AI’s New World Order: US-China, War, Job Loss | Tyler Cowen

America unintentionally jumpstarted China's space program as well as development of ICBMs.

Do I want China to be part of the same hardware network as I am or for China to fully develop its own? It's not clear which one is better.

America's great innovations would have been significantly delayed without its earlier period of espionage.

Get the AIs to think like Nietzsche is harder than to get them to think like John Dewey or John Stuart Mill.

Alexander Hamilton, as Treasury Secretary, was like, clearly it's better just to steal the technology than to do the thing from scratch ourselves.

In a test of 100 questions, O1 Pro already would beat me. Answering the economics questions, it beats me. In areas where I have some background, it's still beating me.

Where do you think the edge is going to be for humans?

More people should study how to operate AI and spend less time learning about the subject.

Peter Thiel said crypto is libertarian and AI is communist.

I don't I don't think that's true. Here's how I would rephrase it.

Deepseek is a Chinese AI model approaching the performance of the best US models, but trained at a fraction of the cost. And the world is in shock. My guest Tyler Cowen was the first columnist to break the news on Deepseek weeks before the markets themselves reacted. And Tyler's surprising claim is that the very US chip sanctions designed to thwart Chinese innovation engendered this innovation. By restricting China's high-end chip supply, the US encouraged China to pioneer techniques in cost-effective training. In the beginning of this interview, we're going to discuss what this Sputnik moment has to teach us about effective policy, US-China relations, and how innovation actually works. And after that, Tyler is going to share his comprehensive forward-looking view on AI, about how it will transform knowledge work, warfare, economics, and everything in between. Tyler believes that AI will radically split the population into two, with a small portion of people who are able to use the technology leapfrogging the rest in all domains of work. A new global super rich class and everyone else. What I found most valuable about this interview then are all the practical tips that Tyler sprinkles along the way about how we can stay ahead of the curve and not be made obsolete. My name is Jonathan B. I'm producing this series on the philosophy of AI as a fellow of the Cosmos Institute, where Tyler is also a founding fellow and board member. If you want to follow along this episode via transcript, the link is in the description. Without further ado, Tyler Cowen. Peter Thiel had this line where he said crypto is libertarian and AI is communist in the sense that AI benefits from this large-scale centralization of data.

I don't think that's true. I'm not even sure that Peter still thinks it's true. Not meaning to speak for him, but we do see that, say, the FBI can now trace a lot of crypto transactions pretty readily. Right. To be a criminal and use crypto is no real way to escape the law in many cases. And I think when Peter said that, the dominant form of AI was surveillance. And to think that that's totalitarian, or at least autocratic, is easy to see. But once you see that AI reshapes workplace relationships, knowledge, freedom of speech issues. It has a lot of further-reaching effects. And again, I'm not pretending to have exact predictions, but it's not really biased toward the status quo. I would say people are seeing that more and more. So if your country is ruled by a status quo and very intent on defending the status quo, uh, they may be seeing more change than countries that are used to things changing more. So, uh, it's quite possible in the future AI— I wouldn't say it's libertarian, I wouldn't say it's anti-autocratic, I would just say subversive. Subversive of many status quos, some of them good, some of them bad.

Almost 2 years ago, US enacted a blanket ban on the most powerful AI chips, restricting their sale to China. And in a recent column, you argued that this backfired. US policy succeeded in hampering China's ability to deploy high-quality chips in AI systems with the accompanying national security benefits, but it also accelerated the development of effective AI systems that do not rely on the highest quality chips. Namely Deepseek.

If I had been President Biden, I probably would have done the exact same thing. But that said, sometimes you need to realize that what you tried didn't work. Now, China might have accelerated its AI development even more with more powerful chips. But do I want China to be part of the same hardware network as I am, or for China to fully develop its own independent hardware network and other innovations? It's not clear which one is better.

Your column reminded me about how America unintentionally jumpstarted China's space program. So there's a Chinese man called Qian Xuesen who was born in the early 20th century. He was born and raised in China, came over to study at MIT, Caltech professor, was a colonel in the US military, worked for DOD, and was eventually a co-founder of JPL. But during the Red Scare, people were accusing him of communist sympathies and they put them under house arrest. And he was eventually deported in a prisoner exchange swap with like Korean pilots in China. And then when he went back, he jump-started China's entire space program as well as China's development of ICBMs. If you look at talent migration between the world's superpowers, America is the biggest gainer of talent.

Including from ethnic Chinese.

Yes, including from ethnic Chinese. And so the general question for you is, when an economic innovator wants to defend their innovations from an economic laggard, there seems to be two general policies. One is to open up so much to get all of their best talents, to make them dependent on your hardware. But the other one is to cut everything off, not just on the hardware side, but also on the talent side. Which strategy do you see most effective in economic history?

I want to make my bet on openness for the most part. We're not very good at choosing the right restrictions and enforcing them, There's always more direct forms of espionage. You're not actually controlling the technology or keeping it in your borders. So you might as well get the advantages of openness. That's my general feeling. I think in particular cases that feeling can end up being wrong. Were we too open with the secrets of the atomic bomb? Well, actually that was a remarkably closed program, truly secret, and it was a big complex out at Los Alamos and America truly didn't know what was going on. But at the same time, those secrets still leaked out to the Soviet Union. So I'm pretty skeptical about these attempts to make it closed.

I see. And in some sense, it's just a lot more fitting for the American system as a whole, right? Because everything around America is built off of openness. So I'm saying even if closeness is the best overall strategy, the cost it would take to turn the entire society into that would be just too great, right?

We're not good at closeness. We're pretty good at openness. In a sense, maybe no one's truly good at openness, but we're probably better than any other country. So play to your strengths and hope that's the right thing to do is sometimes the best strategy.

I see. I want to double down on what you just said there about the difference between protecting, for example, atomic bomb secrets or AI models. There seems to be very different challenges in how you protect these secrets. Like, for example, people are claiming that Deepseek was trained on distillation, right? Which is a method, for our audience, that you essentially have the lower-level AI model quiz the higher-level AI model. And so that seems— if distillation can get you quite close to the leading models, that seems a lot more difficult to protect than the atomic bomb, right? And so that also makes a case for openness in AI, doesn't it?

And just the number of employees at major AI companies It's not like working in the Pentagon. I mean, there are security protocols, but they're nothing close to, say, what the CIA would have. And I'm not sure you could have very high-quality labs attracting the best people if walking in and working there felt like—

an interrogation.

