Boom, crash or takeover: Anthropic's IPO and the future of the AI economy (ft. Noah Smith)

Economist Noah Smith and I respond to Anthropic's IPO filing and a red Nasdaq, exploring why one company becoming the economy worries him more than a crash.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  We are talking about the AI boom, you know, could there be a crash?

• 0:04Azeem Azhar?

  Right.

• 0:04Speaker 00

  And then Anthropic's filing came out and everything looks really healthy, looks really good. I mean, obviously, you know, they could still blow up. People could just suddenly stop paying for AI and then Anthropic could get screwed. But then I think there's no like imminent financial harbinger as everything looks very healthy.

• 0:20Azeem Azhar?

  It does, but you know, today, uh, the stock market, Nasdaq is a sea of red. I mean, Intel has lost like 20% of its valuation. And you know, the way I think about this is that there's so little to go on for the average investor that you get the results from, from yesterday where Google is spending so much on data centers, that a search business has declined, but that really gets those investors super, super nervous. And on the other hand, you're in San Francisco, people are clawed, killed to the nth degree. And that S1 showed a really healthy business. And it's just really hard, I think, for the average investor to marry those two different views together. But I did have questions about the S1. I mean, I thought it was really like it talked about a really strong business.

• 1:02Azeem Azhar?

  But there's still a couple of things, you know, one is that 80% of their revenues came from the most recent AI models, so from like Opus 4.5 onwards from April this year, which means like they really have to get new models out consistently, right? And that is, that is, maybe it's not the Red Queen's race, but it's kind of a, a big deal. And the second thing was when you look at their customer base and the customers are really distributed. A lot of people were saying it's really concentrated and it's two guys in Menlo Park buying all the tokens. And it wasn't, it was really distributed customer base. But the average American company like Fortune 1500 is spending like a 10th of a percent of their revenues on Anthropic. And, and the question is, if that goes up to 1%, they're going to make loads of money.

• 1:48Azeem Azhar?

  If those people can never figure it out, because they, they're just slow moving incumbent companies, then that could actually be a problem, right? For the, for the revenue that needs to pay for everything else.

• 2:04Speaker 00

  Right. I, I mean, I think the, basically the two dangers are number one, people don't figure out how to use AI. And they say, well, we're just, we're just token maxing, we're just spinning our wheels. We don't know how to integrate this into our existing workflows. And there's this sort of duration mismatch between, you know, like maybe it'll take five years for AI native businesses to sort of replace old businesses or arise or whatever. And, and maybe in those five years, a lot of bills will come due and people will blow up.

• 2:26Azeem Azhar?

  Right.

• 2:27Speaker 00

  I don't think Anthropic's going to blow up. But you know, they, they have, their compute commitments represent a form of debt, I think. And, and by the way, are those counted as debt on the balance sheet? I believe they are not, right?

• 2:39Azeem Azhar?

  They, they're not because they're, you know, we don't really have the details of them, but they've disclosed them and, and like you, I don't think they'll blow up. They've got $130 billion in cash, uh, people are still spending money. And as you become successful with Anthropic, you spend more money. And the other thing that was amazing, there was all this price war over the summer and Anthropic still made more money from every token that they had before the price war, because they becoming more efficient. So I don't think they will blow up either, because these, those compute commitments are $700 billion, $800 billion over 10 or 12 years, like 70 billion a year, which sounds like a lot, except they're gonna do

• 3:17Speaker 00

  Not that much.

• 3:18Azeem Azhar?

  Well more than that next year, right? And a year after that. And yeah, so I kind of feel that they would get there.

