that this is an enormously
[39:40] capital-intensive business
[39:41] SPEAKER_00: that replaces labor and
[39:43] SPEAKER_00: And you could get,
[39:44] SPEAKER_00: I mean,
[39:45] SPEAKER_00: I have no idea,
[39:45] SPEAKER_00: but not too far from me are the giant oil refineries of New Jersey.
[39:50] SPEAKER_00: And I presume that they have,
[39:51] SPEAKER_00: not on that level,
[39:52] but they have enormous revenue per employee because there's no
[39:55] workers there.
[39:55] SPEAKER_00: There's just these monstrous capital installations.
[39:58] SPEAKER_00: So is that a...
[40:01] SPEAKER_01: Yeah,
[40:01] SPEAKER_01: I mean,
[40:02] SPEAKER_01: you know,
[40:02] SPEAKER_01: Anthropics,
[40:02] SPEAKER_01: I actually don't have the number here on the top of my head,
[40:05] SPEAKER_01: but it's raised in the tens of billions rather than hundreds of billions.
[40:09] And it had a profitable quarter,
[40:12] which is surprising and a bit ahead of schedule.
[40:17] SPEAKER_01: What we don't know,
[40:18] and we've done a lot of forensic accounting to sort of peel
[40:20] apart the P&L and the
[40:22] SPEAKER_01: balance sheet of these private companies,
[40:24] SPEAKER_01: but the IPO will tell us,
[40:25] SPEAKER_01: and they're going to IPO in the next six to nine months,
[40:28] SPEAKER_01: is
[40:30] How much of that capital is being put into the development of
[40:34] the next model versus
[40:36] the 90 billion is obviously being monetized across previous
[40:40] generations of the
[40:41] SPEAKER_01: model.
[40:42] And we know that you can be much more efficient because the
[40:47] Chinese companies are
[40:50] SPEAKER_01: using much less capital to build models that are nearly as good as well.
[40:56] So I think the harder part of your question is that if every
[41:01] model that OpenAI or
[41:02] SPEAKER_01: Anthropic costs 10 times as much to deploy and develop,
[41:07] but lasts only a couple of years before it's defunct because of
[41:10] competition,
[41:11] what needs to be true for that to be sustainable for more than
[41:14] a year or two?
[41:14] SPEAKER_01: I mean, it's just that to me is a really tricky question as well.
[41:18] You had some numbers or have cited some numbers that are also
[41:23] sort of like
[41:24] SPEAKER_00: intermediate measures.
[41:25] SPEAKER_00: Rather than actually looking at revenue,
[41:26] SPEAKER_00: we look at,
[41:27] SPEAKER_00: on the one hand,
[41:28] SPEAKER_00: lines of code generated,
[41:31] SPEAKER_00: which has exploded,
[41:32] SPEAKER_00: and then sort of actual usable applications,
[41:36] SPEAKER_00: which has not.
[41:37] SPEAKER_00: Is that telling us anything?
[41:40] SPEAKER_01: Well,
[41:40] you know the lines of code generated is is a I don't know what
[41:44] kind of measure that
[41:45] is really I mean normally a line of code is something we can
[41:49] ascribe some value to
[41:50] because a human's written it and now the machines that you
[41:52] write so much more and
[41:55] it's unsurprising that the dramatic increase in in lines of
[41:59] code has not been met
[42:01] by a proportional output in you know the number of applications
[42:06] or productivity and
[42:07] SPEAKER_01: the reason is that that
[42:09] we've made it much cheaper for people to write code so people
[42:12] who are less good who
[42:13] are less determined are now writing code and you're getting
[42:16] this stuff out there
[42:17] but I think that in absolute terms are we getting more high
[42:21] quality code the data
[42:23] seems to suggest that we're getting more high quality code but
[42:26] there's also a lot
[42:27] of useless code that is effectively waste but I would say Paul
[42:31] this isn't the first
[42:32] SPEAKER_01: time that we have seen a useful
[42:35] SPEAKER_01: Paul Krugman paulkrugman
[42:50] and actually it's only the one third that goes into useful
[42:52] energy services so that
[42:53] you know i think there's an analogy to think about slot lines
[42:57] of sloppy code or
[42:58] slop lines of code which is well you know we're used to having
[43:02] these processes that
[43:02] and we that involve waste and we were absolutely happy for more
[43:07] than a century to
[43:08] do that with barrels of oil right yeah i'd say a weird analogy
[43:14] is that the uh
[43:16] SPEAKER_00: I still remember when widespread word processing came in.
[43:20] SPEAKER_00: And one of the things I noticed was that books started getting longer.
