This week, instead of reviewing a recent paper on AI, we go back to a 1993 classic: Michael Kremer’s “The O-Ring Theory of Economic Development.” It’s one of those papers that feels larger than its formal model. The setup is extremely simple: production consists of many tasks, and output depends on all of them going right. But the implications are broad enough to touch development, firm organization, inequality, and the economics of AI. The paper is also a throwback in style: a grand theory paper in which a simple model is asked to explain a wide range of stylized facts. Sometimes that kind of ambition is a bug. Here, it is mostly a feature. The model is elegant, readable, and influential for good reason, even if it occasionally threatens to explain everything and therefore, maybe, not enough.
We start with Challenger, the exploding shuttle, and the literal O-ring that gives the model its name. From there we work through Kremer’s central intuition: when production is highly complementary, weak links matter enormously, and the best workers or firms want to match with one another. That basic logic helps explain why high-wage workers cluster together, why firms may separate sharply by quality, and why richer economies may be able to sustain much longer and more complex production chains. But the further you move from tasks to firms to whole countries, the shakier the fit becomes. Along the way we connect the paper to modern questions about AI and automation, including whether AI reduces weak links by automating them, or instead makes production even more complex and therefore more unequal.
Priors → Posteriors:
Prior 1: Is close complementarity the most important explanation for sorting within firms and productivity differences across countries? Andrey goes in at about 60% for firms and 30% for countries. Seth comes in slightly higher for firms and slightly lower for countries. By the end, both of us move only a little: somewhat more convinced that O-ring logic is a major force within firms, but still skeptical that it is the main explanation for cross-country differences once institutions, natural resources, and other macro factors enter the picture.
Prior 2: Is O-ring logic a better way to think about AI and automation than the standard task-based framework? Both of us come in thinking the answer depends on the level of analysis. At the level of the individual worker, O-ring logic feels very strong: people are often defined more by their weak links than their strengths. At the macro level, substitution and reorganization matter much more. By the end, that view holds up. If anything, the paper reinforces the idea that AI may increase returns for highly capable workers by removing weak links around them, even as it automates more commoditized tasks.
This episode is sponsored by Revelio Labs — a great source of labor economics data for academics and firms. Now available on WRDS.
References:
Acemoglu & Restrepo — task-based automation framework The canonical accessible version: Automation and New Tasks: How Technology Displaces and Reinstates Labor, Journal of Economic Perspectives, 2019.
Gans & Goldfarb — O-Ring Automation NBER Working Paper 34639
Rosen (1981) — The Economics of Superstars PDF.
Becker on Assortative Matching
Hamming — You and Your Research fs.blog.
Philip Trammell — cross-task learning / automation and learning-by-doing: the thread.
Patrick McKenzie — The Sort, the original thread.
Transcript
Introduction [00:00]
Seth: Welcome to Justified Posteriors, the podcast that updates beliefs about the economics of AI and technology. I’m Seth Benzell, always benefiting from the high performance of my closely complementary co-host, coming to you from Chapman University in sunny Southern California.
Andrey: And I’m Andrey Fradkin, always ruining our O-ring production function by not thinking in advance of my introductory catchphrase. Coming to you from New York City today.
Seth: Do we have a sponsor?
Andrey: We indeed have a sponsor, the wonderful folks at Revelio Labs. Thank you for sponsoring the podcast.
The O-Ring Paper and Why It Still Matters [00:42]
Seth: Amazing. More on them later. So, Andrey, interesting episode today. We’re going to be reading and thinking about a bit of a classic about O-rings. So this is a paper from 1993 from Nobel laureate Michael Kremer, appearing in the Quarterly Journal of Economics. And it’s doing theory in a way that is a little bit unusual from our post 2010 econ reading selves, which is it lays out a grand theory of the economy. It shows that a bunch of stylized facts fall out of it and it pieces out. I enjoyed reading this paper, but is it science? Did you learn anything from it? I’m excited to jump in.
Andrey: Yeah. And I think we should talk a little bit about what is interesting here. This paper has been very influential in how people think about production and about differences in productivity across firms and countries. I looked for mentions of this paper and the Nobel Prize. And it turns out that this is not the paper that Michael Kremer got the Nobel Prize for, even though In some sense, it might be more influential than his other.
