Tyler Cowen: The Economics of Artificial Intelligence

The people who are the AI doomers, my suspicion is they like the idea of knowing something very, very important that other people do not know. So I worry about bad humans with AI much more than AI itself. Of course, that's already a risk.

Right now, obviously ChatGPT couldn't do what we're doing, but—

Sure it could. My voice clone is great. And the answers, well, it's for other people to judge, but it would give you good answers. And if you gave GPT-5, and myself a 100-question exam on economics, who would win? Uh, I would lose. It's pretty simple.

So Tyler, you're an economist. You have an ethnic eating guide. You have a podcast, Conversations with Tyler, and a lot of other endeavors going on. And in, in the spirit of giving a through line to what we're about to discuss, involving AI, economics, or whatever else comes up, I wanted to start by asking you about state capacity libertarianism. So what is it? And am I right that it's currently your big picture economic political philosophy?

Uh, that's correct. So I think of libertarianism as the view that capitalism is essentially good, personal liberties are important, Government should not be very paternalistic, and that should be the basis of policy. The state capacity angle is simply to point out that for that to be stable, you need your government to be able to do certain things. So most obvious would be national defense. But at the time I wrote that post, part of what I had in mind was protection against pandemics, namely COVID, that if you were in a situation where the government was not able say, to roll out a good vaccine promptly. Over time, your regime would lose legitimacy. You would probably end up with not something more libertarian, but something more corrupt, more socialistic. And to guard liberty, you need some state capacity. I think it's a simple point. To a lot of people, it's self-evident. To a lot of libertarians, it's highly controversial.

I'll just play the devil's advocate of that. Libertarian you mentioned. And what is the basic argument you'd give against the libertarianism who wants to fight against big government at all costs?

Well, at all costs, that would mean that every road or street would be privatized. I think you'd have big holdup problems. It would mean we would not have an effective national defense. We would not have alliances with other free nations. The United States government would not be supporting science, which I think has had a high return. So I think it would be a much worse world.

Is there a way in which you think about putting a hard line for just how involved the government should be and how much the state's capacities should be?

I don't know that there's a clear line. I would say this: when I meet non-libertarians, I always want less government than they do. When I meet libertarians, I typically want somewhat more government. Than they do. But I don't think there's some point at which the whole curve bends. It's a matter of degree. You take a guess, we're all going to get it wrong. I'm pretty sure, you know, you can either have too much or too little state capacity. You're never going to get it right. State capacity gets used for bad things as well. Surveillance would be one simple example. But in fact, you need some surveillance to limit terror attacks. So it's all a delicate balance.

One area where I think state capacity is particularly relevant today is with regard to AI. And in the past few months, I've had some conversations with people who have very divergent opinions on the topic. The two that come to mind are Eliezer Yudkowsky, who's very much an AI doomer, and then I just spoke to Reid Hoffman, who's very optimistic about the future of AI. And so I wanted to talk about this balance between state and market forces reflecting in AI deployment and the alignment problem. And I know you have views on a wide range of topics. So before we get started on the economics and the politics of it, I was just wondering if maybe we should discuss the alignment problem and existential threats that you may see from AI?

I'm not sure the alignment problem is the correct framing. It assumes there's this one big AI and we've gotta get it to agree with what humans want. I think of it more in terms of building an institutional framework in which there will be checks and balances across different AIs. But I think AI is a good example of the relevance of state capacity libertarianism. Uh, there's a lot of new facets to AI, and we need to build some kind of new legal code to deal with that. So how do AI and copyright intersect? Well, my personal view is AIs should just be allowed to read material on the internet, uh, like a human could read it and use it, as long as they're not reproducing the words verbatim. But not everyone agrees with that. Right now, there's legal uncertainty. For AI to proceed, We need to clear up those uncertainties, have clear rules, uh, and then we'll get a better AI sector. If we just do nothing, there'll be this legal uncertainty forever hanging over AI.

I take your point about the alignment problem, assuming that there's just this one megalithic AI that we need to figure out how to rein in and make sure it aligns with our values. But I guess more broadly, what I'm wondering is To begin with, do you see AI as this major existential threat that needs to be curbed? No. Or are you, you do not?

No, not at all. It's more likely to extend my life than to shorten it is a simple way to put it.

Is there a way or a place in which you think that the doomers are going wrong in their arguments? Because for them it's very straightforward. So Eliezer Yudkowsky's I think the motto for his organization is something like the default consequence of AGI is human extinction.

Well, the rationalist community typically when evaluating ideas, they apply three standards, which I, I favor. The first is look at what market prices tell you. The second is look at what peer-reviewed literature tells you. The third would be something like, Look at what superforecasters tell you. They break the rules in all three cases. So market prices show no sign of a major AI risk at all. Risk premia, in fact, are down even with Trump in office. Uh, superforecasters assign really a quite low probability to existential risk from AI. And then there is no body of peer-reviewed literature that establishes the Eliezer conclusions. You know, they're just these very long blog posts that they write to each other. So I would say try getting that peer reviewed and do what the climate science people did, which did convince most of the world. And on all three, they fall down. So no, I'm not there. I'm not even close. I would say to them, get your act together, do those three things, then we'll talk.

As I was preparing for this interview, I came across a post that you had written a long time ago about something called the fallacy of mood affiliation.

Yeah.

And before I, instead of me saying what it is, I'll ask you what is it and whether you see it at all coming to bear in these debates about the future of AI.