Yeah. So that's another closed versus open decision. We've gone mostly in the open direction. I think I would have done the same thing, but We'll see, right?

Right. I see. So two arguments for open now. One is that it's better fitting to the American cultural mores. And the second one is it's just the nature of AI itself and how it's different from— like, you can have 10 scientists working on the atomic bomb. The field itself lends itself more to openness.

And Deepseek, it's online, it's there, it's embedded in consumer products like the American systems. So it's not like plutonium, not at all.

I see. Let's go back to the CHIPS ban because kind of one of the learnings you drew out is we need to think about the second-order consequences of different policies. What can a statesman do better in understanding the second-order policy effects?

Well, if you're an American statesman, the best thing you can do is familiarize yourself with AI. My sense is remarkably few people in government are at the AI knowledge frontier. I don't mean as builders, but even just as users of the system. So just wake up, you know, pay to access the best systems and spend some part of your day playing around with them. They will even help your daily workflow right now, no matter what you do. So that's a very simple first step, but we're not even doing that. A critical point I make to people, and this may be time dependent, but it is worth spending some extra money to have access to the best systems, whatever they may be at a point in time. If you don't have that access, it's actually quite a bit harder to understand what is going on. So I know that not everyone out there can afford the best systems, but at least consider this as an investment. And the way it stands now, you can access them for only a month for a small amount of money, $200. Just try it for a month. If you're disappointed, okay, you're out $200. You'll at least have learned that you're disappointed. If it's worthwhile, you can continue. So not enough people are crossing that mental bridge of just saying, I need to try the very best. Apart from the direct services or usefulness of the AI, just as an educational experience for themselves, for themselves, I recommend this to everyone but the very poorest people I know. And I would just want to stress this point to our listeners.

I see. Let me give you a quote from one of my favorite essays by René Girard on his philosophy of innovation. Not so long ago in Europe, the Americans were portrayed as primarily imitators. Then, in very few years, the Americans became great innovators. Public opinion is always surprised when it sees the modest imitators of one generation turn into the daring innovators of the next. Until quite recently, the Japanese were dismissed as mere copiers of Western ways, incapable of real invention in any field. They are now the driving force behind innovation in more and more technical fields. At this very moment, imitators of the Japanese, Koreans, Taiwanese are repeating the same process. They too are fast turning into innovators. Hadn't something similar already occurred in the 19th century when Germany first rivaled and then surpassed England in industrial might? Do you agree with the general message from this quote that imitation, copying, is in some sense a necessary first step to a country or a culture becoming a genuine innovator?

There's a lot in that quotation that makes me nervous, starting with some of the details. So if I look at Roger Williams and William Penn, who are important figures for religious toleration and liberty in the United States, they were innovators quite early on. If I look at the American Declaration of Independence or Constitution, those were incredible innovations, the Federalist Papers. Quite early on. So to say we started by copying, it's true a lot of the key American ideas came from Britain, came from the Levelers, came from John Locke. But still, we were both innovators very early and established that precedent. Now, it's interesting, Girard says, oh, the Japanese, they're innovating more and more. That's turned out not to be— not really to be true, right?

Right.

So who's an innovator and who isn't Just be careful where you look. Be open to the fact that innovation can come from places you don't expect. And if you have some overly simple formula— well, it's those who copy first, or it's those who do this first, whatever— you're probably going to miss it. Better to keep an open mind, travel a lot, and talk to as many people you can, and not get too hung up on fixed rules.

I don't think Girard meant to say imitation is a, is a sufficient sort of step to becoming an innovator. But I think what he's saying is that it's necessary, and it's very surprising that the would-be innovators of today were often seen as these, like, lowly imitators. And do you think that's validated in economic history?

Yeah, the Beatles, you could say they were imitating rhythm and blues, Chuck Berry, the Everly Brothers, and in some ways they were. Elvis Presley, the Beatles were the first to admit it. But at the same time, it was quite fresh and bold. And to think that copying and innovation are opposites is maybe the mistake here. Be more Hegelian about it, I would say.

Right. To push back a bit on your thesis about American innovation from day one, that might be true in the political or the philosophical realm. But when you look at the early American economic innovations, they were doing the exact same pirating techniques that America is accusing China of. So like Alexander Hamilton as Treasury Secretary, instituted an entire regime of like forced economic espionage and technology transfer from Europe, especially Britain.

That's right.

And so there was entire incentives, like openly saying, we are going to like essentially bribe people. And so you have people like Samuel Slater who got all the— I think it was the water-powered like cotton mill or something like that. And he had to like smuggle himself out of Britain. Here he's called the father of the American Industrial Revolution. In Britain, he's known as Slater the Traitor. And so, but reading Alexander Hamilton's reasons for why economic espionage, it actually made me like very sympathetic to why laggards steal technology. And his reasoning was just like, look, we don't know how to do this. These guys know how to do it. Clearly it's better just to steal the technology than to do the thing from scratch ourselves, right? So is it almost rational for economic laggards to pirate, to steal, to do espionage?

I'm not very upset by what China has been doing. If something is illegal or unethical, then yes, nominally I would be against it. But in terms of my level of concern, involvement, it's mostly inevitable. It does help China catch up in terms of living standards. I do understand there's a military rivalry. But at the same time, I think it's in the best interests of the world and the US to have a richer China than a poorer China. And we've lived under a poorer China, say, in the 1960s. That was no fun. And it certainly was awful for the Chinese.

Right.

So I think you have to take your chances, engage, make the world with a richer China better. There's a lot of problems you're going to encounter. But try to think about it in as positive some terms as you can.

Right.

It's very interesting to look at a Cubist painting by Picasso and the Cubist painting by Braque right next to each other. Sometimes even experts can't tell the difference. Are those guys copiers of each other? Are they innovators? Well, they're both. Again, let's be Hegelian. It's a dialectic. These two things are not opposites.

I see. Let's move on to the second topic of this conversation, which is regarding your two books, The Great Stagnation and Average Is Over. So in 2011, you wrote The Great Stagnation, and you predicted that it's something that would continue into the mid-run. Okay, so let me give you a quote. Real income growth, widely distributed, of about 2-3% a year. I think it is currently impossible. We don't have the low-hanging fruit to make such a scenario real. What changed?