[43:24] SPEAKER_00: That you could actually,
[43:25] you could just see that it was so easy for authors to turn out
[43:28] hundreds and
[43:29] SPEAKER_00: hundreds of pages.
[43:30] SPEAKER_00: And, you know, and what might have been two volume series became five.
[43:34] SPEAKER_00: It's really, yeah.
[43:36] Yeah but you know on that front I think we also have to you
[43:39] know consider I think
[43:41] we're at you know we're at an enlightenment moment here so
[43:44] let's you know cast our
[43:45] minds back to you know France in the mid-18th century and that
[43:49] battle was between
[43:51] who gets to write who gets to create the truth who gets to
[43:55] express their story and
[43:59] And of course we can go back to those books in the 18th century
[44:02] and they are
[44:03] absolutely remarkable and how these men and women produced
[44:07] works like that with
[44:07] quill pen and encapsulated a world that was just seedling as a
[44:13] little seedling.
[44:14] SPEAKER_01: It amazes me.
[44:16] SPEAKER_01: But is it worse off that
[44:19] we're able to allow for lots more expression and I think it's
[44:24] very very clear that
[44:25] there are places where we are worse off so we are you know
[44:27] worse off when that
[44:28] connect gets connected to a algorithmic recommendation system
[44:32] that just drives sort
[44:33] of slop at us constantly but but I'm not sure we're necessarily
[44:37] and inevitably
[44:38] worse off because what we're doing is we are giving access to
[44:43] lots and lots of
[44:45] people and I just think about you know Ramuna Jan if he hadn't
[44:48] been discovered
[44:50] SPEAKER_01: from his obscurity and got to England.
[44:55] SPEAKER_01: What would we have lost if that accident hadn't happened?
[44:58] And I do feel sometimes with these technologies that we need to
[45:01] step back and say,
[45:03] SPEAKER_01: right,
[45:03] SPEAKER_01: there are problems.
[45:04] SPEAKER_01: Let's deal with the algorithmic slop.
[45:06] Let's deal with all the AI writing that is infesting the world
[45:08] right now in particular ways.
[45:10] SPEAKER_01: But let's also recognize that in reducing these costs of access,
[45:14] SPEAKER_01: we might find amazing people in breaking
[45:18] SPEAKER_01: We might find connections.
[45:23] Maybe there's something in battery chemistry that turns out to
[45:25] be really good in cardiology.
[45:27] SPEAKER_01: Well,
[45:27] SPEAKER_01: it sounds absurd to say that,
[45:29] but we don't know because we've never been able to get those
[45:31] experts to talk.
[45:32] SPEAKER_01: So when I look at this question,
[45:34] I try to break it apart and look at each of those opportunities
[45:41] more discreetly.
[45:42] Okay,
[45:43] I think there's a potential book here,
[45:45] something like The Upside of Slop.
[45:47] SPEAKER_00: But it's real.
[45:48] SPEAKER_01: We can get ChatGPT to write it.
[45:51] SPEAKER_00: Oh, gosh.
[45:52] SPEAKER_00: Yeah.
[45:54] SPEAKER_00: I think we could go on forever.
[45:55] SPEAKER_00: But I mean, it's incredible.
[45:57] SPEAKER_00: I mean,
[45:58] SPEAKER_00: as we said,
[45:58] SPEAKER_00: I mean,
[45:59] SPEAKER_00: this is an unrecognizable scene from 18 months ago when we last talked.
[46:04] SPEAKER_00: So this is, wow.
[46:06] SPEAKER_00: I feel, you know, I started my career
[46:10] SPEAKER_00: writing papers, longhand, alternate lines on yellow legal pads.
[46:14] SPEAKER_00: We are.
[46:15] SPEAKER_00: Amazing, amazing change.
[46:16] SPEAKER_00: Not sure it's all for the good,
[46:18] but thank you so much for talking to me and we'll just keep
[46:22] this conversation going
[46:23] SPEAKER_00: until AI starts doing the conversation for us.
[46:27] Paul Krugman it won't do the conversation for us and you know
[46:30] longhand I write
[46:32] everything with a fountain pen I burn through ink I've got my
[46:35] ink here and I'm
[46:37] writing my new book every chapter is written longhand and most
[46:40] of my research is
[46:41] SPEAKER_01: and it's with a fountain pen
[46:42] and the computer's turned off because I can like the computer
[46:45] the AI does all this
[46:46] boring computer stuff like PowerPoint and emails for me and I
[46:49] now have time to you
[46:51] know apply my brain and my pen and my paper to the things that
[46:54] I want to to think
[46:55] about so I'd be really happy to to continue this conversation
[46:59] in a few months but
[46:59] SPEAKER_01: thank you for for inviting me thanks so much take care then cheers

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