The fallacy of mood affiliation is to choose your views on the basis of a mood, like to think there's a correct mood. Maybe it's optimism or maybe it's doom, Or maybe it's fear. And then from the mood to work backwards to policy conclusions. The people who are the AI doomers, do they do that? I'm not sure I'm in a good position to psychoanalyze them. My suspicion is they like the idea of knowing something very, very important that other people do not know. And that would be a different mistake than one of mood affiliation. I'm not sure they're sort of pessimistic people in general. I don't know that they are.

For my own understanding, is it that the mood is chosen and then the views that align with that mood follow from it?

Yes, exactly.

Or the mood is— well, what I was going to ask is if it's just you're not really aware of your own mood, but then your mood determines the views that you hold.

You may or may not be aware of the mood, but either way, it determines the views that you hold. So in general, I would say be suspicious of people who are always optimistic or always pessimistic. So I would say, for instance, global affairs, global conflict, I'm actually somewhat pessimistic at the moment. Economic growth per se, I'm modestly optimistic. But a lot of people will just be optimistic on both or pessimistic on both. I'm not saying that can never be right, but when the person so consistently lines up with the mood, One should just start being suspicious and focus more on the analysis.

The reason that I made this connection to AI when I was— when I read about— read this post yesterday is I— when I went into this interview with Eliezer, I was in a particularly depressed state. And so I found that his conclusions were very compelling. His arguments were very compelling. And then for weeks afterward, I was just ruing the end of humanity. But then Over the last couple of months, my mood's gotten better. And then when I talked to Reid Hoffman about the same issues, his views were the ones that seemed to be the right ones. And I wondered if it was, oh, just because my mood was better and I wasn't really dissecting the arguments well enough. But a problem there is that I'm not an expert in AI, so I can't really make the evaluations myself about the state of computer science.

I'm not sure the experts in AI are the experts on its social impacts. And speed of diffusion. But I would just ask, have you maxed out on your credit cards yet?

No.

There we go. You're not that pessimistic.

There's a principle in economics. Does it have a name? Obviously, you're the person to ask where, where your views are determined by your actions rather.

Demonstrated preference would be the shorthand name we give to that concept. That's right. Your demonstrated preference from what you say, is not all that pessimistic. You know, how much are you saving? How hard are you working? Other questions? I don't know. But my guess is you don't lead a mega pessimistic life. Not that many people do.

There was a lot of binge eating that kind of indicates letting things go, but that can be GLP-1 motivated, right?

You can eat what you want now. So is the way to undo it.

You said something interesting to me a few minutes ago about AI and copyright. And the reason it's interesting to me is I was recently reading about this band called The Velvet Sundown, which maybe you heard about. Did you hear about this?

The band that was popular and it turned out to be AI music?

Yes. Yes.

Yes. I've heard about that.

One of my best friends is a graphic designer and I hear a lot about this community of artists who is very opposed to AI art because they view what it is doing as theft in that it, it trains itself on extant art and then creates new art. I understand what you're saying, where isn't that how a human works? So maybe it shouldn't be illegal for an AI to operate that way.

That is my opinion, yes.

Is it for roughly the same sort of reasoning I gave you?

Uh, that is a reason, but the additional reason is that there are so many artworks or newspaper articles that to require transactions to make a deal just doesn't seem practical. So in essence, you would shut down AI. Or say you created a government fund, the AI companies pay into the fund. Well, how does the government decide whose images get how much money? That to me also seems like an intractable problem. So I don't know a good way of making that system work.

Copyright is just one of the checks and balances that you think we need across AI as we're buttressing the legal system for what's to come. If the regulations we're developing aren't so much about staving off this existential threat, which you think needs to be demonstrated as a, to be a problem much more than it has been, what are some of these other checks and balances that come to mind that we really need to be considering?

I think it's very, you know, David, who was it? Someone wrote a good tweet lately about the evolution of the AI sector has been quite positive so far. So one thing is there's just a lot of competitive models. It's not all the AI in the hands of one company or one person. There's also models in China. I think that's a healthy development. That's part of the checks and balances. It's no surprise to an economist, but reading some of the AI people, you almost get this vision like there's one model that just keeps on improving itself so rapidly it leapfrogs ahead of all the others. Uh, no reason to think that will be true. A lot of reason to think it's not true. And it's not true. That's good. It's more stable.

You mentioned though, something that is a problem even for Reid Hoffman, I think, which is that, I mean, it's obviously got upsides, but there are all these models. There's AI development in China. It's not localized. It's all throughout the world. And one problem is that putting AI in the hands of everyone produces the capacity, for instance, to make changes to viruses or bacteria and create major existential threats in the form of pandemics that aren't necessarily caused by an AI singularity, so to speak.

Yes, I worry about bad humans with AI much more than AI itself. Of course, that's already a risk. And I think we need to put a lot of work into making sure that the ability of AI to come up with cures just for normal pandemics, for that matter, is more powerful than its ability to harm us. So it's a real issue. We need to devote a lot of attention to it.

At this point, I'd like to shift more towards some of your own works. And in The Great Stagnation, you argued that the US had eaten its low-hanging fruit, but you've also written more recently that we're entering this new wave that we've been discussing. That's precisely about AI or concerned with AI in biomed? And what was it precisely that made you update the stagnation thesis?

Well, I first wrote a piece in 2010 saying I thought the stagnation was over. The fact that we had mRNA vaccines that saved millions of lives to me is a big deal. Millions of lives. And we did that in essence through the power of computation. You can think of AI also as another application of the power of computation. So I think just extremely strong computation is a new general-purpose technology. We're starting to see new fruits from it. Uh, those will continue for quite a while, and we're in a very dynamic age again, but it's just starting. And plenty of sectors are seeing no productivity gains at all, like what it costs to build a house. I see no signs of advancement there.

So you still think that we are in a period of economic stagnation in some sectors?

