Tyler Cowen Is Bullish on AI Education — Here's Why

If you take the very best AIs, do they know more than human lawyers? Can they, on many topics, reason better than human lawyers? Can they find citations better than human lawyers? Can they do it at much, much, much lower cost and more quickly? The answers there are basically yes.

So I decided, in the spirit of what Tyler's going to talk about today, to ask AI to help me basically write the introduction to Tyler Cowen today. So I did not write a single word of what I'm going to read right now. This is 100% AI. I basically asked ChatGPT, help me write an introduction to Tyler Cowen, our first speaker of the year. I'm the president of UATX. So help me build this thing up. So let's go. All right, good afternoon, everyone. It is my great pleasure as president of UATX to welcome you to a conversation with someone who has shaped how an entire generation thinks about economics, culture, innovation, and frankly, about thinking itself. Tyler Cowen is not just one of the world's most influential economists, he is one of the world's greatest, great intellectual energizers. Through Marginal Revolution, Conversations with Tyler, his books, and his extraordinary philanthropic work with Emergent Ventures, Tyler has modeled what it means to live a life of curiosity, discipline, and radical open-mindedness. At UATX, we talk often about building courageous thinkers and leaders. Tyler embodies this ethos. He has spent his career asking the questions others overlook, challenging consensus when it grows stale, and championing talent wherever it appears. In an age of conformity, he has shown how independent thought can still move the world. He's also, let's be honest, one of the most voracious cultural omnivores alive. Whether the topic is inflation, Indian classical music, artificial intelligence, economic growth, or where to find the best dumplings in Queens, Tyler has something original, vivid, and often surprising to say. More importantly, he invites us all into the intellectual adventure with him. And that, I think, is why having Tyler here is so fitting for UATX. This university was founded on the conviction that ideas matter, that truth-seeking requires both rigor and courage, and that students flourish when they engage with minds that expand the boundaries of what is possible. Tyler Cowen's work exemplifies all of that and more. So before I actually welcome him, I asked one more question. I said, OK, this is pretty good. I liked it. Do you like it? I think it was pretty good what AI said here. So then I asked ChatGPT, OK, so if you read Margin Revolution, you recognize this. Is Tyler over or underrated? In almost every meaningful sense, Tyler is underrated. And here's why. This is all ChatGPT again. His influence is far larger than most people realize, 1. 2, Emergent Ventures alone makes him underrated. 3, he's one of the few truly independent thinkers. And 4, his intellectual range is insane. So without further ado, please welcome Tyler Cowen to UATF.