I think the Great Stagnation ended in 2020. If I had to pick a date, it started in around 1973. In 2020, we have mRNA vaccines. Now, I know it took 20 or more years of work to get to the point, but there's a sense in which the vaccine we used was designed in a day. And it worked. And then it was ready in well under a year. At the same time, experts were telling the New York Times the optimistic scenario was it would take 4 years. And we did it in under a year. It could have been quicker yet if the FDA had not messed up a number of procedural things. So that was astonishing. It saved a large number of lives or just limited sickness for many people. And that's a big enough breakthrough. I would say that alone did it. But that's not a standalone event that we could, in a sense, design at the margin the vaccine in a day was due to the growing power of computation. So I think the general-purpose technology that broke the Great Stagnation was a combination of the internet and what I would call computation. And AI is another example of that. So AI and biomedicine are two sectors making amazing progress in the last, say, 5 years. A lot of other sectors are still crummy, stagnant, or worse. Cost of construction, that's a nightmare. It's hardly all going great guns. But to me, those two sectors are big enough, and AI is likely to spread productivity improvements to enough other sectors. I think it's a number of different forces. Some of it is just background Moore's Law. I know it's slowed down by some measures, but you're still getting ongoing improvements just in basic computational power. The Transformers paper, you know, attention, those were key breakthroughs for AI, what was done at DeepMind, then OpenAI being formed with Musk and Thiel and Altman, all of those coming together at roughly the same time. And then the internet just developing and being in more places. So when AI was ready, everyone was so connected to the internet that AI could spread and be adopted really very quickly. The adoption curves for ChatGPT are pretty extreme. All of those things.

In other words, the Great Stagnation was only a symptom. It only appeared Great Stagnation on the application layer, but there were these innovation trends that were continuing. The culture was totally fine. It wasn't a time to ring the alarm bell. It was just we need to wait more before this all comes back into application?

I would say the internet culture was fine. Again, a lot of areas like construction, it's clearly not fine. And I think what I also said in the book, the final chapter, I said the internet has already gotten us out of the Great Stagnation in a sense. We just don't know it yet because the new fruits are not here. But the key thing we need to have the breakthrough was the internet, and the internet's already here.

I see. The name of your blog is Marginal Revolution. Although it seems like what's most critical for her history are the non-marginal moments, right? The transformer paper, Jesus of Nazareth, the invention of the bomb. How do you think about these, these two non-marginal and marginal forces on history?

Well, they both matter. But keep in mind, the big breakthroughs are the result of many marginal changes coming before them. So Jesus comes out of a longstanding rabbinical tradition. Is building upon the Torah, many other earlier writings, other religious traditions that maybe have not come down to us. And he's not appearing in a vacuum. So the marginal revolutions will remain important for the indefinite future to come.

This ties very well back into our discussion about imitation and innovation.

Exactly.

We shouldn't be thinking about these things as completely separate.

That's right.

I want to do— although we're out of the Great Stagnation, it's not all hunky-dory. Let me quote to you your blog. Hardly anyone you know, including yourself, is prepared to live in actual moving history. No one at the beginning of the printing press had any real idea of the changes it would bring. No one at the beginning of the fossil fuel era had much of an idea of the changes it would bring. No one is good at predicting the longer-term or even medium-term outcomes of these radical technological changes. In the absence of the ability for good prediction, How should we live in moving history?

Think about your children. I think psychosocially, it's very disturbing that you can no longer tell people what kind of world they should prepare their kids or grandkids for.

Yeah.

So my daughter, Jana, she's 35. Basically, I adopted her when she was 12. She and I had a lot of discussions, and I felt I could outline to her pretty clearly the world she would be starting work in, and that it would more or less be just like the world when she was 12. There were differences, but you could make the same kind of plans and your plans would be completely fine. And indeed, for her, they were, and it's gone great for her. That's just not true anymore. So just digesting that fact and getting over it— like, if you think about the Obama years, remarkably little happened by most standards. And we're all used to that, at least in the United States. Some countries, a lot happened. And that's not going to be true anymore. We're so not ready for it. We're not familiar with real change. Even if you think it's entirely positive, and that's not my view, but even take the best-case scenario, we will be stunned and shocked and disoriented. In some ways, the purely positive scenario might be harder for us than a mixed scenario, because a mixed scenario, fighting off these challenges gives you something to do, gives you a sense of meaning and purpose. Like, if it's only all positives, maybe that's tougher mentally. So I don't have real guides, but I tell people, like, invest in friends, great peers, mentors, and make sure you have some fun, rewarding, cheap hobbies.

Right, right. And stick on this wave of innovation, right? Always be on the bleeding edge and try—

That's right.

Right.

And you don't have to be an AI programmer or builder. Very few people can really do that. But just use it and be familiar with what's happening. You can do that for free.

Although I immigrated to the West as a teenager, effectively, and it was striking how in China, where I was born and raised, we were living in moving history for at least 4 generations.

You're used to it.

I'm used to it. My great-grandparents lived in the empire. My grandparents lived in the founding of the republic. My parents lived reform and opening, and I immigrated here. So when I look back, all I see is radical change. But then I look at my peers, my friends that I made here. No, no, no. Their dad's lives look like their lives, looks like their granddad's lives, looks like their great-granddad's lives, right? So most of the world is actually totally ready for moving history because history never ended for them.

Much of the world, that's true. There's places that have not grown like China. It's striking to me, I was born in 1962, so I'm now 63 years old. Really at no point in my life did I need to relearn how to operate any device. I learned how to drive, I was 16, common history. I didn't have to relearn anything, nothing. That's stunning. It's quite bad in a way. Now we're at the point where I'm going to need to learn, relearn new things, but that I could go so many years of a life, Well, how to cook. I know more dishes, but how to operate all the equipment? I guess I learned a microwave, never really quite learned it actually, when I was, you know, a young teen. That was the only change. Everything else is the same.

I really think this is a case of grass is greener, because when people who've grown up in moving history, they look at the West and they're envious and saying, this kind of stability, right? The fact that you can know what your grandkids are roughly going to do and plan out this this lifelong plan, some are envious about that.

People prefer it, and probably they should prefer it. I've had a very happy life, and that's the psychosocial problem. But you only get ongoing periods of stability by taking some chances and having big changes every now and then. You can't just keep it in some guaranteed form forever. Like, there is no UBI for humanity as a whole. That's one way to put it.

