Seth Benzell: What Economists Get Wrong About AI (And What They Get Right)

Ben Zweig: The economics of AI has become its own field with its own factions, its own networks, and its own blind spots. My guest today is Seth Benzell, Professor at Chapman University, co-host of the Justified Posteriors podcast, and one of the most original thinkers working on how AI reshapes growth, labor, and the structure of the economy. Here's our talk. Welcome to the Economics of Work. I'm excited to welcome Seth Benzell, Professor at Chapman University, and co-host of my favorite podcast outside of this one, Justified Posteriors. One of the best thinkers and communicators on the economics of AI. Seth, welcome to the show.

Seth Benzell: Ben, thanks so much for having me on. Your podcast is also my second favorite economics of AI podcast.

Ben Zweig: Fantastic. We'll take it.

Seth Benzell: I'm happy to join that fan club.

Ben Zweig: Nice. Top two. It's an elite group.

00:50 The factions of AI economics

Ben Zweig: I want to talk about the economics of AI community. That's your audience. I think to outsiders, economists care about AI, but as an insider, can you break it down? What are the factions? What are the communities? How do you segment the economics of AI community?

Seth Benzell: All right. I'm the dungeon master setting up the RPG, as you're going to align with different factions of the community and building up reputation points. Here we go. Yeah. One major community in economics of AI comes out of an NBER program called Economics of Digitization, which I now believe is called Economics of Digitization and AI. All right. Got to throw it in there. NBER, if your audience doesn't know, it's a big organization like the American Economic Association that organizes economics conferences, tries to promote economicsy things. They're important in this context mostly because they fund some good research and some datasets and they organize economists together. For, I want to say 20, 30 years, one of their subdivisions has been economics of digitization. Questions that they used to be interested in were questions around so-called e-commerce, used to be that selling something on the Internet was a different economic phenomenon than just selling something normally. Questions around digital distributions of products. Then increasingly that would be perfect.

Ben Zweig: Platforms and all that.

Seth Benzell: Those are 90s topics are like e-commerce, then it transitions into platforms topics and social media topics in 2000s, 2010s. Then in that community, there were always people who were thinking about computers as a productivity tool. But honestly, I would say most of that research would have been in labor economics land at that point, 2000s, 2010s.

Ben Zweig: Economics of digitization is distinct or was distinct from labor economics. Yeah.

Seth Benzell: I would say, if you had a paper about computers are substituting for routine tasks and complementing people in non-routine tasks, which is an economic literature pioneered by people like David Autor and others in the early 2000s. You were probably presenting that in a labor economics context rather than this economics of digitization context. Of course, that starts to slowly change over time. Then as AI comes in, AI is obviously the newest exciting thing of digitization. I would say that's community one. Community one is the community that was thinking about like Internet digital platform issues in economics and then comes to this topic. Then you have, as I was just introducing, you have this labor economics community and it's maybe macro labor more than thinking about very micro employment, unemployment issues that is interested in, they used to call it a skill-biased technical change. I always used to be confused, is it skill-biased technological change or skill-biased technical change?

Ben Zweig: I think it's technological change.

Seth Benzell: Dude, I've heard it both ways. I think there's an official way. I say it both ways interchangeably. Sometimes people call me on it. It's not 100 percent clear to me which way, in theory, should be the right way. There's a way people say it.

Ben Zweig: Yeah. Yeah.

Seth Benzell: Skill-biased technical change. That was a literature that was initially concerned with the fact that the college wage premium. There was a phenomenon that was observed that people with college educations got paid more. In your simplest model of how the economy works, it should be that the gap between people with college educations and people, and originally this literature was about high school educations, but then eventually we saturated high school education. College education, non-college education, you would think if they have some complementarity, as you get more college educated people, the wage gap should come down because you're alleviating the scarcity of college educated people.

Ben Zweig: Right.

Seth Benzell: But instead what we found in the 90s and 2000s is that even as we got more and more college educated people, the gap actually got bigger. A series of people, a literature evolved to try to understand what was going on here. That was called skill-biased technical change. We're changing what businesses are doing in some way that is favoring college educated workers. A big theory of that was always computers being used for productivity or industrial robots are also part of that story. Then in the 2000s, that literature evolved to be more about not just college educated versus not college educated, but specific skills. I come back to David Autor because he's the key figure in this SBTC literature. He is the guy who pushed the literature in the direction of, well, actually, it's not like college is magic, it's that being college educated is a sign that you can do these non-routine analytical tasks and computers automate for the routine analytical tasks. We've got more demand for the non-routine analytical tasks. That's a separate continuum of thought that is now feeding into the current economics of AI mishmash. You would consider that macro labor. If you're trying to explain the national wage gap, I would call that macro labor.

Ben Zweig: Okay, fine.

Seth Benzell: But you can call it labor.

Ben Zweig: Yeah, yeah.

Seth Benzell: Big tent. Yeah, big tent. Claudia Goldin. Anybody who's trying to think of that like education in society and the wage gap, I would call that macro labor. If you want to call it just regular labor, fine. I don't have a dog in that fight.

Ben Zweig: Yeah.

Seth Benzell: And then the final big community, and there are other complementary communities I'm going to come to. You've given me because I got to lay out this map of all the communities that are sometimes allies, sometimes rivals. The third big one would be kind of the macro community, macro growth, macro innovation that had always been, always had known that automation is one of the most powerful forces in economic growth. And there had been some papers in, I want to say the late 90s and early 2000s. I'm thinking in particular of an author by the name of Zeira, but other people as well thinking about how would you model a technological change that is skill biased in favor or rather like unskill biased, like biased in favor of capital rather than labor. You would also think of this as kind of being in the Romer growth land. So Romer's model of economic growth is one in which, sure, you get economic growth from accumulation of capital. Right. That's a Solow growth model. But you also have economic growth through the creation of new ideas and new ideas are non rival. So they're awesome. We can always just use more ideas. Every new company is a new idea. The different companies complement each other. So that was kind of one theory of economic growth. And in those models, there's no kind of sense of automation versus like labor favoring. Right. You set up a new business with the same production functions as everybody else. But now you make a vanilla ice cream instead of chocolate ice cream. Right. There's a kind of a generation. There's a in the years after that, people are starting to think more about, OK, what if there's a technology that's kind of favoring physical capital at the expense of labor? Zeira is one early paper there. Then I did some work there in the early 2010s trying to think about how would I think about automation as arising through data accumulating on workers successfully performing tasks. Right. So thinking a little bit ahead to this regime that we are in today around where automation comes from is from seeing humans perform tasks and somehow codifying that. And then I can talk about that research at some point, if you're interested. And then the after that, we get kind of what I would say kind of the defining paper, the paradigmatic paper in this agenda, which is Acemoglu and Restrepo, Race Between Man and Machine, which has both a mechanism where the macro economy and the scientists in it are looking at the price of labor, the price of capital, and either making a decision to innovate in an automation technology or making a decision to innovate in a new task to give to labor. So do we want to invent something that helps labor, invent something that helps capital? Well, the answer is going to have to do with the price of labor and the price of capital. Those I would say the big three communities, I'm happy to take some questions here or go to some complementary communities.