Seth: He won it for RCTs, right? Anybody could have thought of RCTs.
Andrey: He won it for conducting randomized controlled trials in the developing world, which this is decidedly not.
Seth: This is basically the economic opposite of doing RCTs.
Andrey: But nonetheless, it has gotten a lot of attention more recently as people think about AI. How is AI going to affect production? Well, it really depends on what we assume about the production function. So with that, I I think maybe let’s get to our priors, Seth.
Priors [02:45]
Seth: So basically what this paper is going to do is going to roll out a story of how production works in the economy where production is really complementary. In order to get something, a really high quality product, you need everything in your firm to go right and maybe kind of at the economy level if you want your country to be productive, you need all of the elements of your country to go right. And so... He thinks a lot of things fall out of this close complementarity between different inputs in production and like kind of weak links models of production. And two of the things that he thinks fall out of that is the big wage differential between rich and poor countries, and then the positive correlation in wages within a firm. So those are kind of two of the big predictions that come out of this model is that there should be big gaps between countries in wages. And then the second is across firms. there should be big gaps in wages as you get firms with lots of high quality, high paid workers and firms with lots of low quality, low paid workers. So I guess our first prior will be, do you think close complementarity is the most important explanation of those two phenomena?
Andrey: Let’s focus first on firm productivity. Is it the most important explanation? My prior is probably yes. But I think it explains less than 50 percent of the reason wages are correlated within a firm. There are a lot of reasons, and I think this is probably the most important, but
Seth: But there are.
Andrey: I don’t think it explains most of it. So that’s fact one. For developing countries versus developed countries, like differences in GDP per capita, I don’t think this explains it very much unless you take such a broad view of production functions that you essentially fold institutions into production.
Seth: The sheikh is a close complement to sitting next to the oil well.
Andrey: That’s the good one. So yeah, so those are kind of my priors. And I guess I should put some probabilities on those. So this being the most important factor in firm differences, I’d put that 60 percent. And then this being the most important factor in country differences, this being the most important factor in country differences, I guess that’s a little tricky is I put that at 70 percent.
Seth: Okay, cool. So 30 percent chance that it is. I would probably come in with maybe a slightly exaggerated version of those takes. I think at the firm level, if you think about, you know, why do the New York Knicks have five really good players and my pickup, I shouldn’t talk about the Knicks, I should talk about a good basketball team.
Andrey: Yeah.
Seth: I don’t know. You know, when LeBron took his talents to South Beach and he brought everybody to the Miami, why did everyone want to play there? Well, it’s because everybody, the best players want to play with the best players, right? And they understand that they’ll perform better next to the best players. The owners understand that if the best players are all combined with each other, they’ll get a better product. I think sports may be an interesting case because there’s kind of like a max productivity you could theoretically hit, right? You just can’t do better than winning maybe. So. maybe that would count against that particular setting.
Andrey: disagree with you said it. already disagree. I would think like most truly superstar players, they don’t want to play with too many other superstar players because their production function is not the same as the team’s production function.
Seth: No cap, so that’ll count against, so that’ll be a big count against the Kremmer story, because Kremmer’s gonna want all the superstars to be together.
Andrey: Yes, I mean, we observe that’s not how basketball is organized and that superstars are dispersed across teams rather than all being on the same team.
Seth: But maybe that’s how countries work, right? I mean, it seems like most of the superstars are in America if we’re talking about elite CEOs or elite scientists.
Andrey: Well, there might be potentially CEOs and scientists in other countries that are hampered by, like that’s a claim. You’re, you’re, you’re selecting on the sample.
Seth: It’s T-Mobility. Ha ha!
Andrey: I guess basketball is pretty good at getting talent into the system. And then when we see that the talent is allocated, maybe the NBA as a whole is an over-ing production function, but each individual team is not structured to have only superstars.
Seth: would bet that productivity on a basketball team, if you’re talking about like point score, at the point score level.
Andrey: That’s what I’m disputing is that’s the productivity that we care about necessarily.