I'm very pleased and honored to be here, uh, this evening. A few of you may know I'm on the advisory board of University of Austin, which I also consider an honor. I don't feel I give very much advice, or even that I should, This is my first time here. I've been here since 3:30 PM. Everyone has been super nice. Two things that struck me: first is just how amazing the students are. The second, as I remarked to Carlos, I love how you have your space set up. So I teach on a big campus, which is necessary given my school and its 40,000 students, but here you're all tightly packed in. I used the phrase, you know, cheek to jowl. It reminded me of, like, very early trading firms on Wall Street where you want to talk to people, ask them questions. They're just there, you grab them. You know, I saw a chess set before. There's all this comfy furniture. There's spaces where 2 people can meet, where 5 people can meet, maybe where 8 people can meet. It just struck me very positively, the physical space here, which maybe has evolved spontaneously more than it's been planned. But if you're curious, like, what are my first impressions? Those are the two things I would say. One thing I've been thinking about really a great deal since my book, I think it was 2013, called The Average Is Over. It's how artificial intelligence will shape the world. Since GPT-4 came out several years ago, people are now much more interested in that topic than they were in 2013. But nonetheless, we're now well into the AI revolution, and I just thought I would give my current thoughts on how it will affect most of all education. And I also see that as possibly having bearing on the university here. So when I was meeting with the students before, and again, a few of you were there, someone asked me, like, who are the groups that will gain the most and lose the most? I think the gains will considerably outweigh the losses. But what struck me is that if I were looking for the set of people that at least potentially will lose the most, it's actually the lot of you in front of me. Not all of you, and not necessarily. I believe in volunteerism. That is, it is up to us. All of you potentially can be huge gainers. But if I think about the first order income and wealth distribution effects of AI, they will take a lot of jobs and just make those jobs very different. Not worse jobs, maybe not lower-paying jobs, but very different jobs. So I think the simplest example is simply legal jobs, where much of this has already happened. Legal associates, some number of years ago, simply compiled a lot of information. About different legal questions. Even without AI, simply the internet and other developments, but certainly with AI, those jobs are very different now. You can ask an AI, it will tell you many things. There are some technical and legal problems if a law firm poses a confidential question and sends it to an AI. They're actually very reluctant to do that. So a lot of these changes have not been fully realized as of the moment. But if you simply ask the question, if you take the very best AIs, do they know more than human lawyers? Can they, on many topics, reason better than human lawyers? Can they find citations better than human lawyers? Can they do it at much, much, much lower cost? And more quickly? The answers there are basically yes. I know many people not convinced by this claim. I won't debate it much here. I'll just say a few things. First, most people do not use the best and most expensive models. Maybe they shouldn't. Maybe the best model right now costs $2,500 a year. That would be GPT-5.2 Pro for, you know, for legal questions. But even that aside, some of the major companies have models they haven't released yet. They're not cost effective. But if you're judging how good this stuff is by what you see in front of your eyes, it's a big mistake. So Brendan Foody, he's co-founder of Mercore, an AI company. I did a podcast with him. It was published yesterday. In his estimation, he grades and evaluates all these models. That's what his company does. He thinks they're improving about 30% a year. So if you know how compound rates of return work, economists sometimes use like a rule of 70, roughly. Uh, rule of 70 means that if you're growing at 1% a year, you know, you'll double your efficacy after 70 years. If you're growing at 30% a year, well, you'll double your efficacy in less than 2.5 years. So the models get much better very quickly because these returns are compounding. We're not sure for how long, but it's already the case that on most exams, the very best models, which may or may not be released, beat humans in most areas. The models also have high EQ. You know, they're very good at emotions. If you give the models medical diagnosis, They pass the medical exams. They pass the bar exams. On diagnostic tests, when graded by human doctors, the models do slightly better than human doctors. If I run a kind of imaginary economics competition between a very good model and myself, uh, I do not win. I might win on some number of questions, but I don't come close to even beating it on, say, a third of the questions. So this is the world we live in. It is a world where the AI diffuses very slowly. So you can be under a set of illusions as to how good the models are because most parts of our life have not changed that much yet. There's a famous saying, it's cited by Bill Gates, it may even come from Bill Gates, but it's something like, people overestimate how much things will change in the short run. It's very slow, and they underestimate a lot how much things will change over, like, a 10 to 20 year time horizon. And that's the world we're in. So I'm 63 years old. I assure you, I don't have a care in the world. If I have to, I can just, as they say, run out the clock. It's not what I plan on doing. It's not what I'm doing now. But if you're a student here, I suspect you're not 63 years old. You have a long career ahead of you. Furthermore, advances in biomedical science quite possibly will mean you die of old age and not cancer and not a heart attack. So really, I think a lot of you in this room will live to age 97. So in the meantime, that's a lot of years where you have to figure things out. Now, I think in the United States, at any point in time, there's a set of jobs waiting for people These are people who come from good families. I call them the upper upper middle class. I'm not talking about the billionaires, but you're from somewhere like Connecticut. Like, do we have Connecticut people at U Austin? I think we do. I met a few Connecticut people. Or you're from certain parts of California or New Jersey or like suburban Maryland, and you're probably not going to be a billionaire, but you have like good parents that can send you to the best school. I know here it's tuition-free, but these other Schools, they're not tuition-free, right? Parents have to have a lot of money. Your parents are pretty well connected. And there's the expectation you get a job in consulting, in law, in finance, maybe now in tech, or you can fall back, be a programmer, and you too are gonna join the upper upper middle class. That's how things have been. It was a great life for people like, from the kinds of families I suspect many, but not all of you, are from here. And in my opinion— I stress the word opinion— but I think those days are rapidly coming to an end. It's happened first in programming, where now probably 80% of the work is done by AIs. I'm not saying you can dispense with humans or that we shouldn't train humans, but if 80% of the work is being done by AIs, you've got to worry a lot about elasticity of demand. How much work will there be? Well, that's going to vary, but the jobs of the future are for programmers who can work with the AI. In law firms, an example I already mentioned, what will really matter is, can you manage an AI that is itself a repository of legal knowledge and legal reasoning? And that's just a very different thing. Who is good at working with the AIs? Who is good at managing the AIs? The simple reality— again, this is my opinion— but is the people who used to be good at the old things are just a different set of people, partly, from the people who will be good at the new things. The people who will be good at the new things are people who are good at teaching themselves. They don't have to come from good families. They're probably not people who grew up expecting, "I'll be a law partner someday and make my $1.3 million." and have a house in this particular suburb. They're people who are quite ambitious and quite dynamic and quite flexible. And since AI changes a lot and changes rapidly, they'll need to teach themselves these AI skills again and again and again. So people who are lifetime learners in a pretty demanding way that actually many of us will find oppressive. Uh, Again, I think there's some overlap with the people who are going to be successful, but those will be the newly successful people in most areas, those of you who can work with the AIs. Now, the good news is that if you decide to make yourself into that kind of person, you can be far more leveraged than was the case before. So the most successful companies can be bigger. It's just like how even simple email or the internet they allowed managerial empires to grow larger. Sam Altman, who runs OpenAI, he's forecaster, you know, he and I a few times have debated when will we have the first billion-dollar company with one employee? It's a little tricky, like, are there contractors, are there consultants? What if you have two employees? What if it's a husband and wife, whatever? But he and I both think within a few years, Sometimes he'll, if he's optimistic, he'll say within 1 year, we'll have a billion-dollar company with 1 person. I think morale really matters. You need 2 people. You know Midjourney, the image company? It's one of the best image companies. A lot of you played around with it. When they did their peak work to really do the breakthrough technology, they had 7 people. It's not 1 person, it's not 2, but 7. So we've already had a very big company that most of you have heard of and they did it with 7 people. So there will be many more new ventures. Startups will be centered around AI more and more. And people who are mentally or spiritually or temperamentally well-suited to doing new things and taking initiative will far outperform the upper-upper-middle-class kids who were going to work very hard and, like, do great at Brown or Cornell, and become the law partners. There'll be fewer of those people. The ones that are successful will manage these larger law empires and really earn a lot more, but there'll be fewer of them. And then the people who take these initiatives and who build new companies around AI will just be much wealthier, have more opportunity, more influence. And that, you know, in a nutshell, is how I think a lot of the world will change. So as that relates to all of you, I really don't know much about you as a group, but I know you're pretty elite. And it's going to affect you a lot is, I mean, my first core message. So I think very few sectors, if any, will be immune from this. If I look at AIs improving at the rate they're improving or even at their current level, I think it will take 10 to 20 years for them to really fundamentally change the world because institutions are sticky. If I go to my own university and I ask people, I've had these meetings, "Well, what are we doing to incorporate AI?" They'll say, "Well, we have a new cheating policy. We're gonna try to stop our students from cheating with AI." That might or might not be