Absolutely.

But the ones that rely on computation are just going to grow and grow and grow.

And eventually they will spread to improve just about everything, but it will take a long time. So let's say you're pretty optimistic about AI, as I am. Well, how long will it be before the price of a new home falls? It could be 20 years. I'm not sure how long, but I certainly don't think it's in the next 2 or 3 years.

How is computation going to drive that fall in price?

It could lower the cost of building homes. We'd come up with new ways of insulating, better building materials, prefabs, robots. Not soon, but eventually I think they could make building new homes much cheaper. Now there's also the price of the land, but over time, one hopes that, say, through better irrigation, better, cheaper water desalination, we can increase the effective supply of land and make the price of land lower. That to me seems quite long-term, but eventually some set of forces like that will come into play.

These are developments in artificial intelligence that go beyond the LLMs that everyone is talking about. This is, this sounds something more like AGI, something that's going to improve building materials, for instance.

There's a lot of different terms in what you're using. I would just say AIs of many different kinds are advancing, including in the biosciences. AIs as a way of finding new materials, improving the processes in scientific labs. And it's not as if, you know, GPT-5 is gonna make the, the home building cost cheaper. But AI in general will, that will include LLMs though, not exclusive of LLMs.

Yeah, I suppose that LLMs would function or could function still in augmenting the capabilities of humans already working on these problems.

Yes.

Rather than just solving the problems themselves.

Imagine if a construction firm, its accounting and legal departments were more or less free. Right? That would lower cost. That to me seems pretty realistic within 5 years. It's not the main cost behind building a home, but it's something that will kick in. And again, you would mostly credit LLMs there.

I never used to have privacy concerns, but now that I'm convinced the essentially unlimited manpower of AI means that none of our data is safe from prying eyes, privacy is a much bigger priority for me. Enter Surfshark, which is this episode's sponsor and my VPN of choice, and you'll find a link and coupon code in the description. I've had my sister try it while she's been traveling throughout Africa for work. I've had one of my friends who helps with the podcast use it in Asia. Surfshark is actually the only VPN to cover 100 countries, so it's excellent if you're a traveler who needs to stream Netflix or if you just really care about encryption. There are 100 other reasons that you should be using a VPN. Surfshark has a no-log policy, so they never keep any of your data. Their 30-day money-back guarantee means it's risk-free to try for a good chunk of time. And 24/7 customer support means they'll help you out whenever you need it. Again, I highly recommend Surfshark, especially now that privacy is more important than ever. And again, it's risk-free to try, and there's a link and coupon code, Robinson Erhardt, in the description. In July, I think it was July, the Google's AI got a gold medal in the International Math Olympiad. And I think you noted then these rapid capability gains. And I'm wondering, obviously not what effect this will have in, in sectors that are currently stagnant, but what effects will this rapidly developing, like, top-tier cognition have in the labor market?

Well, most jobs will involve working with AI. To a considerable extent. Until we have truly human-equivalent robots, which I think is quite distant, you still need humans to interact with the physical world. So jobs will change a lot, but I think we'll have full employment. You'll just need these new skills the way now today, more or less every worker has to work with the internet in some fashion, but it will be more so.

Somewhat aligned maybe with Reid Hoffman, who I mentioned earlier, who just had this popular piece about new grads' strengths being that they're AI natives. They already have these AI skills that will become really valuable going forward.

Reid and I have discussed this, and I believe I agree with him. We also did a podcast on some of these questions.

So I'm not surprised that employment might— wouldn't change really, and we'd have all the employment we currently have and in an industry like homebuilding, but industries like graphic design where jobs are rapidly being displaced. If you still think that we're going to continue to have full employment, where are these new jobs going to be located?

Well, the energy sector is booming and it will continue to boom. Uh, the biomedical sector will have many new ideas to investigate. It will be much easier to start a project. We'll have a greater chance of terraforming more of the Earth. Uh, there'll be more time, leisure time that people have for entertainment, storytelling, interacting with each other, uh, just care for the elderly. Those would be a few areas.

The storytelling one, to me, kind of goes hand in hand with the displacement of graphic designers and other artists in creative fields, just because I see AI as— I mean, and who knows what's going to happen in 5 years. I mean, today if I look at Instagram, I already can't really tell if some videos are AI generated. And I can just imagine in 5 years going on a place like Netflix and typing in my interests and having a movie generated for me by AI, or going to the Kindle store and getting a book generated for me by AI in the same way that right now, if you need a label for your new tequila, you can just go to ChatGPT and get it to design that for you. And I'm wondering if you think that a lot of these so-called like storytelling jobs, jobs who produce things that produce things that people enjoy during their leisure time. Will be displaced by AI.

Well, that's very likely to happen, that there are AI-generated stories. It's already true, right? But I think people will always want most of what they consume to be human-generated. You want to be a fan of Taylor Swift, not of a robot. You want to watch Humphrey Bogart on the screen, not some AI-generated character. There'll definitely be both. But I think the room for the human will be robust. And also just the work of generating the AI stories. I think that will require human input for a long time. And, uh, that too will be a new source of jobs. They don't just spring up out of nowhere and how they're marketed and which ones they think people will like. A lot of those judgments will be made with AI assistance. But again, I think humans will be in the loop for a long time.

It is hardening that you say that because I've worried about it, and Taylor Swift is the example that I've used for myself, and that I don't see AI displacing Taylor Swift. But for every Taylor Swift, there are tens and thousands of people who aren't Taylor Swift. For every, uh, I, I mean, I can't really name many current famous painters, but for every famous painter, you have hordes of graphic designers who are going to be displaced.