Ben Zweig: That's a good breakdown. I mean, yeah, if I'm trying to think of like, who are the big giants in each field? I would think the first one, this digitization is like Erik Brynjolfsson and co., the second ones like David Autor Claudia Goldin you mentioned, and then like the third, I would think like Romer, Chad Jones style, like so that's kind of like growthy growth people, you know? Okay, so that feels like a good landscape of like the origins of thought here.

Seth Benzell: In the first community I just wrote some names, I would throw like people like Avi Goldfarb, Josh Gans, right, also digitization, right?

Ben Zweig: Yeah, yeah, yeah. Yeah, do you think, yeah, like, we could probably play this game where like, We could play the name names. We could categorize them. Yeah, yeah. Well, I mean, what about you?

Seth Benzell:

11:19 Where Seth fits in

Ben Zweig: Yeah, well, are you, it sounds like, if someone asked me where you sit, I would, my first instinct would be in digitization, but then I would say maybe growth.

Seth Benzell: That's the answer.

Ben Zweig: But here you are on the Economics of Work podcast, so now you're labor, so.

Seth Benzell: I love it all, I love it all, I mean, so.

Ben Zweig: Rounding it out.

Seth Benzell: Yeah, my story is as I started off doing actually kind of macro public finance, right? So my PhD advisor, Larry Kotlikoff, check out his Substack, read his books.

Ben Zweig: Very good Substack, and also he has a good podcast.

Seth Benzell: He has a very good podcast, he just had Daron Acemoglu on the podcast.

Ben Zweig: Yeah, yeah.

Seth Benzell: And he's very concerned with issues around very, I would call them economic growth ideas, like why are some nations growing fast? Why are they able to support themselves through tax revenues? Why are they able to have generous welfare states? What makes welfare go up or down over time? But he thinks about those questions less through kind of the Chad Jones, Romer, the way we usually think of them as like very abstract stylized models, but rather really try to dig into the details of, let's write down all the thousand agents in this model, and see if we can learn something by having them simulate a economy with each other. So a very sort of simulationist framework, a framework that thinks the details are really important, a framework that doesn't start with, let's assume constant population growth, but actually is like, let's go to the UN demographic projections, and see how different it is if we're in the high regime versus the low regime for the welfare of 21 year old Estonians in the year 2052.

Ben Zweig: Yeah.

Seth Benzell: That's the kind of crunchiness he likes, and I take from him both kind of that love of simulation, the importance of thinking about the interaction of growth with fiscal policies and demographics, as well as I just love his sort of hedgehogginess. You know this distinction between hedgehogs and foxes?

Ben Zweig: Isaiah Berlin, like yeah, I guess, yeah, I mean, good at a lot of things. That's the hedgehog.

Seth Benzell: No, yeah, no, he's actually, that's the fox.

Ben Zweig: Oh, sorry, hedgehog is deep, and the fox is

Seth Benzell: He's like-- Hedgehog knows one big thing. So, Kotlikoff, I love hedgehogs. Kotlikoff is a hedgehog. He knows one big thing. He knows that the national debt is a bad measure of America's indebtedness, and that the deficit is a nonsense measure of America's indebtedness, and he knows that the way that you think about long-term fiscal policy is by setting up models that have explicit fiscal policies in them, and I love him, and I think the way that you become a great thinker is not by hanging out with a lot of mediocre foxes, it's by having a lot of mentors who are different kinds of hedgehogs. Anyway, so that's where I started. I worked with him, and we started working, kind of using those ideas to think about economics of automation, right? You didn't call it, in the early 2010s, you didn't call it AI, you called it automation, and AI was

Ben Zweig: Just-- Right, at that point, it was like, you sounded like a poser, if you used the term AI.

Seth Benzell: If it was AI, you mean the thing that doesn't work?

Ben Zweig: Yeah, yeah. Or it was like a placeholder for regression, you know?

Seth Benzell: Right, exactly, yeah, you would call it machine learning. Machine learning was the term of art. And we made some progress in that, and trying to think through some issues that I think are maybe a little neglected in the current macro growth AI conversation. I think we made our points, and we were a little bit ahead of our time. But I got the attention of Erik Brynjolfsson's group over at, he was at MIT at the time, and I kind of, after my PhD with Larry, ended up working at a postdoc in that group. So I like to think that I'm sort of in this unique position of having a foot in macro simulation fiscal policy, and economics of digitization. And as far as I know, I'm the only person at that weird intersection.

15:52 What each camp is focused on today

Ben Zweig: It's a good intersection to be at. Okay, so back to these like factions. So let's say, that's the historical origins. But like, if we think about where we are today-- So then, we've got different, now it seems like there's new kind of, yeah, I don't know how I characterize it. Let's say, there are some people, for example, that really are kind of bought into like super intelligence as like a thing, as something that we should be modeling. Others that are more on like, task displacement stuff, kind of more org-econy, ways of thinking about things. Then there's like, others who kind of think about, bottlenecks, there's different mindsets that kind of go into this, if you're an economist working on the economics of AI, there's like, I feel like there's some segmentation by like, what are the big questions to you? So like, if you had to take a stab at it, what are the flavors of thinkers, in terms of like, what they think is like, the thing to be thinking about right now?