a good idea, but that's a lot of the extent of it. The notion of If you can cheat with the AI, maybe you need to be learning some different skills, and education itself needs to be restructured. People will pay lip service to it, but there's not many schools taking that very seriously, that if all institutions will be changing, that schools need to change also. So one reason I've, from the beginning, been interested in you, Austin, is like you're new and you're small and you have a board. I know some of the people on your board and they're very willing to just do things and experiment. And you are like an experiment, right? You're here. And if you chose to come here, you too are an experimenter. That's great. So like, I'm very bullish on that. But this is a place where whatever should be done with AI, it has a chance of actually being done. So in a given year, I give many talks at universities or go to many sessions or discussion groups, and I almost always feel like I'm talking to the students or the faculty, but I'm never talking to the institution. But here I feel I'm also talking to the institution, that there is some future we can imagine as somewhat different from the present that this institution you know, maybe at least partly can realize, will want to realize, can succeed in realizing. I go to Yale, I give a talk there, like, I'm just shouting into the wind. It's not even shouting into the wind, like, there's no shouting, there's no wind. I'm just talking to the individuals in the audience, which is fine, like, I'm happy to do that. But here's really different, and just don't forget how special you are, I would say, as an institution. And as individuals. That's to me really quite amazing that you exist at all and that you all chose to come here as faculty, you know, staff, students, whatever you may be. So I've been putting thought into how to change university education, and I'll start by outlining the radical description of my proposal, which is like clearly too radical even for here, but I think it's good to outline the radical version because it shakes people up a bit. So my radical version is that, again, I don't think every school should copy all the others. I still want there to be schools, like, maybe that have no AI at all. They ban AI, they ban the internet, you know, they ban the English language, everyone has to speak in French. We have private high schools like that. There's like a French private high school near D.C. People say it's great. It's all in French or mostly in French. The students also learn German. I think it's great we have that diversity. So I want a bunch of universities just to be different, as U-Austin is different. But at least for a bunch of schools, I think a third of the curriculum should be how to use AI. Not like one class, a third of the curriculum. Now, to be clear, that does not mean in this third you learn nothing else. So you learn how to use AI by applying AI to concrete issues. So someone asked me in one of the sessions before, well, do we still read Homer's Odyssey? One of my favorite books, by the way. Probably. But the way you would learn to use AI is to learn how to use AI to unlock some secrets of Homer's Odyssey, for instance. It doesn't have to be Homer. But just be clear, when I say a third is about using AI, you're learning a bunch of other things too. Whatever that institution decides is the priority. I'm hoping it's economics, but you know, whatever. So you're not taking away that third from other stuff, but what you're doing with that third is teaching people the mode of learning that will dominate human existence for the rest of your lives. Most schools I know of, and I visit a fair number each year, as far as I can tell, they're not doing this at all. There may be people doing it kind of privately with their students, but at the institutional level, schools are not turning a third or even a tenth of their curriculum into teaching people how to use AI. So I'd love to see a class, how to use AI to learn economics, how to use AI to learn Homer's Odyssey, to learn Shakespeare, again, whatever. With coding, this is already done. In fact, everyone now knows it's what you have to do. I'm just saying what's already done with coding, we should start doing kind of with everything. So a third, that's a big, big amount, a third. Faculty hate this. I'm faculty, by the way. The incentive for faculty is to do what they've been doing and teach the same way, and I think that has to change. So one reason this is a radical proposal is it goes against the interests of faculty. And furthermore, I see often that the students are better at AI than the faculty. Not always, and I think it's changing. I think the last year there's actually been really a lot of progress on the faculty side. But still, it's a big adjustment for faculty who have to stop doing what they've been doing, you know, maybe for decades. So there's another two-thirds that's left. Again, this should vary by school, but I— of that other two-thirds, I imagine a third just being writing, that you learn how to write to learn how to think. I'm not convinced the demand for human writers in the future will be that large. I'm sure it will be there. We want to read Proust and James Joyce. Maybe AI can never do that. We might want to read a columnist. I want to know what Bari Weiss thinks, whether or not I think the AI is more correct than she is. I wanna know what Barry Weiss thinks, so Barry Weiss will write the column or give the talk, whatever. You're listening to me, you're not listening to an AI, right? So that third will be writing, and I think some of that has to be writing locked in a room, kind of pen and paper. Some of it could be oral exam. You need to learn how to speak, of course. But it's a mistake to think we need that whole third to be everyone locked in a room a third of the time. If you think of this in terms of statistical sampling, you know, right now there's massive cheating. Not here, but everywhere else, right? Uh, and no one is doing that much about it. If you lock people in a room for, say, 3% of the time and test their pen and paper writing, and all the other papers are take-home, and then you compare the two grades, and if there's a discrepancy, you simply investigate more. You don't kick them out, you don't throw them in prison, but you investigate, maybe using AI. I'm not sure you can convict people with certainty, uh, but you would monitor the situation. And I think you can get the rate of cheating down a lot just by having an explicit program that compares what people do locked in a room to the other papers they hand in. You don't have to make it all locked in a room, you just do sampling. This won't eliminate cheating by any means,. But I think it will get it way down. This is what I advocate. I think it's still super important, whatever the skill may be. I'm talking about writing, but it could be arithmetic, could be all sorts of other skills. People need to have it, if only to learn how to think. And I still think there will be some human writing. Every now and, you know, I write columns for different media outlets, and I've tracked AI over the course of the last few years. And I never use AI to write the columns. I'm not allowed to, but I don't want to do it anyway. Like the writing for me is the fun part. But a few times I've like written a column and then when it's done, I'll go to the AI and say, you know, write a column on this. And a lot of the time it's better than what I did. I'm not saying I agree with it. Like by definition, I agree with what I said, but I look at it honestly, like who wrote the better column and say, eh, you know, I don't want to always make a final judgment. Sometimes the AI is kind of artificial sounding. But again, you can always train it more. It's improving rapidly. There's better still unreleased models that are not yet cost-effective to release. So like, I see the writing on the wall. One thing I've done personally, I go around much more, I travel much more, I meet people face-to-face much more, I give many more talks. Because the AI can't do that at all, right? You're not listening to me even on Zoom. You want me, the real me here, and I'm here. So I've at least 2x'd how much of that I do, and I write a bit less. That is me rationally responding, even in age 63, where I'm tenured, I'm pretty secure in a number of ways, don't have to do it, but I want to actually be, you know, add more value. ,and that means I should go around more, meet people more. It's also the case when you ask, like, what are future jobs going to be like? It's going to be based way more on who are your mentors, who can recommend you, what's your network like. Everyone could now submit a perfect application with a perfect cover letter, and if there's cheating, the grades also matter less. So how do companies hire people more and more? And CEOs tell me this, I have this conversation frequently, They look to who is personally recommended to them. So if I meet more people, I can help more people. Also, more people can help me by recommending me. So I just go around much more. So another thing universities should teach— they do a lot already, to be clear, even through things such as fraternities or football teams. I know you don't have all of that here, but just to a bit more explicitly teach people how to network How to find a mentor, how to be a mentor, how to be a good mentee. Super critical skills. You can bundle them in with other classes. That's fine. But I think all that needs to be taught a lot more. So there's like the one-third on how to use AI. The other two-thirds, a lot of that is writing, thinking, numeracy, statistics, and anything particular you want to learn. If you can be taught how to learn it through the AI, I mean, that's great. So I asked before, how many classes does U Austin offer? And someone estimated 35. I think that has to be roughly accurate. That's great for a new school, but if you ask the simple question, how many potential classes are there? I mean, you could say infinite, but practically speaking, it's somewhat less than that. But I still think it's thousands. Classes that someone might want to take. Like right now, maybe the answer is yes, but does U Austin offer a class on Tudor England? Does anyone know? No class on Tudor England. It's not that obscure a topic. England, important country. Protestant Reformation. Henry VIII. You know, there's Shakespeare coming up. It's a big deal. So say you want to offer a class on Tudor England. Just show of hands, you know, maybe you'll surprise me. How many of you here are just dying to take a class on Tudor England? Show of hands. That's pretty good. Like, maybe you can offer a class on Tudor England. But I can get a bit more obscure, and like maybe one hand will go up. And again, even in a much larger school, how are you going to accommodate that, right? There's overall declining enrollments, faculty cost more, tenure's a weird screwy system. And so on. So what I thought I would do is just come up with a simple model of how to offer a class in Tudor England. Let's have this come up on the screen. This is a very mainstream model. I need to scroll up here. I'm not sure this is the right way to teach an AI-generated course at all. This is just the way an AI-generated course would copy a regular course. You might think, well, if AI is so amazing, we should actually change the model more radically. And I'm sympathetic there. But just think, we can start with copying how a class is taught now. So I think we're almost at the top. I pulled out my trusty ChatGPT 5.2. This is paying for