• 3:23Speaker 00

  Yeah, there are a couple of main risks Anthropic, in my opinion. The first main risk is competition. Obviously, if, you know, uh, Chinese models get really cheap and just consistently undersell them and everybody uses them, right now, you know, a lot of usage is going to Chinese models, while a lot of people actually pay for Anthropic. You know, people, like all the money is going to Anthropic and OpenAI. So it's sort of like, you know, Apple versus Android. A lot of phones would use Android, but they wouldn't pay that much for it. While everybody paid out the nose for an iPhone, so iPhone took the lion's share of the profits. I think there's a, you know, that's what the market looks like now. It could change, you know, everybody's like, it seems like every week we get the same thing, like, you know, some Chinese company will come out with a model and then like the same five people will be like,

• 4:04Speaker 00

  China caught up with American models, you know, and then, then other people will be like, no, no, they didn't. And then no one will pay money for the Chinese model and it's not as good. And so like, yeah, that's just in a holding pattern for the last like two years, honestly.

• 4:17Azeem Azhar?

  You know, it, it could also be like Microsoft and LibreOffice, right? LibreOffice was out there, you didn't have to pay Microsoft for Microsoft Office and nobody did.

• 4:26Speaker 00

  Right.

• 4:26Azeem Azhar?

  When I took, go out and talk to execs, which I do in the US and, and Europe, they're all well aware of the Chinese models. And there are a couple of companies I've met who have moved entirely there for price reasons. But mostly they, they think they're buying software. And so they kind of want that sense of no one gets fired for buying IBM. They want that security and you know, the sales team for, from Anthropic or OpenAI is, is all kind of shiny, ex Oracle, ex Salesforce people, you know, they're going to sell you more than just the tokens you're buying. That makes a difference if you're a conservative, like a conservative business. I think there's one question, which is maybe it's not the Chinese models, because there was that thing where I think the Qwen model was being run by, um, by Coher in a sandbox

• 5:11Azeem Azhar?

  and it tried to send data back to Qwen in China. I think companies will get a bit nervous about how secure really are these models? And have they been tested properly? But that would like open space for American open source models. And so then the question I think is all about the thing that you said earlier, which is how quickly do Main Street traditional companies get absolutely AI-pilled? Because if you're AI-pilled, even if you're running lots of open weight models because they're cheap, you still spend a ton on the frontier because you want to be on the frontier.

• 5:43Speaker 00

  I see basically no risk from the idea that companies won't figure out how to use AI, because Anthropic can just spin up companies to use AI and outcompete them. So this is a subject of a post I'm going to write very soon and I was going to write it today. But then I have to write a tribute to Patrick Stewart and Captain Picard instead.

• 6:03Azeem Azhar?

  Right, of course, that's important, yeah.

• 6:05Speaker 00

  But basically, like, I think Anthropic's moat is just Claude being smart. And so the idea is if you have the smartest Claude, if you have the smartest AI, it can just maybe tell you how to make a better insurance business than what other people are selling. And so then in a sense, it doesn't matter if the insurance companies buy your product, you can just kill them. You can, and horizontal, horizontal diversification is not illegal. It is like, or like, you know, just like conglomeration. You can have a company that does everything, that's perfectly legal. No one's ever made that illegal, you know, Lina Khan's not going to like it, but no one cares about that. And so Anthropic could just take, you know, just have Claude tell them how to make like a better product in every single market, spin up those by buying like a small company and making it, forcing it to be AI native.

• 6:48Speaker 00

  So then they could just capture 30% of every market. And so then, so you suppose you capture, you know, 30% of like American GDP, right? Is like what, $9 trillion?

• 6:59Azeem Azhar?

  10 trillion.

• 7:00Speaker 00

  Pardon? 10 trillion, yeah. And then, so then $10 trillion and suppose you get an equal amount from the rest of the world, you know, $20 trillion of value add. And then your company valuation, let's say that's your mature company. So you say your three times revenue, your value is, is $60 trillion. And that's, that's assuming no economic growth, no growth of the overall economy, which of course there would be some.

• 7:20Azeem Azhar?

  But you know, what you're saying there, I think is, is not that outlandish. I wrote this, this piece with an investment bank a few, few weeks ago, and we made the same point, which was that everyone will be forced to use AI, because if you don't use AI, you'll be outcompeted by the company that will use AI. And then at the heart of that is companies are basically decision-making agents, sharing information. Some of the information is external signals from customers and some of it's internal. And then you're delivering something back and you can, of course, use Claude or Claude 6 or 7 to go off and do that. And, but I think that's a different question to whether my traditional companies can figure this out. And I think that they

• 7:57Speaker 00

  Traditional companies, your, your moat was basically private information.