Well, people who produce muzak for a living, that sector will shrink in terms of employment. I think that's a good thing. It's liberating them from doing something horrible. Maybe they'll be more creative doing something else. So it's the lower tier, most routine forms of the art. Become done by AI. That's great. More people can try to be the next Taylor Swift. But it's not only Taylor Swift at the top of this pinnacle. So there are thousands of creators with smaller audiences where you still want some kind of relationship with the person. Like, I write a blog. I'm not anywhere close to as famous as Taylor Swift. But if I put up some GPT-generated output, as I occasionally do with transparency, people get upset. They're like, we want you. It doesn't matter that GPT may be as good as I am or better. That's heartening in a way. I think it's also somewhat irrational, but it is how the world works.

No, it is interesting. I, I obviously, I have a show too that you are on right now.

Exactly.

And right now, obviously ChatGPT couldn't do what we're doing, but sure it could.

My voice clone is great and the answers, well, it's for other people to judge, but it would give you good answers. I would say better than the median human podcast guest. Whether it's better than me is for others to judge.

I, okay. I think, I think you make a fair point, but in order to displace us, let's say we enjoy, we're listening to a, a comedy podcast. And maybe comedy isn't even a good example because maybe jokes made by AI, if we know that they're from AI, we won't react to them the same way that we would from an ordinary human. But say somebody's listening to a comedy podcast with real-looking humans speaking in a realistic way that are saying the absolute funniest things imaginable. I wonder if that would displace real comedy podcasts with real humans? Or if, let's say, you were just having an economics interview, a real one compared to one with ChatGPT-10, where they're much better, much faster than any living economist, would people rather listen to the real thing or the fake thing?

I definitely think we'll have both. We do already. But, you know, take GPT-10, whatever that means. GPT-5 already knows more economics than I do. So if 10 is somehow way, way, way better, the listeners are not able to appreciate the improvement. So they might as well listen to me. That, those are the cold, hard facts of life. That's part of my protection here. So we're already at margins where in many, many areas, We cannot so readily perceive quality improvements in the AI, for better or worse. So they're going to listen to me. I'm not worried about GPT-8. I'm worried enough about GPT-5. 8, in that sense, will not be perceptibly so much better. Maybe in comedy it will be, because right now they're not that funny, but not in economics.

When you say that ChatGPT-5 already knows more economics than you, What do you mean by that? Because I've been thinking, as I have taught undergraduates here in philosophy, I've often wondered, does ChatGPT know the philosophy much better than me? And in certain ways I want to say yes, but in other ways I want to say no.

I think that's the correct answer. But it does know more than you do. And with good prompting, which of course requires your input, I mean, it will beat you. Like a simple question. If you gave GPT-5 and myself a 100-question exam on economics, who would win? Who? What would win? Whatever. I would lose. It's pretty simple. Now there's more to economics than just answering questions, but still, I think that tells you a lot. I would lose. And you would lose in philosophy. No slight intended, to be clear.

I would. Absolutely lose at a certain sort of exam. If it were historical questions like, what view did Kant defend in X or Spinoza in Y? GPT's read absolutely everything. But if you asked me about something that I happen to maybe know a lot about and you really got granular on it, that's when you would see ChatGPT start to crumble and say things that don't really make sense. And I wonder if it would be similar in economics for you.

I wouldn't say that it crumbles. I would put it this way. There's a lot of areas where if you asked one of the top 5 humans in the world about that question, you would get a better answer than from GPT. But that's the top 5 humans. You could still ask like a tenured professor at Harvard or Stanford in that area, the same question. And odds are GPT would win, but they wouldn't beat the top 5 people, especially as you got into more and more depth. But man, that's a pretty thin reed to hang on to. And like how many people in the audience will even know that the top 5 expert is ahead of the GPT? Some, but like not 5,000 of them.

How do you see this impacting education? This is something else that I talked with Reid about because up until relatively recently, my expectation had been, well, I'll, I'll just be a professor. That's, people have always done that. It's a very old profession. There's a pretty, it's difficult and I don't know if I'll be able to do it. I'll get there, but it seemed like a very viable career pathway, but it really does not feel that way anymore.

I don't think we know. So I view our institutions of higher education as unable to change. No one is actually delegated with the kind of power to make any real change. Forget about AI. So they're just stuck. And in the meantime, people like you and I, we freelance on our own, and use AI and learn things. We kind of privatize that part of our education, and it will be that way for quite a while. Now, you're young, and let's say you live to 98 because AI, you know, cures your cancer. So what, you're 30 now? So that's like 68 years or like 60 years, 50 years you've got to worry about. I can't tell you it will be intact that whole time. Maybe it won't be, but I think it will be intact for longer than many people are expecting. So I'm not sure you should give up the plan.

Here's my thinking, and don't interrupt me if you think I'm wrong. I'll just give you a variety of thoughts I have. My understanding is that graduate students are brought into universities largely because they offset teaching load for very little money relative to what a full professor would have to do. And as I was teaching ethics, uh, last quarter, and ethics is not an area in which I'm an expert at all. This was an area where if my students asked me a question, I could tell them my opinions, my understanding of the text, but ChatGPT would give them a much more canonical answer about something regarding Kant than I could. And I just see a future in which there may be humans poking their heads in here and there to supervise things, but my students would get a lot more out of getting their questions answered by ChatGPT than by me, getting the material explained to them by ChatGPT than by me. And to the extent that writing is something that's going to be valued in the future, I can't spend nearly as much time as ChatGPT can giving very good feedback on essays, especially as ChatGPT continues to improve. So I just imagine that in a place like a philosophy department, where unlike in, say, chemistry, you don't need human hands to manipulate lab equipment, this is just going to be— this labor is going to be displaced relatively soon by AI, especially in the current environment. Of cuts that we're undergoing. So that's just where things start to begin in my thinking.

