Tyler Cowen on the Risks and Impact of Artificial Intelligence 5/15/23

Today is April 19th, 2023, and my guest is Tyler Cowen of George Mason University with Alex Tabarrok. He blogs at Marginal Revolution. His podcast is Conversations with Tyler. This is his 17th appearance on the program. He was last here in August of 2022 talking about his book with Daniel Gross titled Talent. Today we're going to talk about artificial intelligence, and this is maybe the 9th episode on EconTalk about the topic. I think the first one was December of 2014 with Nicholas Bostrom. This may be the last one for a while or not. It is perhaps the most interesting development of our time. And Tyler is a great person to pull a lot of this together, together, as well as to provide a more optimistic perspective relative to some of our recent guests. Tyler, welcome back to EconTalk.

Happy to be here, Russ.

We're going to get your thoughts in a little while on whether our existence is at risk, a worry that a number of people have raised. Before we do that, let's assume that human beings survive and that we merely have ChatGPT-5 and whatever comes next to change the world. What do you see as some of the biggest impacts on the economy and elsewhere?

I think the closest historical analogies are probably the printing press and electricity. So The printing press enabled a much greater circulation of ideas, considerable advances in science. It gave voices to many more people. It really quite changed how we organize, store, and transmit knowledge. Now, most people would recognize the printing press was very much a good thing, but if you look at the broader history of the printing press, it is at least connected to a lot of developments that are highly disruptive. That could include the Protestant Reformation, possibly wars of religion. Just all the bad books that have come out between now and then, right, are in some way connected to the printing press. So major technological advances do tend to be disruptive. They bring highly significant benefits. Uh, the question is, how do you face up to them? Electricity would be another example. It has allowed people to produce greater destructive power But again, the positive side of electricity is highly evident, and it was very disruptive. It also put a fair number of people out of work. And nonetheless, we have to make a decision. Are we willing to tolerate major disruptions which have benefits much higher than costs? But the costs can be fairly high.

And this assumes that we survive. Uh, which would be a big cost if we— if it's not true. But just starting with that, with that and what we've seen in the last, shockingly, few months— we're not talking about the first 5 or 10 years of this, uh, innovation— where, where do you see its impact being the largest?

These would be my guesses, and I stress that word guesses. So every young person in the world who can afford, you know, the connection will have or has already access to an incredible interactive tutor to teach them almost anything, especially with the math plugins. Uh, that's just phenomenal. I think we genuinely don't know how many people will use it. It's a question of human discipline and conscientiousness, but it, it has to be millions of people, especially in poorer countries. And that is a, a very major impact. I think in the short to medium run, a lot of routine back office work will in essence be done by GPT models one way or another. And then medium term, I think a lot of organizations will find new ways of unsiloing their information, new ways of organizing, storing, and accessing their information. It will be a bit like the world of classic Star Trek, where Spock just goes to the computer, talks to it, and it tells him whatever he wants to know. Imagine if your university could do something like that. So that will be significant. Not that it will boost GDP to, you know, 30% growth a year, but it will be a very nice benefit that will make many institutions much more efficient. So in the shorter run, those are what I see as the major impacts.

I'll give you a few that I've been thinking about. You can agree or disagree.

I would add coding also, but this we already know, right? But sorry, go on.

Yeah, I was good— coding was my first one, and I based that on the astounded tweets that coders are tweeting where they say, I've been using ChatGPT now for 2 weeks and I'm 2 to 3 times as productive. I don't know if that's accurate. Let's presume what they mean is a lot more productive. And by that, I assume they mean I can solve problems that used to take me 2 or 3 times longer in a shorter period of time. And of course, that means at least in one dimension, fewer coders because you don't need as many. But it might mean more because it can do some things that are harder to do or were too expensive to do before. And now there'll be auxiliary activities surrounding it. So do you have any feel for how accurate that transformation is? Is it really true that it's, it's a game changer?

I've heard from many coders analyses very much like what you just cited to me. They also make the point it allows for more creative coding. So if a GPT model is doing the routine work, you can play around a lot more with new ideas. That leads to at least the possibility the demand for coders will go up, though coders of a very particular kind. Think of this as entering a world where everyone has 1,000 free research assistants. Now, plenty of people are not good at using that, and some number of people are, and some coders will be, some economists will be, but it will really change quite a bit who does well and who does not do well.