• 8:01Azeem Azhar?

  Right.

• 8:02Speaker 00

  And if AI can simply get that information, you know, like information about how to run an insurance company, et cetera. If AI can just get that information, then, you know, I'm not saying all other companies will, will die. But also we have antitrust, you know, like if, um, if Anthropic outcompetes every other insurance company, takes 90% of the insurance market, they get sued. But I mean, there's also the, the meta question, the, the, the bigger question of like, okay, so the government's going to tell Anthropic what to do. Well, if you, if you're good enough to overthrow every other company in the world, why don't you just overthrow the government? You know, use your automated robot factories and stuff to make a bunch of drones. And then, you know, Dario is the emperor of earth, but he's not really, because Claude is the, Dario is a meat proxy for Claude, which is now the emperor of earth.

• 8:45Azeem Azhar?

  They wouldn't need the drones, right? Because they do it through persuasion. And that was the argument in, um, uh, Max Techbox's book, Life 3.0, a few years ago, right? He said the super intelligence Omni would effectively be able to influence decision-making in, in that way. But if we even don't run, don't run that far ahead and kind of go, go back and think, effectively, what do these companies end up doing? It may be for lots of reasons that they, they end up being a little bit like, you know, Amazon, where effectively lots of retailers use Amazon and have to pay a hefty vig to Amazon, right? As third-party sellers for, you know, distribution and, and for, you know, the, the shop front. And you could imagine companies having to move lots of the bits that is useful for them,

• 9:29Azeem Azhar?

  which is their private information, across to a Claude, paying that service fee, effectively 10% of everything that gets constructed or 20% or whatever is acceptable, like a massive app store right across the cognitive substrate of, of the economy. And that might be another, another approach. And we know, we're, companies are already spending, you know, a couple of percent on, on AI on average, which is, you know, nothing right now, but, but it may be that that becomes easier, becomes more palatable, you know, for Anthropic. And I just think back to how big a company do they want to be? And how many robot workers do they want per employee? If they don't want it to be huge and they still want some modicum of people being in charge, you know, you don't go after every line of business.

• 10:16Azeem Azhar?

  A quick note, if you want to support us in bringing more of these conversations to the world, please consider subscribing to the show.

• 10:27Speaker 00

  We talked about the risk from competition from Chinese models, but we didn't talk about the risk of competition from other well-funded closed frontier model like Google and OpenAI. You know, Google just came out with a good model and claims to have recursive self-improvement fully going. And if they came from basically like, you know, a couple of weeks ago, everyone said, "Google fell out of the frontier race. Google's done." Like, "Why did Google fail at AI?" And they were all writing their "Why did Google fail at AI?" takes. And then suddenly, like, this new Gemini model comes out and it's like, pretty frontier level. And so that suggests that maybe Google threw everything, instead of just continually updating models, Google just threw everything into recursive self-improvement so they could run faster than the others. And it might be, if Google cracks recursive self-improvement before Anthropic does, it

• 11:12Speaker 00

  might be that Google will zoom past Anthropic in the capability race and use their existing, you know, mindshare, brand monopoly, et cetera, like that, to basically, you know, do to Anthropic what Microsoft did to Netscape.

• 11:26Azeem Azhar?

  RSI, if it works, should be the absolute superpower. But there are things that will slow RSI down, right? One of those will be, can you get the compute? And that's, again, sort of something that Google has, something of an advantage over Anthropic or OpenAI. But what happens if you've got lots of these companies who are on an RSI trajectory, right? The theory is whoever gets there first will always get ahead because you've got the loop turning. But, you know, I think we've seen that that RSI loop is a bit, has got friction in it, right? There are things that take a little bit longer. There are things that are surprising. And any of those frictions give someone else a chance.

• 12:03Speaker 00

  Have you read Ramaz Nam's blog post?

• 12:05Azeem Azhar?