Seth Benzell: Okay, great question. I think the economics of digitization people, and I think that community is also a little bit closer to the business side, right? So one variation riff on that community, I didn't quite mention was, even though a lot of those people are straight up economists, there's also some blending into that of people in the business school who might be in an information systems group, or in sort of an organizational science group. And I would say that's the group that's kind of closest to the businesses. And I would say one issue that they're really interested in is questions around job reorganization, organizational restructuring. In other words, how hard it is, how long does it take? What are the best ways to upgrade your business to take advantage of these new technologies, right? Organizational capital, right? So there's of course, friend Daniel Rock with Erik Brynjolfsson has his framing of the J curve, which is this idea that when a new technology comes around, you have to make investments in order to make your company take advantage. And so there may be what looks like a productivity dip before the productivity explosion, because those sort of unmeasured investments won't be picked up in GDP measures, right? GDP only picks up investments that are, I bought this factory, it doesn't pick up. The CEO spent half of his day thinking about how to reorganize the jobs, right? I would say that's big question number one for that community. Big question, and oh, I guess job reorganization is also a big question for the labor economics community. I would say those two guys are interested in that. The macro growth guys aren't so interested in job reorganization, right? If you look at the models like Acemoglu and Restrepo's Race Between Man and Machine, or Anton Korinek, or Chad Jones, those models, there's like a big lump of labor, right? And that's an exaggeration, but to the extent that it's not a lump of labor, it's like I show up and I've got like two units of analytical labor and three units of like routine labor, right? And it's all just kind of getting added up in the economy in a way that doesn't really care about whether the skill is within one individual or whether a job definition or an organizational definition, it's really at a much higher level of abstraction. Although one of the key parameters in those models is gonna end up being diffusion frictions, right? How hard it is it to spread the new technologies? So in some sense that micro question is inescapable.

Ben Zweig: Yeah.

Seth Benzell: A question that I think kind of cuts across all three of these communities, I think is a problem that, it's kind of a problem that none of these communities has like taken ownership of it. And to the extent that you see it, you see it kind of in antitrust land or industrial organization land, where you have to admit every part of economics is getting touched by AI these days, but I didn't list those communities in my top three.

Ben Zweig: Right.

Seth Benzell: But I think this is an important cross cutting question is questions around, maybe this is AI finance, is like who wins from the AI, right? Who's actually gonna make the money?

Ben Zweig: It's distributional questions.

Seth Benzell: Distributional questions, industrial organization questions of like, does it make sense for the labs to vertically integrate, right? Is that the only way they're gonna be profitable? Questions around what is competition between different models going to look like? My co-host Andrey Fradkin has some really fascinating work on this recently published in the Journal of Economic Perspectives, which is like the economics, like cool lit review journal, and also on our blog, Justified Posteriors, which all your guests should always go to. Definitely. Thinking about how does competition between models work with the end result, like that's gonna tell us something about is the big frontier lab gonna win it all and be worth $10 trillion company, and Dario Amodei is gonna be right, there's only gonna be one private company on earth and it's gonna be Anthropic, right? That's one extreme where the only thing you care about is who's got the smartest, smartest, smartest model, and then they win everything. And then you have got another extreme where it's like, no, being a little bit smarter doesn't help you that much. These open source models are a lot cheaper. In a world where open source models are good enough, the frontier labs don't make any money, and the money instead goes to the people who are making the chips and running the cloud infrastructure or maybe you think, well, actually, there's gonna be plenty of chips and that's a more competitive market people will enter, that'll get saturated. You have setting up a cloud provision, setting up your data center, that's gonna be a commodity. People are just gonna make their marginal costs. The real benefits, the real winners are gonna come from the people who are able to organizationally adapt and be the nimble incumbents that are able to take advantage of these new technologies. Or if you're like my friend and co-author, Andrew McAfee, you'll say, you know what, these big conglomerates, these lumbering Volkswagens, they're never gonna be able to adapt to the new technologies. The real winners are gonna be the superstar entrants that are able to explode in growth because they're gonna be able to take advantage of the cheap open models, the cheap chips that are commodities, and they're gonna disrupt the incumbents. So I think that's another big kind of cross cutting question.

23:15 AI believers vs. skeptics among economists

Ben Zweig: Yeah, yeah, that's a good lay of the land. I mean, so what do you think, yeah, let's say like, let's say outside of the economics of AI, like do you think there's a separating line between like the AI-pilled and like the non-believers to some degree? Yeah, and I'll give you a reason why I ask. I mean, I mentioned earlier, I was talking to Ben Castleman about this question and he had said, oh yeah, economists aren't really bought into AI. And I was like, what? Like everyone I know is bought into AI. And he's like, well, I guess we're talking to different economists. And I got to imagine that's true. I mean, I'm pretty much talking to like young labor people and like maybe he's just talking to like Greg Mankiw or something, like not to I don't know what Greg Mankiw believes. Well, like economists that are like less involved in this. And I don't know, do you think there's like a separation between like economists that have like embraced AI as like a thing that really matters and those that are just like, yeah, I'll believe it when I see it in the productivity statistics.

Seth Benzell: Interesting framing. I mean, I think there's definitely a bit of a phenomenon of people in economics of digitization selects for people who think digitization is exciting, right? I think that's kind of inevitable. I also think that sort of the Chad Jones, Paul Romer style of very abstract theorizing. Interestingly, this doesn't really affect Daron Acemoglu so much, but like certainly Anton Korinek is that crowd appeals to the highly, how would I put it? Like the super rational, right? This kind of we always talk about, what do they call it? Is the rationality movement, right? There's a certain kind of intellectual style that puts kind of the abstract reasoning as like the highest form of doing science.

Ben Zweig: Right, like they'll answer questions like, oh, what is the likelihood of extinction? And like, is that worth, is stuff like that?

Seth Benzell: For example, and Chad Jones writes those, I love Chad Jones. We're not trying to make fun of those papers.

Ben Zweig: Yeah, yeah.

Seth Benzell: But I think that intellectual style is attracts the kind of person who thinks that, oh, a thinking machine that can think as well as anybody, this is the most important thing fricking ever, right? Whereas I think someone who's kind of in the Autor sort of messy jobs, work, oh my God, businesses have to keep track of thousands of different things, Hayekian world. They are more sensitive to the idea that if you solve one problem, you still got 999 problems left, right?

Ben Zweig: Yeah.