It's funny, I find this whole topic fascinating, as listeners probably have come to realize. I, I— it's probably the case that there are listeners to this conversation who have not tried ChatGPT yet. Just for those of you who haven't, in its current formation, in its current version that I have, I have the unpaid version. From OpenAI, uh, there's just a field where I put a query, a question, a comment. I want to give a couple examples for listeners, give them a feel for what, what it's capable of doing outside of coding. Uh, I wrote a poem recently about what it was like to take a 14-hour flight with a lot of small infants screaming, uh, and, and try to put a positive spin on it. I was pretty proud of the poem. I liked it. I posted it on Twitter. Uh, I asked ChatGPT to write a poem in the style of Dr. Seuss. Mine was not, uh, but in the style of Dr. Seuss about this issue. It was quite beautiful. Then I asked it to make it a little more intense, and it made a few mistakes I didn't like in language, but it got a little bit better in other ways. And then for fun, I asked it to write a poem about someone who is really annoyed at the baby. I wasn't annoyed. I thought— I tried to put a positive spin on the crying, uh, And it was really good at that. And, you know, of course you could argue that, you know, it takes away some of my humanity to outsource my poetry writing to this third party. But that's one thing it's really good at, is writing doggerel, rhyming, pretty, pretty entertaining and sometimes funny poetry. The other thing it's really good at is composing emails, requests for a job interview, a condolence note. I wrote— I asked it to write a condolence note just to see what it would come up with. You know, for a friend of mine has lost a loved one, write me a condolence note. It writes me 3 paragraphs. It's quite good. Uh, not maybe what I would have written exactly, but it took 3 seconds, so I really appreciated it. Then I said, make it more emotional, and it did. And then I said, take it up a notch, and it did. And it's really extraordinary. So one of the aspects of this I think that's, that's important— I don't know how transformative it will be. For— but for people whose native language is not English, uh, and I assume it'll eventually— maybe it already does talk in other languages. I use it in English. It's extremely helpful to avoid embarrassment, as long as you're careful to realize it does make stuff up. So you do have to be careful, uh, in that area. I am under the impression it's going to be very powerful in medicine in terms of diagnose— diagnoses and And we thought this before when we were talking about, say, radiology. There was this view that radiologists in the United States would lose work because radiologists in India, say, could read the X-rays. That hasn't, as far as I know, taken off. But I have a feeling that ChatGPT as a medical diagnostic tool is going to be not unimportant. The next thing I would mention, and I'll let you comment, the next thing I would mention is all kinds of various kinds of writing, which are the condolence note or the job interview request is just an example. I met a technical writer recently who said, I assume my job's going to be gone in a few months. I'm playing with how ChatGPT might make me a better technical writer because otherwise I think I'm going to be in trouble. And of course, then there's content creation, something we talked about at some length with Eric Hall. Content creation in general on the web, especially for businesses, is going to get a lot less expensive. Not going to be very interesting in the short run. We'll see what it's capable of in the medium term, but the ability to create content is now exploded, and those of us who try to specialize in creating content may be a little less valuable, or we'll have to try different things. What are your thoughts on those issues?

Just a few points. First, I have heard it can already handle at least 50 languages, presumably with more to come. One of many uses for this is just to preserve languages that may be dying. Or histories, or to create simulated economies of ways of life that are vanishing. There's a recent paper out on medical diagnosis where they ask human doctors and then GPT, and they give it a bunch of symptoms, report from the patient, and then there's a GPT answer and a human doctor answer, and the human doctors do the grading and GPT does slightly better. And that's right now. You could imagine a lot more specialized training on additional databases that could make it better yet. So we tend to think about, you know, America, or in your case also Israel, but think about all the doctor-poor parts of the world, including China, which is now of course wealthier but really has a pretty small number of doctors, very weak healthcare infrastructure, obviously many parts of Africa. Uh, it's really a game changer to have a diagnostic instrument that seems to be at least as good as human doctors in the United States. So the possibilities on the positive side really are phenomenal. Oh, by the way, you must get the paid version. It's better than the free version. It's only $20 a month.

Yeah, I thought about it.

Best investment you can make.

I thought about it, except I didn't want to advance the destruction of humanity yet. I wanted to think about a few more episodes. So maybe at the end of our conversation, Tyler, I'll upgrade. Uh, the other thing to say about diagnostics, of course, is that, you know, what happens now when you don't feel well depends where you live and how badly you feel, um, how poorly you're feeling. So, uh, here in Israel, you know, I can get appointment anytime. I don't pay, it's already included in everything. I can get a phone appointment, I can get a, a time to see my doctor, uh, and it's not a long wait. So at least for me so far. In America, there are a lot of times I thought, you know, I'd like to see a doctor about this, but I, I think it's probably nothing. And so I'm going to just hope it's okay because I don't want to pay the, the fees. And I get on the web, I poke around. And of course, most, most of us have symptoms every day that are correlated with all kinds of horrific conditions. So people who are hypochondriacs are constantly anxious and their main role of their doctor— and this is me sometimes— is to say, you're fine. That's— we pay lots for that. It's a wonderful, not unimportant thing. If ChatGPT or some form of AI diagnostic could reassure us that what we have is indigestion and not a heart attack, because it's not just looking at the symptoms and looking at simple correlations, but has some actual— knows what questions to ask the way a doctor would and the follow-ups and can really do a much better job. That is a game changer for personal comfort. And especially, as you point out, for places where you don't have access to doctors for whatever reason in easy or inexpensive form. I have a friend in England who says— I was telling him about some issue they're having, and I say, "Have you gone to the doctor?" "What's the point? They're just gonna say come back until it's, you know, until it's an open wound or until you pass out." But for— if you think about that, this is not an unimportant part of the human condition in 2023. Is anxiety about one's health. And I think the potential to have a doc in your poc— a doctor in your pocket— is really extraordinary.

Yes, but as you know, legal and regulatory issues have arisen already, and we have to be willing to tolerate imperfection, realizing it's certainly better than Google for medical queries, and it can be better than a human doctor, and especially for people who don't have access. And how we will handle those imperfections is obviously a major open question.

That's an excellent point, but I would add one more thing that I think is quite important, uh, and this is something you learn as your parents get older and they go to the doctor. You tell them what they, what they should ask the doctor and they forget or they don't know how to follow up. And so even if, uh, some of these, some apps that we might imagine will not be, uh, easily approved, the ability to have access to something that helps you ask good questions, which I think ChatGPT would be quite good at— I have these symptoms, What should I ask my doctor? What else should I be thinking about? Just gloriously wonderful.

Absolutely.