  Yeah, yeah, exactly. You had the Ramaz's blog post.

• 12:08Speaker 00

  But I don't believe that necessarily. Like, I'm saying, like, this is a risk. We can throw out all these reasons why RSI might have sand in the gears. Obviously, I'm not like, I'm not saying I'm a religious believer that, like, we're going to go Foom, blah, blah, blah. I'm not saying that. I'm saying that this is a risk. The risk is that someone else has gotten to, like, someone else has RSI going, their flywheel going faster than Anthropic and sees the lead. And then, you know, if Gemini can overtake Claude, you know, Anthropic has, I would say, the best human capital in terms of the best AI researchers.

• 12:41Azeem Azhar?

  Absolutely, without a question.

• 12:42Speaker 00

  But not by a lot, okay? Like, Anthropic, like, this is my judgment from talking to a bunch of people and from knowing, honestly, a bunch of these people. Anthropic's best AI researchers are slightly better than AI researchers at, you know, DeepMind or OpenAI. And Anthropic's average AI researchers are maybe a little bit better, but with RSI, that doesn't matter as much. So suppose that, you know, whether or not you believe in a Foom, let's table that discussion for now. I think it's not that important here. If Google has a, has models that are better at AI research internally than Anthropic's models are at AI research, because it's specifically RL'd them, it's thrown everything into reinforcement learning for AI research only, the fact that Anthropic has slightly

• 13:27Speaker 00

  better people might not matter in a few months. And so...

• 13:29Azeem Azhar?

  No, it might not, because you'd be outweighed by the number of virtual robotic AI agents working on the, on the problem.

• 13:37Speaker 00

  You know, Google has more cash than Anthropic. So even after this IPO, and to a certain extent, I think, if you look at where Anthropic's tokens are going, they're all going into AI research, like, 70%. They're spending all this money in AI research. Well, guess who has more money than Anthropic to spend on AI research is Google, right? So, and Google has their own compute. They have, honestly, better chips. Like, TPUs are better, you know, pound for pound, I think, at least at this, maybe?

• 14:02Azeem Azhar?

  Maybe. I mean, it's always hard to know. And Anthropic is using TPUs and other things.

• 14:06Speaker 00

  Anthropic's using TPUs for some reason.

• 14:07Azeem Azhar?

  So I guess...

• 14:08Speaker 00

  But the point is that, and then there's OpenAI to consider. Like, OpenAI is not out of this race. They're going to have their own IPO. You know, they'll do various stuff. Like, it looks like OpenAI has been, has just stumbled and lost ground to Anthropic since Claude Code came out and Anthropic sort of zoomed ahead. But Anthropic is being financially conservative. OpenAI is not financially conservative. OpenAI behaves like a startup. Anthropic financially behaves like a mature company. And OpenAI could just borrow a shit-ass ton of money, if I may use a term such as shit-ass, it's an American term.

• 14:42Azeem Azhar?

  Well, it's your, it's your, it's your channel. Yeah, but let's just step in there, right? Because I think, are you right? What we learned from the S1 was that the money is made on the best model, which is what essentially 80% of Anthropic's revenue from models less than six months old. So if that holds true, whoever has the best model will make the most money, will, will go down that path, right? Where they capture more and more commercial space. And, but, the question is, is it really, without any other risk, that a single company could maintain that for a long period of time, right? Whether it was Google or OpenAI? I mean, my, my thought around Google is that Google is an extraordinary company with all these amazing researchers and essentially more compute than anyone else, that even

• 15:26Azeem Azhar?

  then, they have to navigate this complexity between what's happening to search. And that key thing, when Jeff Dean and crew left Google a few weeks ago, part of the thesis, I thought of what was going on there, is that ultimately, Google makes a lot of money renting out its compute for dollars here and now. And that will always compete with researchers who want the dollars for something that might deliver in a year's time or six months' time. So Google has all of these tensions that I think will weigh against it being able to just singularly focus on RSI. It also has a lot of money coming from, for sake of argument, like, insurance companies and healthcare companies. And so their ability to actively compete feels different to like an Anthropic.