Seth Benzell: So I would say, so that's my initial read is like the super abstract people love AI and the super digitized people love AI love in the sense of very optimistic about the economic impacts. Yeah. Whereas kind of maybe the more messy empiricists would tend to be a little bit more pessimistic. And then, but I think I would push back on the idea that like there's this giant divide amongst economists. Cause I think if you looked at economists versus like the general public versus San Francisco, the way that conversation works is like San Francisco thinks in the year 2040, we're going to be gods living in cyberspace, right? Yeah. And the normal person is like, I sure hope social security is still paying benefits in 2040, right?

Ben Zweig: Yeah.

Seth Benzell: So like that's like the social divide. And I think for economists, the divide is much more between like at the low end, I think this might mean a couple of fraction percentage points of productivity, like so Daron Acemoglu's The Simple Macroeconomics of AI, not to pick on him. Cause apparently he's getting picked on these days, but you just give him like a low ball estimate says I don't think we're going to get one additional percentage point of GDP growth over the next decade because of AI, right? And he's, he's since said that like you would update those numbers, presumably you would update them up. But like that's like on the low end, people are saying, well, what's 1% of international GDP? That's still tens of billions of dollars, right? So that's like the low end estimate. It's worth thinking about. Yeah, I mean, that's like an important thing, right? Whereas kind of the, I would say like the medium high end economist is thinking we might get one to three additional percentage points of economic growth a year. So maybe the kind of towards the more optimistic end of this range, if we've been averaging two and a half percentage points of GDP growth a year recently, maybe this doubles that to 5%, right? Which would be to an economist that's like astounding, mind blowing, world historical success.

Ben Zweig: I can't even fathom it.

Seth Benzell: Cannot, well, I mean, you can fathom it, right? So like what's another 50 percentage points of GDP between now and 2040? It's a lot, the world's a better place, but it's not the Flintstones versus the Jetsons. It's it's Denmark versus Ukraine, right? It's a difference, but it's not like we're in a different reality, right? Which is the San Francisco land. And now there are a subset of economists who are taking the super crazy, super crazy, I don't mean to be pejorative, the really unbounded prognostications seriously. People who come to mind are Anton Korinek, Phil Trammell, who are big friends of the show, who think there is a percentage chance to take seriously of singularity above 10% GDP growth rates a year world 15 years from now isn't 50% richer. It's unrecognizable, right?

Ben Zweig: Yeah.

Seth Benzell: So those people exist, but I guess my takeaway here is, yes, there is a divide, but I would say kind of, it's a lot narrower than within kind of the general conversation.

Ben Zweig: Yeah, yeah, that makes sense. I mean economists are involved in measurement and like have to write things down and like think carefully. So it makes sense that like the bounds would be narrower. Okay, good segue into this conversation about like bounds of belief.

29:58 Justified Posteriors and the shape of priors

Ben Zweig: So your podcast is called Justified Posteriors. So I want to talk about priors. Ooh. So like, all right, so you've got, so let's say you personally have a set of priors, Andrey has got a set of priors so you guys have like your own like distribution of priors. Do you think that is like, do you think your own priors and by your own, I mean yours and Andrey's, I guess, jointly, like, do you think those are kind of like representative or are you guys weirdos in your own way?

Seth Benzell: Are we weirdos within the economics of AI community? That's a great question.

Ben Zweig: Yeah.

Seth Benzell: We're definitely weirdos. Sure. Yeah. I guess that's what that's saying. Are we outliers, I guess, along this dimension of optimism about, about--

Ben Zweig: No, I guess I'm thinking that there's like a whole bunch of dimensionality. So like you could have like a whole distribution of priors, like a whole distribution of distributions where you have like beliefs in all these different facets of the importance of AI, what it means, how it affects this or that, like, like there's all these things we can have priors on. I mean, you have a different prior in every show. And I guess, I guess like, is the, are the shape of your beliefs roughly representative of like of the community or do you think you guys like are thinking outside of the bounds of like the general public in any way?

Seth Benzell: I see what you're getting at. Okay. I would say along a lot of dimensions, I would say we are taking normal economic logic and applying it to AI. And if we're, if we end up somewhere that other people aren't, it is the path that they would naturally follow if they had taken this path, right? Yeah. So for example, Andrey's work on thinking about decreasing marginal returns to intelligence is a super natural question and a super natural framing that an economist would bring. Right? This a model that is one unit of intelligence seems to be this useful to a business. I wonder if a model that is two units of intelligence is twice as useful, or we would usually guess there's diminishing marginal returns. And the fact that he's the first person that go out and try to measure that isn't, doesn't, it makes us weird because we're first, it doesn't make us weird because we're thinking about this in a way an economist wouldn't think about it. Right. It's a very economist way. The San Francisco way of thinking about it is twice as much intelligence, oh my gosh, singularity explosion, right? Once you reach some threshold, the end of the, like it's a very, it's a very economist way of thinking about it. Just the way I would say. Yeah. Then there are questions where I would say,

32:49 Can we direct technological change?

Seth Benzell: I think we have a view that is fair to say is a little bit different than I think some of the macro growth guys. I would say that one area we diverge is I think there's a little bit of over optimism in that community around the idea of directing technical change, technological change, right? So if you think about Daron has written about this, David Autor has written about this, but also there's this whole field of economics of innovation that is about how does innovation respond to price signals? How does innovation respond to beliefs? And they find results, I'm not trying to question any particular paper, but there's kind of an optimism that comes out of that, that is of the form. Well, it is a social question, whether we want to develop technology A in healthcare or technology B in agriculture, right? Or whether we want to, or this even more so, whether we want to invent pro-worker technologies or whether we want to invent, I guess, presumably anti-worker technologies, right? I am less optimistic about this paradigm for a few reasons. Reason number one is kind of implicitly, the idea here is that like, that rewards to investing in thing A have to be like sufficiently close to the rewards of investing in thing B, such that reallocating revenue is going to like ever make sense. But I think perhaps you're familiar with Sutton's bitter lesson.

Ben Zweig: Yeah. Yeah. The idea that basically like you're, you just like the more general model is, is just like going to dominate in every sub domain. Is that, is that my getting there?