What do you think about this issue of content creation, and, and do you think there's any chance that, that we're going to be able to, uh, require or culturally find ways to identify whether something is ChatGPT or not?

Oh, I think GPT models will be— already are sufficiently good that if you ask them for content the right way, it's very hard to identify where it came from. If you just ask it straight up questions, what causes inflation, you can spot a GPT-derived answer even without software. But again, with smarter prompts, we're already at the point where, you know, I sometimes say the age of homework is over. We need to get used to that. And for a school system, you know, some of the East Asian systems that put even more stress on homework than the US system, they're going to have to reorganize in some fundamental way. And being good at a certain kind of rote learning will be worth much less in labor markets. So the idea that that's what you're testing for will no longer be as— I'm not sure how meritocratic it ever was, but it can be fairly meritocratic. But it won't be anymore, and it will be hard for many nations to adjust to that.

Yeah, I view that as, you know, a lot of people are anxious about the impact on teaching and grading, essays, exams. I think it's fabulous.

I agree.

I, I think, you know, the age of homework is a bad age. So if you're right, uh, I think that's a pretty much unalloyed benefit. Other than math, I think the, uh, It may end up that we, that we do our math homework in class, uh, where we can't use, you know, secretly use ChatGPT to help us answer, and we use home for something else, you know, some something like that.

But our educational institutions are often the slowest to adapt, and you as president of a university, you must be facing real decisions, right?

Oh yes, Tyler, it's, it's so, it's such a, it's such a burden No, we're not, because we're a small seminar place. We don't have lectures. There's no way that, that the papers and essays that our students write could be done by ChatGPT, at least at anything remotely like the current level. Oh, are you—

I wouldn't be so sure about that. It's not that GPT can write the whole paper, but in my law and literature class now, I've required my students to write one paper with GPT, but then they augment, they edit, they shape. Those have been very good papers. So you're getting— you're going to get work like that now. Yeah, that's true.

And I don't have any problem— what?

You're going to have to make a decision. Do you allow it? Do you recognize it? What are the transparency requirements? Do you grade it differently? This is now. This is not next year.

Yeah, no, I, I— my view on this, and I, I— it sounds like it's very similar to yours. Let's start with the condolence note. Okay, so I write a friend a condolence note. Uh, you, by the way, people talk about putting a watermark on ChatGPT. That's not useful. I'll just recopy it. It's silly for in this kind of setting. Maybe in a 40-page paper, maybe. But, uh, so I, I write a condolence, condolence note to a friend, say, and, uh, and I go through various iterations that I mentioned earlier, and I picked the one that I think sounds most like me. Is there anything wrong with that?

I think it's fine, but to the extent that intersects with institutions for certifying people, ranking people, assigning, you know, different slots in universities and awards to people, it does mean a lot of other practices are going to have to change, and they'll have to change from institutions that are typically pretty sticky.

But surely whether my friends think I'm an actually empathetic person might even be more important than whether I certify someone as skilled in economics. I think the— I think the— there is something lost when I outsource a condolence note. You know, I mentioned it briefly elsewhere, I don't think on the program, but the play Cyrano de Bergerac by Edmund Rostand is— that's what that's about. It's about a person who's gorgeous, a young man who's gorgeous, who falls in love with a beautiful woman, and he's inarticulate. And he gets a very unattractive person to whisper the sweet nothings into his ear that he can pass on as if they were his own. And that turns out not to have the best set of outcomes. A beautiful play, by the way. If you haven't seen it, it's been adapted in movie form, uh, in various, various ways, uh, and they're, they're all pretty good.

How you behave in real life—

sorry, go on, say that again.

How you behave in real life might matter more. So how you behave in textual life, anyone can now fake. So your charisma, your looks, how well you express empathy, probably those premiums will rise. And again, that will require a lot of social adjustment.

That's very well said. I think, yeah, the fact that probably get to a point where we can adjust those too, you know, with the way my eyes look and how much I smile and, you know, who knows. But certainly for a while there will be a premium put on authentic face-to-face interaction that can't be, uh, that can't be faked. And of course, when you write a book or an essay, forget being graded, when you write a book, I, I don't know about you, Tyler, I ask friends for comments. And you know what, I take them sometimes, and I thank them, just like people I think will for a while maybe thank ChatGPT. But it's not— is it that much different? I, I that, that you run your draft through a ChatGPT version or, and then augment it, change it?

ChatGPT gives me good comments as well. But again, I do think there's a genuine issue of transparency. If someone is hiring you to write something, what are they getting? What requirements are they imposing on you? What is it you need to tell them? I don't use GPT to write, you know, say, columns. It just seems wrong to me, even though it might work fine. I think I shouldn't do it, that readers are, you know, reading for like the Tyler Cowen. And well, there's all these other inputs from books, other people's blog posts, and the input from GPT is for the moment somehow different. That's arbitrary, but that's the world we're living in.

Well, I'm not going to name the columnist, but One columnist recently wrote a piece I thought could have been written by ChatGPT. It read like a parody of this person's normal writing. And of course, while I am interested in the real Tyler Cowen, sometimes the real Tyler Cowen is actually doing natural ChatGPT on his old columns. Not you personally, of course, Tyler, but, but I think a lot of columnists get in a rut and it's a lot to see what happens there.

I have the mental habit now when I read a column I think to myself, what GPT level is that column written at? Like, oh, that's a 3.5 or oh, that's a 4.0 occasionally. So, oh, you know, maybe that's a 6 or 7, but a lot of it is below a 4, frankly.

Yeah.