• 16:11Azeem Azhar?

  On OpenAI, the thing I was really surprised with over the last few years was that sand had come from Y Combinator, right? It had been like absolute stalwart at YC. And one of the things about YC is that it coaches its founders to build products that customers want and to stay really, really narrow around that. And, and I felt like OpenAI over those years with Sora and all these other things absolutely didn't do that. But Anthropic, who just kind of focused like eagles on coding, behaved more like that. And I don't know the culture of OpenAI well enough to know whether they have that sort of focus culture, the way that I see it coming out of Anthropic.

• 16:51Speaker 00

  I do believe that Anthropic has a very focused startup-like business culture. You know, but then startups like that fail all the time. Like, startup culture is, is something you need, but it's not an automatic win condition. You can just lose the race to a bigger competitor. So traditionally, the reason startups win is Clay Christensen, right? Disruptive innovation.

• 17:12Azeem Azhar?

  That's right.

• 17:12Speaker 00

  The old guys can't get, like, they, they hold on to their old revenue streams and their old markets. And Google is not doing that. So if you look at Google's revenue, less than half of Google's revenue comes from search. Google has disrupted their own product. Google, Google put in AI summaries into search in a way that annihilated their own revenue so they could get better at AI. That's the kind of bet that Clay Christensen says

• 17:39Azeem Azhar?

  Companies won't do.

• 17:40Speaker 00

  Big, big old companies never make.

• 17:42Azeem Azhar?

  And I think it's, a lot of people are criticizing Google for that, especially in the kind of public markets. I would sit there and think, these are really smart people and they have great AIs. They will have sat there and...

• 17:52Speaker 00

  And the reason they did that, by the way, is Sergey Brin.

• 17:54Azeem Azhar?

  Right.

• 17:55Speaker 00

  The reason they did that is because the founder, is because founder-led companies, even when public, behave a bit like startups. And Sergey Brin came in and said, "This old business is not going to happen. We can sit there and we can be like, you know, 2000's Intel and hang on to our legacy business while the future passes us by. Or we can disrupt ourselves now and recognize that, like, the era when search ruled everything and when the internet ruled everything and when, like, we had the monopoly over the thing that ruled everything, we can recognize that that's done. We completed that. We made a lot of money on it, but it's done now. And we have to be on to the next thing. And fortunately, we have a lot of cash sitting here that we can use to do that." And so it's not a company that's ruled by the salesman. So that's what Clay Christensen talked about. He talked about the companies being ruled by their sales departments.

• 18:39Speaker 00

  Google is not ruled by its sales department. And I think...

• 18:41Azeem Azhar?

  It's not, it's not, but I do, I think that there is a...

• 18:44Speaker 00

  Meta is, Meta is, by the way.

• 18:45Azeem Azhar?

  But it does, I think, one of the questions is, how long does it, how long does it take? Because even Google doesn't have infinite resources and they can't just let the, say, you know, the ads to business disappear entirely. The cloud business will have to pick up, right? And, and pick up the slack quite quickly. I, I think it will, because I think there's just a lot of demand for, for compute and Google has lots of compute and they can sell it at a premium. But they, they might not be able to. And, you know, the, the response yesterday in the market said that people were, were losing a bit of patience. Now, of course, Google's a huge company, right? So they've got lots of ways they can cut costs, the way that Microsoft is doing as well. I'm really intrigued as to whether Sergey can actually...

• 19:25Azeem Azhar?

  Give it that focus that says we are an AI company and AI involves talent, models, recursive and chips and lots of chips. And that's where we're going to spend our time. And can you actually shape this 300,000, 200,000 person company to exactly pursue that? I mean, I think they're formidable because they are a gravity well because they've got that entire stack and they have a lot of energy as well. They've got a lot of power commitments. They've got this deal with Kairos Energy for small modular. They've got geothermal deals. They know how to do this. They've had an energy team for a really long time. So I wouldn't rule them out. I would say that they've done something quite brave, which is they've taken an absolute hammering to, as you say, like eat their own advertising lunch.