Seth Benzell: Right? Exactly right. That's exactly right. And Andrey and I, when you talk to technologists, they say that the framing that I just had of like, why don't we just spend less of our revenue on military AI and more on healthcare AI? Like their answer is like, that is dumb. That is not understanding the technology. The bitter lesson tells us that the general system you might, you can, that's not saying there isn't reward to make developing the specialized system and adding in rules and adding in harnesses. And of course that's a giant role for the economy and individual businesses, et cetera, et cetera. That's not saying that stuff isn't important, but the bitter lesson suggests that the highest, most generalizable returns are always going to be building the best, biggest generalizable model. Right. So if you take the bitter lesson seriously, which Andrey and I do, that makes you more pessimistic about directability, at least at this very high level. Of course there's directability of like, are regulations going to let us apply this to this context? Right. Okay. So that's the first thing.

Ben Zweig: And I guess, I guess if you do believe that like the, the actual AI the, the fund, the foundational technology is not going to be the one that produces the, the economic output. If you believe that's going to be due to the firms changing, their business processes and whatever, then like you're going to have a thousand applications of this, of this like core AI. And those presumably are going to be engineered just like any other. So maybe that's the kind of directing that, that these, that these people have in mind.

Seth Benzell: I mean, I think that's a steel manning. I think that's a steel manning. I think they would say, oh, take resources away from the frontier labs and instead give it to something else.

Ben Zweig: Yeah. Yeah. You may be right. I mean, that's like, yeah, I think you're probably right. Like, like I think, I think even, even when I, when I hear that argument, they, they sometimes say like, oh, we can impact the way that people think at Anthropic or whatever it is. That sounds, it sounds dicey to me too. Also, I mean, yeah, it's hard to imagine like creating a set of incentives that are not like distortionary in like all sorts of undesirable ways.

Seth Benzell: Right. So that's the policy instrument question. So maybe that's the reason number two to be pessimistic. And just to get this out, the third reason I'm pessimistic is I'm not sure we know enough about elasticities of substitution and complementarity across different organizations and sectors of the economy to really have a good sense of what technology would be truly pro worker versus anti worker. Talking to these people, talking to these erudite and really, really smart economists, they will, if you, if you pin them down on this question, they will admit, yeah, creating new tasks is pro worker in the sense of if you give a worker a new thing to do, you will increase demand for that guy. But that if that guy that you've just made more productive is somehow substituting for another sector that's even more labor intensive, well, a net that's an anti worker technology by at least we mean by wages going up or if we mean by labor share going up, that's an anti worker technology, even though it is helping this worker that you have now enhanced with this AI or this computer and vice versa. There are plenty of examples in history of an automation technology that automates 90 percent of a worker's job that ends up incredibly demanding boosting demand for that job. That's a Jevons paradox you might have heard of. So, yeah, so I'm pessimistic. I would love it's so freaking cool sci fi as of psychohistory. Let's point the direction of the tech tree in the way that's good for people. I would I would I'm just pessimistic about it.

Ben Zweig: Yeah, yeah, I mean. It's interesting. I mean, I think I think I share your perspective that like that, that is very hard, maybe impossible to control. But I don't know if I'd characterize that as like necessarily pessimistic. I mean, if anything, I feel like that like, like, I feel like I view it a little bit optimistically that like, oh, there will be this technology and like capitalism will do its magic and people will reconfigure and like society will find its best uses and like everything will be fine. I mean, that's like an exaggeration.

Seth Benzell: No, no, I like I would say that's I am also optimistic about society, even as I am pessimistic about our ability to direct technology.

Ben Zweig: Yeah, yeah. Right, right. So so but I mean, I think that sort of stems from a belief that like, even if we could direct technology, that would maybe be a bad thing. So like the fact that we maybe can't is like maybe good.

Seth Benzell: We're on a ride. I think I think the best model of technology is that technology takes us on a ride. I know that's very. And it's maybe teleological. I'm not really I'm not trying to get woo here, but I think that things like the bitter lesson are examples of sometimes you're digging in the bedrock of ideas and you find a rich vein and you like you're going to mine out that vein before you go looking for new veins.

Ben Zweig: Yeah, yeah, yeah. I mean, I mean, it makes me think that like in economics, generally, I feel like there's a tendency to just like treat technology as exogenous mostly. And we stumble upon whatever gets stumbled upon. And like that drives the next paradigm of like economic prosperity. And it happens to have been a I maybe in a different universe. It would have been quantum or whatever like, like it just like we are. We are just like we are just like you said, we're being taken for a ride in like the just like randomness of like what technology gets discovered and landed on. It's I think it sounds kind of right to me, but I mean, I think the reason why it sounds right is like a little bit dissatisfying is because like we can't really we don't have a good theory for like why these things come to be in the order that they come to be.

Seth Benzell: May I suggest it's just like, yeah, you understood the technology well enough to predict the next technology would have just already invented the next technology.

Ben Zweig: Right, right. I guess that's true. Although I mean one way to like abstract a level before is this like idea that like neurons that fire together wire together. You know what I mean? That like if you're if you're sort of working at the intersection of like two things, those will sort of fuse and you'll just like innovate at that intersection. So maybe there's like some who knows like what that innovation is going to be. But like it's probably safe to assume that like we will innovate where we spend our brain power. Right.

42:21 What moves Seth's priors

Ben Zweig: Anyway, fine. So so just back to priors for just a minute. So like when you're when you're shifting your priors like on the show. Yeah, which I love. I love that.

Seth Benzell: It's our famous gimmick. Although you got it. Andrey, if you're listening, co-host Andrey, beloved co-host Andrey, we've been doing a lot more interviews. I love interviews. But we got to bring we need some more papers because people want to see our beliefs getting updated. Listen to me, Andrey.

Ben Zweig: Yeah, yeah. So so we want to we want to see these beliefs get updated. But like you could you could shift your priors a lot in two ways, right? Like you could have especially weak priors that just like shift a lot or you could have like especially strong evidence. And I guess if you're like doing a retrospective on like when your priors have shifted, do you think that's coming more from just like your beliefs being weakly held? Or do you think that's coming more from the evidence being especially good?