Seth: My yes, so then there’s this other productivity which is like adjusting for compensating differential of you the stars getting each other’s ways, right? And of course that’s going to be challenge with all of this is bringing these hypotheses to the data given that in the background there could be all sorts of compensating differential things going on, To you. No, no. That was good. That was good feedback. So I guess at the firm level, if you were at 60 percent
Andrey: Yes, yes. Anyway, sorry.
Seth: I will come up to maybe 70 percent that this is the most important factor in correlation of wages at the firm level of all of the top people wanting to be together, is that they’re multiplying each other rather than substitutive. At the country level, I think you said 30 percent chance this is the most important thing. I think if we look empirically at the world, if we were just like weighing all countries equally, Right. You’ve got the issue of like these big resource rich countries, which has like a whole ‘nother story about why they have such high GDPs. I think there’s an issue around, you know, just like policy, like there’s a lot of reasons that countries might have different total productivities. And I think as we go on, we’ll talk more about the fact that even though this is a kind of a beautiful story in this model that can work kind of at the individual level, at the firm level and at the country level. I think my intuition is that this story is kind of like the most powerful at like the individual level. It’s pretty powerful at the firm level. And as we move to the society level, you can imagine so many margins of adjustment that maybe close complementarity isn’t the only way to think about the economy and that there are opportunities for substitution. So I would maybe come in, you said 30 percent chance is the most important thing, just below that 25 percent chance.
Andrey: And to be clear, just, you know, it’s the most important, but it could still be like only explaining 5 percent of the variation and at the country level, right? Right. So I think that’s kind of the way in which it’s tricky. I don’t think either of us think that it explains more than 50 percent of the variation.
Seth: For firm level, I might be convinced this is more than 50 percent of firm variation.
Andrey: could be convinced, at the country level I don’t.
Seth: Country level, it’s hard. There’s so much going on.
Andrey: Yeah. Well, what about the second part?
AI Prior: O-Rings vs. Task-Based Automation [10:13]
Seth: the second prior. my gosh. Second prior. So it’s an Econ of AI podcast. So I got to ask you an AI question, Andrey. Yes. Which is, you know, the foremost way of thinking about how AI will impact the economy is this Acemoglu-Restrepo task-based CS framework where you think about the firm or the economy as
Andrey: fire.
Seth: taking in a whole bunch of labor inputs of different kinds, combining them with a CES production function. And then the way that automation shows up is it shows up in some of the range of tasks that you need to combine. So you’d have a large range of tasks that you need to combine and then use a little bit of capital or you have a small range of tasks that you need to combine. And then you have more capital that represents capital substituting for labor. There’s kind of some sense in which that model Automation makes the economy in like this O-ring sense, like less complex because you have fewer labor inputs. If I replace, you know, imperfect Seth with, you know, pure heartless robot, there’s in the sense of the model we’re about to read that would make things like kind of less complex. There have been other authors who have kind of come at the question of automation from another direction, including I believe it’s Avi Goldfarb and Josh Gans.
Andrey: They’re in fact both on a paper.
Seth: Both on the paper. Avi will be able to talk to soon. And maybe, Andrey, you talk to us a little bit about the way that they think about how O-rings could apply to automation.
Andrey: I don’t want to get too deep in their paper, but they’re thinking about essentially that time allocation of an individual person. So as podcasters, we do various things. We read the papers. We record the podcast. We do the show notes. We read the transcript. Now, what if a wonderful piece of software automated the transcript correction for us? Or maybe that’s even a human. that gives us more time to focus on other parts of the podcast. How is that going to affect the overall productivity?
Seth: really don’t like this map. The implication I would not immediately take all leisure, but also the implication that if I’m thinking about basketball, so in current basketball, you need to be agile and you need to be tall. And like, let’s say we move to a version of basketball where for whatever reason, tallness no longer mattered, right? It’s not like I can reallocate my skill points from tallness into agility, right?
Andrey: So I think you’re setting up a straw man of their argument. I don’t know. like, I think my straw man. even read, which is, know, haven’t we all. But I think the idea here is that like, there’s something within a person that needs to happen about certain production. And if you can, if one part of that all of a sudden gets automated, but the rest you’re still doing, there’s a sense in which Parts getting automated, cache increased your productivity. So that’s kind of the argument.