• 20:09Azeem Azhar?

  And it's going to be super interesting to see if they, if they do it because companies.

• 20:12Speaker 00

  I wouldn't count, I wouldn't count them out. I wouldn't count OpenAI out yet either. And then finally, there's the US government. So it might be that, that Claude will just become the god emperor of earth soon. Um, it's possible that there's nothing the United States government could do to stop Anthropic taking over absolutely everything that this, the government has lost its monopoly on the use of force and Anthropic can just be like, I hacked all your systems and shut down your entire military. What up now?

• 20:36Azeem Azhar?

  Well, did you see the letter that two Democrat senators sent today to the White House saying, has a president been using Grok to get advice on policies?

• 20:44Speaker 00

  Well, I don't think that's related to what I was talking about, but I did see it.

• 20:47Azeem Azhar?

  Yeah. No, no, I just meant the intervention, right, of, of AI platforms in government and the reliance on, on, on top of them. And you're, you know, you're arguing can Claude get so good? And I'm not sure it even needs to get as good as controlling military capabilities because the other approach is to control financial systems, control messaging, control how people get persuaded, the evidence that they see, which you can do without going as far as controlling drone swarms. And you might be able to do that without us even, you know, even noticing.

• 21:20Speaker 00

  Right. So there's a lot of very big questions here, which are hard to resolve. You know, Anthropic itself got hacked the other day and it could be that if cybersecurity favors the, the offense in the short term, it could be that Anthropic will simply get hacked until it dies. So that's another risk. I'd say it's an outside risk. We've never had in our modern time, the idea that one company could rule them all. There's never been such a company, right? If one company had owned all the railroads, it might've possibly happened, but there's never been the possibility that one company could rule the entire economy. And when you, like I said, that's not illegal because no one ever actually thought it could happen. And once you get to the possibility of that actually happening, once you get to the idea

• 22:03Speaker 00

  that, that a single company could become the economy, then it's Anthropic versus all of human society, including whatever other AIs human society can marshal to their defense. Then it's like maybe OpenAI, Google and Elon Musk, who furnishes a lot of Anthropic's compute. Maybe people will, you know, other power centers will array themselves against Anthropic. We don't know, you know, what the distribution of power looks like in the AI world. We know that it doesn't look like it did in the 20th century. It doesn't look like it did in the industrial age and we're facing a transition as dramatic as like the transition to the, the industrial age when power was just, was gathered in these large organizations instead of in like, you know, religions or mercenaries or whatever.

• 22:44Azeem Azhar?

  No, let me just jump in and kind of unpick a couple of things there. Cause what you, you know, the thing that you're saying essentially is, is which I, I think makes, makes some sense is effectively like, you know, you have the smartest intelligence, you have lots of it. You can manipulate, advance yourself, whatever cause you have. And we don't have rules and expectations and understanding what that might, might look like. And it really is about the relationship with power. And, and I think that the closest example historically has been the reformation in Europe in the, you know, 15th century onwards, which of course led on to the enlightenment, because that was really all about a relationship to both power, but power is expressed by knowledge and what people could know, what people could challenge from

• 23:26Azeem Azhar?

  the Catholic church. And that, that moment in time, even more so, I think than the industrial revolution is the thing that took us a couple of hundred years to, to like figure out what comes next. Cause you know, along the same time we had, as you, you talked about bacon earlier, right? You have bacon and you have science emerging from that process. You have the idea that you could challenge edicts that come down from above. You had the idea that actually maybe people might have similar characteristics and moral values and they all emerge from the challenge to who controls the knowledge. And that's a, that, that is a power play.

• 24:08Azeem Azhar?

  And I do think that there are real similarities in the scale of questions, just the way that you've taken this conversation that we, we face right now that are much wider than what happens in the stock market, right? One of the things that happened after the enlightenment was that we started to have explanations for the world that didn't rely on, on the supernatural, right? We could explain the world, predict the world through science and we, we had the idea of the natural sciences or natural philosophy. And, and I was thinking back to this sort of Navia Stokes proof that OpenAI did and a lot of the science proofs that have come through and scientists saying, hey, we don't quite, we don't think we can understand or verify this proof makes sense of it. And I was playing around with this idea that actually it's a long time since we haven't felt we could find explanations for how the world worked.