Seth Benzell: I hope this isn't an indictment of economic research when it comes more from the beliefs being weakly held. Right. Any individual study, especially empirical studies. It's hard for them to move your beliefs a lot just because we know there's so much. I'm not accusing any particular researcher of this, but it's inevitable. There's garden of forking paths issues around model design. There's all sorts of data collection and selection issues where even if you're meticulous, you're only looking under the streetlight of what's available. There's issues about extrapolating out of context because we go hard on ourselves. Like most of our priors are not like in this paper, will they estimate a parameter of greater or less than two? We want to ask the question of, OK, this paper is going to try to tell us something about complementarities between humans and robots in writing code, right? So, yeah, sure, they're going to be looking in this context, but I want to know, is this going to change my beliefs about the complementarity between humans and robots and code writing in the economy as a whole for the next 10 years? Right. Those are the questions you really care about that hopefully these kind of empirical studies give you a piece of evidence towards. And so, yeah, inevitably, every individual paper can only move you so much. That being said, there are papers that have moved us. So I just gave one example that we did recently that I felt my beliefs moving significantly, which is. Both through theory and empirics persuading me that at least current generation automated coding tools are complements to code writing, at least at some levels of the hierarchy, right? If a lot of people suggest that, wow current generation tools are so good at writing code, we won't need coders anymore. Right. Because you say what you say Andrej Karpathy says into his microphone, write me code that automates the next generation of Fable five. Right. And then boom, it happens. Right. So why would you need coders in this very interesting paper, which I'm sure Ben will link to in the show notes. What the authors do is point out that the additional lines of code that you're writing at these different open source communities explodes. But the number of releases only goes up 30 percent. So it's like 10X increases in the amount of code written, but only about a 30 percent increase in releases. Right. And the way that a first of all, they do a really good job of identifying the effect of a there. So that was a very persuasive part of the paper. And then they give you kind of this theoretical framework, which I had some critiques on the podcast. Maybe it was a little over complicated to make its point, but persuade you. Well, what that's got to mean is that there's some step in the process where the human is the bottleneck. It's not the writing the 10 X lines of code because we're able to do that. So there is some human bottleneck there. So there is some complementary territory there. And so that's an example of I might still think at the macro level and that's the challenges. How do you extrapolate out from that? Right. So my takeaway was at least at the level of like the code writing project, there's still a big role for a human bottleneck there. Even if you think at the macro level, AI is a substitute for humans because, OK, what comes out of software? Well automation technologies. And so like this other sector of the economy we haven't talked about at all is going to have a lower labor share because we've just boosted the productivity of coders in this first sector. Right.

Ben Zweig: Right. Right. I think I think I've also been a little bit more compelled by this idea that like at least in this moment, AI seems to be more of a complement to human labor generally. I mean, that's like obviously a big, big statement. But but I think. One, I don't think we've seen like great evidence that like that, like there's a big like substitution effect. And I think there's some evidence that there's like a complementary effect from this paper and like others. But I think that has like changed my mind a little bit.

Seth Benzell: Yeah. Yeah. Yeah. Same place. Right. My intuition was always this is a huge automation technology and then you look at it and then there's all sorts of human bottlenecks everywhere. Now, maybe it's just a matter of time.

Ben Zweig: OK, so yeah, thanks for taking me through all that. One one final question for you.

Seth Benzell: Final already. You're going to let me go? I'm having so much fun, Ben.

Ben Zweig: Yeah, yeah, we could go on forever.

Seth Benzell: I haven't even told you about like the marriage alliances between the different factions of the AI econ land. All right.

48:18 Overrated: pro-worker AI and seashell collecting

Ben Zweig: Yeah, we got to we got to model that out. But but it's a big one. I mean, what do you think is the most like overrated conversation in this field and the most underrated?

Seth Benzell: The most overrated term is pro worker AI because I don't know what pro worker AI means if.

Ben Zweig: Yeah.

Seth Benzell: OK. I mean, I think that if you in a model, you can write down in this model, this sort of technology boosts wages or lowers wages or increases labor share or decreases labor share or increases wage polarization or decreases wage polarization in a model. That is a term that makes sense. But I think in the abstract, it is not a helpful term. Another thing overrated. And this is and I'm in a business school right now. Well, actually, I'm in a college of business and economics. Right. So I'll take a potshot at kind of the people who are in the information systems, organizational science, kind of adjacent business fields, is you see a lot of work that I might you might call. I call it a seashell collecting. Right. Maybe you're familiar with this metaphor from Newton that like there's this idea that. OK, so like back in the old days of science, before, like we had modern science, people would have these cabinets of curiosities. Right. And be like, here's a bunch of interesting rocks that I found. And some are crystals and some are fossils and some are like a relic of a saint. Right. Yeah, that eventually became science. Eventually, people were like, hey, all these fossil things kind of look like each other and we all kind of find them in the same place. Right. So the collecting of interesting rocks can be an eventual step into being a science, but it's not quite there yet. Right.

Ben Zweig: Yeah.

Seth Benzell: I think a lot of papers that I see sometimes are business X used an AI tool and an increased Y by three percent without like a framework to help me understand why this is important or why there'd be any chance of this result extrapolating anywhere else. It's seashell collecting. It's interesting rock. Yeah. And I'm not saying that can't ever be an important input into a theory that then comes up and says, I have a theory about why this business did well and this business did badly, but it's like it's so far. It's so many steps away from the interesting thing.

Ben Zweig: I feel like that's the that's the big critique on like RCTs and development in applied micro and pseudo experiments generally. I mean, it's like interesting nuggets that like tell us something interesting locally, but don't generalize. Is that like a microcosm of your complaint about like,

Seth Benzell: I think the best RCTs, the best RCTs think really, really hard about external validity before they get set up. Right. A hundred percent.

Ben Zweig: But external validity is like maybe I mean, it's a cousin to like generalizability, but it's not like you may have like a very specific treatment that like does have external validity, but maybe doesn't tell you something like a deep truth about the world. Right.

Seth Benzell: Yeah. So anyway, I think you're right. Like the best RCTs do think about that. The best RCTs do think about all of these challenges. But you see a lot of work, which is, oh this cool business allowed me to see some of their data and I was able to run a diff and diff on they used AI and something happened. Right. And I don't I don't know how to think about that because it's certainly not super interesting to me seeing the one paper that does that. I do think there is a role for a business academic in teaching firms. This is how you do diff and diffs the right way. This is how you A.B. Tests the right way. Right. So if the way you want to think about this is like this is a thing a business should be doing on its own in order to understand itself. And I'm just like, this is just an interesting case study. Right. Not opposed to that. It's not interesting to me when I put on my economics hat.

Ben Zweig: Yeah, I saw a critique about this about RCTs and the way they characterized it, which was like I thought like a little harsh way they spoke about it was like we love RCTs. They're super valuable, but not for like Nobel Prize winning. It's like that's being done by like a World Bank staffer.