Seth: Right. the idea there is we should be really automation. We should be thinking about within individuals who have tightly complimentary components rather than thinking about kind of at the firm or the economy level, you bring in some automation, you dump your automation into your, know, your stack of labor inputs and you know, more output and maybe it’s substitutes for labor. you drive wages down, right? Yeah. I mean, Andrey, do you want to take the first swing at this one?
Andrey: I think both are going on. This is going to be a typical Andrey answer. I’m very against mono-causal. I think there are definitely parts of production that are exactly O-ring. I think for certain individuals, it must be the case.
Seth: Two-handed, the famous two-handed economist. there’s also a sense in which the economy can reorganize your job to break up the O-ring, right?
Andrey: So they can and that and that kind of goes to the question of like, what is differentiated about your job and why isn’t the economy able to trivially substitute for it? And podcasting is a great example. think like, I will not use our system example. Let’s use Dorkash as an example.
Seth: Never heard of him. Is he a podcaster?
Andrey: A favorite of the AI crowd. Let’s say that he’s able to spend more time researching and preparing for his podcast and is able to recruit better guests because he no longer has to do some of the stuff that he was doing when he was starting out, like manually doing his own editing, hustling to get noticed and so on and so forth. Presumably that podcast becomes a lot better. And since people, he’s already built a relationship with a certain group of people, it’s going to be very hard for the economy to substitute for him.
Seth: I don’t think you even need the time reallocation part if you automate his weak link He doesn’t even need to get better at other stuff, right?
Andrey: Yeah, he does. Yeah, he could. He could just take that as leashed for sure. Right. I think that’s a good point. But he could invest it even to the extent that like, it is an O-ring production function. He’s getting such great returns now from if now his audience is millions of people and before it was thousands, then the returns to quality under are probably a lot higher, right?
Seth: Income effects as well, right? A richer person works less.
Andrey: Yes. I think both of us are in agreement here. if we posited other production functions, then his returns from investing quality would be a lot lower because he’s kind of already good enough.
Seth: Okay, so. Okay, so I guess this is kind of a vague prior, but I guess the question is, like, gun to your head if you have to choose between O-Ring and Osamuoglutus strappo, or how do you want to frame this one? Or what percentage is relatively more important? Or we could talk maybe more about like, at what levels of analysis?
Andrey: Yeah, that’s kind of where I was thinking about going. think the more micro, the more O-ring, the more macro, the more substitute. But even like at the macroeconomic level, I think there are parts of production processes that feel O-ringy to me. So that I’d say maybe 20, 30 percent at the macro.
Seth: So 20 to 30 percent of the macro level at the individual level 80 percent. So it depends so much across occupations. you’re shoveling shit, I don’t know, your O-ring matters for you.
Andrey: I know Seth had to get in a curse word just for our special listeners. Shove it. Doodoo. We’re kid-friendly as a podcast. Yeah, I think that’s about right. What do you think?
Seth: du-du. I think that’s exactly the right way of thinking about it, which is I would think both in terms of scales and timeframes. So at the individual level, and this isn’t really gotten into in this paper, so maybe we’ll talk about it a little bit later, I think it’s inarguable that abilities are complementary within an individual. If you have an individual who’s smart and strong and lazy, they’re a bad worker. If you have a worker who’s strong and energetic, but not smart, they’re a bad worker, right? I think people are weak link people, right? They’re kind of defined by their weaknesses in some ways more than their strengths. If you think at the firm level, you imagine firms optimizing around this problem to try to minimize it, but obviously there are cases where you can’t minimize it. If you’ve got a fab where you’re making computer chips and you do a thousand steps to go exactly right to make billions of dollars of computer chips, You better believe that you’re going to think hard about every part of that process, right? But as we get to the macro level, we’ve talked about this. There’s all sorts of substitution that can happen and reorganization. The part that I would add to what you just said is across time, think this matters too, right? So we know the famous, I think Tyler calls this Le Chatelier principle, which is actually a chemistry principle. But the idea that things get more substitutable over time, yeah, I mean, Right now, in 1990s, the only way to get to outer space was with a space shuttle. And then in that space shuttle, the weak link was the O-ring. We’ll talk about this in a second. But now there’s like two or three different ways to go to outer space, which might have different weak links. And if you can choose between different production functions that have different weak links, all of a sudden, weak links start mattering less, even if every process has one.