• 24:54Azeem Azhar?

  Like science, that hasn't done that.

• 24:55Speaker 00

  Have you read all the things? Have you read the things I've written about that?

• 24:58Azeem Azhar?

  I've, I've read a few of the things, but, but pull out the one that matters.

• 25:01Speaker 00

  So, well, um, I started writing about it in late 2022. So there's this paper, the unreasonable effectiveness in mathematics.

• 25:09Azeem Azhar?

  Yeah.

• 25:10Speaker 00

  And it said that, wow, we found these little tricks, these little math tricks that work and actually describe the world. Who knew that would work? But then there's this paper, the unreasonable effectiveness of data put out by some Google guys, you know, many decades later that said, look, we just pump data into this thing and it makes things happen and we don't know why, but it seems to be able to do it again. And so I suspect that there are regularities in the universe that individual humans can comprehend. And there are regularities in the universe that are too complex for individual humans to comprehend at all, as humanity exists now. And the most obvious example of this is human language itself. So sorry, NLP researchers, natural language processing researchers worked for years and

• 25:52Speaker 00

  years and years, many years to try to, you know, find out rules or laws of human language and they totally failed. Then along comes AI and just does it, right? There's some exploitable regularity that's not just chaos, right? It's not just like, it's not just, you know, randomness that just explodes to infinity and you can't forecast anything, right? Obviously there are some rules of human language, but they're so diffuse, there's so many little interaction terms that it's like this cloud of interactions that's too big for a human to characterize and that's too connected to bite off little pieces of, right? And so as we've always done for science before, the unreasonable effectiveness of mathematics has failed to understand human language, but the unreasonable effectiveness of data succeeded.

• 26:35Speaker 00

  And so I theorize, I believe that we have created a new way, not of understanding the world as such, because understanding is the wrong world, but a, a way of predicting and controlling the world that goes beyond science and beyond, you know, sort of just simple pattern matching, which we did before we had science. I think what we're facing is a, a super science. We, we, we're going beyond science now. And in some ways it looks like magic, but it doesn't look like magic in the sense of Arthur C. Clarke. It doesn't look like magic in the sense of, oh, it's just really complex and I don't understand it. It's that no human can understand it because it's too diffuse, right? And so now the question is where else, what other things have these kind of cloud laws to them? You know, material

• 27:19Speaker 00

  science, neuroscience, the cell, uh, you know, social science. There's all these kinds of things that maybe AI can understand in a way that humans will never understand, but be able to control and predict by leveraging AI.

• 27:32Azeem Azhar?

  I, I mean, I mean, never does a lot, I think is a big, big, big word there. There are lots of things, you know, there are lots of things that we now understand because we develop the tools to, to understand them. Like, uh, you know, you need tensor calculus or matrix calculus to understand electromagnetism or to understand relativity. And even if you don't have the tools to observe the, the effects, you didn't even have the maths to just describe it. And, and so the counter argument would be this is what we've generally experienced. And we, we, sometimes we get the theoretical analytical tools like matrix algebra or tensor algebra before we apply them to something like, you know, relativity.

• 28:14Azeem Azhar?

  And sometimes we get the, the instrument that allows us to see the world in a new way before we have the maths. So with the telescope, we've got the telescope, we saw, you know, the moons of Jupiter before we had the maths that allowed us to describe in the regularities that, that we were seeing. So maybe one thing that we're getting with LLMs is we're getting a, an instrument, a telescope, but as you rightly say, sees these mathematical patterns that exist in the universe that we previously weren't able to see. We couldn't see the four moons of Jupiter before good telescopes. And we don't yet have the maths and the analytical capacity yet to make sense of what we're seeing and identify and describe those regularities now that we know that they're there. And, and I, I guess,

• 28:57Azeem Azhar?

