CEOInterviews.AI
Start App
Get an email when Tyler Cowen gives a new interview
One email per new interview. Unsubscribe in one click.
Tyler Cowen
Holbert L. Harris Professor of Economics; Chairman, Mercatus Center, George Mason University

Why Is Tyler Cowen Optimistic About AI?

📅 Aug 19, 2026 Tetragrammaton with Rick Rubin 146 MIN 15925 VIEWS 408 SEGMENTS · 3 SPEAKERS
Tyler Cowen is an economist, author, and professor who holds the Holbert L. Harris Chair of Economics at George Mason University and serves as chairman and faculty director of the Mercatus Center. He co-authors the long-running economics blog Marginal Revolution with Alex Tabarrok, co-founded Marginal Revolution University, and hosts the interview podcast Conversations with Tyler. Cowen’s books include The Great Stagnation, Average Is Over, and Talent; his newest project, The Marginal Revolution: Rise and Decline, and the Pending AI Revolution, is a free interactive book on the history of marg...

Questions asked in this interview

12
  1. 1:59Do you think it will continue evolving at the same pace or is it going to plateau at some point?
  2. 13:12Were there any that would have been better off getting out into the world sooner?
  3. 22:12And what are your thoughts about it?
  4. 36:02What are some of the worst regulations that you've come in contact with?
  5. 55:32Has censorship been imposed by the AI companies on models? Currently?
  6. 1:10:32Can you either read the piece or have your device read the piece?
  7. 1:21:24What advantages do Singapore and UAE have?
  8. 1:31:50What are the other forms of wealth that are not under capital control?
  9. 1:41:54Are there too few or too many people on the planet?
  10. 1:49:47Is there any part of it where you feel like it's good for you to like these things?
  11. 2:02:08Isn't there an argument that all of the material in the frontier models essentially is stolen material?
  12. 2:13:49Do you work on a lot of things at the same time?
Tyler Cowen 0:18 ↗
So there's new models which we think will be released if they receive approval, and I think some of them will, at least, that will be very powerful first of all as cyber defense, but also as cyber hacking, and those two are two sides of the same coin. So pre-AI is a paper I looked at recently, and it estimated the costs of cybersecurity for the US economy are right now a trillion dollars a year. That's a 2018 estimate, has nothing to do with AI as we know it. So that's a trillion dollars of you get whacked or you invest in defense. Trillion dollars is a lot of money. In the short run, that number is likely to go up. It's probably up already. But the nice thing about AI is you can run checks on your system and find out what the vulnerabilities are and patch them. But in terms of what is that cost, most of all in terms of expertise, it does cost you some money in the form of tokens, but the number of people who can do it and attend to your problems and fix them, that's very scarce. And now all of a sudden there are these AI agents out there that can do more and better hacking. So temporarily the balance of power has shifted in favor of the hackers. This will be a big problem 10 years from now. I think the balance of power will definitely be shifted in favor of defense. The more time passes, the more it moves in the favor of defense. But with laggards, some people, institutions are just laggards. But it's not just that. Like how many experts are there out there who can patch up your systems? It's just not nearly enough. The people aren't trained. These new AI models, uh, Fable 5 is the most famous one, have only been out on the market for some number of weeks. So, we're just not ready.
Interviewer 1:59 ↗
Do you think it will continue evolving at the same pace or is it going to plateau at some point?
Tyler Cowen 2:06 ↗
Well, at some point, I think it will plateau at some point, but in the meaningful sense, it seems to be accelerating for the foreseeable future. And the people who work in the sector at the major labs are reporting that it's accelerating with no obvious end in sight. But yes, at some point I think it will asymptote, but that's not the relevant decisions we face now, that there's new and much better models every month or two, and that's because we're using the AI to improve the AI and do the programming, and that pace seems to be accelerating. So whatever problems we have, at least for a while, they're going to multiply at a faster rate. And again, we're simply not ready. Governments aren't ready for this. And you can say, 'Oh, we're not going to let these models be released.' But again, there's models from all over. Most of all, China is the relevant alternative. And if you don't build out your models, they'll use the other models to hack you. And those other models might be 15% worse than what you've got, but they're still good enough to hack you, probably. So, it's a kind of race where you've got to stay ahead of the bad guys. And we're playing this race. I think we should be playing this race. We've got to win. But it's nerve-wracking. And here we are in the middle of a race that we didn't really plan for.
Interviewer 3:21 ↗
If you were to guess when the plateau would be, do you imagine it being in 5 years, 10 years, 20 years, how long?
Tyler Cowen 3:29 ↗
I think it depends really on the area. So if you ask an AI, what is the meaning of life? You'll get an answer slightly better than what you would get from a very smart human. But it's not that good an answer. It's just not a very well-defined question. And I think 30 years from now, you ask the much better AIs, what's the meaning of life? The answer won't really be any better yet. So on what's the meaning of life, we've already plateaued. It's not a very legible sort of answer that can be spelled out in words. But if it's proving math theorems or trying to cure cancer or coming up with new ideas in chemistry or material science, the plateau does not seem at all close. I would think more than 10 years, 20 years, 30 years. I don't know. I don't feel that I'm qualified to have an opinion, and I'm not sure anyone is, because it's all so new.
Interviewer 4:24 ↗
What you're describing is the materialist worldview which is based on reason and science.
Tyler Cowen 4:29 ↗
Right.
Interviewer 4:30 ↗
So is as far as AI can go just more reason and science?
Tyler Cowen 4:36 ↗
I know quite a few people who think the AIs will become sentient and conscious and morally valuable beings, and they will also have a literal understanding as humans have an understanding of their world. And if you believe that, to be clear, I don't, but those AIs would be much more profound than what I would call our current crop of glorified cash registers. So that's a hypothesis. Like Jack Clark, who runs Anthropic, he thinks that there's a 10 to 15% chance that right now they're in some way sentient. I find that very hard to believe. I definitely disagree with that proposition. But look, he's very smart, and other people agree with him, or maybe would give you a higher percentage yet. I think there's something about biological substratum that is necessary for consciousness, and we will at some more distant point integrate current AIs and biological systems. That to me seems quite distant, and maybe then they were to be conscious entities. I don't think anyone's qualified to judge how distant that is, but it's not soon.
Interviewer 5:43 ↗
How do you think about AI versus how, you know, the mainstream views AI?
Tyler Cowen 5:49 ↗
I would say there's not a mainstream, or there's multiple mainstreams. If you mean ordinary America, for ordinary America, it's fun but not a game-changer. And I think they're quite wrong in that regard, that it will change most of what we do in institutions, in businesses, in government over the course of the next 20 years. Now there's a large number of people in the AI sector, most prominent at Anthropic, but all over, who have what to me are two extreme views of AI. It will become a sentient being. It will become almost like a godlike force that oversees our world. It will enforce rewards and penalties. Our world will become unrecognizable. Say in 20 years, we might have economic growth rates of 50, 100% in a year. I wouldn't say that's mainstream, but there's a whole group of people who believe it.
Interviewer 6:41 ↗
It's the most powerful AI company in the world that you're talking about who believes that.
Tyler Cowen 6:45 ↗
Not the company, but at least 10% of the people who work there, I think, believe some version of that. And then a lot of people, especially who live in the Bay Area, they're sometimes called the rationalist community, which a funny word, what a funny way to describe it, uh, effective altruists. And a significant percentage of them, and they're mostly not working for Anthropic, they believe versions of the story. And some of them have yet the other version of the story. It'll be so powerful it will just kill us. It will wipe us out like we might step on ants. I don't agree with that either. I wouldn't call it mainstream, but within a particular layer of society, in the Bay Area it is mainstream. So now in China it's a very different cultural attitude to AI. They think AI will be very powerful, but they don't think it will metaphysically change the nature of reality. So my view is actually pretty close to the Chinese view. You would say I'm in the Chinese mainstream, but is that US mainstream? No. So it really depends where you look. The European Union, opinions vary, but it's very often seen as dangerous and it needs to be stopped or halted. I would not be in that mainstream. I'm closest to the Chinese. Why do you think there's so much fear around AI?
Interviewer 8:00 ↗
Why do you think there's so much fear around AI?