Even if I agree with it and it's entirely correct, it's like, eh, 3.4 for that one.

Yeah. Well, that's why there will be a premium, I think, for some time on novelty, creativity, to the extent that ChatGPT struggles with that. It's pretty it's pretty somewhat sterile still right now. So we'll see. It's going to get better at some point. It may be very soon. We'll talk about that too in a little bit. Let's turn to the— is there anything in this what we've talked about so far that you would regulate, try to stop, slow down, or we just say all full steam ahead?

I think that's too broad a question. So I think we need regulatory responses in particular areas, but I don't think we should set up like the regulatory body for AI. That to regulate a thing doesn't work well. Modular regulation as the world changes, that in turn needs to change. So if, say, a GPT model is prescribing medicines, which is not the case now, not legally, that needs to be regulated in some manner. We may not know how to do it, but the thing to do is to change the regulations for prescribing medicines, however you might wish to change those. That to me makes more sense than some, you know, meta-body like regulating GPT. So I think the questions have to be narrowed down to talk about them.

Do you think there's any role for norms? Now, you, you just confessed to a norm that you would feel guilty— and I'm trusting you on this, Tyler. For all I know, you've written your last 18 columns with ChatGPT. But, but is there any norm? Is there any role for norms to emerge that constrain chat, that constrain AI in various imaginable ways? I can't imagine someone saying, well, I could do that with ChatGPT, but it probably is a right, so I won't do it. And that would be one way in which— and not just that, but I could develop a version of ChatGPT that could do X, Y, Z, but I don't think humanity is ready for that. That seems a little bit harder for people to do. Do you think there'll be some norms around this that will, that will constrain it in some way?

Oh, there's, there's so many norms already. And to be clear, I've told my editor in writing that I don't use GPT to write my columns, uh, just to make that clear. Here's one example. There are people using dating apps where the texting or the content passed back and forth is generated by GPT. I'm not aware of any law against that. It's hard to believe there could be one since GPT models are so new for this purpose. But it seems to me wrong. There are norms against it that when you meet the partner you've been texting with, they'll figure this out. They ought to hold it against you. I hope that norm stays strong enough that most people don't do this. But of course, there's going to be slippage. Getting back to Cyrano, right?

Yeah, yeah. It's like people being honest about what their age is online. It seems to be a norm that it's okay to not tell the truth. But I don't know, when you uncover that, it's a pretty unpleasant surprise, I think, for some people. Well, let's turn to the issue of so-called alignment and safety. We recently had Eliezer Yudkowsky on the program. He is very worried, as I'm sure you know, about AI. You seem to be less so. Why do you think that is?

Well, let me first start with the terminological matter. Everyone uses the phrase alignment, and sometimes I use the word as well, but to me it suggests a social welfare function approach to the problem, that there's one idea of social good, as if you might take that from Benthamite utilitarianism. And that you want, you know, the programs, the machines all aligned with that notion of social good. Now, I know full well that if you read, you know, LessWrong, Effective Altruism, Alignment Forums, plenty of people will recognize that is not the case. But I'm worried that we're embodying in our linguistic practices as a norm this word that points people in the Kenneth Arrow, Jeremy Bentham direction Oh, everything needs to be aligned with some notion of the good. It's a— instead, it's about decentralization, checks and balances, mobilizing decentralized knowledge, that Hayek and Polanyi should be at the center of the discussion. And they're all about what are the incentives. It's not about information and knowledge controlling everything. But again, it's about how the incentives of decentralized agents are changed. And too much of the discourse now is not in that framework. But I mean, here would be my initial response to Eliezer. You know, I've been inviting people who share his view simply to join the discourse. So they have the sense, oh, we've been writing up these concerns for 20 years and no one listens to us. My view is quite different. You know, I put out a call and asked a lot of people I know, well-informed people, is there any actual mathematical model of this process of how the world is supposed to end. So if you look, say, at COVID or climate change fears, in both cases, there are, there are many models you can look at, including, and then models with data. I'm not saying you have to like those models, but the point is there's something you look at and then you make up your mind whether or not you like those models, and then they're tested against data. So when it comes to AGI and existential risk, it turns out, as best I can ascertain, and, you know, the 20 years or so we've been talking about this seriously, there isn't a single model done, period, flat out. So, you know, I don't think any idea should be dismissed. I've just been inviting those individuals to actually join the discourse of science. Show us your models, let us see their assumptions. And let's talk about those. The practice instead is to write these very long pieces online, which just stack arguments vertically and raise the level of anxiety. It's a bad practice in virtually any theory of risk communication. And then for some individuals, at the end of it all, you scream, "The world is going to end!" Other people come away, "Oh, the chance is 30% that the world will end, the chance is 80% that the world will end." A lot of people have come out and basically wanted to get rid of the US Constitution, uh, get rid of free speech, get rid of provisions against unreasonable search and seizure without a warrant, based on something that hasn't even been modeled yet. So their mental model is so much, we're the insiders, we're the experts, no one is talking us out of their fears. My mental model is there's a thing, science, Try to publish this stuff in journals, try to model it, put it out there. You know, we'll talk to you. I don't want to dismiss anyone's worries. But when I talk to people, you know, say, who work in governments, who are well aware of the very pessimistic arguments, they're just flat out not convinced for the most part. And I don't think the worriers are taking seriously the fact they haven't really joined the dialogue yet. Now, on top of that, I would add the point, I think they're radically overestimating the value of intelligence. If we go back, as I mentioned before, to Hayek and Polanyi, pure intelligence is not worth as much as many people think. There's this philosophy of scientism that Hayek criticized. And the people who are most worried, as I see them, they tend to be hyper-rationalistic. They tend to be scientists. They tend not to be very Hayekian or Smithian. They emphasize sheer brainpower over prudence. And I think if you take this more Adam Smith-like Hayekian worldview, you will be less worried. But, you know, we still ought to recognize the costs of major technological transitions as we observe them in history. They are indeed very high. I do not want to have a Pollyanna-ish attitude about this.