Seth Benzell: Oh this is the this is the Kyle. Sorry, Kevin Bryan. Right. I said, yeah, in his maybe you even heard this on our podcast. Oh, maybe Kevin Bryan, formerly of the amazing micro theory blog A Fine Theorem writes in the Esther Duflo, Banerjee Prize, and Kremer. I think it's Banerjee, Duflo, and Kremer. Yeah, they win a prize for running RCTs in Africa. And Kevin comes out strong. He's like he says exactly what you said. Like, I don't know why we're giving out Nobel Prize. Like, sure, the World Bank should figure out whether its programs are good. But this is it's economic dentistry. It's not it's not the high thought that some of us aspire to. And of course, program evaluation is important. I'm not trying to dis program evaluation. It's does it. And the best kind of program evaluation can advance the field, but often it doesn't.

53:53 Underrated: capital accumulation

Ben Zweig: Right, right, right. OK, so what about underrated? What do you what do you wish people were talking about more?

Seth Benzell: Here's two underrated topics. I'll give you one that's kind of my hobby horse. And the kind of both things that I care about. Underrated topic number one is capital accumulation dynamics. Right. So if you look at a lot of the kind of the very abstract macro growth land, they're either assuming that OK, first of all, I'm going to give you a little bit they're either assuming that OK, first of all, let's talk about why capital accumulation might be important. OK, yeah. So all right. In a Solow growth model of the economy, the economy gets bigger over time because we accumulate more factories and more capital. Right. And at a certain point, economic growth peters out because we're investing as much in making new factories as the at the rate that the factories depreciate. Right. And there's a decreasing marginal return to factories because you need workers in the factories. Eventually you have so many factories and not enough workers and the marginal factory doesn't help you and it just decays after you make it. Right. That's the Solow growth model in a nutshell. In the simplest version of a what happens when the economy automates model, that's called an AK growth model. So you've got your it's called the AK growth model because there's a for the technology level and K for the amount of capital. Right. No labor. No labor because I mean, and they give you like plus labor on the side.

Ben Zweig: But yeah, yeah, yeah.

Seth Benzell: Right. But that's not the contributor to growth. Exactly. Plus labor in that model. What determines the growth rate, the indefinite growth rate forever is the saving rate and the technological growth rate. Right. It's how fast does K go up and how fast does a go up? But much ink has been written on a going up. Right. Much much. Right. And there was a point in time in which economists were very much obsessed with what makes K go up. Right. This was a question around Solow's time in the 80s where this was a major fact, focus of macroeconomic policy. How do we boost investment? What sort of taxes crowd out investment versus boost investment? And actually we're kind of seeing an echo of OK, I'll come back to that point. Right. This tax policy aspect.

Ben Zweig: Yeah. So why you could also say this is like being forgotten about in trade policy.

Seth Benzell: Absolutely. Well, a lot of things are being forgotten. That is probably like a flip side. But yes, you're exactly right. The flip side of a being a net importer is that you're a net investor from abroad. Right. The way that you get investment from abroad is by running a trade deficit. Right. So capital surplus. Yeah, that's the capital. Yeah, right. So. Um. Why is this so this is the most important thing in an automating world, in a not fully automated world, it is still essential. If you think about a model where we're so imagine a model, this is like I'm kind of riffing on a tour not sorry on Acemoglu and Restrepo race between man and machine here, you get a shock that we allow the automation of a bunch of jobs. Right. So capital is going to be stolen, stolen or crowded out from being invested in like factories that humans work at and they go into the fully automated sector. Right. To me. Yeah.

Ben Zweig: Sorry, just to just to cut in for the whenever I hear of that model and I see this a lot in literature, just like a little bone to pick, it is the Heckscher-Ohlin model like it is like you have like allocations of capital and labor and the price of goods. You have like this two factor model. And like I feel like that gets thrown around a lot without like really framing it as like we're going to repurpose Heckscher-Ohlin toward this toward this. But that like is kind of I think what it is.

Seth Benzell: I think there's definite echoes, definite echoes. OK, but OK, so I love this. It's I understand why it's the paradigmatic model because it does a lot of stuff that is intuitive for economists. And then we can talk about breaking it. Right. OK, so you've automated a sector. Capital is going to flow into that sector in the short term. That hurts workers because you've got less capital going to complement them in the in the not automated sector. Right. And there's a fixed amount of capital, fixed amount of investment happening in the long run. What determines whether those workers who are initially hurt are better off. Well, it's the long run capital accumulation. If we say enough again, well, then they get factories again and their wages come up again. If we don't save enough. Well, now you've crowded out tools that hurt that help workers forever. And now you've now permanently hurt those guys in the race between man and machine model. This is solved with a very one framework for thinking about capital investment over time is something called a representative agent model where you have this one guy who represents everybody in the economy, who's trading off and is investing now versus investing tomorrow, blah, blah, blah. One sort of the takeaway from this class of models is the long term interest rate is fixed. Right. So you're all if you if the interest rate goes up, you're going to start saving again to bring the interest rate down. If the interest rate goes down, you start dissaving to bring the interest rate up.

Ben Zweig: Yeah.

Seth Benzell: So in other words, there's kind of this very simplified saving dynamic which makes sure that in the long run, automation technologies don't hurt wages in the long run. And you always save up again such that the worker gets the factory. Right. If I was going to tell a story about why automation might hurt workers in the long run, it would be a story about capital accumulation, it would be one in which workers can't wages go down in the short run, but then they don't save and reinvest in the future. And if you think about why people save over time, they save according to life cycle dynamics, right. So young people go into debt to go to college. Then they then they save to pay down that debt. Then they go into retirement and then hopefully they have their savings in their retirement to eat when you have a technological change that shifts incomes from workers to capital owners. The numbers here really matter, but it's very plausible that the capital owners, if these are mostly retirees, have a higher marginal propensity to consume. You lower the overall saving rate of the economy, lower the overall investment rate of the economy. And that's a story in which automation hurts workers in the long run. It's because it's not through the technology directly. It's through this negative knock on effect of your harming capital accumulation. And so and there's also lots of interesting dynamics. I don't know if you've had Basil Halperin on the show, but I love the guy. You got to have him on.

Ben Zweig: He's so brilliant.