the pro model, $2,500 a year. But I tell you, if I had to pay 10 times as much, I would without hesitation. And I put in a prompt. I just asked it— if you now look at the top of your screen, we're done scrolling— please draw up a reading list for an advanced undergraduate class on Tudor England. Oops, sorry about that. It should consider political, economic, and historical issues of import, religion and the Reformation, It should cover the rise of the nation-state, the later rise of liberalism, and the rise of the West. This is all you, Austin, right? This is the right prompt, more or less. The reading list should focus on books and articles an undergraduate actually could access, and no books or textbooks at exorbitant prices. Prompting is very important as a skill. You want to put something like that in there, otherwise what you get is not that useful. It takes you too literally, in a sense. You want it to be things people can get at and don't cost them a fortune. So that was my prompt. It then gave me a long and detailed answer. Core books, primary sources students can access without paywalls, 12-topic reading list, primary reading, secondary readings, the court, the council, the king, the break with Rome, and so on. I sent this to a good friend of mine. She is a leading historian of Tudor England in Great Britain. And I asked her, like, what do you think? She used the phrase, astonishingly good. She's a very well-known author, historian of Tudor England. She was just astonished how good it was. But if you think this isn't good enough, or if you think it has errors, you know, just fact-check it through Claude, through Gemini. And make it better yet. So it took about half an hour of thinking to come up with this list. By the way, during that half an hour, I don't have to sit there watching it. Like, it took really no time. And you can scroll down. I'll send the people who put this talk out the link so you can all have the link. I say, please turn it into a ready-to-teach syllabus for a class of 14 weeks. I didn't ask for that up front. What it gave me was a bit sprawling. Assume students have JSTOR access. And then it does that. Tutor England. Logistics for 14 weeks. Each week the readings, lists learning objectives, required books, how the readings work, assignments, grading. Again, you can change this. You're not a prisoner to it. But if you imagine the poor students denied the class on Tutor England, They go to the AI and they create this for themselves. And then now I asked it next, you know, you have to grade people in some manner. We're just scrolling here, scrolling, scrolling, scrolling. Oh, how did I ask? Now draw up a multiple choice quiz, an essay quiz, and identify this quiz based on the readings. And it can grade your answers. And so on and so on and so on. Now, I'm not saying that we need no humans in this process. I think some students can do this with no humans. My guesstimate, my very subjective guesstimate, is that's like 15 to 20% of students, even at a very good school. Ah, but the role of the faculty now is just to be a coach. It's to teach students how to work through this process how to make the process work, how to come up with a manageable syllabus, how to grade oneself in a meaningful way. So the faculty person, I don't know if they can know nothing about Tudor England, but I feel like I could be that coach, and I'm far from an expert on Tudor England. It could be Mayan Mexico, something in chemistry. Again, so many topics. And what the coach does is direct, guide, motivate, tell the student when the student is somehow off course or not making any progress. But the actual teaching is through the AI, which is an infinitely patient textbook. Depending what plan the institution is buying, but at the margin, doing this costs zero, literally zero. There's some upfront cost. Which is still pretty modest. And voilà, you have an AI-generated class. Again, I think you can do a bit of this without any human teacher, but that's not my primary model. My primary model is some human assistance. You have a coach who checks in with you every 2 weeks, teaches you a bit more how to prompt. If you get stuck, they help you with something. If you can't find a reading, maybe it's not actually in JSTOR. They tell you what a good prompt might be to get some other reading. And as students do a few of these classes, the coach, I think, on average can do somewhat less because the students get how it works. Now, it's interesting to me, there was a paper published recently, I think about 3 weeks ago. There's a company called LearnLM. They take large language models and they spruce them up to make them even better tutors. So this is just basic GPT. But there are people building models that have built-in features to make tutoring with them easier. And there's a paper written by the LearnLM team, Google, and something called AEDI. And this was a randomized control trial done in the UK. So randomized control trials, they're not perfect, but they're pretty good. This is secondary schools. So 13 to 15-year-olds. So 13 to 15-year-olds can do a version of this. There's 165 students, and this is for math. I would say on average math is a somewhat less pleasant topic for people. Not everyone. A lot of people love math. But if it can do it with math, I think you're not picking the easiest case. It's not like, oh, it can teach people basket— actually, basket weaving's really hard too. But yeah, it does it with math. And the results are those who receive tutoring from human tutors alone, you know, we're doing well at 60%. Those who receive tutoring from the enhanced LLMs at 66% in a randomized control trial. So the people taught by the AIs are doing better. It's only one paper. It's only one study. You should never overreact to a single paper. But again, the models are getting better. Our knowledge of how to use them is getting better. I just think we're in a future where education will be very mixed. There's a lot of topics you can learn on your own. This model I laid out where it gives you a reading list, I suspect that's not even the best model. It's just a model we're all used to. We know how it works. So you can see that we can replicate it in some way. But a lot of these readings, it's often better to use the LLM directly than to do the reading. At least I think so. Not everyone may agree, but you can always have this fallback option, which is sort of copying what we do now. I suspect, and again, I use that word suspect in a very subjective sense, But ultimately what you want to do, again, with some coaching, is just let the individual humans loose on the AI and just let them do stuff. And the grade at the end is maybe not a letter or a number, but you just ask the AI at the end, well, what do you think? And it's a summary digest of analytic strengths, weaknesses, where they're good at reasoning, where they're not patient enough. And you can produce a 10-page version of that, a 1-page digest. You could turn it into a number or a letter. Maybe it depends like what the dean wants, what employers want. Like no one knows, right? But you can have a much richer evaluation scheme just by letting people do their thing. And I think the LLMs learn a lot about your work habits if you spend a lot of time with them. And you write the correct prompt and you ask them like, how can I improve my work habits? You know, don't be so fawning and positive about me. I know you're selling your own product, but I want some negative critical feedback. Where do you see weaknesses in my workflows? They will tell you. And it's amazing how right on they are if they have enough of your data. So again, the copying the class, that's great. It's like an intermediate step we can take. But longer run, it may just be you deal with these things and they produce reports on you. There's always an issue of cheating, as there is now. But I think what you need to do is spend some of the time with people, you know, locked in the cage or locked in the room, and they're using the LLM in the room, but you're making sure, like, they're not using Claude to manipulate GPT or using Gemini to manipulate Claude. If you get 3% of their output proven as genuine and you just ask the LLMs, the AIs themselves to compare, like on the rest of the stuff, is that really the person or are they using Claude? It's not a perfect method. And I'm not saying you send people to jail if there's a big difference, but I think it will deter cheating a lot more than what we're seeing under the status quo. So, I mean, that's my vision. My vision is we can teach people a lot more, a lot more diversity. We can do so at lower cost. I don't think faculty per se are losers, but they really have to change their roles in some, not all cases. Like, you might just be the writing instructor who keeps a job as a writing instructor, and you teach how to use the LLM to, like, grammar check, But you more or less do the same. But a lot more faculty have to become inspiring mentors, coaches, guides. They have to stay on top of AI, which is a big job. It changes quickly. It's not a you-learn-it-once kind of thing like calculus. You know, taking a derivative, you learn it once, you've learned it for your whole life. You just have to remember it. Learning AI, you have to keep on learning. And I don't mean once every 5 years. I mean, like, once every 3 months, things typically change a fair amount. That pace at some point will slow down, but right now we're in the middle of it. Uh, so that's what I'm thinking now. I'm just thinking no one really knows what's going to work. It's changing all the things it is touching now. No reason to think education is immune. We do need to keep human touch and motivation at the center of education, but a lot of what we now call learning is being revolutionized. When I teach myself things, like I run a podcast, it's my own thing. I'm Tyler, it's called Conversations with Tyler. Like, I can do what I want, no one tells me what to do. It's kind of, I'm not paid to do it, but it's like my profit center. It's my outreach, it's my career. It's a big part of what I do. So I wanna do it right. Before current AI, if I'm prepping for a guest, I might have to buy like 20 to 30 books and read them, which is A, costly, and B, slow. And then there's all these used copies coming on Amazon. Has it arrived yet? Do I trust the third-party distributor? You're keeping track. It's a pain. Now maybe I'll buy 4 to 5 books, read those books, start asking the AI questions. I can prep at about 2x the speed. I used to. I spend much less on books. I feel I'm learning more. I'm not sure the podcast is better, but it's not worse. That's like at least doubled my productivity, plus it saved me some money. So that's like my learning, which is internalized on me. And I see like I've doubled my productivity, saved myself some money. Like, it could be, well, that's only me and this kind of formula won't work for anyone else. Like, that's possible. But I guess, like, I just don't believe that. I think education can be different, should be different, will be different. The groups that will do it first, I think, are the startups. There are so few startups in higher ed. To me, that's depressing. As I said at the beginning, it's one reason why I'm so excited about the startups we have, this being one of them. So anyway, those are just my basic remarks. There's plenty of time for questions, but that's the vision I wanted to outline. I believe in it. I'm quite sure we don't know how to do it properly, but I also strongly believe we need to start experimenting. Thank you all for coming and listening.