I definitely agree with your basic description of what you said, but I'm not sure how much the world cares that the GPT answer is better. And keep in mind, you're at Stanford. That's a highly atypical school. 80% of the sector is state schools. So much of the sector, students don't really learn anything. They learn how to be in society. They interact with different groups of people. They learn something quite intangible by interacting with humans. And the fact that the GPT answer is better, I don't know. I think it takes a long time for that to chip away and arbitrage down what might end up paying for your salary. I think eventually it will, but eh, it's a very slow process.

I agree that it might not, like, the students might not think, wow, ChatGPT is answering my questions better. I'd rather have a conversation with ChatGPT. What I am thinking is that while ideally the university is this bastion of learning, the reality I've discovered is the university is a business. I don't quite understand why universities are businesses because Nobody owns them. That might be something interesting to talk about because I've been curious about it, but money rules the university.

Of course.

I'm not, I'm using the university quite the term to cover Harvard, Yale, Stanford, all those big places.

Right. Land state schools. It's completely true. But that's also a reason why it's not easy to dislodge the key piece whose fall would wreck the system. Would be if employers started saying, I don't give a damn if this person went to School X, you know, let them debate with the AI or take a test with GPT and see how well they do. That's a good enough credential for me. And then the whole thing really collapses. I'm just not sure when that will happen. Again, I don't think it's soon.

I can't say that I think it will be soon because a name like Harvard just has a lot of oomph to it.

Well, those schools, it will never happen. If nothing else, they're a dating service, right? But, you know, take away the top 10 or 15 schools, when will it happen? I think that's the key question. It'll never happen for Harvard.

Even Harvard, to again play the devil's advocate, a lot of the students that get in there aren't getting in there because of academic merit. And maybe that's always been the case.

Always been the case.

Moreover, inflation is— great inflation is absolutely undermining—

No, no one cares. That's the whole point. Stanford average, I've read it's an A-. No one minds. It's not a problem. It's a sickness that it's not a problem, but in fact it's not a problem.

I'm going to compound what is not a problem and just say that ChatGPT makes grades even less sensitive to merit, of course, I suppose. But you're saying that you think that people still aren't going to care. It's just the Harvard name or, or the Yale name that will get people employed.

Well, the networking— it's not only the name. The networking is critical, and I don't see how that ever goes away. But again, a lot of schools, the networking value is much smaller. So I'd still draw this big distinction between the very top and lower down.

Okay, back to this question that I, that I raised a few minutes ago about the university as a business. What I don't understand is that, say, take a public corporation. There are shareholders, there are people who own the corporation, and so it is very important to these people that certain metrics like money in the bank or goodwill or sales or profit go up because that increases the value that they can appoint to their shares when they want to sell it. But nobody owns the university, so nobody is personally— I mean, there are some high salaries among administrators, but If the endowment goes up, it's not like somebody's net worth is going up by billions of dollars and they highly value it. So I don't understand why universities are so governed by money.

People get paid and paid off by the system in different ways. So look at a simpler case that we understand well, hospitals. There's for-profit and there's nonprofit hospitals. No one owns the nonprofits, right? But you look at data, you compare the profits to the nonprofits, they hardly behave any differently. They do a bit, but you can't just walk into a hospital and tell. So the nonprofit has incentives, it has to raise revenue, meet market demand, raise money in capital markets. It's just not that different. It is somewhat less efficient, but the thing will continue. And we see it with nonprofit hospitals. You could say like the output of universities, what they really do, it's somewhat mysterious, but not hospitals. We know what they do, right? They operate, they fix people, they kill people, whatever. And that's how it works with nonprofit hospitals. So I don't think that's such a big mystery. You expect the world to be too rational or look at churches, uh, driven by revenue. Most people in them are also sincere. But they know they have to raise revenue to pursue the mission. They stick around, they have outputs, people donate, it, you know, works reasonably well. And there you go. We have churches.

I guess it would make more sense to me in a way if I believed that administrators at a university really cared greatly about education and the pursuit of knowledge. We do.

But the word greatly is doing a lot of weight in that sentence. I would say they care about it, but at the end of the day, you face a revenue constraint and to keep your operation going at all, you've gotta meet the revenue constraint and that shapes your behavior. How much you like care about Jane Austen up here in your little head, you can debate that. But at the end of the day, you've gotta raise the revenue.

Stepping back again to where we started. You don't think that becoming a professor is a totally non-viable career path, but how do you see AI changing higher education? And I'm guessing though, you said, you said that higher education is something that is going to be very slow to change. It never changes, but how do you see it changing, if at all, from AI?

I think over time there will be more and more alternate means of certifying talent, which will include AI, but it doesn't have to be AI, you know, Thiel Fellowship, or just you take a test or, you know, start a company when you're 19 and the labor market will reward that. And that combined with declining enrollments, which is already in the cards because of birth rates, nothing to do with AI. Will really squeeze the sector and the sector will become worse and more good people will leave it and that will make it a bit worse and there'll be the slow painful downward spiral. So I'm not sure I should encourage you to join a sector like that. And furthermore, you're clearly smart enough to know that some chunk of what you're doing is a kind of fraudulent theatrical stand-in for what GPT could do better. And that might demoralize you. So if I were your age, I think I would not do higher ed. But that said, this notion that it will all just be gone in 7 years, I'm quite sure that's wrong. I think it's an option you can consider, but I would want you to be very clear on the basic economics with declining enrollment hitting before AI does, right? And that's just a fact. Like it takes 18 years to produce an undergrad. No way around that. I don't care if it's GPT-14.