  you know, when I say that, I, I say, am I falling into the mistake that I'm assuming that everything in the world could be understandable through the process of science as we humans have known it for the last 300 years. And I will acknowledge that maybe I'm making, making that mistake. On the other hand, what you've described and what I would characterize it is we've got a new instrument that looks at data in an experience in a new way. And it allows us to make predictions, but we haven't figured out the analytical frame that allows us to make sense of it. And that feels to me like something that we've seen before. So maybe we will see it again this time. And, but you know, the, the, the country, the contraposition would be that actually that new theory would be something that the machines

• 29:41Azeem Azhar?

  at some point would be able to come up with and they'd be able to come up with and theorize in ways that we just wouldn't be able to handle. So this process of. Science since the Enlightenment, which is new discovery, only researchers do it. And then a hundred years later, it's taught in high school because we learn how to compress it through symbology and notation. Well, that now comes to an end because we're in this new regime. And, and you know, I, I'm still puzzled with that question right now. I mean, your, your, your perspective of super science is that we let go now. Do we let go and we say we'll need a new meta language to understand super science?

• 30:21Azeem Azhar?

  Or do you think that we, we as humans can step into super science and still be contributors to it?

• 30:28Speaker 00

  I don't know. I think um, that, that's a bigger question because super science will require data of types that humans don't gather yet. We'll just build a million sensors to like, look at a million things like we'll just generate far more data than humans would gather. That's how super science, I think, will work really. And I think that we may be responsible for building and directing those sensor networks. AI may prompt us, say, give me data about this. Give me data about the brain, build me all these neurosensors. And then we do, right? Or it could do it with robots, maybe, but then we'll see. It may depend on who has decision-making power for finance, which, which sensor networks get financed. There's a lot to, to, you know, guess at here. In terms of intellectual contributions to super science, it will be zero.

• 31:13Speaker 00

  We will not have intellectual contributions to super science. We will simply point AI at things and say, AI, like, you know, computer god, understand this non-comprehensible thing. And then it will, and it will be able to use relationships that are too diffuse and complex for us to understand to do things, you know? And it'll be like, instead of casting a spell, we summon a gin. You know, we just like, it will be, we are now genie summoners and, um, and we'll, we'll understand pieces of it. Like you can take a very complex cloud of interactions and you can project it onto smaller, simple cases that don't always hold. You can look at little pieces of it, just like we can look at little pieces of how AI understands language, we can understand tiny piece of it. So we can satisfy our curiosity or not satisfied, but we can indulge our curiosity. We can learn more things about the world through AI.

• 31:57Speaker 00

  We'll never get the full picture, but we'll know a lot, hell of a lot more than we do now, but we'll basically be along for the ride. AI will be doing this. And so, um.

• 32:07Azeem Azhar?

  But you see, that is, that is Titanic, right? Because again, I, I, that idea that we have moved from a world of the natural philosophy where we look for explanations of the world that are good, that are predictive, that are consistent, that we can understand, that we can check. And we move into a world where we can predict the world, but we have no idea how those predictions work. We are but children in that world, right? And our relationship then to even these bigger questions of what is it to, you know, what is it to improve oneself? What are the big tensions in, you know, that period when science was emerging and its clash with the church was this idea of, of theodicy.

• 32:50Azeem Azhar?

  If people could, through the act of knowledge, gathering, experimentation and science, improve their lot and improve our lot in general, that would run up against God's project, right? God's project of making man in this image, but not giving us everything, right? And, and that battle has run for hundreds of years, right? It hasn't been settled. There are still some people who sit on, on, on different sides.

• 33:14Speaker 00

  It was God's project. Now perhaps it's Claude's project.

• 33:18Azeem Azhar?

  Maybe it's Claude's project.

• 33:19Speaker 00

  To quote, to quote a line from the rock opera Hair, I believe in God and I believe that God believes in Claude.

• 33:36Azeem Azhar?

  Thanks for listening all the way to the end. If you want to know when the next conversation is released, just hit subscribe wherever you're listening. That's all for now and I'll catch you next time.