That's very well said. Um, I want to start with the point you made about modeling. I, I don't demand a mathematical model. I don't think you—

I do, to be clear. But go ahead.

You do or you don't? I do.

And again, I'm not saying the model will be good. I don't know. But if it's not good, that's one of the things I want to know. So the COVID models, I would say they weren't very good, but I'm delighted people produced them because we all got to see that. Again, opinions may differ, but that's part of the point of modeling. Not that models explain everything. That's like the bad defense of modeling. The good defense is let's see how this is going to go.

Well, let me put on my— channel my inner Eliezer Yudkowsky, which may not be easy for me, but I'll do my best. I think his argument— he has a few arguments, but one of them, the one I find most interesting, I did not find it completely compelling. And I would not call it a model. I would call it a story. And I find stories interesting. They can help you understand things. They can lead you astray, just like a model can. It's mathematical. But his story is, is that in the course of what we would— you and I would call emergent processes. Example he uses are— the example he uses is the creation of, say, a flint, a flint ax, a hand ax that natural selection works on inexorably on, on the fact that people who make better ones are going to leave more genes behind. And so that relentlessly, and for that technology, pushes it to improve. And no one's asking the technology to improve. There's no designer other than perhaps God, but there's no human force or will to push that process. It's the natural incentives of an emergent process. His claim then is that out of that will come other desires akin to our desire to go to the moon, which has nothing to do with directly with leaving more genetic copies behind or improving our fitness. It's just something that comes along for the ride. And his claim is that as intelligence grows in these models, these— and this is what I think is what is called the Orthogonality Thesis. I like the word orthogonal. I can say it. Orthogonality is much harder. But the Orthogonality Thesis is that you're going to get these emergent goals for these intelligences, whether they're conscious or sentient is irrelevant. Now, that's a story. It's not a model. Be really hard to model it. It'd be really hard to model it in a way that would be, I think, compelling for, for a skeptic. But I think it's worth taking seriously. Or do you disagree?

No, we absolutely should take it seriously. I think we should try to build those models. Think of an ecological population model where species co-evolve. It could end up like, you know, me and my dog, where we get along great. Humans select for dogs that do their bidding. At the same time, the dog is pretty good at manipulating you, right? You feed it, you take it for walks, you spend time petting it. I'm not convinced that's the outcome, but it seems to me at least as likely an outcome. There are checks and balances in the system in the short run, at the very least, maybe the medium and long term as well. There's a lot of pressures for humans to recommend systems that make them happy. It could be the real danger is a kind of Aldous Huxley world where you're just entranced by your AI because it's so nice to you. Again, I'm not sure that's the case, but that seems to me more likely than the scenario of it killing us. I would just make a general point about risk communications. The way you get better outcomes when you talk about issues is to give people positive, constructive steps they can do in the short run, not panic them, and not turn an issue into something that will be polarized. A lot of the doomers, it seems to me, they do the opposite of that. They should be much more constructive. Teach us how to flex our regulatory muscles in sound ways consistent with the US Constitution. I would be very enthusiastic about that. But also, I mean, maybe this argument's a bit of an ad hominem, but I do take seriously the fact that when I ask these individuals questions like, Have you maxed out on your credit cards? Or are you short the market? Are you long volatility? Very rarely are they. So I think when we're framing this discourse, I think some of it can maybe be better understood as a particular kind of religious discourse in a secular age, filling in for religious worries that are no longer seen as fully legitimate to talk about. And I think that perspective also has to be on the table. To sort of deconstruct the discourse and understand it as something that doesn't actually seem that geared to producing constructive action or even credibly consistent action on the part of its proponents. Another general point I would make— here's another thing we need to consider: if AI is truly very powerful, is the chance that it will save us from other existential risks higher than the chance it will kill us? Now, I genuinely don't have a good idea how to even go about answering that question, but the doomers are so convinced the chance it will kill us is much higher than the chance it will, say, save us from an asteroid. Again, that, that just feels to me like the pessimistic doom is something they want to build into the view. I don't think the substantive or even modeled or empirical arguments for seeing one is larger than the other are very strong. So I just think we need to step back and say to people, like, wake up, don't be entranced by the emotionality and mood affiliation of this discourse. We need to look for sane, sober, constructive step-by-step things we can do in the short term, not polarize the issue and think more constructively.