You bring up another aspect of the problem that I hadn't raised, which is the demoralizing nature of ChatGPT's improving performance, because I saw that you, you've written a paper with Derek Parfit, which is pretty amazing. And maybe it's the case that just like with, with storytelling, philosophers will only want to read papers written by philosophers and not AI. But for people who are interested in, let's say, philosophy or, or history, whatever it is, to the extent that they're just interested in getting the answers to their questions, it does seem like in the future AI will be significantly better at providing those answers than humans. I agree. Now, if somebody's very interested just in, let's say, the, the very act of doing the research, of solving the math problems, then that's maybe rewarding and a, a luxury in a way. But it doesn't change the fact that ChatGPT could much more easily provide the output of the research.

That's right. That's why if I were you, I would not become an academic. But I think it's a, it's an option worth pondering. You have to think about how do I feel about that fact?

You're not only doing economics for the love of economics, you're doing it for getting the answers to the questions. And if the answers can be had so cheaply, it's not worth doing the work, or it's just no longer going to be economically rewarded.

Well, I'm age 63, which is older than you are. I'm also much better known. AI has massively changed how I allocate my time. I spend much more time going around meeting with people. You could call it evangelizing, in large part evangelizing for AI. And I think I have a high value doing that for now. But if we're in some kind of steady state and everyone has figured out the value of AI, and we're at GPT-9 and it's just better than I am flat out, uh, I would in fact retire from doing economics. And it's quite possible that comes about. It's a kind of race. There's my age, the speed of improvement. I'm not really sure, but I can definitely imagine saying, well, I'm only going to do the mentoring part of this job and I'm not going to do the producing economic answers part of the job. I can easily imagine that. I kind of expect it. But I don't have to sustain the 50, 60-year career that you do. So, like, I'm home free. I can run out the clock and just stop and not need another job. So I'm in a different position.

Some sectors of the economy are stagnant, some that revolve around computation are moving much more quickly. They're no longer stagnant. But earlier this year, you still wrote that the AI takeoff looks very slow, even as these capabilities leap. What is it about the takeoff that is slow versus what we expected?

I mean, the sarcastic answer would be humans suck, but there's this new paper out of MIT. I've only read the abstract. Tyler Cowen: But it basically says, in current corporations— and this is looking at a pretty, like, exalted set of companies— in 95% of cases, AI has not yet boosted profits because people don't really know what to do with this stuff. So I think you'll need something close to a whole generational turnover of companies where new companies arise, being built upon and centered around using AI. And the current companies just aren't that good at the transition. And that could be right there, 20 years. Now, not in all areas. Programming, it's gone very quickly. Graphic design, chess, it happened some while ago. But that's what I think it looks like. And we see it already.

The idea maybe is that, sure, in a law firm, ChatGPT is speeding the rate at which people are writing perfunctory emails, but it's not really creating meaningful increases in profits. And corporations need to really restructure their operations around AI to unlock its benefits.

Yes. And I think there's enough status quo bias in human behavior that in companies full of 47-year-olds, some will adjust, some have adjusted great already. Uh, but basically you'll need a new set of law firms or whatever, medical practices, whatever, before it really sticks as fundamental to the enterprise rather than just, oh, the doctor talked with me, the AI took notes, which is great, right? But it's not that revolutionary at all. It hardly saves much money.

Okay. And I'm guessing that if you had an idea of just what law firms would need to do to unlock this revolutionary capability of AI, then it probably would've been done by now or would be doing. So is it more just you have a hunch that AI will do these things or you have an idea of how they will be done in a, just like a law firm, for instance?

Well, I know some people in law firms and they tell me that right now, for legal reasons, the firms insist they do not send any confidential material into LLMs. And I understand fully why. So the first thing you need is for LLMs to be good enough that when they're housed on the law firm's hard drive rather than in the cloud, it's still super good. And we're a ways from that, and it won't even get started much Until that happens, I don't know how far away that is, maybe 5 years, not, not some eternity, but that's what you need even to get started. 'Cause you can't just dump your confidential information into the cloud. So that's a huge barrier and we're not about to break it.

Well, since law has now come up, I'm starting law school here at Stanford next week.

Great.

And one reason for that is my existential dread about academia.

But it's the same. I mean, law is actually a competitive sector, so I would worry more about law than academia.

Could you tell me more about why that is? I know that paralegals are really seeing their work reduced, but why law firms or why lawyers?

What you might call senior partners who can bring in clients will be 10 or 20x more valuable than they are today. And they're already the most valuable asset because the legal work can be done by GPT. I already use it as my lawyer. It does great. Uh, my wife's a lawyer. She's impressed by it, but partners, partner-client relationships, that will just become so important. So if you can end up doing that, Wonderful. But exactly how you get to that without having climbed the previous ladders that we've had before, that's what I think we don't understand. And I don't have a good forecast for how that will go, but I would think that's the critical variable.

Something interesting that's coming out of this conversation for me is that it sounds like in addition to the obvious, which would be having skills for working with AI, one of, or the most valuable skill people can have for making it in this coming age is people skills.

Correct. I said this in my book, Average Is Over, in 2013. Charisma, being personable, connecting people, understanding other people's talents, just so much more important. And it's why I'm not worried about me not having a future job. I can always do some mixes of those things, which I have a lot of experience at. I don't need economics per se.

Why did you write about this already in 2013?

Well, I, this will sound immodest, but I saw it coming. I wrote a whole book called Average Is Over. And I think the predictions of that book have held up quite well.