So I wanted to ask about how you think that social signaling or status signaling will affect the progress of AI and the rate at which it takes over various institutions. I think it's possible maybe that people's preference for real art as opposed to AI art can partially be accounted for by a status signaling signaling that you know about this artist or that you can afford this particular artist's work. So do you think that there will possibly be more status associated with real education versus AI education, more granted to one and not the other? And kind of class division based on that?

It's a very good question. You know, there's a long run and there's a short run. So in the short run, to look at the corporate setting, you can easily imagine a corporate CEO who is unwilling to ask the AI what it thinks of his or her decisions. I get people, they write me emails all the time. In many cases, I've started writing back, "Ask the AI, it knows more than I do." But that's like me saying I have crummy status. I'm willing to do it. Like, I don't, in a sense, need them. They're not paying me. They're not doing me any favor. But maybe I lower my status by not being like the guru of answering their question about Tudor England, because frankly, I don't know the answer to their question about Tudor England. I might recommend the Lucy Wooding book called Tudor England, which I do know about, and it's a good book. But it's actually a better answer to just say, AskGPT, but make sure you're paying for the good model or ask Gemini, which now the good model is free for now. It may not stay free, but there's all these other further order effects. One might be that things done totally before AI take on higher status. I'm not sure that's true in the longer run. Like we have recorded music, which at first was lower status than live music. Maybe for classical music it's still lower status, but a lot of what people listen to is just recorded music through earbuds, and no one thinks that's terrible at all. It's not really lower status, it's the popular music. So I think medium term there'll be a big shift of status into AI. But there'll be this transitional period where things, people, institutions that use AI a lot, the rest of the world will feel threatened. They'll think like, eh, it'll be lower status. I'm reminded of the famous Peter Thiel comment The way that he— he was sort of half kidding when he put it this way, but he also meant it. Like, the way to do really well in life is to invest in things that are very low status. It was tongue-in-cheek. But you look at crypto, at first it was very low status. It's still highly mixed status, but it's far more established. I think AI will have this long and bumpy road ending in high status, but again, with many bumps along the way. That's one reason why schools aren't so keen to do it, because they worry about their status.