Tyler Cowen 8:00 ↗
Well, I don't think it's people being crazy. Even the most positive visions for AI, it will bring a great deal of change. People fear change. It will require us all to make significant life adjustments, which may be good for us, right? And we might live to a hundred, but you need to think about a whole new set of issues. How to change your job and how you live, how much you should save. If I tell you, Rick, you're going to live to 100 on net, you're happy, but it's like, hey, I'd better save some more, right? Or earn some more or something. So, you've got to re-evaluate all the decisions you're making. That's psychologically uncomfortable for people. So, when people are hostile or upset or worried, I don't try to talk them out of that. I think they should reframe it more in terms of the benefits, but it will be for many people traumatic, even if it's entirely good for them. And another set of issues has to do with children. So most people's children, I think, will have much more opportunity than before. But you won't know in advance what your kid should do or could do. There used to be all these tracks like my kid will be a lawyer. I'll save some money. I'll send them to Cornell. They'll get into Harvard Law. They'll become a partner, and you just paint by numbers, fill in the dots, and it's if your kid is smart and works hard, I wouldn't say automatic, but there's a formula, and a lot of upper-class people with the right class markers can follow the formula. I think those old formulas are becoming rapidly obsolete. On average, people will learn more, but it might be different people. You don't know in advance what the right path is. And if parents, grandparents are upset by that, I think it's entirely understandable. Though again, I think it's a good thing, and in class terms will open up a lot just to the whole world.
Interviewer 9:48 ↗
You wrote recently a piece called, 'Is school worse for your kids than social media?' Tell me what was the thesis of the piece?
Tyler Cowen 9:57 ↗
Well, I see a big chunk of the world going crazy over kids on social media, that it makes everyone depressed or ruins their lives. When you look at the data, the connections are unclear. I do think there are quite a few kids who are worse off because of social media, but there's quite a few who benefit. They learn things, they meet other friends. I don't think we know the net calculus. But kids who go to school, that's bad for a very large number of them. It helps many others as well. But just for mental health, it does seem that for kids, mental health goes up when school is out. And I thought to present that point as a kind of shock would wake people up a bit about social media. I'm pro homeschooling when it's possible, but it's not always possible. And when it's not possible, I do think people should send their kid to school. But the general point that whatever you do with your kid is actually pretty fraught, I don't care what it is, uh, puts things in a useful context.
Interviewer 10:56 ↗
Was there a reason you chose to formulate the question with school before social media instead of posing the question with social media and getting to school?
Tyler Cowen 11:07 ↗
The headlines are written by the free press, to be clear, but I very much liked the headline they chose because I think it creates a bit of the shock effect to put school first, because school is such a holy, sacred thing that everyone should just bow down and worship to. And in fact, it's not. As a kid, I was bored in school for the most part. I didn't feel it tortured me, but it was quite an unsatisfying experience. So, I don't feel I have extreme delusions about school, but it's centered our culture as something you barely are allowed to criticize.
Interviewer 11:39 ↗
What are some of the other things in the culture that you're barely allowed to criticize that you feel comfortable criticizing?
Tyler Cowen 11:45 ↗
Well, going to college, it's just an extension of this, the grade school point. It's taken as a given that we ought to get as many people as possible going through college. I doubt if that's a good idea. Everyone is on board with this. It's true the people who go to college on average do better, but that doesn't mean college made them do better. They're just better prepared to begin with. And a lot of people who are thinking of going to college should go to trade school or just do something else altogether, or maybe immigrate. Who knows? But the education lobby has pushed this idea: always more government loans, government funding, get everyone going to college. I think it's been a big mistake, and student debt has become so high. I think you're now allowed to criticize it. But for decades, it was the sacred cow.
Interviewer 12:31 ↗
It's also funny coming from you being a professor.
Tyler Cowen 12:34 ↗
But I see my marginal students, the worst ones, they're not ready for school. It doesn't help them. I'm not sure what they should do, but they either don't show up for class, or maybe they cheat, or maybe they flunk out, and they're paying money. And you ask, why are they doing this? I recall one student, he came into my office, and I wanted to help them, and he said, 'No, please don't help me. I want to flunk out. This way my parents can't make me go to school.' Now again, I don't know what he should or should not do, but when some percentage of the population has that attitude, I think you need to reassess the whole project somewhat.
Interviewer 13:12 ↗
What about for some of your better students? Were there any that would have been better off getting out into the world sooner?
Tyler Cowen 13:18 ↗
Probably. I think that is more true at top, top schools such as Harvard, where you have people like Mark Zuckerberg. He drops out, what, at age 18, obviously goes on to do Facebook. That was the correct decision for him. At George Mason, there's fewer people like that, but I still think some, they should just go start a business. It may not be as successful as Facebook, but something like run a food truck or start a restaurant or just start a small business enterprise. Quite a few, I think.
Interviewer 13:48 ↗
Do you think you could learn as much starting a small business as you would learn in school?
Tyler Cowen 13:54 ↗
Well, you're probably going to learn more. Now, there may be this longer-run penalty that when you try to interact with other people, you won't have the class markers they will be looking for to certify you as someone worthy of being a manager in a larger enterprise. Those barriers get higher and higher in what to me is a bad way. So, there's a long-run cost even though you're learning more. And that strikes me as another way in which things are screwed up.
Interviewer 14:20 ↗
Are class markers as important as they once were?
Tyler Cowen 14:22 ↗
They're much more important. If I think of my father, who was very successful, he ran a pretty large chamber of commerce, never went to college. He never could get that position today because he didn't go to college. He was better off not going to college. I think he would have flunked out for one thing. And he slowly acquired class markers. They never came naturally to him. He was always not of the college-educated class. And I think that limited his ability to get promoted, maybe to larger, better-paying chambers of commerce. But today, again, he wouldn't even have a chance. What is the use case for AI in education?
Interviewer 14:59 ↗
What is the use case for AI in education?
Tyler Cowen 15:03 ↗
We don't know yet. AI is the best tutor we've ever devised. It can teach you close to anything. It is tireless. It is inexpensive. Even the free version is very good for most questions and topics. But if you pay for the best versions, it's way cheaper than a human tutor and better. There's not the emotional dividend of working with a human tutor. So I think human tutors, actually, the demand for them will go up. It'll be a job of the future is to be a human tutor working with the AI, teaching kids how to use AI to learn things, but also giving the whole experience a human face, that when you show up for this session, there's a human being there where you sort of feel you ought to have done your homework, that kind of thing. And I think many more people will do versions of homeschooling with AIs, but also with human tutors who might only come by, you know, once or twice a week. And I think that's very promising. If you live in, say, Kenya, where you may not have good schooling at all, I don't know how many Kenyans will do this, but again, you can teach yourself anything, and it's surely better than what they had before. You want legal advice, business consulting advice, healthcare advice, it's there now, people are doing it.
Interviewer 16:16 ↗
Can AI be taught to be a full-rounded tutor where the human tutor would not be needed?
Tyler Cowen 16:24 ↗
Somewhat. So, you could imagine AIs that have human-like avatars and appear on screens, or eventually they're 3D holograms of some kind. How much more will that motivate people? I don't know. I still think we'll need the human in the loop because we're biologically programmed to respond to other humans and their warmth and their presence and their touch. So, how far can we get with the AIs? I think further, but I don't think we can replace humans in that way. We learn a lot from books, and yes, they're written by humans, but you don't have that connection to a face or a teacher when you're reading a book. But so often you read books at all because some human told you to, and you found that human inspirational or persuasive. Or someone buys and reads your book. They may not know you, but they know of you. Something in you resonates with them, and then they read your book. So, there's still the human element, and I don't think that's going away. It's one reason why I don't think we'll have mass unemployment.
Interviewer 17:26 ↗
Have you experimented with the new open-source Chinese model that outperformed all the pioneer models?
Tyler Cowen 17:32 ↗
To get a really good version of KI3, you need to have more hard drive than I have and put a fair amount of time and effort into setting it up. Now, here's the thing with AI. If KI3 were the end of it, I would go ahead and hire someone to do that for me, buy more hard drive, but you know, in a month something better is coming along. So, I'm not going to do it now. I'm just going to wait, and at some point I'll do it. But things are advancing so fast, it's often not worth it to learn every new thing that comes out. But that said, I don't think KI is better than either the best versions of Claude or GPT. I think it's close, but not better, and still somewhat worse. And it's somewhat distilled from Claude. We believe the pace of change being disorienting, I see, is a major issue. So, as you've probably read, there are all these new advances in mathematics that have come from the recent model from OpenAI. Those are very likely to accelerate. Now, okay, there were 10 new results that came in, and we have enough mathematicians to look at those results and tell us it's for real. But if you imagine, say, 50 years of math progress condensed into a single year, which now is quite plausible it will happen, we don't have enough mathematicians to digest what is done or to verify it or even tell us what it means. So, we're rapidly approaching that point where we're just overloaded with new knowledge and don't quite exactly know what to do with it. That's one thing that will slow down the rate at which AI changes our world. We're just not smart enough to absorb it.
Interviewer 19:06 ↗
In some ways, you can make the argument pre-AI that it's not an AI issue. Historically, breakthrough ideas are really met with resistance.