Andrey: Minor point of clarification, please. 1980s.
Seth: 1980s. No, but the shuttle was still going into the 90s though.
Andrey: happened in the 80s. I, so I think like a couple of other thoughts here, know, various folks have written about production functions in interesting ways that kind of reflect this, like, Richard Hamming wrote a lot about what it takes to be a good researcher. And a lot of that, you know, his essay, you and your research, you know, talks about kind of weak links in the production process of being a researcher. And I certainly think that We’ve all seen various versions of that failure mode. A classic version is if you’re a bad presenter, that could really sink your scientific career. And that seems pretty O-ringy. On the other hand, I do think there’s something about the modern, more advanced economy in which certain things that would have been O-ring production functions historically are becoming a lot less so. To give an example, with chat GPT, maybe lack of ability in a particular language is less important for doing scientific research than it was previously, right? Because then now you can write in English, even if you’re not a good writer in English.
Seth: That’s a really good one, right? Exactly. So one way to think about that is automating a thing that is a weak link for a lot of people so that they can now just focus on the thing that they’re good at. Yes.
Andrey: And then the other thing to think about, and think Phil Trammell has written about it, is there might be some sort of benefit from doing all the tasks in a production process because of learning. Like, it’s hard to do good transcript editing unless you’re actually the one who had the conversation in the podcast because you kind of know what you’re trying to say. That’s kind of a trivial example, but you can imagine many such examples. so then... You want to you’re tempted to get rid of the O-ring by, you know, having the computer do certain of the tasks. But then that ruins your learning and your ultimate ability to produce.
Seth: All right, well put, Andrey. Are we ready to move into the evidence?
Andrey: Yes, the evidence being a theory paper. Exactly.
The Challenger Disaster and the Origin of the O-Ring [22:12]
Seth: All right, so the evidence. Well, in I think an underappreciated and underutilized way of starting an economics paper, he starts with an explosion. More economics papers should start with big explosions, Andrey. Particular, he’s talking about the Space Shuttle Challenger disaster, which is famous for, you know, we used to have these space shuttles that would go up, come back.
Andrey: True.
Seth: You can Google them if you’re interested in them. In particular, there was this one event where they had a high school social studies teacher on the spaceship. It was going to be the first civilian in space. And she was the first civilian in space for about 30 seconds before the whole thing explodes. Right. And so the government, Reagan, convenes this blue ribbon panel to try to figure out why did this spaceship explode. And there are so many cool people on it. Neil Armstrong’s on the panel, Sally Ride’s on the panel, and of course, iconic iconoclast, Richard Feynman. And so Richard Feynman is working with his team trying to figure out why did the spaceship explode? There’s all this sort of bureaucratic, you know, finger pointing, hey, what went on here? Sally Ride gets an anonymous tip from a NASA engineer showing this table that said, Hey, when temperatures get low, there seem to be this one part that has a higher rate of failure. So in particular, it’s this thing called an O-ring. It’s kind of a gasket. It’s a rubber gasket. And if you think about the space shuttle engine, it actually kind of has a bunch of kind of chunks in it. And the hot gas needs to flow between these chunks. And so this rubber gasket went around the connections, keeping the hot gas between the chunks. So what went wrong? Well, Feynman famously showed in one of his committee hearings that they were having, bringing in evidence and talking to experts about what could have gone on. He famously kind of opens up a model of the space shuttle, takes out a bit of O-ring that was of the same material, dunks it in a glass of ice water. Then at the end of the hearing, he pulls it out and he’s like... clearly the things like about to break, it’s the wrong shape. And he’s like, I think this may have something to do with our, we’re discussing today. So, love a Feynman moment. Yeah. And so the committee ultimately decides that this, you know, the shuttle program as a whole was worth over, costed over $280 billion in 2026 numbers. This whole program eventually, you know, this particular space shuttle of five was blew up. And after this disaster, the program released, it’s the beginning of the end of the program, all brought down by a single, what, couple hundred dollar large rubber gasket. So that’s the O-ring. It’s the weak link that destroys your amazing operation. And so what kind of production function does that inspire, Andrey?