I see. Let me give you another quote. New AI developments pose a great conundrum. We don't know how to respond psychologically or for that matter substantively. And just about all of the responses I'm seeing, I interpret as copes, whether from the optimists, the pessimists, or the extreme pessimists. What are the different responses to AI in your view and why are they copes?

Well, a few of the responses I see all the time, people will just say, oh, it hallucinates all the time. It's a fun app on my phone, but it's not really important. That's a cope. I don't doubt that it's often sincere. These people are not just lying, but they don't put in the effort to look further into it, and thus they stick at the cope. Another cope is to just say, well, it's going to kill us all in X number of years' time, and there's nothing we can do about this.

It's fatalism.

Fatalism. I also think that's wrong, but it's a cope. Right?

Well, because it saves you from having to engage with it seriously.

That's right. Now, I know a very small number of people who act consistently with that belief, max out on their credit cards or start taking too many drugs, do crime. But most people don't. It's just a compartmentalized view that they don't really act on for the most part. Then there's other layers of cope which are sort of intermediate They're actually reasonable positions that could turn out to be true. It's like, well, we need to change these few things and then we'll settle into some new version of the world we knew. Those could be true. I guess I think the changes will be bigger and more radical than that. But I think partly they're held for motives of cope, even when they might be true. The people just don't really want to be bothered having to learn this new thing. When they haven't had to learn how to drive some new kind of car. And so they stick at that. Like, oh, I need to teach my class how to use AI, which is not a bad view to hold. I try to do that. But the notion of maybe your class won't need to exist, it's being held at bay by the, I'm gonna use my class to teach my students how to use this.

Right, right, right. I'm going to drive my car to my horse buggy dealership. Right, right. Yeah, right. I see. So there's another kind of— I wouldn't call it cope, but perverse psychology from a lot of people working in AI, which is to— again, this is going to be unfair, but it's something like exaggerate the risks to make themselves feel special.

That's right.

And so, you know, before GPT-3 came out, I had acquaintances who worked in OpenAI, and these people were claiming, "Oh, 2 years, nothing left for people to do. Let's just meditate." And like, that's not cope, right? They're working in this. But it's a kind of grandiosity.

That's a Nietzschean view you're putting forward, right? It would have been harder to get that from the AI.

Yes, right.

And I agree, that's another psychological reaction you see. You know, guys who want to boast to their dates about how they were mugged once and almost killed. There's some sense of power they get from relating that story. And that pops up in a lot of different ways.

What do you make of the existential risk movement in general?

Well, I think they're wrong. I mean, it's interesting to me how many of them are the people who have accelerated AI. So it's been counterproductive from their own point of view. But I would just say this, there is no peer-reviewed literature supporting what they say, nor do they seem to make any attempt to create one. That doesn't mean they're wrong, but I've issued the challenge. I say, well, the global warming people were worried, they built up an extensive literature basically showing their worries are validated. I say you do the same. If you do it, I'll be very interested. And I've even offered to referee some papers for free and so on. So far, none of this has happened. So it's not supported by peer review. And there's nothing, and I do mean nothing, in market prices to support this fear. Again, that doesn't mean it's wrong. But our two best sources of information, you could call peer-reviewed science and markets, they're both putting up a big zilch. And I'm like, I'm your audience. Like in the Harry song, Jesus Christ Superstar, like, oh, I'm listening. Here I am. You know, your 17-page hyperventilating blog post, like, that's obsolete. You've got to do better. Right. And I'll stick with that challenge.

And here I think the psychological explanation makes a lot of sense, like why they are claiming X-risk and accelerating at the same time. It's kind of a look at me, right? Like, I can't remember which theologian said this, but something like admission of sin is to claim credit for the wrong.

Very good.

In the sense that the thing I'm working on is so important that we're all going to die if we don't listen to these things, right?

Yeah.

I see.

More Nietzsche. Theology and Nietzsche in the same sentence.

There we go. One of the genuine risks about AI is the fear of war. When you look at human history, technological innovation, it often comes after major technological innovations that people try to use it in warfare. How do you think about that? I think much more legitimate risk.

This is by far my biggest worry, simply evil humans or misguided humans doing destructive things with the AI. I mean, one point to make is this is a different kind of AI than what people are talking about right now. It's not per se large language models. Large language models may evolve into something that's more of a destructive danger. It's quite possible. But say the AI embedded in drones, Drone warfare could be very bad if it favors offense more than defense. Technology, with or without AI, it's always changing. If the world is doing okay, you're basically rooting for the defensive technologies to win. I would say that's the case right now. The more you remix the arms race, the greater the chance you're just not going to stay put with the defensive technologies always winning. I'm extremely worried about that.

You think the latest developments of AI have favored defense over offense?

I'm not sure. I tend to think that drones will favor offense.

Yeah, yeah.

Now you could say so far, you look at the Ukraine war, Ukraine was defending itself. Ukraine, I accept that data point, has favored defense. But the notion that a large open society has so many targets and defense is pretty hard, I think in the medium term it will favor offense.

Yeah, yeah. We can't even figure out what's going on in New Jersey.

Exactly.

Yeah, right. Let's talk about the future of work as you described in "Averages Over," because the trends you, you foresaw there have just been magnified with the latest AI models. What is the future of work going to look like?

Many, many jobs, you'll just be working with the AI, and that's already the case. So people who are good at working with the AI will be far more productive. Sam Altman predicted a while back that pretty soon we'll have these billion-dollar companies maybe just with 1 employee. I think 2 or 3 is more likely because 1 person can go crazy. You just need a bit of support and you'll be able to afford it, doing the work of a whole pretty large company and having a lot of it done by the AIs. So people who are what you'd call project managers, they'll just be way, way more project managing. And those are the people who will learn the most, our great project managers.

Right. What's really interesting is that AI seems to favor getting rid of the white-collar work first, that robotics is lagging a lot further behind, uh, sort of, uh, symbolic manipulation.

Right.

Let's put it that way. Right. Which is interesting because people want the white-collar jobs usually.