It sounds like there's a thesis in the title, Average Is Over. What is the thesis of the book?

If you're able to work well with AI, you can do incredibly well. That's a sector open to everyone. It's genuinely open, right? Like everyone's GPT or Claude is really good. Something stopping you from being great with it. But a lot of people, they're not motivated or they're caught up in other battles. And those people, if they end up having to compete with the AI, they won't do so well. And I think it's a volunteerist question. How many of us will rise to the occasion? Learn how to work with AI and do really well? I'm not sure what the answer is. I would say overall, I've been heartened by the rapid rate of adoption in the United States. I think it's a very good sign.

Earlier you said that, I mean, it was a joke, but Harvard is kind of like a dating app.

It's not a joke. I mean, it's a joke, but it's a true joke.

Not that Harvard's value wasn't always in part its networking role, but that going forward, that's just going to be its dominant function.

You got it. For networking, Stanford also. Now Stanford, you network in different ways. You're more likely to network with the startup world than you are at Harvard, but you know, different schools are different. They're both incredible for networking.

How do you advise, I mean, young people who are going into this world going to college now, do you tell them, look, spend less time studying math, spend less time studying computer science, worry about joining the frat, the good frat, making good friends, maybe interning at the right businesses? Like you say, that's, that's what's really important now going forward.

Well, I think being interesting at math or computer science is the best way to network with other people. Who are very, very smart. So I don't tell people to study less. You may be studying with a slightly different path in mind, but also studying is fun. It's your chance in life to learn those things. It is, I think, of high intrinsic value to know certain things about how the world works. So I don't discourage that at all. In fact, it may be a luxury that more people can afford in a world with a lot of strong AI. But I do tell people you need to network much, much more. So more networking, but I don't want to say less math, more of both.

Do you have an opinion about— I mean, obviously one's individual preferences are going to be king in the end, but whether or not it would be more worth somebody's while to study music or literature than math or computer science, because like we said, nobody's going to replace Taylor Swift, but computer scientists, just entry-level programmers or mathematicians, are being replaced on some level and will be more so.

Well, Taylor Swift did not study music in the academic sense. I know this sounds corny, but I think you should actually study what you love. Now, if your parents are immigrants, you came to this country when you're 8, your family has no savings, maybe you just need a very practical major. You know, you can't study what you love. There's a lot of cases like that. Of course, you just need to be practical. But if you're in the position where the choice is a luxury at all, I would just tell people the old school corny advice. Do what you love. Work it out. You know, it'll try to work it out later. Uncertainty right now is really quite high. You can't figure it all out in advance anyway, and do what makes sense for you.

One of the big large-scale economic problems or worries I have, or lots of people have, is widening wage disparities in the coming AI revolution. And I'm wondering, how you would— well, one, what your attitude toward that is right now, whether it's just inevitable and where the one where the capital is going to coalesce, but also how people who aren't already wealthy can still achieve mobility in that arena going forward.

I don't expect widening wage disparities I think gardeners, carpenters, plumbers, people in the energy sector who do physical work, they're going to earn great. It doesn't mean every, you know, manual labor job will do well, but I think there are some of the biggest gainers and the people who are, you know, word cells, a lot of the upper middle class maybe should worry some more.

What do you think the upper class positions will be going forward? I'm not talking about old money, but people trying to create new money.

People who start new businesses with a super small number of workers and hardly any dilution, and the AIs do 85% of the work, they will earn a ton and be very high status, and there'll be a lot of it.

It sounds like going forward, entrepreneurship is going to be an extremely—

as it has been. Yeah.

More so than before if professions like being a lawyer, which have been the stable professions going forward that everybody's wanted their child to be, are no longer viable.

And rather than your tech company having a payroll of 4,000 people, you know, maybe it will be 80 people and you'll earn a lot more money until you sell your company.

But again, I guess that's when you make the money.

Yeah, sure. But that's not new. And again, all these new projects, I think the people who make it work on the ground, Uh, they will earn more. Your janitor.

You recently wrote in a piece that what we're writing, everything we're writing, especially on social media, the primary audience is now going to be AI, or at least it's going to be AI as much as regular humans.

Sure.

What are the implications of this?

You don't have to be as clear a writer because the AI understands your background context. So you could talk about, you know, Frege, say whatever about Frege, and not have to explain to your silly human readers, you know, why Frege was important. You can just say Frege this, Frege that, and the AI knows. It makes it easier in a way for people who like to get to the point.

This makes sense to me if you're typing directly into a, into the terminal to ChatGPT, you don't have to be as specific, but if you're posting on your blog, for instance, and you expect the blog itself to be read by humans, then it seems like you would probably be writing the exact same way you always have.

No, I've changed somewhat. And I also know the humans can ask the AI. If they need to learn who Frege is, well, they had Wikipedia, but I think asking current AIs is better. So again, I can take some more shortcuts and I do.

It is funny. I, in, because law school is starting and I know that my life is gonna be kind of over for a couple of years, I've been reading a lot of, uh, books I've wanted to read and I'm reading The Sound and the Fury right now. Right. And today, I was, I had to like copy a bunch of passages and put them into ChatGPT because I had no idea what was going on. And I wanted to say, nobody's gonna put your words into ChatGPT and ask what this means, where I'm already doing it today.

So, and people email me all the time like, oh, Tyler, you wrote this. I put it into ChatGPT. Here's what it said. What do you think? I get emails like that multiple times a week.

Do you notice with your students that writing quality is going way down?

Well, my particular students, I tell them to work with AI when they write their papers, so the writing quality is going way up. I grade them more, more hard, more harshly because I know they have assistance, but they're way better than they were 3 years ago.