The O-Ring Production Function [25:20]
Andrey: Yeah, so hence the O-ring production function. It’s quite simple in that you imagine that there are n tasks in a production process, and each of them is done at a certain quality level. We can think of this as a probability of success. So it could be 0 out of 1. And you just multiply that quality together, and that determines how much is produced. There are additional terms there relating to overall productivity and the amount of capital. but let’s just ignore that for this discussion.
Seth: Yes.
Andrey: It’s in some sense, like strikingly simple, right? That’s, I think, one of the reasons that this theory is so influential. But you get these implications out of it that are quite interesting in that for production, for example, you want, you know, if you have one high quality component, then you’re getting a lot more out of having the other tasks also be done by a very high quality person. And then you’d have longer chains. Obviously, you want longer chains to have higher success probabilities. And so he just derives a competitive equilibrium. That’s kind of the first thing that he does in such a simple model.
Seth: And you just point out the big thing that comes out of the equilibrium is assortative matching. Yes. In equilibrium, everyone should have the same skill level. And just to pull out a quote on that, since profits are zero for all firms given the wage schedule, firms are indifferent as to the skill level of their workers as long as their labor force is of homogeneous skill. Equilibrium holds when firms demand the number of workers of each skill available in the population. Well-behaved problem, the competitive equilibrium is optimal and unique up to reassignments of workers of equal skill. firms enter, every firm is going to have a worker of exactly the same skill level, right? Or rather, yeah.
Andrey: Yeah. And this has like a flavor of some of these labor market matching models, you know, going all the way back to Becker, where you think about like, do high skilled workers match with high productivity firms, and you got very analogous results under certain assumptions. So he does this and this is kind of like a true, you know, pretty trivial grad school exercise. And then he goes into applications, which is kind of cool to do. It almost reads in many ways like a sub stack post, honestly.
Seth: I have to say, Andrey didn’t want one of our priors to be about theory today versus theory in 1993. But guys, if you want a recommendation about one of our papers that’s actually super readable and not too long, this is a very readable paper. The math doesn’t get crazy. The math is exactly as complicated as it needs to be. And all of these interesting results fall out of it that seem to stylistically hold. Andrey, wish I could write. I wish they let me publish theory papers like this. I wish there was more. There we go. There we go.
Andrey: Some sec they do. Well, you know who I view writes in this style? Friend of so Tom Cunningham, who writes very influential blog posts on his website, which are essentially written like this.
Seth: Yeah, this is a truly epic blog post that we’re unpacking, guys, for you today. You said that this had the flavor of matching markets. You know what this kind of had the flavor of, to me, is kind of Rosen’s Superstar Markets. And Rosen does get cited a bit here, right? So you can think about, in the context of superstar markets, the richest guy wants the best surgeon for his heart disease, right? So that’s kind of another example of this kind of close complementarity and matching you would get.
Stylized Facts: Development, Firms, and Sorting [29:09]
Andrey: Okay, so let’s go through some of the facts. There are a bunch of facts. So maybe we’ll be a little
Seth: Yeah, I think let’s go bang. Yeah, the first one I have is wage differentials between rich and poor countries as large He starts with kind of one of the most tendentious ones
Andrey: Yeah, well, I think there obviously, you know, we already started talking about this, but so many other reasons why there are differences in wages and productivity, know, institutions is certainly something that for me is very important, but also natural resource endowments, as we discussed. I guess that’s even related to some of the geography explanations that we can talk about. Institutions are really broad, goes, know, specific laws, could be, you know, the levels of corruption. And then we get into these kind of intermediate things where like, well, if the education system isn’t good, is that something about O-rings or is that something about something else that’s important?
Seth: No, I was going to agree with that. I’ll just give an example. If your economy doesn’t have any natural resources and so energy is very expensive and that’s why wages are low, is that no ring story? In some sense, no, because o-ring is about uncertainty and inputs. I’ve just told the story about, there’s a compliment that’s expensive.
Andrey: Yeah, well, all right. So uncertainty and inputs is not something we’ve talked about yet, Seth. let’s table that for a second. But is the O-ring the first component of many of those explanations? No, it’s not. I think like a counter to that, let’s steel man this a little bit. Well, you do see that Immigration flows to high income countries from like very highly skilled people in those countries. So you have like the brain drain effect that’s suggesting that like an individual level people are choosing to
Seth: There’s a sort of matching, but do you the opposite? Do you see like low skilled Americans moving to like?