But the white-collar jobs are in more short-term trouble. Those are the people who will lose status. That will be socially very disruptive. If you're a decent gardener, probably you're in great shape. Over the medium term, you'll also need to learn the skill of reading out someone's AI-generated output as to what the garden should look like. I think most gardeners will be just great at that. But it is a new thing. They'll have to figure out how it works, or maybe they'll end up doing some of the work with the AI themselves. A lot of physical labor is just going to earn a higher wage. If you're physical labor in the energy sector, my goodness, demand for energy is going to go through the roof. We already see this happening. There's a question of which localities allow it to be built, but wherever it's built, a lot, lot more jobs being created. Medical trials, just running the routine elements of them. Of course, the doctors, the researchers, yes, but just routine processing of everything that's being done, including on the legal side, there'll be a lot of new jobs because AIs will generate a lot of new ideas for medical devices, drugs. A lot of them won't be good. We'll have to test them. That will be an enormous growth sector, I think.

Let me give you a quote from that book, Average Is Over. I imagine a world where, say, 10 to 15% of the citizenry is extremely wealthy and has fantastically comfortable and stimulating lives. The rest of the country will have stagnant or maybe even falling wages in dollar terms, but a lot more opportunities for cheap fun and also cheap education.

And healthcare, right? And legal advice, many things. So people will be doing great, but it's not going to show up necessarily in the form of normal wage growth. You get a raise every year. Just so many things will be free or near-free once you've paid for some basic subscription, right?

I see. But if democracy requires a strong middle class, um, and this is fundamentally an aristocratic force —right? It's about separating the top 10% from the bottom 90%. Do you think democracies will be challenged? This is again circling back to our initial conversation by AI.

Well, I think democracies are always challenged, and the Founding Fathers understood this, and they had no notion of AI. So they will continue to be challenged is the way I would put it. I think people will have happier lives overall. That's no guarantee of democracy going well. But I don't predict some kind of slide into chaos or totalitarianism. It'll just be a very different world. We'll have to do a lot of coping, adjusting of expectations. I'm not convinced it will go well, but we have the pieces to succeed with it. It'll be up to us to do it, though. Right.

However, it doesn't have to go to tyranny. It could become something like the Roman Republic, where you have a very small minority senatorial class, and maybe these are the 10% technologists who can actually wield AI, and then you have the plebeians. That seems like a political model that's a lot more fitting for the economic picture you described here.

Well, I don't know. Let's say the 10 to 15% of people, let's say they have the net wealth roughly equivalent to having $100 million today, which basically means you don't have to worry about anything. A lot of Americans have that already. Right. I'm not sure the exact number. It's not 100 million, but like net worth in the millions that is unimaginably wealthy to most of the world. So you look at the federal budget today, most of it is quite popular with the median voter. You have Medicare, Social Security, Medicaid, defense spending, interest on the debt. It's almost all of it. Whether or not one agrees, it's what most people want, and we have it. It's pretty democratic. I don't see why that has to change.

Yeah, I guess my intuitions here are a lot more with Marx, that the political system is an outgrowth and must, at least in the long run, match the economic structure in some way. And it's really difficult for me to see a world in which, you know, you have 10% or 5% of 100 millionaires, everyone else is like a plebeian who doesn't really have a say in the economic system, and there'd be no sort of political change because of that, given where we are and the strong middle class we have today?

I would just say I don't have a clear prediction. So recently, income inequality went up quite a bit. The guy who's been elected president twice, almost won in the middle, is Trump. At first, like, hardly any wealthy people wanted Trump to win. Now that he's been reelected, that's changed. But he was not the candidate that all the rich people were putting up. Very much the opposite. Unless you think there's some kind of coup d'état where just voting is shut down. Again, I don't have a prediction one way or the other, but it doesn't seem rational to expect it. It's not priced into markets. It's not in prediction markets. It's going to be a lot of big challenges. I just get to them and let's do the best. We're going to have these challenges anyway. I see.

Let's move on to the next topic of this conversation. Which is regarding AI and authorship. How do you view your intellectual production as both a scholar and a public intellectual differently now with AI?

Well, as you know, more and more people query AIs, whether it's Deepseek or OpenAI or Claude, Meta, Google's Gemini. The more people do that, the less they're directly reading humans. You want your ideas to be input for the AIs. You're writing at least as much for the AIs as for the humans. One thing that means is you want to be doing something open source that the AIs can read. That's important. You'll have more influence. But I think also you want to produce content that the AIs find appealing. We're not sure what that is, to be clear. But I think there's a premium on originality. If you're just repeating what many other people have said in a routine way, I doubt if you'll shift the opinions of the AIs very much. And the AIs, to me, they seem pretty positively oriented and pretty friendly. You could say a little submissive, or they apologize a bit too much. Does that mean they're going to prefer humans who are the same way? I don't know.. But you should at least think about that question. And maybe by being a little nicer, including nicer to the AIs, the AIs will like you a little more and listen to you a little more.

At least consider that as an option. Suggestions are nicer, potentially to other humans and AIs, open source so that they can read your work. Is there some way that you structure the ideas that you're making it like the format? No, I don't.

I assume the AI is smart enough to figure out what I mean. You could say it's my deepest and most sympathetic reader, my reader with the best memory, my reader with the best set of background knowledge. So take that into account a bit. But I don't have some formula for how to write for them. So it's not like, oh, if you put everything in a poem, the AIs will love that. Like, if that were true, I'd love to know it. But I've never seen signs that we can figure that out.

But once you frame it that way, other than making it open source and maybe being nicer to the AIs, hiding your more doomer opinions about the AIs, It doesn't seem to be that different than just writing a good book for humans, right?

Well, that's not what most people do for other humans. You could say that properly understood, the two enterprises are pretty much the same. But I'm trying to get people to properly understand it. Like if I'm writing a book for you, sentence one should not be, "You're an idiot," right? I figure you'll get more out of the book if I'm just generally open and nice. We're not sure the AIs are like that, but they are, in a sense, built in our image. They're trained on us. They're taught to be nice through reinforcement learning. So I'll just take as a premise that I can't prove, be nice and be nice to the AIs.

I see. Nietzsche is notoriously not nice, yet it seems like he has a pretty big influence. When you query stuff, people often raise Nietzsche. So that seems to be a counterpoint.

They know Nietzsche a lot, but I think to get the AIs to think like Nietzsche as their first-order response is harder than to get them to think like John Dewey or John Stuart Mill. I don't know that. I'd love to see that measured, but I suspect it's true.

But that's not because there's something about the nature of AI that likes niceties. It's because of how people are training them. It's because of the OpenAI and their more left-leaning attitude towards life. If we had a fascist AI, I'm sure it'll sound a lot like Nietzsche, right?