That's such a fascinating attitude to take because I hear from other educators, especially like one of my cousins who's teaching high school, that their students' writing is pretty deplorable. And like you, I note, hmm, my students' writing is terrific because I know that they're all using ChatGPT. But I take that as a bad sign because it indicates that maybe they're not developing personally the way that I would like them to be developing, but I suppose they're just developing in a very different way.

I agree there's a bad side to it. I think we need to restructure all of our education so there's quite a high percentage of time spent writing in the classroom with no aids. So you learn how to write as a way of learning how to think. We don't really do that. We focus on assignments. And I think that utterly has to end. There should be writing on your own. We should like 5x that more. And then writing with GPT, we should teach that too. And we've gotta do both right now in this way. It's a halfway house where we don't really do either.

You've recently told universities that they should devote chunks of their curricula to teaching students how to write and work and learn with AI?

Of course. And I do this in my class. It's one reason why I feel I'm still adding value.

How do you do it?

I teach them prompting. I teach them strengths and weaknesses of different models. I teach them when hallucinations are most likely. I teach them how to stay current on a sector that, as you know, changes very rapidly. Uh, I just teach them, you know, different tricks.

What's very funny to me is that this is so clearly highly valuable.

Yeah.

And I've never considered doing it and nobody has ever suggested that I do it, or—

And I say to my students, is anyone else teaching you this? Everyone says no. Like it's not even mentioned. It's crazy.

I guess I'm a Luddite and I, I'm kind of hostile, even though I begrudgingly use ChatGPT today with The Sound and the Fury. It was kind of like an experiment. But I feel this hostility toward AI. Maybe it is deep in my subconscious, this resentment that it might be coming for my future career or coming for my friend as a graphic designer. But I do see how willfully ignoring it and not trying to learn how to work with it could be in the long term very detrimental, like an ostrich sticking their head in the sand.

Average is over, like I said, right? What side of the distribution do you want to be on?

You must have thought about this a lot more than I have. I know that food is a big part of your life. It's a big part of my life. Have you thought much about the impact of AI on the food world?

I don't think it's that much in the short run.

In the same way like housing, it's just not an obvious avenue for AI deployment.

In the short run, food will be modestly more expensive because energy costs are up. So in the short run, there's a penalty on food. Some more distant long run, we'll have more robust crops and seeds and genetically modified this, that, and the other. Food will be more nutritious. We'll improve GLP-1 drugs so you'll be able to So eat irresponsibly, not pay the penalty. A lot of long-run effects, but they seem somewhat distant to me.

The GLP-1 drugs. I'm curious about that. Have you read or are you aware specifically of people who are trying to improve them with AI?

You know, AI is not very well defined. If you mean tools of computational biology, This is happening all the time, just like mRNA vaccines are designed with computational biology, but not with LLMs. And this is already underway. And supposedly next year there'll be a version of GLP-1 that's just a pill. For a lot of people, that will be less painful. In other ways, it will be better, minimize side effects. Like, is it correlated with thyroid cancer? We're not sure yet. We'll learn that, probably come up with a way to fix that if it's true. So a lot of improvements can be had, but I think they'll all take time.

Maybe I'm not up on this, but the GLP-1 drugs are the ones that inhibit appetite, right? Or suppress appetite.

That's one way of putting it. They do much more than that, but that's a true statement. Do they inhibit compulsion or do they inhibit appetite? Like some of the drugs, there's two different channels. One is that anticipatory pleasures more generally become weaker. Another is that your stomach feels like it's fuller for longer. And there's two different channels. So it's complicated, but some version of what you're saying is correct.

It doesn't sound then to me like these drugs help people enjoy food. Maybe that's not what you said. Maybe I misheard or am misrecalling what you said. It sounds like people are just eating less food. I'm wondering if these new variants or new drugs would actually enable people to eat more food and not worry about gaining weight. So I used to have a show that was called Robinson Eats where every day I would eat a pint of ice cream on the show. So I'm, I'm very in favor of eating copious quants of quantities of food, and I would love it if there were some drug assisted by the aid of computation that were available to help me eat more without suffering the consequences.

I don't know what we'll get, but I can say with the drugs we have now, I mean, you're currently thin, no issue for you, but say you were older and not inclined to be so thin, you could eat all the ice cream you would want and you wouldn't put on weight. That's the case now. Will you be able to eat more of everything? That's much less clear.

But why ice cream?

Well, it— anything you could eat more of, anything you really craved without putting on weight. And how is that status quo? Because you eat less of other things. It may or may not be good for you, but it won't make you fat.

Yeah, but if you wanted to eat like 10,000 calories of ice cream a day, it would make you fat.

Well, you couldn't eat 10,000 calories, but you could have a big serving of ice cream every day and still lose weight.

Yes, that is, that is true.

Yeah.

But I guess I, in a fantasy world, I mean, I always, I've heard that there was once this product called Olestra. Does that sound, I've heard of it. I think it was some artificial fat that was added to food that you wouldn't digest. It ended up just being like a laxative, so it didn't work. But I, if, if AI is going to do one thing for us, it would be helping to develop a really good synthetic sugar and fat that they could put in ice cream and other foods that we could just eat without obtaining calories from. That would be my dream ask.

Someday. But when you think about processes of approval, discovering it, Again, it seems like 20 years away or more.

I could wait 20 years for that.

You will.

That'd be great. 20, 20 more years of working out wouldn't be too terrible.

Yeah.

Okay. Tyler, thank you. This, this was a really, a much more fast-paced conversation than I'm used to, but we covered a lot of ground and it was really a pleasure to have you on. So thanks a lot for joining me.

Great chatting with you. Let me know how things go.