Thank you.

All right.

Hi.

I had two questions. Maybe I can ask you the other one afterwards. My main question is, like you mentioned, the students are already probably better at using AI than their teachers. I know lots of people here who are using it every single day in order to understand text, to do side projects. And frankly, like, using AI is not really that hard. Like, prompting is not an incredibly difficult skill. Sure, it takes some getting used to, but it's not something that, for instance, you would have to force students to spend a third of their time doing. From a libertarian perspective, people are already using it, making the free choice to use it and engage with it. Why would you force them to engage with it?

There's always better. And I think If it's simply prompting with questions, that's not so hard, as you say. But if it's organizing your workflows to truly take advantage of AI, like reorganizing an institution so that AI is more central and not just an add-on, that's very hard. I mean, you could say no one or hardly anyone knows how to do it. Some programmers are learning. I asked Sam Altman, this is on a public podcast recently, You know, when will it be that the finance department of OpenAI is mostly run by AIs? And he didn't know. And that's Sam Altman, like he's hardly bearish on AI, right? So if he doesn't know, that's significant. But those are the skills we need to teach people, not just prompting for simple queries, but how to actually use it to write, validate, referee research papers. I don't think that's simple at all. And teaching people how to stay on top of AI is not simple because it changes so quickly. So I maybe, you know, I say a third to be dramatic. If you want to bargain me down at 29%, I'll do it. But if you're like, oh, my university has a class on how to use AI, I'm like finger wagging, nope, you're not there. It's the big thing and it's really central to what humans will be doing for the rest of your lives.

Okay, thank you.

Thank you for your remarks. I run the Austin Union Society, which is the debate society here at UATX, and last year we had one of our founders, Sir Neil Ferguson, come give quite the opposite address that you gave.

Oh, I know Neil, and I've read his piece on this.

Yeah, that's right.

And, uh, well, I wanted to ask you, first of all, would you be willing to debate him? And second of all, would you say to his thesis that these large language models threaten to atrophy student minds, that the best way we can be prepared for employment is to work on ourselves as much as possible.

Well, I think if you devote one-third of the curriculum to writing and some other basic skills—debating could count, oral presentations could count—skills won't atrophy at all. In fact, we're freeing up faculty from a lot of rote presentation of information to actually getting people motivated to truly learn those basic skills, which they are not. Did you see that recent study like entering students for UC San Diego, the percent of them that had decent math, I forget the numbers, but even I was horrified. And I've been a professor my whole life. So I would gladly debate Neil. I'd gladly have the AI debate Neil. I think the AI would beat him too. But in part, it's because of atrophying that we need to direct faculty effort to truly making sure that actually a third of your time is things like writing, analyzing, debating, being graded both by humans and AIs. And we can do it much better, and we can do it all at lower cost. Thank you.

As you start automating so much of work because of so much of it being codified with AI, what's the one thing you think we're at risk of overlooking? And I think you kind of already touched on this, but how would you actually design schools So that children and students actually develop those critical skills to succeed with AI rather than just becoming dependent on it? Because a lot of the new AI schools are using it so much. And then there's kind of the question of like, how do you ensure they still have the judgment to actually discern the value of their ideas or like how pertinent it is to like serving humans?

Well, one nice thing about AI is you can test people along the way as much as you want at very low cost. This actually makes schools uncomfortable. Because if schools are failing, we can all learn that very quickly. So, I think the key is experimentation. No one knows exactly how it should work. Like, I'm giving you my vision, but I understand fully well we mainly want a lot of experiments. But be willing to test along the way and receive results that indicate my approach is failing. And it's very simple to do. What we're lacking is just the will to actually do it. Which is not really in the interest of schools. The interest of schools is to pass people, have grade inflation. Happy to see that's not how things are done here. But we need to go further and just measure all the time, are our methods working? But I believe if you truly devote a third, or more if you want, a faculty effort to those basic skills, and you test all the time, all the time, you'll get much better results than we've ever gotten in the past. OK, thank you.

Thank you for your talk. So the description of the teachers or professors just becoming coaches who don't even have to be experts in the topic, just coach you while you look to AI seems very lonely and very individualized. And how do you prevent that from just kind of intensifying the silo effect that we have in schools where everyone's learning their own thing and we're getting more and more separated from each other? And we can't actually have like a common intellectual basis to discuss things from.

I don't think it has to be that way. If a school wants, it can force people into larger units. But when I asked, "How many want a class on Tudor England?" I think I saw 7 or 8 hands. Whether or not you will hire a faculty member to do that, I don't know. But those 7 or 8 people can do the class together, learn from each other, be less lonely. I think you'll see a lot of groups of twos and threes and fives. If the group is 13, you know, just have a class, right? What number— at what number can you afford to have, not have a class will depend on how much money an institution has. But I think the school can facilitate people coordinating. Oh, here's 3 people, they want to do Mayan Mexico, 7 people who want to do Tudor England. And if it's so desired, you could actually have more interaction on a more regular basis. What people want, I'm not sure. A lot of people seem to have become these weird loners. I don't know how to fix that. You can't force them to do all this. Like all these people, they don't even go out on dates. You read the stuff, it's like, what has happened to these people? I don't have the answer to that. But I think at least the basic structure of AI education can be quite group-friendly. But if you really want to learn something that no one else cares about, I guess you're on your own. But look, you could have a WhatsApp group and you do it with 7 people across North America also. No reason why you can't do that. So there's always other stuff you can try. Thank you.