Well, I'm gonna come back to that probability it'll save us as well as possibly harm us. We'll return to that in a little bit, but a couple thoughts. Uh, one of my favorite moments really in the history of EconTalk is when I realized that on air— that I think it was on air, certainly said it on air— that Nicholas Bostrom's view of artificial intelligence is no different than the medieval, uh, vision of God, the Aristotelian vision: all-powerful, omnipotent, omniscient, will know what you're going to think next, can manipulate you in ways— can't that are, that are, uh, and make you do its bidding. Um, uh, he didn't like— I asked him anyone ever mentioned that, he said no. He was rather taken aback by it, is in my memory. You can go back to listen to that 2014 episode. But I think that I'm not accusing anyone of looking for alternatives to religion, but I do think there's a theological aspect to this akin to religion for how people— the role this can— we can think about this in our lives. It certainly is a end-of-days kind of discussion. I'm not accusing anyone of, of any kind of irrational thinking or at all, but I do think it has things in common with it, as you're alluding to. The thing I would— a couple things I mentioned. One is, uh, this is a G-rated program, but when you mention that, uh, people may find it comforting to, uh, talk to ChatGPT, I do think that application is going to be quite, pronounced and, uh, extensive. The people turn to it for solace of all kinds. And finally, you mentioned the dog— when you mentioned the dog analogy, I think it's an opportunity for OpenAI or others, instead of naming their things like Bing and Sydney and, and Bard, they should name it Fido and Rex and Lassie. That will encourage people to think they have a, a sort of, uh, co— I don't know, cooperative relationship with their with their virtual pet?

I think as a, you know, a general point of view, I found it very interesting to speak with clergy, priests, rabbis, and also national security people about AGI existential risk. They're both groups of people that have a lot of practice thinking about existential risk. And they take it seriously. They don't dismiss it. In a sense, you could say it's their profession. But they've also heard a lot over long periods of time about existential risk. And I find they frame it in saner ways. They definitely do not have Pollyanna-ish worldviews, but they are able to, you know, step off down from the ledge of doom and recognize these are arguments that have recurred throughout human history on a very regular basis. There are these very serious risks, but there's also a millenarian tendency in human thought that tends to rise in volatile times. And we have to think about how to deal with those millenarian tendencies. They are themselves a form of risk, right? That our reaction to an event can be worse than the event itself, even if the event involves some very high costs.

Yeah.

So when I see commentators arguing, well, we need to have the government take control and monitor everyone's computer drive and have the option of shutting it down, That was in a recent blog post. You know, based on something that hasn't even been modeled yet. It's not just that I disagree, but I think there's something about how the discourse has proceeded where you need someone to just raise his hand and shout a bit to everyone. Something has gone wrong here. You all need to wake up. And my own role in this debate, I think, largely has been that.

Well, that's also what the other side's yelling. Something's going wrong here. You need to wake up. And you're saying, why don't you make the case a little more persuasively for me? I'm not an idiot. Do a better job convincing me and do it in ways that we would, as you say, follow the, the norms of discourse that have been productive in the past for improving our safety, our security, our well-being, and so on.

Right. Right. So if someone says, well, here's my proposal for how we should let or not let GPT systems prescribe medicine, one proposal for the US, maybe a different proposal for a poorer country. I mean, I'm absolutely all ears and that will actually help our ability to address the larger issues. If you just kind of start off with, oh, here's this big AGI thing and in 17 years it's going to kill us all. That is one of the best ways you can imagine not to make progress on a problem. In fact, it's moving backwards.

Yeah, it certainly induces apathy rather than, than zealousness in trying to cope with it. But it's going to be a fascinating time. I would mention, I think it was yesterday, Uh, kind of spoiling all the fun, uh, Sam Altman of OpenAI announced that, well, we've kind of exhausted all the benefits from this larger dataset strategy. You know, it's everybody's mind, it's, wow, look how much better ChatGPT— we need a better name for it, by the way— uh, ChatGPT, uh, 4 is better than 3, and then 5 is so much better than 4. Oh my gosh, think what 14 is going to be like. And the answer is kind of like, well, we're stuck at 5 for a while, folks. We're gonna have to find some new techniques. So there is a tendency to think that the rocket taking off, the pace of improvement is therefore like Moore's Law to be just extrapolated endlessly into the future when not only will it write a great condolence note, but fill in the blank. I don't think that's, it may take a little longer than people were thinking.

As you know, iPhones, at least for a while, have plateaued. That was some while ago. I don't personally know whether that's the case with GPT models and LLM models more generally. There is the advantage now that they're good enough, in essence, to grade themselves, right? But some of the issues are, well, if right now it gets 100th or 99th percentile on a math test, you know, with plugins, well, it can't get 110%, right? So then you bump into the question, well, how creative can they be? And they are somewhat creative already, but that may be a very different problem than can they get 100% on the test. So I don't know, but I would just say whatever anyone says from a, you know, a company, I wouldn't take it face value. They may not know, their competitors may do something they haven't foreseen. This is the market process. There is competition in place in a very real way. And I certainly don't rule out very rapid and very startling advances even from our current place.

Do you have any unease about the profit motive in this process, or does that give you comfort?

Well, I'm not sure it always is the profit motive, for one thing. So I don't want to, you know, offer comments on particular companies, but very large companies often behave as bureaucracies, and they're not profit maximizing, and they face regulatory constraints which limit often the good and also the bad sides of profit maximizing. So we're not seeing profit maximizing right now. And on top of that, OpenAI is a nonprofit. Now it's affiliated with the for-profit. It's a very complicated structure. Uh, I don't pretend to understand how it works, but I would just say I'm not sure how to model the current, uh, AI competition in the United States. And I think it's a very complex problem people should work on much more. Now, if you ask what do I think of the regulatory process, who or what exactly in government— and I do mean exactly, I would like to hear much more on this question from the doomsters— who's capable of making the problem better rather than worse? Is it the FTC? In my opinion, no. All their doom saying about, you know, concentration in social media clearly has been shown to be wrong, and they're still at it, trying to take down those companies. One Biden proposal gave some authority to the Commerce Department. They may be good in the sense they might not do that much, but their expertise in this area seems to me highly limited. They tend to be quite mercantilist, looking for national champions. That might be good if you're relatively positive on these developments, but it doesn't actually seem like a good match is something substantive needed to be done. So I think we should focus much, much more on the question, who exactly should be doing whatever? And there's not such a good answer when you look at the details.