Andrey: Everyone wants to move to... Yeah. Yeah. So there’s that going on. think like a very controversial theory that’s out there is an IQ driven theory of productivity. think, is it Garrett Jones that...
Seth: Right, you should care about your country’s average IQ, right?
Andrey: Yeah, yeah. So that kind of seems to relate to some O-Ring production function. So if you have like very high ability people and all the important positions in the economy and society, then you’re going to do better than if than if you don’t. mean, I’m not an IQ, you know, truth or anything like that. So, you know, I wouldn’t like lean on that too much. But if we think about something about ability. then there’s naturally a genetic component to it and then there’s some sort of.
Seth: Also education of course.
Andrey: And like even just, you know, even more so maybe than education is just like the upbringing environment, whether your certain norms are instilled in you.
Seth: at that time. So like leisure preference, Protestant work ethic. Is Protestant work ethic part of the O-ring production function? At a certain point, this concept can expand so big that you get everything.
Andrey: Exactly. You know, you do have countries where they do grow a lot. And so then there’s a question like, well, does the increase in the ability of the workers explain the growth of the country’s economy? I’m not sure. I mean, it is plausible that it’s a contributing factor. think the other, obviously, the opposite direction is something that we see where like the Soviet Union had extremely high quality education and worker abilities and yet was unable to produce things very effectively.
Seth: Yeah, Soviet Union is kind of a counter example because that’s an institution story for why things don’t work out, not a complementarity story. So it seems over determined, could have been a lot of stuff going on. All right, next one. This one’s a little bit, I think we’re going to be agreeing a little bit more with, firms hire workers of different skill and produce different quality products. And there’s a great quote here.
Andrey: Yeah, all right. Yeah, it is a bold claim to- Paper, yeah.
Seth: In many industries, different firms hire different qualities of workers. Restaurants, for example, come in a range of quality levels. McDonald’s does not hire famous chefs. Maxim’s does not hire teenage waiters. Charlie Parker and Dizzy Gillespie work together. So do Donny and Marie Osmond. Some of those references quite live. I have no idea who those dancers are or singers.
Andrey: Yeah, I mean, this is a warning about putting dated references in your papers. There needs to be like an AI that opts that like auto up. You know, McDonald’s does sometimes hire famous chefs, I think.
Seth: Nervous. Right. And I think that’s actually an issue here. Right. And we’re going to come up. We’re going to come to this in the next bullet, which is this sorting is far from perfect. Right. So I definitely do believe that within a production process that’s a weak link production process, you do have to think really carefully about investing in a quality level such that you don’t blow up the thing at the last step. Right. That makes sense. That’s in this paper. However, a firm isn’t a production function, right? If you think about a conglomerate that has one business that’s doing financial lending and one business which is building jet engines, there’s no close complementarity between the banker and the financials decision and the engineer and the jet engines division. I agree with the concept, but I don’t think that concept is a firm in real life, right?
Andrey: Yes. Well, there are kind of like other reasons why firms might hire workers of similar abilities. mean, I think one version of that is cultural norms is certainly one, Sure. And this kind of explanation also for like wages of workers across different occupations within the firm. But like if you have very wealthy partners at your firm, then even they might feel bad paying a secretary low wages. And then at the same time, given the high wage of the job, you’re going to get the best secretaries to join. There’s kind of like a reverse causality that, call the factor, it might be something about cultural norms, the fairness within the firm, rather than the all-ranginess of the...
Seth: Right, a related idea is just an institution of profit sharing, right? So a lot of like places have a norm that if you have a windfall year, everyone should get bonuses, right? And that might not be like optimal behavior that could just be an institution.
Andrey: Yeah, or just another like a brilliant. Yeah, like let’s say a firm is very cool, you know, then lots of talented people may want to work there and potentially even at lower wages because of, you know, all else equal because of compensating differential. They all want to say that, you know, they want to work for the NBA, right? Even if, know, if you’re not a basketball player, maybe, you know, you don’t need as much skill in that job, but it’s cool. It’s a cool job. Right. So.