So it seems to me as the teachers that would become these coaches don't really need to know much about the subject that they had been teaching. So what's stopping students, for example, from becoming coaches in their own right?

Well, they do it all the time. I mean, I meet people and talk with people almost every day, and I mean young people. You know, I, I run this thing called Emergent Ventures. I have Zoom calls mostly with younger people. I'd say I have like 10 a week and a very strong preponderance of them. These are elite learners. They're not all from wealthy families. They might be from poor countries, but close to all of them do this now. So this is happening anyway. The real question is, should schools be part of it or should they be excluded from what Like smart people are finding the most effective way of learning. I say be part of it, but if some of what you do is give people the ability to do this on their own, well, maybe in the longer run, like a college education is 3 years instead of 4, I'm fine with that. Again, schools may not love it, revenue, I get all that, but I think it's probably an improvement and we would be headed that way in any case. And the second, third order effects are going to be significant. So, yeah, people can do this on their own. You know, what the coach needs, I think we need experimentation. But if you're teaching a class on Tudor England, I don't know that you have to be an expert on Anne Boleyn, but to have a general sense of what good historical reasoning is like, I do suspect is important. So people who have some training in history, some training in working with students, How we should train our trainers itself, they're more like coaches, you have to teach them that skill. You know, I became a professor, no one ever taught me how to teach. I have a PhD from Harvard, no one ever spent 3 seconds teaching me how to communicate, how to motivate, how to inspire, how to whatever, nothing. To some extent I was taught how to write research papers, and that is important, But the rest was at a zero, and that too should change. It's not really what I focused on in this talk, but, you know, second and third order changes, you want some, that people become a certain kind of generalist and they have this skill of like, what other people or sources should I put this student in touch with? That's a pretty intangible general skill, but I think that is becoming more important also. Thank you.

If you were to recommend 4 courses for people our age to take, like using AI, what would those be, like the most important ones?

Well, it really depends what you want to do. Can I ask you like what you think you want to do provisionally? I'm going to say economics no matter what, but that's just my bias. And when you ask a question, it's like you clearly ought to ask it, well, how would Tyler Cowen answer this question? No, I'm not joking. I mean, you can customize. You can have John Maynard Keynes teach you the class, Milton Friedman. I find a lot of AI models, they give you especially good answers when you do a compare and contrast. So if I ask what causes inflation, I get a good answer. But if I ask it, compare and contrast Friedman and Keynes and their theories of inflation, whichever you agree with, maybe you disagree with them both, but you get a much better answer. That's like a prompting trick. It depends on the model context. So yeah, like one thing I like about AI, this is just me being selfish, it knows me incredibly well. I've written an online blog for 23 years, totally open access. I guess I've done 600 podcasts, open access. I've written, I think, 18 books. It's read all or most of them. So there's like a whole group— I used to write a dining blog for the Northern Virginia area. There's a whole mini subgroup of people out there. They want to go out to eat. They say, like, "Oh, where are the— what's the taquería in Austin where Tyler Cowen would want to eat?" It's an excellent answer, at least for Tyler Cowen. I'm not saying you have the same taste, but think imaginatively about how you can use it. And filling in persona, you know, it could be Anthony Bourdain, it could be, well, a person from rural Mexico, a person from Mexico City. The more creative you are, like the more you learn about prompting and the better the food you will eat. So yeah, sky's the limit. But AI has been super kind to me because I've produced a lot of output over many years and there are not that many living humans that knows better than it knows me actually.

Thank you for your time today. My question is, which AI company will win the AI race in the next 5 years?

And why? I don't know. I visited the major companies multiple times. I'm very impressed by all of them. I think the models right now are roughly equal in quality. I would say this for images, Gemini is clearly the best. For what I call aesthetics, I prefer GPT, but that's a more subjective judgment. For tracking down data sources, I recommend Gemini. The best writer probably is Claude. I use GPT the most. It seems to know me the best and it has all this memory. You could say I'm a bit locked in, but I think the three main are very, very good and they're all good enough. I wouldn't agonize too much over which one to use. Now, who's going to win? I guess I think they're all going to do well, would be my prediction. But if, you know, one of them is bought up by one of the others, you know, it wouldn't shock me. That happens all the time. So I think they all have incredible pools of talent. And I know we're spending all this money on data centers, but big tech companies, their profits are still slightly, ever so slightly higher than their CapEx. So it's not being debt-funded at the level of those big companies. We'll see. Thank you.

Hello. My question tonight is, how should AI and LLMs be used to support a religious education? There are some students who choose to go into seminary or rabbinical school. Is AI a danger to this kind of education since religion is so deeply entwined with who we are as humans? Or is there a specific application that it can be used for in this type of education?

I would start by distinguishing religions that are text-intensive and those that are not. So Judaism, for instance, is quite text-intensive, and AI's I found super useful for unlocking, say, different readings of the Hebrew Bible. Christianity, obviously the New Testament is extremely important, Old Testament depends on what kind of Christian you are, but it's somewhat less textual than Judaism, but still super useful. And for Islam, all those religions should be much easier to learn. And there's one rabbi, he's a good friend of mine, he's just super impressed by how people are using AI to learn Judaism. And I've asked him a bunch of times, like, what do you think of this answer? He's a very well-known rabbi. He's written about 20 books, you know, on the Torah, and he's just blown away by how good it is. So if now what you're asking me is Buddhism going to benefit? Well, knowledge of the Buddhist texts will benefit. Can you use AI to make you a better meditator? Like American Buddhists love meditation. It's not that central to other forms of Buddhism. I just don't know. But yeah, my intuition is it will favor text-based religions over the others to some extent. But in absolute terms, it's much, much easier to learn about all of them. So, uh, should be good for religion. There's this other question: will some people strangely start treating the AIs as gods or deities or oracles? I think we're seeing that. Maybe that's a matter of concern. Uh, we'll see, but it wouldn't surprise me if it becomes more of a matter of concern. Thank you.