In time we will get that. And you need to ask yourself, which of the AIs are you writing for? Because you're writing for the whole future of AI. Now, I would rather write for the John Stuart Mill AIs, but I understand different people will make other decisions. But it's all the more reason why you want to make the John Stuart Mill AIs smarter and better.

Right. That's interesting. So properly writing for humans and properly writing for AIs, in your mind, converges. Mostly, yeah. Mostly, yeah.

The humans, you have to explain things too much. I don't like that. The AIs you take for granted, they know all the background.

I see. One thing you mentioned about writing for AIs is that this might be your chance at immortal glory in the sense that only so many authors will be read, not just 5 centuries from now, 1 century from now, but in some sense, all of your writing will always be read by AI. Is that right?

There's a very good chance that's true. We don't know it for sure. Just as current AIs are not trained on 13th century manuscripts, at some point in the more distant future, they might decide people such as myself are totally obsolete and so cut them out of the training canon. I think we're quite a ways away from that. Uh, and the AIs will be interested in me, if nothing else, as an early theorist of AI. People always want to know their family tree, their family history. I had someone do that for me. I was fascinated. Maybe those were not interesting people at all. What do I know? My great-great-grandfather, but I wanted to know about him, what he did. So I think the AIs will want to know a lot about Isaac Asimov, for instance, and a more than average amount about me. I wrote a book and many articles, blog posts on AI. So I think in their view, I kind of check in higher than a lot of other people, maybe higher than I deserve, but I'll take it, right?

Other than nicety versus meanness, are there other ways that you are adjusting your content for AI? It's like, are you, for example, sharing more of a broad scope of your own personality?

Of course, more subjective, more— you could call it biographical, more like a memoir, more personality-driven. And that's for two reasons. One is to stand out to the AIs, but the other is for your human readers. Since at least the current AIs— and Deepseek is somewhat of an exception— but they're often bland as a first-order ask. You can get them to be more exciting. And if you sound too much like them, I think you just get ignored today. So you need to change your own mode of presentation. Right.

It's something like, even though AI writing is not on par with the best human writing yet, the fact they can kind of write like humans cheapens bland writing as a whole. That's my view, yes. And so you need to inject excess personality.

Correct. Catherine Boyle tweeted something like this today. She heard my podcast with Roth Douthat and has decided she needs to let herself go more in her tweets. And I think that's exactly right.

Right, I see. Until the AIs start doing that as well.

And yeah, well, you have to stay ahead of the curves, right? But, you know, start with one step and see how things go.

But that's why I think the podcast and the lecturing medium and rhetoric as a whole, I think, is making a comeback over writing because rhetoric is embodied speech. The video AIs are a lot less advanced than the writing AI is in terms of generative AI. That's why I'm doing this recording.

Now, partly without AI, we'd have nothing to talk about. But as a separate thought experiment, if it wasn't for AI competing against my writing, the number of videos, podcasts I would do, I would cut by a factor of 3 or 4. And do more writing. And do more writing. But no, we're doing this. This is better. This is better and more important, and it will have more impact. I see. And the AIs cannot copy it yet.

Yet. Yet. Yeah.

But I even think that's further away. So I think there's something about video. So say the AIs could copy our mannerisms, how we speak. There's already a clone of my voice. It's very good. There'll be a clone of my mannerisms within a year or two. If you feel like you're watching the AI clone, I think there's something disappointing about that. In a way where if you just ask, you know, ChatGPT a question and you read text, you're used to Google impersonal internet. I don't think people are put off, but they would be put off by the doppelganger of me built by AI. Yeah.

In other words, the bar for the uncanny is a lot like— or like—

Even if it's perfect, people want to see for now the real you, the real me. Oh, I see. That may not last. I see. So there is an uncanny valley problem. It's not perfect, but say it's perfect in 3 years, I think the real humans will still have an advantage.

Right, in the same way that we still prefer to watch humans play chess.

You can go to musical concerts now, I think ABBA has done this, some other performers, where it is literally an AI-built clone. The sound quality, the music is probably better than like ABBA concert performances ever were. There's a market for these, I think they're pretty popular, but I don't think they come close if you could, actually seeing, you know, live ABBA in its prime in terms of marketability. So we have that already, and there's a niche for it, but it's not as popular. Right.

So you have a double claim here. One is that it's very hard for AI to make it not uncanny. But the even stronger claim is that even if it was perfect, even if people couldn't tell, people still want to see the real deal.

That's right. And in print, I think that's different. Right. I just want often the right answer. So if I have a question about trade theory and if Paul Krugman writes the answer, that's fine. But if ChatGPT writes the same answer, that's fine too. I don't feel gypped. I see. Or tricked.

You mentioned that because of AI, you're thinking about yourself less as a producer of content and more as a networker.

Well, that's still a producer of content. I think it's different content. I'd say I'm a producer of content more and more. And it's not just videos and podcasts, but personal appearances. I have way upped what I do, and that the AI is not at all close. It may copy us on video, but for it to show up and shake hands, that's, you know, past my relevant time horizon. So that's what I do more of the most. I see. And that networking is part of that. I see. But to think of personal appearance with charisma as the fundamental unique thing that is surviving.

Absolutely. About 4 years ago, almost at this point, your Emergent Ventures gave me the first grant that I received to produce a series of lectures on Girard.

Yeah, those are great. I recommend them to everyone.

Thank you. And so I had spent about like 3, 4 years reading all of Girard's canon at that point, and I wanted to be exactly precise in everything I say. So I sat down with my friend David Perel, another, another EV grants winner, And we did the lecture multiple times. And I wrote the entire script down. Like, I recorded and then edited our entire transcript of these early, like, real lectures. And I projected a teleprompter behind David's head and read off it. And the initial idea was that, no, no, I want to trade off as much roboticness of the delivery for the accuracy of my words because I'm trained as a philosopher. But then I realized that's completely wrong trade in modernity. That's right. You want to gain the authenticity, even if it means losing some accuracy. I very much agree.

When you hear Peter Thiel on video, he's on YouTube a lot, and I always enjoy listening to those. He is not a perfectly smooth articulate speaker, and that is better. So like his podcast with Joe Rogan, which was extremely popular, There's some times where Peter just appears stuck, but those are some of the best moments, actually. You sort of see, hear, feel Peter thinking, and that's amazing. Better than if you got this perfect Mr. Spock-like answer with no pauses, no bumps, no whatever. Right. Interesting.

So it's almost like the surplus of AI perfection is making these human imperfections all the more and more valuable to the audience.