That'd be interesting to challenge the industry to create an advisory council of, um, people from within the industry, people outside. I would put you on it, Tyler, in a second. Uh, I might even put an Eliezer Yudkowsky on it. Uh, in a minute, because it would be good to have that voice where some of this conversation could take place, maybe in an organized fashion, uh, and, and consider guidelines. But of course, even that model leads sometimes to the worst kind of collusion and other things that we wouldn't, you know, restraints on innovation that might actually be harmful.

So, but I would favor an antitrust exemption to allow that if one were legally required. And I can see quite possibly that it might be. In the meantime, you know, I would like the people working on this issue as outside commentators to just state outright that they favor the First Amendment, that they favor constitutional protections against unreasonable search and seizure of your property, and just to see, you know, who in the debate is willing to make that move and who is really looking to undercut the United States Constitution because they've so talked themselves into a theory that is not validated that they're willing to do quite extreme things. And, you know, a group like the effective altruist people who I think have been an enormous benefit to discourse in almost every way, but we're now facing an issue where they have to choose. I mean, is there loyalty to the US Constitution or is it to a Benthamite utilitarian notion, you know, as would come out of their discussion boards. And I hope they choose the right way. Uh, so, you know, I'm willing to say I think the United States Constitution has proven more robust than the rationalistic arguments you see in LessWrong blog posts. It has held us back from any errors we might have committed out of excess enthusiasms or excess rationalism. And I would just like to see that reaffirmed as a framing of the debate. So far I'm seeing the opposite.

Well, if I remember correctly, and I think it was in our conversation or in a blog post, Eliezer, and I'm sure he's not alone, has suggested that, you know, we would have to monitor GPU, that is computing capacity, and places that exceeded that capacity, which would be suggesting that they're training, uh, models on larger and larger amounts of data that this kind of process requires would be subject to a war, you know, very— called—

yeah, data centers. So yeah, it's a violation of international law as I understand it. But again, I think that's illustrating the danger of people talking themselves into things on the basis of very long, vertically stacked arguments that are actually not resonating with most other people in the world. And again, we just need a wake-up call to have a reframing, not let emotion and mood affiliation get the best of us. Look for positive, constructive steps.

Just to make it clear, AGI, which you have mentioned, is artificial general intelligence, and LLM are large, large language models. Large language models, correct? Referring to the large datasets of trillions of words that these projects have been, have been trained on. Tell our listeners what, what you mean by mood affiliation.

Mood affiliation occurs when you become attached to a particular mood rather than a substantive conclusion based on models, data, and analysis. So there are people who just are pessimists, and they look for pessimistic interpretations of phenomena. There are people who are just optimists. That's also a fallacy. You're not choosing across moods. Your mood should be a conclusion. But people, when they're in moods, they find it hard to get out of those moods. So not to do your thinking based on a mood-first approach, which everyone will say they don't do. But in this debate, I'm seeing quite a good deal of it.

Let's close and talk about intelligence. It is, in my view, a suitcase word, a word meaning that people cram a lot of different things into it, and they— when they use the word, they don't always mean it the same way as someone else uses it. It can mean knows a lot of things. It can mean can solve problems. It can mean can take knowledge I've learned in one context and apply it in a different one. It can mean solves all problems effortlessly. And, and for me, when people say, uh, they treat intelligence like a scalar, a single number— IQ would be one measure that is a scalar, it's just a number that, you know, if your IQ is higher than mine, you're, quote, smarter than I am. Uh, which probably true, Tyler, but I don't think that's the measure— any of those measures. First of all, they're all different, and I don't think having more intelligence in the way that ChatGPT, say, 5 has than 3, solves all problems. Most of the problems of the human experience are not solvable. They involve trade-offs. I come back to our classic the, the dictum of our profession, uh, no solutions, only trade-offs. And trade-offs therefore require judgment, and ChatGPT will never provide that unless you believe in a social welfare kind of approach that you're— you alluded to earlier. And so I think the, um, the belief that smarter— quote, smarter and smarter, uh, computer tools will help us solve more and more problems is simply incorrect. It will solve many problems, you know, and some of them will be quite important, potentially, uh, avoiding an asteroid. But many of the problems of the human experience are not due to a lack of intelligence. And I think that's— I think that's an understanding that you and I are trained in our bones from being economists as both students and teachers over the years. And I think it's very alien to the computer science community. What's your reaction?