Tyler Cowen 19:16 ↗
Sure.
Interviewer 19:17 ↗
And usually the only way that the paradigm changes is when all the people who believe the old paradigm die, right?
Tyler Cowen 19:23 ↗
Even though 40 years earlier the people in the know, the way we still teach it, is totally obsolete.
Interviewer 19:31 ↗
But rarely has change come at speed. Like take popular music. I would say like 1964 through 1971.
Tyler Cowen 19:39 ↗
You could debate the years, but you know that period very well. It's incredible how fast things change. Every year it's different, has a different sound. Imagine the whole world changing on that basis. It's incredible. It's remarkable. It's mostly positive, but again, it's disorienting.
Interviewer 19:55 ↗
What is the role of open-source in the AI ecosystem?
Tyler Cowen 20:00 ↗
It's becoming more and more important. So, as cyber risk from the released models becomes more of an issue, some people have said, 'Well, we should stop putting those models out.' But the Chinese open-source models, they're out. They're not restricted. You can put them on your own hard drive, manipulate them as you want. So, it basically means AI progress cannot easily be stopped. The open-source models are a bit worse. Some of them are much worse, but they're there. They're getting better all the time. And it's a constraint. It's a competitive force that keeps the best American proprietary models on their toes and progressing rapidly. But it limits what policymakers can do. It's like you've got to run the race whether you wish to or not. But also the open-source models tend to be cheaper, and a lot of companies, and I mean US companies, of course Chinese companies too, they just use the open-source models for a lot of their basic operations where the open source is good enough. It'll probably be that way for the foreseeable future. It makes the total bill much lower. And they use the Anthropic and OpenAI models for more complex tasks. And that sort of division of labor, I think we'll see for some while to come.
Interviewer 21:14 ↗
For most people, are the open-source models perfectly good for ordinary use?
Tyler Cowen 21:20 ↗
Absolutely. But that said, if you think about myself, and I'm not a mathematician at all, but I just want simple answers to questions like, well, I'll be in Europe for three weeks this summer. What are the classical music concerts I might want to see? The best models still give a better answer than the open-source models. The open-source models will be pretty good, but I might miss the one concert I really want to see. So, I'm still going with the best models.
Interviewer 21:47 ↗
Have you tested that?
Tyler Cowen 21:48 ↗
Yes, and it's more thorough what GPT or Claude will give you. It's more double-checked and it leaves out less. Now, if you ran three different open-source models and had them check each other and asked it, you know, for more thoroughness and stacked them up, maybe you could replicate that, but again, you can just ask GPT and it will do it. So, why not?
Interviewer 22:12 ↗
Since we last spoke, agentic AI has really taken off. Have you experimented with it? And what are your thoughts about it?
Tyler Cowen 22:19 ↗
I've experimented with it. It is the part of AI that will do the most to change the world. It still needs some improvements, but it's good enough to do a lot of the work in science, like say cleaning data or running statistical regressions or producing charts, helping people think about the data they've collected, doing things in mathematics. So the frontier area is to get a bunch of agents and have them talk to each other and have the agents decide how to solve a problem, and have like 18 agents working for you at once, and you don't even tell them that much. They just figure out by talking, quote-unquote, back and forth what to do. And that does incredible things. But my life is very simple. Like the number of things I do is pretty simple. So, they're useful to me in the sense that other people using them gets me better results, but I don't have a need to use them myself.
Interviewer 23:14 ↗
Do you think we need a new language for the things AI does? We say talking to each other, they're not really talking to each other. We say they're thinking, they're not really thinking, and we say they hallucinate, and they're not really hallucinating. Might it be helpful to come up with a new lexicon of terms?
Tyler Cowen 23:31 ↗
We shouldn't even call it artificial intelligence. We should call it machine intelligence. But I fear that's all too late. We anthropomorphize as humans, and we think in terms of a world with humans and animals, and we even thought, you know, the thunder came from Thor. Maybe it does in fact, right? So I think we're stuck. But yeah, they're all misleading terms.
Interviewer 23:51 ↗
Since originally becoming interested in AI years ago, how has your relationship to AI changed?
Tyler Cowen 23:58 ↗
Well, when GPT-4 came out, that was really the watershed. So in 2013, I published a book on AI predicting it would be the next big thing. But that was quite speculative. There wasn't really hard evidence that that would be true. Whether that was my foresight or luck, you can debate. But there was nothing useful I could do with AI other than in the world of chess, where AI was already much stronger than humans for some while since 1997. When GPT-4 came out, I just fell in love with it immediately. I had advanced access. I knew what was coming. And I just thought within 10 seconds, this is going to change the whole world. And I thought, well, kind of working on this, learning about it, following it is the most important use of my time, and I'm going to do it. And I had already done a lot of thinking about AI with the earlier book.
Interviewer 24:50 ↗
Do you think your relationship to playing chess against a machine set the stage for you to be interested sooner?
Tyler Cowen 24:58 ↗
Oh, not probably. That's definite. And it's interesting to me, the other chess players I know like Kasparov or Ken Rogoff, they all were clued into AI early on because a lot of outsiders think chess is about calculation. So if a machine can do calculation, that's not new. But people who play chess at a higher level, they know chess is mainly about intuition and understanding. So when the computers could play chess early on, you're like, 'Oh, they can do intuition and understanding.' And you get that. I understood that for decades. So you think, 'Well, it's only a matter of time, and we need some better or whatever before they can do intuition and understanding in these other areas.' And of course, it happened. So, chess players were the best qualified to have a sense of what was coming.
Interviewer 25:40 ↗
I'm just shocked to hear you say that AI has intuition.
Tyler Cowen 25:44 ↗
Well, again, it's this human term that's misleading. We don't understand how a lot of these models work, even the people who built them. But they see that it works, and they keep on grading them, kind of rewarding them for better answers, and it works better and better through the quote-unquote neural net. And I'm laughing as I say this.
Interviewer 26:04 ↗
It's amazing how well all this works.
Tyler Cowen 26:07 ↗
And most computer scientists, you know, the neural nets approach was neglected and looked down upon. 'Oh, that's a dead end. People, oh, it's all going to be neurosymbolic AI.' But no, that approach worked. Geoff Hinton was right, which shocked many people in the field.
Interviewer 26:24 ↗
Will AI allow more startups with less overhead to make more things that we like?
Tyler Cowen 26:28 ↗
It's already happening. There's data. Some of it's from the United Kingdom. Some of it is from the payments company Stripe. There was also a, I think, Wall Street Journal article that showed there's just more companies run by one person or maybe two or three people that have pretty high revenue. They seem to be staying in business. They're growing at rapid rates. We don't see inside those companies, but the fact that this coincided with the rise of AI models, it's pretty clear when you speak to people that the company is working at the small payroll scale because they're relying on AIs. So, we're going to see much, much more of this. I think maybe some enterprises will become bigger, but the median business firm will be much, much smaller. Like, why not outsource a lot of your payments and finance and accounting to AI agents? And people are doing it. It's a slow process, and it's hard to know where to go to ask how to do this because we're all learning it as we go along. But it just seems obvious it will happen at some pace.
Interviewer 27:28 ↗
That's a positive story even for the people who would have worked at those companies, is now there'll be many, many, many more companies.
Tyler Cowen 27:35 ↗
That's right. And it would be possible to start a company from more places. So I've been struck how many people I've spoken to lately from Turkey. You don't think of that as a high-tech country. Teenagers who are starting their own businesses using AI agents who otherwise never would have had access to the capital or the learning or the education, whatever they might have needed, and they're doing it themselves. It's not that all these will succeed, but the places that were the most locked out are sometimes where people are jumping in the most because for them it's a new opportunity. If you're a North American, Mark Zuckerberg did Facebook at 18 without AI agents. As an 18, 19-year-old, you can go to venture capital. They might not fund you, but they'll hear you out, right? The rest of the world, that's typically not been the case. So, I think a lot of the innovation will actually come from outside the United States, from teenagers, people who decided not to put most of their time into formal schooling and just play around with this stuff. And like, there's no one at Harvard to teach you this. So, if you're not at Harvard, it's a blessing, not a curse.
Interviewer 28:41 ↗
So would you say AI is a DIY movement?
Tyler Cowen 28:47 ↗
You need to learn from someone. So you could be in a chat group. You might have teen peers.
Interviewer 28:53 ↗
YouTube.
Tyler Cowen 28:54 ↗
YouTube. I don't know that it's on YouTube yet. It'll get there, I would think. And the models change so quickly and improve so quickly. Whatever was on YouTube two months ago, it's not good anymore. So, it's just this running a race idea, and the people who A, take initiative, and B, are good at running races and always retraining themselves, they're the big winners, not the same people who would become law partners by going to Cornell and Harvard. Some will be the same, to be sure, but it's a very different set of winners and losers. And I think the premium is on initiative and retraining and not getting exhausted by the ongoing speed of it all.