I wouldn't call them imperfections, but I agree with your point.

Yeah, I see. Is the era for blogs now over?

Well, it's been over for a long time, but what's even a blog maybe was never well defined to begin with. But if you're writing, you need a brand name or you need a truly unique product. Otherwise, you're done. Now, a key thing about marginal revolution, much of it follows the news cycle. So some AI follows the news cycle, but the very best models do not. They're trained in advance. So if you're writing about what happened yesterday, like, should— how should we reform USAID? I have no worries that we can produce some unique content there that the AIs are not catching up to. And something like Perplexity, which is updated all the time, so to speak, uh, that will pick up on what we write. So I'm very happy doing that, but I think it does mean there should be fewer generic advice posts, you could call them, and more things that follow the short-run new cycle. And again, we're doing that. It's going fine.

So we talked about how you position yourself relative to AI by expressing your personality, just all your human aspects essentially. But let's talk about the realm of ideas and especially scholarly production. In what ways do you think AI is already better than you as an economist?

Well, answering questions, it's better if we're talking about the L1 Pro model, So you say AI, there's many different AIs. I'm still better than most of them, I'm happy to say. But you've got to compare yourself to the best, and progress is rapid. This is now February 2025. The O3 model, some people say, will be out in 3 or 4 weeks. So any specific claim will be obsolete. But in a test of 100 questions, O1 Pro already, I believe it would beat me. I don't really have much doubt. Right. So yeah, it's not better at asking questions. There are certain matters of judgment where I'm better, understanding what's important, certain things about how to prepare a research paper for publication where I'm still much better than O1 Pro. Uh, I could list some other things, but again, the improvements will be coming, and I see the writing on the wall. And just answering the economics questions It beats me.

Right. Even in terms of getting out original answers?

I don't think I have original answers, so I wouldn't say it beats me on originality. Right. But I don't beat anyone else on originality either. Right. Now, when I say it beats me, there is a qualifier here. Like, even a worse AI will sort of beat me on breadth. So if you ask about the Turkish economy in the 13th century, I know literally nothing. So when I say it beats me, I don't mean it knows something and I know nothing, though that's also true. I mean, in areas where I have some background, it's still beating me. And if you just ask a question about, you know, say, the incidence of corporate income taxes in a model where you change some different assumptions, how does the incidence change? I would say on a question like that, which I've studied pretty well, I think I would fight it to a tie. But just the more questions you ask, it's going to beat me.

I see. For new scholars who are like training in grad school right now, Where do you think the edge is going to be for humans?

We need to see how good the models get. And I don't know, and I'm looking forward to O3, which will be quite soon. But my guess is more people should study how to operate AI and spend less time learning about the subject, though you need to know a fair amount about the subject to know how to operate AI. But at some point, the human as double-checker of the AI won't make that much sense. It still does. When it makes some mistakes, I can catch them, ask it again, go fix that. But I don't think 2 years from now that will be so significant. Wow. So we need to rethink the whole enterprise. And the human is more like a kind of dog trainer. How do you point it in the right direction? Doing something like writing the prompt, giving it the proper context, having it work with other AIs, other humans, A lot of things maybe haven't even happened yet, but it will be human work to figure out how to make them happen. And we'll have to improvise as we go along. I don't feel I have any upfront recipe. Well, here's what you're going to be doing. But just stay on top of working with the AIs and the competitive process will figure it out over time. But it won't be like the old system where if you say got a PhD in economics from a good school, you are almost right off the bat employable at like a top 30, top 50 tier 1 research university. It seems to me that will be on the way out over some time horizon. But there'll be more and better research. I'm not like being pessimistic about the products here.

So scholarly production as a whole in more fields other than just economics might look completely radically different.

That's correct. Now, there's a separate question: how quickly will universities adjust? A lot have big endowments. Many more have what I would call sticky tuition revenue. People are just going to go to school for the fun, for the credential, whether or not they learn. That gives them a financial base where maybe they don't have to adjust for very long. And maybe like the Department of Botany, 11 years from now is not in fact that different, but actual progress in botany will then be coming from other places. So the actual speed of these sluggish nonprofits. I'm very curious to see what it will be, but it could be pretty slow.

Maybe I'm just coping here, but I wonder if my discipline of philosophy is different, at least in two ways. In the first way, a great deal, at least why I read philosophy, is not true in the academy, is I want ideas that I can live my life through. Like, I want to figure out how I can live my life, which requires me to do all the reading to understand, you know, Heidegger and Hegel and Nietzsche and all the greats. And so I don't think until I have that button where I can download things in my brain directly, I don't think my time there would be wasted training in that tradition. But the other thing is, it seems like at least for a certain kind of philosophy, it's a lot about the subjective experience of man in a certain period of history. And so I wonder if there's just an epistemic barrier about what AIs can do in that domain.

Those are two very good and separate points. The first one I think is a bit of cope. Yes, you can read Heidegger and Nietzsche and many people will for historical reasons. Not quite yet, but I think in a few years, you could just say to your AI, create a dialogue for me with Heidegger, Nietzsche, and me. I want to interrogate them and have them challenge each other. I suspect pretty soon, not this year, but pretty soon, you'll learn more that way than just reading Heidegger and Nietzsche. I've read Heidegger, I looked at every page. I'm not sure I learned anything. I learned what it's like to look at every page and feel baffled. And I'm glad I did it. But we're— AI is going to beat that pretty soon. Now, the second question, also important but quite different. Here's how I would rephrase it. I think when AI is truly strong and quite omnipresent, many more people will want to spend tourist time in more antiquated societies. Is it rural Mexico? Is it Laos, wherever, because they won't be so tinged with AI. And it may just be 2 weeks a year, could be a month a year, but it will be especially fascinating. So reading the real Heidegger, Nietzsche will be a version of that. There will be many other things you could do, like maybe playing the board game of Parcheesi will make a comeback. It's a terrible game. I played it as a kid. Like, no way on the grounds of meritocracy should you play Parcheesi. But I feel it will be a thing, the marginal value of a certain kind of extreme nostalgia, deliberately retro. I think there'll be more of it. It's just a guess, but I see. Yeah. And it will create human jobs also. Right.

In a recent talk at Yale, you described why AI benefits those who study humanities more than STEM.