Hi. Thank you for your talk. I wanted to ask you if you think that AI will narrow or widen our world. And the reason I ask that is because the internet was supposed to expand our world so much, which in so many ways it did. It gives us access to so much knowledge and things that we would not have access to before. However, another thing that the internet did is that it siloed especially people of my generation. Instead of living in a common world, we live in a world of our own algorithms. We have— some people have described as a schizophrenic culture, infinitely fragmented and infinitely self-centered. And this made me think of that when you started talking about, you know, people trying to decide design their own university courses, for example. Because I think that's kind of almost antithetical to what a university is supposed to be, which is an institution which kind of gives you— puts into you the universal knowledge that has been passed down for generations. And almost like regardless of what you think the knowledge that you should learn is it gives it to you and it's trying to like build a world for you rather than you going into it, building your own world. And I think that the use of AI, although it can be very useful, at least, you know, as a very advanced search engine, it seems to me that it would further silo people because it's only responding off of the things that you think to ask. And the things that are in your own head. So how does that affect, I guess, our sense of reality or education?

There's a number of issues in what you raise. One question is just how bad a problem is this? But I would just say to the extent it's a problem, you can easily have a university here just say, well, if you wanna do a class on something, you need 3 people to work as a group. Fewer than 3 won't do. Set the number at 5, 4, I don't know. Again, experiment. And then it would bring people together. I would also say the third of the instruction on how to use AI, I would like to think we teach people how to use it to make friendships, human relationships. We never did that with the internet. Might not be easy. I'm not sure. Right now we don't even try to teach that. I want to see us try to teach that. I don't think it has to be isolating. I don't think the internet has to be isolating. Most of the people I've grown to know over the last 20 years, many of whom I'm now quite good friends with, or like modestly good friends with, I've met, I think it's possible I've met all of them through the internet, or close to all of them. So it's what you make of it. You can use it to bring people together. But if a university needs to nudge us into groups, you know, by all means let them do it. Thank you.

What can humans do that AI can't?

One view that I've heard a lot is that AI works purely input to output, whereas humans have this faculty of creativity, which is to create something out of nothing.

AIs can do that. You know, they're solving mathematical problems that humans hadn't solved before. They write poems that humans have not written before. There's no reason to think AIs cannot become much more creative yet. We're training them to be more creative. We grade them on creativity. I've listened a fair amount to AI music. I visited the company Suno not that long ago. I don't yet like that music. It's way better than it was at first. And I think a year from now there'll be a creative new AI song that I quite like. And I'll enjoy it. There's tricky issues involved with liking it too, I get that. But yeah, AIs are creative, no doubt. Just wait, they'll be more creative. Thank you. Dr.

Cowen, thank you for coming tonight. I think his previous question actually teased mine up relatively well. I'll play devil's advocate on the idea of AI being creative because ultimately I see it as somewhat of a manipulator of existing corpuses of text. So my question would be more so focused on the frontier of knowledge. Is there a concern to you of potentially running into an issue where AIs are trained on AI-generated knowledge and where it just becomes a recursive effect and potentially the output of, let's say, publishers or authors or whatnot just becomes more or less what an AI would generate anyway, or it would sound the same or whatnot. And so then these poems or creative works would just all kind of start to sound the same.

Well, insofar as that's true, there's plenty of room for humans to be creative. So if you imagine as a thought experiment, train the AIs on everything written before 1580. There's some people doing this, by the way. I haven't seen their results. But you have an AI with a mind like an AI would have in 1580. Can that AI then write the works of Shakespeare? While the experiment is not finished, I'm pretty sure the answer is no. Right? So there's this unique role for humans in this future, and that's great. Human outputs will compete more with AI outputs. Not everyone likes that. I think that's fine. But keep in mind, human creativity is putting together pieces. You look at a Shakespeare play and you read up on the sources it came from. Great to do that through the AIs, by the way. You know, Measure for Measure, Shakespeare had his sources and he manipulated them and he takes different things from different places. And there's a creative element. But like most humans aren't that creative to begin with. And a lot of human creativity is synthetic. And there are unique elements, and I think readers, listeners, whatever, will demand uniquely human elements forever and ever and ever, and that's a nice feature about our world. But, you know, can the AIs— you know, I think right now I would argue the AIs can write a poem as good as the median Pablo Neruda poem. Now, to be clear, Neruda wrote a fair amount. I don't think the median Pablo Neruda poem is great. I think it's okay. The AIs, in my judgment, are already there. No reason to think the AIs are going to stop getting better right away. So they'll write better poems yet. I want to read those. I am ultimately more interested in Tennyson and Shakespeare because they were humans, but I want to have this bigger mix and more choice, and I think I'm going to get it.

Thank you. Hi, thank you for your talk. I was very interested in the ideas you suggested with how to prevent cheating both with AI versus pen and paper and with AI versus, well, AI inside the smaller room. And I guess my only question was, where did you get the 3% number in terms of 3% of the work we do should be sort of cloistered, I guess?

Depends how many classes you have and how many tests you have. I just think of it as a problem in optimal sampling theory. So if someone is cheating, what percent of their real work do you need to look at to see the difference? I actually think it's very small. So say we're doing this class in Tudor, England, and someone cheated the whole thing using Claude. If you just pulled the person into a room and asked them one question, What do you know about Elizabeth I? My guess is with one question you could figure out if they're cheating. Not with certainty, not that you would kick them out of your school. Maybe if you punish them, you know, they could sue you, depends on the laws in your state. But you just sit them down and you'd say, look, like, you couldn't tell us anything about Elizabeth I, come on. And then let them try some other classes, and if this just keeps on happening, Well, you'd let them make the case. Oh, I just get so nervous. You know, if a human asks me, I can't answer, but I'm great in front of the machine. I mean, that's a life problem they need to deal with. And you're gonna have some cases like that, but I think you know very, very quickly. Again, even with like one question if it's well posed. So if you want 3% for more certainty, that's fine. But I'll say one good question, basically you'll know.