Interviewer 29:32 ↗
Tell me about finding the best people to hire.
Tyler Cowen 29:35 ↗
Well, there are some companies that are using AI to tape interviews, and we need years to see who ends up being successful. So, we don't have results yet. But say 10 years later, you will see what from the interview predicted success. And once we learn that, we'll know a lot more about how to interview people and what to look for. So, the science of evaluating talent will change radically. We have this already in sports. There's analytics. You watch people play in college, high school. There's tape. You run the tape through computers. This is a burgeoning science. The pieces are not all there yet, but it mostly works. And there'll be something like NBA analytics for many, many jobs. It just takes a while. You need to measure who are the people who ended up succeeding. Uh, but the other thing that is possible, this is again speculative, but I think it will happen. Let's say you're in Kenya and you're very smart, but you're not really well connected at all. Well, even in Kenya, but you play around with your AI, there'll be a kind of service you can hire. Maybe it's for free. And you just say to the service like, 'I want you to send a report on how smart you think I am to some clearing house.' And if you've done well,
Interviewer 30:46 ↗
What's a clearing house?
Tyler Cowen 30:48 ↗
It doesn't exist yet. There'll be a clearing house like a jobs bank. You know, here are the 500 people in Kenya who were the smartest with their AI, asked the best questions, interacted with it most fruitfully, and that information will be sent maybe automatically to employers, and people will message these kids in Kenya, maybe hire them from a distance. They'll have proven themselves over years interacting with their AIs, and it'll be easy to find them. Right now, those geniuses or just hard workers, whatever they are, they're very, very hard to find. They're going to become very easy to find. Again, this is speculation. It does not exist yet, but I don't see any technical obstacle to it happening.
Interviewer 31:28 ↗
It sounds like for the fast-growing AI companies, it's just data they're not looking at. But they have that now, and they could have first-mover advantage of seeing who those people are asking the best questions.
Tyler Cowen 31:38 ↗
That's right. It's like everyone gets a free audition for Capitol Records, or you know, in the old days.
Interviewer 31:44 ↗
Where's regulation helpful and where does it get in the way?
Tyler Cowen 31:47 ↗
We're going to have more regulation of AI where new models are first run by the government, not by a regulatory agency, but there's a process involving the executive branch, national security people, and the other AI labs, and they're forming a kind of consortium to get approval for saying this model is okay to release. Now, as we're speaking, it has been announced that a process has been approved and agreed upon, but the exact nature of the process is not yet public. But I strongly suspect these early reports are true. It will not be like the FDA or SEC, that it's an ordinary regulatory agency, but it will be a weird mix of a consortium, national security, the executive, what the companies themselves say. Very patchwork, but again, when things are changing this fast, it's pretty close to the best we can do, who I suspect.
Interviewer 32:40 ↗
Can the different AI companies work together in that way? They seem to hate each other.
Tyler Cowen 32:46 ↗
Well, there's a lot of gossip that the CEOs dislike each other, but the actual employees live in the same town, go to the same parties, maybe date the same partners. I don't know that there's animosity across most of the people, but when you have to make something work, you can make it work. And I think we'll more or less make this work. There's always going to be a lot of error. Like I said before, pre-AI, the cybersecurity bill was maybe a trillion dollars. I think that number is going up under all scenarios, but you don't want it to go up too much. And I think we'll end up with something more or less workable.
Interviewer 33:23 ↗
And maybe over time it'll come down with the help of AI.
Tyler Cowen 33:26 ↗
Oh, absolutely. But that's more several years out or more. And if you're a major AI company, there's issues with your rivals. You're also worried about antitrust law, right? There's legal issues with these constructions when companies are getting together exchanging information that actually, if it's done the wrong way, can violate some of our laws. So a lot of things need tweaking. But as a company, if you're told, look, if we don't do this, there's going to be more formal regulation, and you've got to wait four years to get an answer, that's not workable, right? So I think we're going to make this work.
Interviewer 33:59 ↗
Beyond AI, where's regulation helpful and where does it get in the way?
Tyler Cowen 34:03 ↗
Well, I'm considered what you'd call a market-oriented economist. I'm generally pretty skeptical of a lot of regulation. I think we should have much less of it in the American economy. When we got rid of regulation for flying and airfares, flying became safer, airfares fell, there were a lot more routes, things got better for consumers. So, I would do that to more parts of our economy. We have crypto as an active entity or sector right now because we deregulated it under the Trump administration. So, I'm pro-deregulation. I do think we need a lot of government regulation to keep certain things safe.
Interviewer 34:40 ↗
Like what?
Tyler Cowen 34:41 ↗
If you want to put out a new drug on the market, we take way too long, but I think we need some regulation where that should be approved and have to go through some tests. If not, people lose faith in the whole thing. But if you're in a world where it takes 10 years and costs a billion dollars, which is often our current world, you're keeping a lot of drugs off the market. My colleague Alex Tabarrok has referred to the invisible graveyard: people who die, drugs that never come to pass, they're never invented or discovered in the first place because everyone knows if I do this, it takes 10 years and costs a billion dollars, and you're stuck. Now AI will force us to reform regulation whether we want to or not because there will be so many submissions. We'll need to streamline our processes, but there'll be this intermediate period, I think, where it gets worse. Like the number of lawsuits right now is rising rapidly, and someone did a study of the filings, and a significant percentage of them were written by Claude. No surprise. Oh, you can sue someone. It used to be a big pain in the neck. Now you just tell your friendly AI, you speak to it for an hour and say, 'Oh, write up the thing and file it,' and it does it. That's going to break our legal system. Now, there's plenty of good defenses using AI itself, but the people filing the suits are quicker on the offense than our government is on the defense. And we're again likely to have this intermediate...
Interviewer 36:02 ↗
period where things just don't work very well until the defense catches up. But again, long run, defense has all the advantages. What are some of the worst regulations that you've come in contact with?
Tyler Cowen 36:15 ↗
Well, I just think the way we treat high-skilled foreigners who want to immigrate to the United States, they go through all kinds of indignities and uncertainties. There's a lot of people you definitely want. We can be way too lax on people we don't want by not securing the border, and then way too tough on people we do want. And remedying that and getting it right, that would be the first thing I would do if they put me in charge.
Interviewer 36:36 ↗
Tell me about the course that you filmed.
Tyler Cowen 36:41 ↗
It's a principles course, principles of economics. There's microeconomics and macroeconomics, and my colleague Alex Tabarrok and I worked for about seven years and spent several million dollars on this. They're YouTube videos. They're on YouTube now. You can also go to sitemu.org, and you can take a complete course in micro or macro with quizzes and gradings and exercises for free. It's online. Large numbers of people around the world do this. And the videos are not us at a blackboard droning on and on. They're typically five- to seven-minute videos, professionally produced. A lot of them look and feel like documentaries. Some are better than others.
Interviewer 37:24 ↗
How are they labeled?
Tyler Cowen 37:26 ↗
Like, what are the laws of supply and demand? How do you learn supply and demand? When do price controls work or not work? What are the economics of the minimum wage, of pollution? How does the Fed work? There's a video for all of these. You just use the search function, which we're trying to upgrade with AI.
Interviewer 37:40 ↗
By the way, if I were to go on YouTube and search, would I search Tyler Cowen?
Tyler Cowen 37:46 ↗
Just go to YouTube.
Interviewer 37:47 ↗
YouTube.
Tyler Cowen 37:48 ↗
And put in Marginal Revolution University. We have dozens of other courses, but a lot of those are lower-quality videos just like we talk. You don't have to give us any personal data. It's a complete free lunch, the thing economists tell you does not exist.
Interviewer 38:02 ↗
And that would be the same course that a student might take at university.
Tyler Cowen 38:06 ↗
And indeed, I teach a course at George Mason which is online only. And the students are given all those videos and a text. They can use whichever they prefer, or both. And they take the class and it's graded. And at the end of it all, if they pass, they get three credits. And every year I teach 700 students this way. So the whole world can do this. This gets to my point. Institutions are very slow to adjust. The whole world is not doing this. And I would readily grant a lot of students need the presence of a human professor to do the work. But a lot don't. And it's a big world. Let's say there's only 10, 20% who don't need the human there. They can all get education for free right now. Many are doing it, but many still are not.
Interviewer 38:49 ↗
How much of what you've learned over the course of your life was from humans versus other sources?