Why is that? I don't remember exactly what I said. I think it's more of a Barbells answer. So people who are very, very good at STEM will do phenomenally well because they'll work with the AIs. And people who are phenomenally good at the humanities will do remarkably well because they'll have charisma or understand how AI fits into some kind of human equation, human picture. That would be the way I would put it. If I had a chance to say it again. I see.

At what point do you start writing with AI as a helper or as a co-writer or something like that?

It depends on what the world allows and how good the AI is. I think if I did it right now, early 2025, it would sound too bland and the world would be biased against me. And a lot of my outlets, like I write for Bloomberg Opinion, they wouldn't allow it. And I agree with that policy. So I'm just not doing it. But I think in less than 2 years, this will all be quite different. I'll just have to see how the environment is. I enjoy the writing myself. I'm not looking forward to AI being the better writer in the style of Tyler than Tyler is. But again, I can see it coming.

I see. So your latest book, Goat. What's really cool about this book is that you trained a model on it, right? That's right. You can query the model and ask it questions, right? Kind of what you were talking about like the Heidegger and Nietzsche example, in the sense that it might be faster for a student just to query the model instead of reading your entire book if they have a specific question. Tell me about the process about how you trained that while you were writing the book.

I worked with Geoff Holmes, who was also producer of my podcast, Conversations with Tyler Cowen, and I sat down with Geoff and we figured out what we wanted it to do. What you would call the hard technical work was done by Geoff and a consultant he hired. I oversaw the process and had the vision. For what it should be, but credit for the work goes to Geoff. I see.

What has been the reception of not just the book, but specifically this model where you can query a companion reader, essentially like a TA? What has been the reception?

Oh, I think it's been great. People are intrigued. It's the first real book I know of that was just written in GPT-4 and the other models, you could say. OpenAI had an educational event in New York City a few months ago. Hundreds of people were there, and for every person, they handed them out a card that had a QR code. And if you scan the QR code, you get the book. That doesn't mean they endorse the content or anything, but I was so delighted to see them do that. It just meant, like, the book has been recognized by OpenAI, and that was, like, far better than what I wanted. But the most important fact, getting back to our earlier chat, The AIs have read the book, right? We know this. And whatever there is to learn from the book, they've learned it. So I reached my most important readers. Wow. Not every author can say that.

I see. Anything else in your worldview that you think we should have covered around AI?

AI raises the value of travel to other places, to very different places. So YouTube, to some extent, can make like a trip to Madagascar come alive for you. I don't think AI does that very well. Not yet and not in the next few years. Substitute more into things that the AI cannot do because during the regular day you'll be doing more AI. If you ask what is that? I think it's travel to places that are very different, places that are less saturated by AI. Those places are often quite affordable. There's nearby versions of those places in Latin America, but wherever you live, there's probably a nearby version. And to take more seriously that you should do that, and also to see parts of the past world, worlds, past centuries, millennia, before they disappear. You will regret it if you don't do it. So I regret never having seen Syria before it was, I mean, literally physically destroyed. I regret not having gone to Russia more. I have been twice. It's great, but I could have seen other things. Some of my biggest regrets in life are things that are gone, will not come back. So if you're a young person, you have a chance now to experience some of those things and just consider doing it in some manner at the margin, more than what you're doing now. Yeah.

Having grown up in all these global cities, it's striking how few places feel actually foreign and how rare they are. And for me, that was Nepal. I practiced in the Tibetan monastery there. Great thing to do.

Exactly what I have in mind.

These are kind of small, last gasp, and they're usually religious. That's usually why they've resisted modernity. These are small sort of communities in the world that have sort of pushed back heavily against modernity. And those, I think, have always been my most precious travel experiences.

But there is— I mean, I'm all for what you did. It sounds tremendous. It's exactly what I mean. But there are very simple ways to do it. So we live in Northern Virginia. Every day there's a direct flight to El Salvador. It's what, 4 hours plus? It's closer than California. Get out of the main capital, which in fact I like, but it's not going to do this for you. Go to the number 2 or number 3 cities, which are extremely safe right now. Just camp out there. It's not like a monastery. It's just a normal city. It's some version of like something from 1950 in a way. But with people having mobile devices, live there for a week. It'll be changed. There's very cheap, simple versions you can do. You don't even have to change your dollars if you go to El Salvador. Eat pupusas.

What are you most hopeful and optimistic about given all the developments with AI?

That humanity will understand much, much more about the world. I think that's intrinsically valuable. In addition to its instrumental uses. And I'm just very excited for all this to happen. I'm making a deliberate attempt to stay in as good health as I can, exercise more.

You want to see as much of it as possible.

I want to see as much of it as possible. And say 5 years ago, I don't mean that my life was a wreck then. I was not a drunk in the street, but I'm consciously working much harder toward optimizing my lifespan.

And this is why. You know, one of my philosophy professors told me the exact opposite relative to the growth of AI. And he told me, He's like, "Oh man, I'm ready to die. I don't want to see how the world has changed." And I find that—

How can you say that and be a philosopher? That seems— Yeah.

Well, many of the philosophers were very anti-innovation, like Plato, who was very against this idea of innovation, especially in the good state.

But even if I hated it, I'd want to see it.

Because you're curious.

I'm curious. You know, like Star Wars movie, Empire Strikes Back, it ends on a very sour note.— but you're dying to get to the end of the movie. And some of the— I don't think it's going to be so negative at all, but I'm saying even if— I'm curious.

But I do think this is such a rare quality, especially in academics. A lot of them who are conservative— I don't mean politically, but in the expression that my professors— Oh, most of them are, yeah. —are like, I don't want things to change. I want the peer system, the tenure system to stay the exact same, especially when people get up there in their age. They get even more conservative.

Believe me, it's all true. It's worse than you realize. That's all I can tell you.

How did you keep the fire alive?

Some of it travel, but I think like most of our traits, they're at least 60% heritable. So I guess I got lucky with that part of my genes.

I see. Well, thank you, Professor. Thank you for such a fascinating interview. My pleasure.

Great chatting with you. Thanks for watching my interview.

If you like this conversation, I think you'd also enjoy my discussion with Michael Wooldridge about the history of AI. It's a comprehensive introduction to all the major developments of the field in the last century. Now, these interviews are part of an AI series that I'm producing as a fellow of the Cosmos Institute, a nonprofit focused on researching philosophy and artificial intelligence. You can find links to Cosmos's website, the Wooldridge interview, and everything else we cover today in the description below. Thank you.