Absolutely. Everything is coded into a program in one worldview, in what you might call the Austrian economics worldview. Decentralized and often inarticulable knowledge, as outlined by Michael Polanyi, is critically important for just about everything that happens. GPT models do not in any direct way access that. They're not trained on it. There is a sense in which they digest a version of the outputs from the process based on decentralized and inarticulable knowledge. So they have a good deal of cognitive oomph. But the notion of AGI to me is not entirely well-defined. Getting back to this point of multidimensionality, it's not just prudence or wisdom or judgment, but even very crude notions— how you realize your intelligence through physical actions in the world in a way that, say, LeBron James does, a phenomenally smart human being in both the intellectual sense and also the physical sense, and integrating the two. That's yet another dimension we haven't considered. So this idea when people say, oh, AGI is X number of years away, that doesn't make sense to me. If your notion of AGI is a particular kind of smarts, you could argue we're already there. I mean, the GPT-4 has been measured at an IQ of 130, GRE scores, you know, very high, sometimes close or at perfect. But it's a device. And again, we need to keep the broader perspective. And I'm hoping that you in particular with your training in Hayek-Smith, don't let the doomers sell you the emotion. You know, the correct attitude here is to take lessons from history. History is one form of knowledge. It is not always generalizable. But there's a lot you can learn from it. And one thing we learned from it is that the options and possibilities for progress are true and real and can make our lives much better. But the disruptions are very true and real as well. Typically, The world doesn't end. The real problems of mankind tend to be foreign conquest, pandemics, war, environmental issues. That's likely to remain the case over the future. You have to ask, especially in a world with hostile powers, is American artificial intelligence likely to help with those or hurt with those? I think to have the intuition, you know, again, subject to critical scrutiny and discourse, that it's more likely to help with them is a better starting point than a lot of what I'm hearing. So just a lot more historical reasoning in the debate based on these broader frameworks that are not just computer science.

Well, I think it's a little bit tricky. I, I, I'm not sure history is— I think we should learn from history, and we should certainly invoke history, and we should certainly understand emergent things, but it is a little more complicated than that. I would not I don't think that my newgrown sympathy to the concerns here is an emotional response, but rather actually the role that emergence plays in some of the doomsday stories. It's intriguing to me. It's not— it's not persuasive. I find it thought-provoking. But I want to say something about your point. And try to avoid making it too personal for any players in this conversation beyond you and me. If I were stuck on a desert island with my survival at stake, put into a hostile environment, say in a place where, say, my kind, whatever kind that is, is not smiled upon, I think one of the last things I'd want is a high GRE person. I, I just— that using that as a measure of power, brain power, intellectual power, problem-solving power, just strikes me as absolutely wrong, absurd. It is an incredibly uninteresting measure. It's— sorry, there's an interesting aspect to it. It's kind of a dog-at-a-time legs kind of thing, that it can be done, that, that a computer program can pass a bar exam or get a good score on Brian Kaplan's econ exam. But it really is not so useful in the kind of context that human beings find themselves other than applying to college or doing well in college. And so part of that is your, your point about the physical world. I don't just want something to tell me smart things. I wanted to be able to act in those settings. But we understand that in 2023, much of life takes place outside of those physical settings. It takes place in our virtual world, in our imagination, online, in various different forms. But even there, I do not think the ability to do well on a standardized test is— or an IQ test of any kind— is remarkably— is remotely— excuse me— is remotely connected to intelligence. I just find that bizarre. Do you agree?

One of my predictions is that GPT models will raise the relative and indeed absolute wages of carpenters. So if you're a very young person, or if you have kids, and they're thinking, well, what, what should I do for a career now? Uh, you face some very serious decisions. Now, it might be that you're a wonderful general slash manager of GPT models and you do incredibly well managing your 1,000 research assistants. If so, great. But if you're just producing fairly routine word-based content in whatever form, you probably need to give a fairly serious rethink to your career plans. And I think there'll be a lot more science, a lot more ideas, a lot more projects. And like very good gardeners, very good carpenters, people with synthetic abilities who can make things happen, the kind of person who in a lab actually helps build the fusion reactor, say, rather than just writing about it, is already becoming much more valuable. And it's going to disrupt many things. So you and I are like broadly of the same generation. It won't change my life that much. Even for me, I'm not sure I'll write another book. My plan is to go around and actually give more talks. I think that will be more rewarding. Uh, it's not that I think GPT can write as good books right now, but I think people will be playing with their GPT models rather than reading my book at some point.

So, and they'd love to look into your eyes and shake your hand after a talk and chit-chat for a minute and a half and have a human experience than read another paragraph. If you possibly—

we all need to rethink what it is we're doing, reevaluate our professional lives, It's a very serious matter. It will disrupt many of us, often in good ways, sometimes in bad ways. But I would just here, you know, urge against complacency. Please take this seriously. Give it a rethink. Look for things you can do now to make this a better future rather than a worse one.

But I think it definitely puts a premium on human interaction, and that's going to be— people who are good at that, uh, whatever form that takes, I think will, will remain incredibly important. Uh, or we're all going to retreat into our bedrooms and play on our phones all day, which a lot of people are. But there's also backlash against it, even, even among young people. So I think, uh, I think that part is going to be extremely interesting. But it is going to force us to think about what it is to be human, which is not a bad thing. And, um, that's what I find myself thinking about when I think about that condolence note or interacting with an avatar of Tyler Cowen rather than the real thing. And of course, I'm doing this over Zoom with you, Tyler. It's It's an inferior form of interaction, but it's extraordinary that we can do it at all across 6,000 miles or 7,000, whatever it is. So I think that's pretty cool.

It will be a fascinating future, very weird in many ways. It may not feel weird to the people born into it, but we should all be ready for it. And I'm very glad you're doing these episodes on AI to help get us ready.

Uh, well, I appreciate your wisdom, which I'm very confident if I read the transcript is not ChatGPT. Uh, Tyler, thanks for being part of EconTalk.

Thank you, Russ.

This is EconTalk, part of the Library of Economics and Liberty. For more EconTalk, go to econtalk.org, where you can also comment on today's podcast and find links and readings related to today's conversation. The sound engineer for EconTalk is Rich Goyette. I'm your host, Russ Roberts. Thanks for listening. Talk to you on Monday.