Tyler Cowen 38:55 ↗
I think I learned very well from non-humans. I might be an outlier, but even in those cases, it's often the human who clues me in to where to look. So I interact a lot with humans, and the things I learn from humans are often not the things they think they're teaching me. They'll say blah blah blah blah blah, and at the end of the blahs, they mention a book, and I read the book, and I've learned from the book and not from the blahs. And that's fine. Like, that's a great human from my point of view, but they just think it's about the blah blah blah.
Interviewer 39:25 ↗
Tell me about the strengths and weaknesses of loading books into AI.
Tyler Cowen 39:32 ↗
There's good and bad ways of reading books using AI. The bad way is to ask the AI to summarize the book. It's not terrible, but you tend to get a stupider version of the book, and it's also very bland. It won't inspire you to keep on learning more. The best way to use the AI and the book is to be reading the book. As you're reading the book, there'll be a lot of things you don't understand. Keep on asking the AI. Keep in mind, the AI probably has been trained on the book you're reading. There's another radical step you can take, which may be illegal depending on the book. But if you upload the book in the form of a PDF, if the book is still in copyright, and literally upload it into the AI and then interrogate the AI about the book, that's the very best method yet. It often will be illegal, but it is extremely effective.
Interviewer 40:19 ↗
Why would that be illegal?
Tyler Cowen 40:21 ↗
If it's a book under copyright, for you to have a PDF of that book, it depends how you got the PDF. If you do the classic New Jersey Sopranos Mafia line, it fell off the back of a truck, I don't know if that's illegal. Maybe the PDF did fall off the back of a truck. But if you got it by visiting a site with illegal downloads, you too are breaking the law. People have different opinions of that. I don't do it myself, but again, there's plenty of wonderful books out of copyright, and that's completely legal, or research papers, PDFs, again, typically fully legal. Upload them into the AI and just keep on asking, and that works very best of all. You just keep on learning. It's incredible.
And it also means the research paper, or for that matter the book, will soon be evolving into something new. So it used to be you would write a paper on some topic, and the paper is a static thing that's just out there and doesn't change. There's no reason why that should be the case. So if you write a paper on the Federal Reserve system, fine, you write the paper, but three years from now, someone should be able to take the paper, upload it into their AI, and say to the AI, 'Update this paper with the last three years of data,' and the AI will do it. Now I think the AIs are already good enough to do this, or if not, they will be, say, within a year. I don't quite know that we have all the infrastructure connected to do it, but I don't think there's any additional technical problem. We will have papers and books that you can update on command, and in many areas this is very useful. 'Update this for me,' or 'just tell me what has been written about this topic since this came out and add that as an appendix,' or whatever you want, and it will do it. So the things we create will be much more fluid.
Or like the last book you published, someone could say to it, 'Well, since this book came out, Rick has done all these podcasts, other work, including putting out music. How do you think his ideas would change based on what he's been saying?' And it will be speculative, but it will be highly intelligent. And it's not saying it's exactly you, but it will take a digest of what you've been saying and modify the book if someone wants that. So the idea of a fixed text will change.
Interviewer 42:31 ↗
If another person has the same data points that you have, and if you write a book explaining those data points and they write a book explaining those data points, I'm guessing the books will be different, right?
Tyler Cowen 42:43 ↗
You'll be able to marry the two books. 'Oh, turn these into one book.' Now, how much people will want to do that, I'm not sure. Like, there's many books. Let's say it's Jane Austen. I don't want the book, quote-unquote, improved. I want to read the same Jane Austen that everyone else is reading so I can talk to them about it, and that's more important than, say, improving something in the book that might, now you know, have been a mistake. But a lot of research, I mean, you just want the updated version, right? You care about the decision. So I think this will be a big thing. All of science will change. This gets back to people being disoriented. If you're a scientist, again, a lot more opportunity, you'll be much better informed, but what you used to do, create these static, immobile papers, that's done. You've got to learn this new thing. People don't like that.
Interviewer 43:31 ↗
I've read many different translations of the same book, and the translations are totally different from each other.
Tyler Cowen 43:37 ↗
Yes.
Interviewer 43:38 ↗
So I wonder about getting any information from a book from AI, if it will always have that lost-in-translation issue.
Tyler Cowen 43:48 ↗
It can tell you what biases it's introduced. Current translations don't do that. I'm sure you read things in translation all the time. You still find them useful. You'll still have the option of reading the original. But again, especially for practical areas, you're reading papers on how to build a bridge that won't fall down, you just want the updated data. If you're reading the Bible, I think probably you want the original Bible, whatever you might take that to mean. But in fact, there's at least two versions of the original Bible, Hebrew and from ancient Greek. And which one is the true Bible? They're pretty close, but they're not the same. AI can at an instant explain to you where there are differences, where the differences come from, which one scholars might think is the correct translation, which not. There'll be a lot of neat things you can do, even with the Bible, where clearly, I think almost everyone wants the actual Bible.
So we're already in this mess. It's just easy to ignore. But when AI allows you to manipulate the mess, it becomes harder to ignore. And again, that's more disorientation. Like, oh, you thought you knew the Bible, but until you've been through it with these new tools, turns out you didn't know the Bible. Again, more of an opportunity, but a kind of psychic burden for people too. Like, you've got to relearn everything you felt you knew.
Interviewer 45:06 ↗
That doesn't sound like a bad thing.
Tyler Cowen 45:08 ↗
I think it's pretty awesome. But again, people are rebelling against it, and I get that too. It's like, say you had to learn how to play the guitar all over again. Like when Stanley Jordan came along and he did it with the two fingers, a lot of people did learn how to play it all over again. A lot of opportunity was created, but it didn't help everyone either.
Interviewer 45:28 ↗
What have you learned about prompt writing?
Tyler Cowen 45:31 ↗
Prompt writing changes all the time. You need to keep on relearning the art of prompt writing. So in what you'd call the early days of AI, quote-unquote, two years ago, how you wrote the prompt was very important and would change the answer a great deal. People now have made the mistake of thinking writing the prompt doesn't matter that much because the AI does know what you want, especially if memory is turned on. So there's a sense in which for superficial answers, the prompt definitely matters less. That's a correct impression. People who say that, they're not just bullshitting. But that said, the AI is now so much better and so much more powerful that if your better prompt improves it by 2%, improving the more powerful thing by 2% is really a big efficiency gain, especially for agents. So in this funny way, prompt writing has become both less important and more important at the same time.
Interviewer 46:22 ↗
What are some of the tricks of prompt writing?
Tyler Cowen 46:24 ↗
Well, the best trick is to ask the thing itself how to write the prompt. And I find this useful. If you have a question, and this goes for more than just prompts, if you have a question about GPT, ask Claude. If you have a question about Claude, ask GPT. They know each other better, by hypothesis, unconfirmed. If you ask Claude about Claude, it's a bit programmed to give you a certain kind of stock answer. Same with GPT about GPT. It's sort of like you're talking to the PR department all of a sudden. But if you ask one about the other, it will be super smart. And you can always cross-check by asking the same about the same. And that's by far the best way to do prompt writing. And the things could improve a lot, but that as a starting point will never fail because the AIs are getting better also. And then from there on out, just experiment, see if you're getting what you wanted.
Interviewer 47:14 ↗
Let's talk more about AI and intuition. When you're reading based on your context, you get something different from what you read than someone else might. And if there's something you're interested in, you might read several different accounts that are different, and you may feel like this third account is probably the one that resonates most with me. With AI, you don't get that opportunity. You get an answer. Are we missing our ability to interact with the information by using the AI?
Tyler Cowen 47:52 ↗
But AI is awesome at this. You can ask it, take this biblical passage, how would an Orthodox Jew, a Catholic, and an Eastern Orthodox read it differently, and it will give you all those, and it's amazing at that. So if people are only getting one answer, that's their own choice. It's never been easier to get multiple perspectives, and often you should. In a way, it's an embarrassment of riches, like, where do you stop? Well, what would an LDS person say, or what would an evangelical say? And presumably this would take some work. It'll become easier over time. You could instruct your agent to go around and ask a bunch of AIs and write a report as to what they said and what the differences were, and then read that report. I'm not sure that's worth the trouble because you can do a version of it directly. But again, your ability to customize things is so extreme already and only growing. People haven't really internalized this in their daily routines for using AI. They tend to be, here's my question, there's the answer. Yeah, and that's what they do. That's fine. But you can customize.
Interviewer 48:56 ↗
It makes me feel like that ability that we have to parse different information, if we're not actively practicing that on a regular basis like we do now, if we go with the AI answer, that we can lose the ability to do that. Seems like a scary road to go down.