Seth Benzell: He's so interesting. He's got a riff on this theory, which is the idea that not only should you expect automation to raise interest rates, you can learn about the market's beliefs about automation in the future from interest rates today. Right. And he's got an argument saying that if you thought the singularity was coming, you would have to save today because things are going to be free in tomorrow or the world will be destroyed tomorrow. So I should say, right.

Ben Zweig: And so I mean, sort of cuts the other way.

Seth Benzell: So I think I they're mutually reinforcing, dude. Because anything that he's got a story in which the anticipation of automation leads to dissaving, which raises interest rates. Right. Yeah. I've got a story in which automation leads to disaving due to factory allocation, which raises interest rates and then bad things happen. Right. So I think we actually have sort of aligned stories. Yeah. But anyway, I'm not saying that's going to happen. I'm saying that if you wanted to tell a story about automation being bad, that would be the direction I would go. Right. And so I think that's underappreciated.

1:02:05 How VCs really allocate capital

Ben Zweig: Yeah. Yeah. I think that I think that sounds that sounds interesting. I mean, I like these like interest rate thought experiments. Just something like kind of I mean, this is like taking off my hat as an economist and putting on my hat as like an entrepreneur. Do it like I've been surprised by like how capital allocations are like so weirdly unpredictable. But like, I mean, I when I first started Revelio, I came in thinking like, all right you start a successful company, you like it's going well, you raise a seed round thing, raise a series, a series, b series, c, whatever. And like I think in the eight years or so that I've been running is I feel like the VC landscape has become so thematic like five years ago, I think all of the big VC funds were just like, you know what? We're going to stop looking for a successful company. We're going to put it all on blockchain. And now they're just like, we're putting it all on a like they don't want to miss the trend. And and I feel like there's like a lot of friction in like just like not like companies that are not on theme. They just have really different like access to capital. It's a little popularity contest aspect. Yeah. Yeah. And I wonder if that like, like I haven't seen that like models in any in any rigorous way, but I don't know, to me, I feel like I no longer sort of trust models of capital accumulation and like capital allocation rather. No, that's a yeah. Allocation.

Seth Benzell: I was talking about it.

Ben Zweig: Yeah. Allocation. Not accumulation.

Seth Benzell: Fair enough. But this is an aspect of that. Yeah.

Ben Zweig: Yeah. Yeah. But it relies on this sort of assumption that like capital gets reallocated toward the toward the sectors that have the highest return on capital, which like it has to. But but I guess I would assume now that happens with a lot more like noise than I would have previously assumed.

Seth Benzell: Well, I'm glad to have your real world insight away from my law of one price. All all investments must have exactly the same rate of return.

Ben Zweig: Yeah, I mean, you can you can sort of like bridge them with some like noise element or like capital messiness diffusion parameter or something. Yeah, sure.

Seth Benzell: But anyway, just to close out that idea before I give you the second under appreciated idea, yeah, this is also an optimistic story. Right. Which is I think a lot of models, they under appreciate the fact that as you automate the economy, you move us closer to AK land and in AK land, growth is unbounded. Right. That's not necessarily a singularity. Right. That's that's merely exponential growth, which we have already, which we already have. But you can imagine a kind of a turbocharged exponential growth. Right. Yeah. So that's that's maybe a reason to even be positive when you think about capital accumulation dynamics. OK, second underappreciated idea.

1:05:14 Underrated: data as a factor of production

Seth Benzell: This is I'm starting to see some papers come out at this. One is by a Dr. Koh, K-O-H, who is a student who recently graduated out of MIT Econ. I believe he's working for one of the Frontier Labs right now. And I've seen kind of variations of this in different places, but trying to think really seriously about. Data from people completing tasks as a factor of production in making AI that's able to do those tasks, right? So I think as we get better and better at kind of the general purpose AI. I still think I guess I'm not 100 percent bitter lesson pilled in the sense that I do think the what you pour into that giant AI is going to determine the tasks that it's good at first and the tasks that it's good at later. And so I would love to see more work on trying to measure the elasticity of task performance to the quantity of data available on humans doing that task, right? That's something that I think may be happening a little bit in the labs at a very abstract level, but I would want to see this kind of like here's a hundred different companies and over time we fed it more and more of our data of the humans completing the task and this is how much more productive it got over time. I think if you understood that you would be on a much better path to understanding are we on the path to a singularity versus something that concaves off, right? That's what I would want to know better.

Ben Zweig: Yeah, I think you'd also be able to like make more micro level predictions because you could say, oh, all you need to do is invest in these like 10,000 investment banking interviews and then you have investment banking models. And something like that, like you would, yeah, you would know you would be able to project like some

Seth Benzell: Tasks much more precisely about this is when this task is going to get automated. This is when this job is going to get automated. There's also like a splitting of rents issue here. Right. Which is if I am creating the code that's automating my job forever, that should give me some bargaining power. Right. So this is in some ways even a pro worker frame. Right. Well, yeah, to the extent that it's scarce if I'm if I'm selling myself today, I should get less than if I'm selling myself today plus some percentage of that efforts forever into the future. Right. And so I think that framing could be helpful to workers.

Ben Zweig: Yeah. Interesting. So much to chew on. Thank you so much. Yeah, we can go all day and we'll have to do it again sometime.

Seth Benzell: We are sure to do it again.

Ben Zweig: Yeah, for sure. So any listeners follow Justified Posteriors, follow Seth's work and anything else and oh and plug that you got a book coming out at

Seth Benzell: Some point. I cannot tell you the title. I cannot tell you the subject.

Ben Zweig: Tell us everything you can.

Seth Benzell: What everything I can say is I've got a trade popular book that we're working on with Erik Brynjolfsson and Andrew McAfee and keep posted. Hopefully I'll be able to say more this spring. But if you liked any of the topics of this episode, I'm sure you'll find that book fascinating.

Ben Zweig: All right. Stay tuned. We'll give updates every week on the status of the book.

Seth Benzell: And of course, the best place for that information is Justified Posteriors.

Ben Zweig: There you go. All right. Thanks. Thank you so much, man. The Economics of Work is brought to you by Revelio Labs, workforce data for research and benchmarking. If you enjoyed, please remember to rate, review and subscribe. It'll help us reach other curious listeners. I'm your host, Ben Zweig. Our producer is Cole Wagner. Thanks for listening.