Tyler Cowen 49:16 ↗
Yes. I think in a funny way we're treating them too much like humans, that there aren't many humans where you could ask what are five different interpretations of this biblical verse. So we haven't emotionally internalized that the AI can do that easily. And if we would think of it as true machine intelligence rather than like a human or a humanlike entity, we'd use it better. But it's hard to make that leap because our whole lives we've dealt with humans and not so much with machine intelligence.
Interviewer 49:42 ↗
I love the idea of calling it machine intelligence and thinking of it that way. It seems really helpful.
Tyler Cowen 49:47 ↗
Yes. And like, what about it is artificial? Like humans, we're natural. I mean, we've tamed nature. We pull oil out of the ground. We build nuclear power plants. We fly supersonic jets. Whatever you think of all that, it's not exactly natural, right? So we created them. If we're natural, so are they. If they're artificial, so are we. But they are machines.
Interviewer 50:14 ↗
C.S. Lewis said that he thinks that the automobile destroyed our connection to time, because before we understood distance in terms of how long it would take to get there, either walking or by a horse, and once you introduce automobiles, it cuts that connection to understanding our place on the planet.
Tyler Cowen 50:38 ↗
Trains may have done the same. And with an iPad, it's worse yet, because you fly on the plane, time flies. You're quite absorbed, maybe talking to your AI, whatever, talking to your machine intelligence, excuse me. Oh, that was a 12-hour flight, and it flew by like the wind because your iPad now is so much fun.
Interviewer 50:56 ↗
What are your thoughts on AI doomerism?
Tyler Cowen 50:59 ↗
I don't think AIs are going to kill us all and end the world. These are very speculative claims. I sometimes put it this way: if we as humans cannot handle an additional dose of intelligence, we're likely to do ourselves in by other means anyway. And that machine intelligence is more likely to extend our lives than to shorten it. But the point that the future is very uncertain, none of us are good at prediction, you need to be prepared for contingencies. Safety is a real issue. We see this already with cybersecurity. We'll see it in more areas. All that's correct. I'm very glad there's a large number of highly intelligent people working on AI safety, and we do need to get it right.
So we're still in a world where 40,000 people die a year because of cars, it's the US alone. It's going to be much greater across the world. Now, you can say it's worth it, but you don't want a future where machine intelligence is killing 40,000 people a year and we're all sitting around debating, is it worth it? I don't think we're going to have that. But again, you need to work hard to make it safe.
Interviewer 52:03 ↗
AI safety has a lot of different meanings. Tell me some of the different ways you think the term AI safety is used.
Tyler Cowen 52:10 ↗
Well, the most extreme people, they just want to pause or they want to ban machine intelligence. Eliezer Yudkowsky has talked about blowing up the data centers. He may have been tongue-in-cheek, I'm not sure, but he said we should make a deal with China, and if they build a really advanced data center, we just bomb it. And that might start World War III, but that's a lower risk than letting them build the data center. Now, that to me is crazy, literally crazy. But that is a view on the extreme point of the spectrum. But you know, within the main companies, there's a sense of models can do things you don't want them to do. You need to train them properly, as you would a pet. The smarter they get, both the higher the return, but the higher the risk. That's definitely true as well.
But I have a different perspective on this. My main worry is not what the machines will do on their own. I think they're mostly like poodles. If someone says, 'Oh, there's a poodle walking down the street,' you're just not that terrified. The poodle might create some problems, but it's not that big a deal. It's when you have humans commanding pit bulls, evil humans commanding pit bulls, that you get really afraid. So if you think the real risk is evil humans giving commands to powerful AIs, you then want the non-evil humans to have very strong, capable AI to defend or stop the evil humans. So to me, being concerned with AI safety involves a kind of acceleration of progress from the good people.
It's very tricky because, like, who are the good people? You like to pat yourselves on the back. Oh, we're the good people. But it's not that simple either. Like, the good people, is that the US government? Do you feel the US government's always the good people? Like, what percent of the time? Well, okay, maybe not US government. Do you trust the Danish government? How about the New Zealanders, the Chinese? This debate can go on and on, but when there's not an unambiguously set of the good people, do you still want to accelerate? That's the difficult decision.
Interviewer 54:07 ↗
What about the idea of dangerous information?
Tyler Cowen 54:11 ↗
Many more parts of our world will be transparent. So right now, the machine intelligence models we have, they're designed that they won't tell you people's secrets. You can try to jailbreak them, as it's called. But if you just go to a jailbroken model and say, 'Oh, famous person X, give me the three women he's most likely to be having an affair with,' if he's having an affair, the chance that the AI or machine intelligence can figure it out is pretty high. It'll give you a list. It's speculation on the part of the machine, but you've gone down the road of violating someone's privacy.
And how well we can jailbreak future models. There'll be open-source models that will be designed as jailbroken to begin with. So this will happen. So I think people actually will have to be more virtuous in this future, or there could be a situation where just everyone is exposed and standards more generally fall. Like if we all learn X% of these people are having affairs, we could either get tougher or that's just how things are, a bit like how things might be in parts of France. I'm not sure how that will evolve, but it will change human mores because there'll be more transparency into many parts of your life. It may be stochastic transparency.
Interviewer 55:32 ↗
Has censorship been imposed by the AI companies on models? Currently?
Tyler Cowen 55:38 ↗
Censorship is a tricky word. They steer what the models tell you. So they will have humans grade the answers of the models, and the future evolution of the model in turn responds to the grades. Is that censorship or is that quality improvement? It will depend on your point of view. It's very hard to get the models to say, for instance, things that might be considered racist by different groups of people. But there's different ways of interpreting, is that appropriate or is that censorship. People argue about that. I would say overall, I'm pretty happy with the models, and there's certain things only Grok will tell you, and if you want, go to Grok. It's great that there's diversity and competition, and if Claude won't tell you what Grok will, and we have both Claude and Grok, to me that's, you know, the best of all possible worlds. So I get less upset by that than many people I know do.
Interviewer 56:29 ↗
How do you think you see AI differently than people who have not experimented with it?
Tyler Cowen 56:33 ↗
Well, if they haven't done it at all, it's going to depend on who it is they're speaking with. And that's a little hard to forecast, but they're probably going to either greatly overvalue or greatly undervalue it. And again, we're all experimenting, trying to figure out what the real value actually is. So I'm not in some privileged position where I know and they're just guessing. I'm just guessing along with them. And I may know more, but there's some situations where the people who know less make better guesses because they sort of know what they don't know and they're not too arrogant about what they do know. I wouldn't dismiss, you know, their estimates too quickly.
Interviewer 57:11 ↗
What is looping?
Tyler Cowen 57:14 ↗
Looping is setting up a system where the thing gets feedback and then keeps on going, so you don't have to intervene and keep on giving it instructions or asking it questions. And it, you know, an agent will do things for you if you set up the loop. Now, it can still take a lot of time, effort, and attention to set up accurate loops, which again is not a personal benefit to me. It's an experimental thing, but I think in a year, at most two years, this will all work much better, be much easier. You'll just be able to talk to it and say what you want, and you won't have to know anything, and it will just work.
Chris Best recently wrote a piece on Substack about how they've added an AI detector to articles written for Substack. I don't think this will last. So I'm glad they did it because it's a wakeup call to people, showing them how much is written by AI and sometimes how well AI can write. As people read more AI, I think especially Claude, a lot of people have decided they hate the style. I feel this way myself, weakly, not strongly. I don't prefer the style. I don't really like it. But again, I think in a year's time, once the companies set their mind to making the style whatever you ask for, but if you want more human, get it more human, this will happen quickly.
In a year or two from now, Substack, the tests it's using, they won't be able to tell is it AI or is it you. The best test will just be like, how smart is it? And if it's really smart, you might think it's not a human, but it's not always going to be a good test because there are plenty of really smart humans. So written material on the internet will be, it's already verging on, what, 50% AI. It'll be well over half AI. It will be very smart. No one will be able to tell. I don't know how much people will care. They might desert parts of the internet en masse, or they might just say this is it. Like, do people care that, you know, Paul Simon ripped off some of his most famous tunes from melodies by Bach? They don't even know. And those who know maybe enjoy the fact. I don't think they care in a negative way. We might be in a future where people stop caring and they would rather not know.
So I think the Substack thing is temporary. It will disappear, but in the meantime, it's a great way of teaching people what's going on. So I'm all for it, because if you tell people over time like two-thirds of my product is fraudulent, people desert you, right? Not just those particular Substacks.
Interviewer 59:45 ↗
If AGI is smarter than any human, wouldn't we want to read what it writes?
Tyler Cowen 59:51 ↗
That is my opinion. But stylistically, there are certain things I want only from humans. So there's something about the fact that, say, the Beatles are real human beings, or were real human beings, that matters to me. I'm receiving a communication from a person, and that makes it more meaningful, and there's a life and a history behind it. And I think for a lot of what we do, that really matters. Reading the four Gospels, well, they were written by four, maybe with collaborations, but at least four people.
How much of our world is like that? I don't think we know. I think we're going to find out we care less than we like to think, but there will still be this ineradicable level where we care deeply. So I think when it comes to writing, people won't care that much. I think when it comes to podcasting, people will care a lot. So I've been doing more podcasting because I feel my writing will be competitively undercut by AI smarter and better than I am. And they'll be able to mimic me if they want. They can't do this yet, but they're getting closer. But podcasting I feel is secure, and certainly in-person talks. I've been doing like 2x those or 2.5x of what I used to do. I'm really quite sure those are safe.
Interviewer 1:01:08 ↗
In March you released a new book, The Marginal Revolution, and you released it free online, and it has a Claude companion that lets you search and summarize, examine and debate, or pull the thread. Tell me about what it is and what the thinking was of presenting it in this way.
Tyler Cowen 1:01:26 ↗
I think a lot of books in the future will have AI companions embedded in the book, in something like a Kindle, but a better technology. Kindle's pretty clumsy and now primitive. And you can ask questions, ask what does this word mean, ask for follow-up. The way I do it now is I have the book open and the AI on my laptop over here, and I'm like with two different, switching back and forth. It can be one integrated system. Publishers are not ready for this yet, mentally or emotionally, but I wanted to be the first person to do this. I did two books in this style. They're both online and free. The other one is called Goat: Who's the Greatest Economist of All Time? They're both on my homepage, tylercowen.com.
Again, there's no catch here. There's no ad. There's no data you've got to give over. You can just grab them, turn them into PDF, send them to your friends. Totally free. Upload it into any other AI and interrogate away. I think that's the future. I want to be part of the future.
Interviewer 1:02:24 ↗
Let's say we ask it to search and summarize.
Tyler Cowen 1:02:26 ↗
Yeah.
Interviewer 1:02:27 ↗
Have you vetted the search and summary that comes back?
Tyler Cowen 1:02:30 ↗
Absolutely. I played around with it. My assistant Jeff Holmes played around with it even more, and we kept on iterating on the prompt until we were happy.
Interviewer 1:02:39 ↗
How much work was it to get it to be good?
Tyler Cowen 1:02:42 ↗
You know, we did two books this way. First time it was a lot of work. Second time it was easier because we had done it once and the AIs themselves are better. But we wanted it to sound somehow both substantive and playful, and that's a hard balance. I think we got pretty close. It doesn't sound like me. I don't think it should sound like me, but that was a decision.
Interviewer 1:03:03 ↗
Yeah.
Tyler Cowen 1:03:04 ↗
If there's some more distant future where it can sound just like me, I would experiment with that. But I think it should be an outsider.
Interviewer 1:03:11 ↗
I think so too, because it isn't you. So if that's right,
Tyler Cowen 1:03:14 ↗
if it's not you, it shouldn't sound like you.
Interviewer 1:03:16 ↗
And they get enough of me already.
Tell me about the OpenAI lawsuit from a couple of months ago. Who were aligned on the two sides, and what came from it?
Tyler Cowen 1:03:27 ↗
There's a whole bunch of lawsuits. The most famous one is the Apple lawsuit. So OpenAI supposedly is putting out a device of some kind which is due to come out in 2027. It's for your home. Some rumors have suggested it will be a flat thing that you put on a table and can speak to, and maybe can both answer questions, be voice-activated, and even control some of your home systems over time. To be clear, that's rumor and speculation, not fact. And Apple filed a lawsuit against OpenAI claiming there was a theft of intellectual property. OpenAI put out a document saying the lawsuit is totally unfounded and based on a confusion of two different people's names.
I have not read through the OpenAI response, so I cannot have an opinion as of the moment, but I believe from what I've heard secondhand, OpenAI is claiming that the lawsuit is simply unfounded. It doesn't sound to me like a lawsuit that will succeed. Like, you might think an employee of a major company stole from a company they came from, but it's very risky for the company itself to be party to that. So you might have something against the employee, but I would be skeptical of huge conspiracies to steal things from other major companies that can defend themselves with top-flight legal departments.
Interviewer 1:04:49 ↗
There was another lawsuit that was actually tried more having to do with the creation of OpenAI. Do you remember that one?
Tyler Cowen 1:04:57 ↗
Well, I would say I remember too much about it because everyone I know wanted to tell me their take, and it's become a blizzard of confusion. Or I feel grateful I don't really have to have a specific opinion on what happened. But OpenAI was set up as a nonprofit with the original mission being it would be not-for-profit AI that would be used in a safe way to help the whole world. It then turned out, according to OpenAI, that if you did things that way, you were not able to raise enough capital to actually make the AI useful. So OpenAI tried and basically succeeded in giving up its earlier status as a nonprofit. It split off a thing which is now called OpenAI Foundation, which received an immense amount of money. It's by far the largest charitable foundation in human history. And OpenAI Foundation, which is a very new thing, it's only been around less than a year, has some version of the original OpenAI mission, and it is empowered to spend the money to fulfill that. And then you have OpenAI, which is a bit more complicated, but basically operates like a for-profit company.
Now, personally, I think that's a good arrangement. I think as a nonprofit, it couldn't have raised enough money. But if you're asking who lied to whom and what did Elon, what did Sam say in this email, all this came out, I heard from all my friends this, that, and the other, I feel liberated by not having to take positions on all the specific issues.
Interviewer 1:06:29 ↗
I'm not asking you so much your position, but tell me both sides.
Tyler Cowen 1:06:32 ↗
Well, Elon says that Sam and others ripped him off. They took his money under the guise of it being a nonprofit and then turned it into a for-profit for their own gain.
Interviewer 1:06:41 ↗
OpenAI, they did take the money under the guise of it being a nonprofit?
Tyler Cowen 1:06:46 ↗
I believe so. I'm not the one to vouch for this, but I believe so.
Interviewer 1:06:50 ↗
The OpenAI people say that Elon also wanted it to be a for-profit controlled by him.
Tyler Cowen 1:06:55 ↗
Mm-hmm.
Interviewer 1:06:56 ↗
I read a text of an email that seemed to indicate that.
Tyler Cowen 1:06:59 ↗
But again, this is all out of context. I don't consider myself the adjudicator here. It all seems to me like a big mess. Many institutions have messy origins. The United States Constitution was passed in a big hurry without necessarily telling everyone.
Interviewer 1:07:15 ↗
Is that true?
Tyler Cowen 1:07:16 ↗
Absolutely. We had this thing, the Articles of Confederation, right? Where there was a lot more power in the individual states, but it didn't work out well. There were all these fiscal crises. You had tariffs across the states. It was a mess. You could say they couldn't raise enough money to produce a strong nation. Like a nonprofit, OpenAI could not raise enough money to produce good AI. So a bunch of people got together and more or less behind closed doors recruited the people they thought would support them, and they did the Constitution and voted on it and pushed it through. And I'm glad they did that.
Who said what to whom back when, and what was truth? What was a lie? What was someone being misled or manipulated? Again, you can read the history books or ask your machine intelligence. It will tell you. But at the end of the day, I'm glad we replaced the Articles of Confederation with our Constitution. And for really big changes, I think you need a lot of tolerance as to the origin story.
So OpenAI has charges against Elon. Elon has charges against OpenAI. Like, let's say everyone's guilty. I'm more focused on did we end up with something workable? I think we did. The products are great. It's a successful company. It still has to earn more money to keep going, but it has a chance to do that. It's raised a lot more money. OpenAI Foundation has this huge stash. You can look at the exact number, and I think they are planning on using it to support AI safety. Will they be effective? I can't say. But the world goes on.
To me, it's more important that OpenAI has real competition with Grok, Claude, Meta, Google, Gemini, China. When you're concerned about safety, the competition, in some ways, it's a negative because there's a race. Everyone has to hurry, hurry, hurry. But it's also a guarantee that if your product is not safe, people don't want it. And I look more to the competition than to anything else. What Elon said, what Sam said, what was the original plan, who misled whom, I'm an economist. I look at the competition. I'm pretty happy with how things have ended up.
Interviewer 1:09:29 ↗
Did you read what the new pope wrote about AI?
Tyler Cowen 1:09:32 ↗
Yes. It's interesting that a lot of what the Vatican put out failed the tests and it seems was written by AI without anyone admitting it. But that's a good wakeup call for people. It seems to me the new pope is very smart, very constructive, and he sees the world is going to change a lot. He was non-committal in a wise way, but he told everyone in the church in a way that no political leader has managed to do, like, 'Hey people, pay attention.' I think it's a great thing he did. It's the most happy I've been about a pope that I can remember.
And he didn't talk about banning AI, but he said, 'Look, we need to do this right,' real leadership. And I think he hit just the right tones.
Interviewer 1:10:17 ↗
It's a lot of generalities, to be frank, but sometimes that is what is useful.
Tyler Cowen 1:10:22 ↗
It creates focus either way.
Interviewer 1:10:24 ↗
Exactly. If you're too specific and committal before we know the details, you end up looking like a fool.
Can you either read the piece or have your device read the piece?
Tyler Cowen 1:10:35 ↗
The Pope's?
Interviewer 1:10:36 ↗
Yeah.
Tyler Cowen 1:10:38 ↗
Okay. Here's the first sentence. One dot. Humanity created by God in all its grandeur is today facing a pivotal choice: either to construct a new Tower of Babel or to build the city in which God and humanity dwell together. Each generation inherits the task of shaping its own era, of guiding history to become a place where the dignity of every person is safeguarded, justice is promoted, and fraternity is made possible.
So it starts off fairly critical, like Tower of Babel, that's the world collapsing. But the Pope goes on to say we can avoid this future if we do it with prudence and wisdom, and in his view, inspiration from religion, God, perhaps the Pope himself, which I do think we will need. And that there's a possible future where humans and machine intelligence dwell together on Earth and dignity is safeguarded. But we need to work very hard to get there. So the first sentence is a concession to the worriers, but he circles back to the possibility of a more positive vision. But I read it all as ending by reminding us hard, hard work and wisdom and piety are demanded from us to make this work. And again, I think it hit all the right notes. I'm not Catholic, to be clear. I'm not even religious, but I was very pleased by this.
Interviewer 1:12:05 ↗
Okay, let's talk about data centers. Can we see far enough into the future to...
Tyler Cowen 1:12:12 ↗
understand that the things that we're building are actually what we need going forward.
Right now, the data centers are doing fine. They work. They create some number of jobs. They pay a lot of tax. The environmental damage is minimal. Now, Elon has the view, you need to put all the data centers in space. We'll see if that's possible. Many other people are skeptical, but as Peter Thiel likes to say, never bet against Elon. I'm not sure about that one. But if Elon is correct, we need to prepare for all the data centers being in space. And he's working toward that end. In space, there's no local government that can stop them. They have in essence free power because they're solar powered, but there's no atmosphere diluting the force of the sun, and they would beam power back to Earth, which is itself a tricky matter. But at least in principle, if you've seen enough James Bond movies, you might feel this can work. Something that worries me is anything powerful you put into space can be a weapon, and it may not be intended as a weapon, but data centers in space, if they can beam power to Earth, can be potential weapons, and that's a reason to be wary of the idea.
Interviewer 1:13:26 ↗
Why would a city want or not want a data center close by?
Tyler Cowen 1:13:31 ↗
I live in Northern Virginia, maybe half an hour away from the world's capital of data centers, which is Loudoun County. Loudoun County was once, if not poor, not wealthy. I believe now per capita, it's the wealthiest county in the United States. It's perfectly fine. The data centers are not noisy. They don't create trouble. They don't even bring in so many kids where you have to pay for their schooling, and they pay about half of the tax for Loudoun County. It's not quite a free lunch, but it's gone great for Loudoun County. Now, a lot of parts of America don't want the same. Some of that is misinformation. Some of that is the data centers are a stand-in for a more general unease about AI, and they want to strike back against AI. So, many, many places are passing restrictions on data centers. But as you know, we have a country of 50 states, and it will happen somewhere or another, and a poor place like Louisiana, which has agreed to do a big data center with Meta, is going to do it. And we're going to do it somewhere. I'm pretty sure it's going to work.
Interviewer 1:14:37 ↗
Why do data centers need to be all over the world?
Tyler Cowen 1:14:40 ↗
They don't, and they probably shouldn't be. You should concentrate them in particular places that learn how to deal with them. It's somewhat more efficient if data centers are in colder climates. That won't always be possible, but people talk about Norway and Finland as two countries with special potential for data centers. That may happen for Europe. Canada doesn't seem to be running with that ball, but they have the potential as well. Some people have talked about Iceland. And also in the north, there's just plenty of empty space. You could put these things far from people, and it would still all work fine.
Interviewer 1:15:13 ↗
What are some of the negatives?
Tyler Cowen 1:15:15 ↗
Jasmine's son just published an excellent piece on this. She went all around what you'd call middle America and interviewed people. I haven't read the whole piece yet. I've read parts of it. She seems to be suggesting it's not about the data centers per se, that they are a stand-in for people's broader worries about the future of the country, what AI will do to lives. That's just something visible and immediate they can fight against and strike at, and that if you just lecture people data centers are fine, they don't use that much water, dot dot dot, that you're missing the point. She did the work, not I, but I suspect she's correct.
Interviewer 1:15:54 ↗
So, it's really more general fear of AI than what the data center brings with it.
Tyler Cowen 1:15:59 ↗
And just the future, you know, we're in 2026. If someone feels a lot of things have not gone the way they were promised, that's a pretty reasonable reaction to a lot of global events. And then you tell them, well, there's a new thing, and this new thing is going to go the way that will be quite positive, but we can't predict much about it with much clarity. But trust us, again, if they're skeptical, that that's hardly crazy on their part, even if I, Tyler Cowen, might disagree with them.
Interviewer 1:16:26 ↗
If there was an island in Alaska, uninhabited, if you could build enough data centers there to do all of the world's computing, could it work in one part of the world, or is it necessary to be distributed around the world?
Tyler Cowen 1:16:43 ↗
For some economic reasons, it's necessary because transmission is an issue, but there's political reasons why it's necessary. So if you're a European, you trust the United States less than you used to, and you don't want your data center to be dependent on the possible whims of an American president. Now, it's not as easy as the Europeans think. They think if they have everything all in the EU, their problem solved. But in fact, the EU so often disagrees amongst itself. And if say the main valuable items are in France, well, the balance of power in the EU is shifted, and France maybe wherever has a lot more say in things. So they're just transferring the problem to a different level, their own governance, and they might get along less well once one of them has a lot more power. So I don't think there are easy solutions, but it seems to me obvious that it's not acceptable for the world to put it all in one country, no matter what the country. And we're especially now at a time where the US, whether one agrees or not, is disliked more than it was say 10 years ago.
Interviewer 1:17:46 ↗
What about if they were all in the Antarctic where nobody can go?
Tyler Cowen 1:17:49 ↗
Well, then how do you set them up? How do you get the power? How do you do repairs? That's an issue in space. So, if they're up in space, which in some ways is like the Antarctic, there's a lot of damage done to things up in space. Meteors, for instance. Oh, and you need to repair the data center in space. One vision is you send up a robot. Another is, oh, just send up more stuff and let there be a high rate of attrition. Hard for me to judge, but you can see those are major issues, right?
Interviewer 1:18:17 ↗
What makes reasoning models different than the models before them?
Tyler Cowen 1:18:22 ↗
We're using the human analogies here, but maybe there's no better way to proceed. The reasoning models can actually think things through and reason. They're not just what we used to call stochastic parrots. So they can solve problems for you. If you give them an economics problem that no human has seen before, a reasoning model typically will get it right. And the earlier, say GPT-4, which was not a reasoning model, would not get it right. So o1 was the first reasoning model. o3 was the first good reasoning model. Now it's embedded in everything. And it was a huge breakthrough in my view. When we had reasoning models, that was AGI, if you're going to insist on when that term should be applied, because they reason better than humans do, but not perfectly.

185 more exchanges in this transcript

Sign in free to read the rest of this interview. No card required.

Sign in to read the full transcript

Cite this transcript

APA, MLA, BibTeX
APA

Cowen, T. (2026, August 19). Why Is Tyler Cowen Optimistic About AI? [Interview transcript]. Tetragrammaton with Rick Rubin. CEOInterviews.AI. https://ceointerviews.ai/interview/3203166/

MLA

Tyler Cowen. "Why Is Tyler Cowen Optimistic About AI?." Tetragrammaton with Rick Rubin, 19 Aug. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/3203166/.

BibTeX
@misc{cowen2026_3203166,
  author       = {Tyler Cowen},
  title        = {Why Is Tyler Cowen Optimistic About AI?},
  howpublished = {Interview transcript, Tetragrammaton with Rick Rubin. CEOInterviews.AI},
  year         = {2026},
  month        = {aug},
  url          = {https://ceointerviews.ai/interview/3203166/},
  note         = {Speaker-attributed transcript with timestamps}
}