Seth: onions
Andrey: So once again, you’re going to select for high ability people who care about this common factor about this firm. And that’s going to create correlations.
Seth: I think I want to push just a little bit harder on that kind of like cultural lining up idea, which is just the idea that you could have a theory that firms end up with homogeneous groups just because like that’s how management works. It’s like it’s easier to manage a homogeneous group. And then it like it just so happens that homogeneous groups are of equal skill levels and demand similar wages. Right. So you could get there otherwise. A closely related hypothesis is there is a positive correlation among the wages of workers in different occupations within enterprises. This is just kind of a different way of saying the same
Andrey: Yeah, I agree. I think it’s kind of the same thing. Yeah. I think one relevant piece of evidence there is there’s this literature on domestic outsourcing where essentially like functions of the firm that were like done inside the firm, like janitorial services get removed from the firm and then get contracted out to a different firm. And that allows there to be a cultural break in the wage paying of the janitors. Essentially, it’s a way for firms to get much They would say much more efficient janitorial services.
Seth: I think the way to say it would be commoditized, right? There’s huge boundary of the firm issues here, right? So I gave the example of conglomerates, right? That’s kind of like stuff that’s in your firm that shouldn’t be. And now you’re giving like the other example, right? Which is if you outsource a thing, you know, maybe you outsource the high skilled thing to the consultants, you outsource the low skilled thing to the janitors. I mean, it’s not necessarily the case that you would only outsource the things that aren’t at exactly your skill level. Maybe you’d be more likely to. but you outsource things for all sorts of reasons. One thing I wanted to ask you about here, Andrey, is I guess two things. So point number one is I think a corollary of this that I don’t think is really teased out in the paper but immediately falls out is the large firm wage premium. Maybe he does mention that at some point. And that seems to be a really well established fact. Larger firms pay more, I’ve looked into that. But then I actually think about, okay, so what are the largest firms? And I think about something like Walmart. Like it’s a little bit hard to convince me that Walmart is the biggest firm because it only hires the strongest workers, right? So help me think about what’s going on in the Walmart case.
Andrey: Yeah, mean, Walmart has, I’d say, like a management system that is very good at taking in, let’s say, lower skilled workers and making good use of them. And that for Walmart, I think explains a lot of what’s going on with the firm size, right? Because most of Walmart workers are not corporate.
Seth: Is the way to think about this is almost like a conglomerate of two production functions, right? There’s the managerial production function where we’ve got lots of super high performing people who are close complements to each other. And then there’s like the store level, essentially commoditized labor that isn’t a close O-ring with everything else. And we should almost think about two production functions in the firm.
Andrey: That seems probably right, yeah.
Seth: Right. It’s just a challenge of bringing these concepts to the real world, right, given that firms don’t mean firms, right?
Andrey: Yeah, yeah
Seth: See what I’m saying? Firms don’t mean production function and he wants to talk about production function.
Andrey: Yeah, I mean, there’s a version of that. So we like subset to like a particular like banking or tech, isn’t it? Those are both interesting cases to walk through. Like it is true that big tech firms like, you know, I’m thinking about Google here, like have a very high talent level. If you’re in tech and you’re a very good firm, you’re presumably a growing firm. So eventually you’ll get to a large size. In finance, it’s a little trickier, I’d say. I don’t think very many people would argue that the average quality of an employee in corporate investment banking is better than at a top hedge fund, which probably has way fewer employees. But I’m sure there’s a positive correlation between size and finance and productivity. it’s very easily, you can think of very strong counter examples amongst subsets.
Seth: Right. Even at the firm level, can eat. said that this like the story works the strongest at the firm level. It’s to work strongly at the firm level. But even then, you can tell stories where it doesn’t so well. The next bullet was firms only offer jobs to some workers rather than paying all workers their estimated marginal product. This one I was kind of an eyebrow raise about. Right. I know in equilibrium, everybody’s going to end up at the same tier in his model. But like, you know, if I had a really big compensating differential and wanted to work at a firm, that was like way shitty, way less good than I am capable of working at. Like fine, I don’t see why this is ruled out. If in the background, you know, workers have compensating differentials, it seems like there’s no reason you shouldn’t make an offer to everyone at their marginal product.
Andrey: I guess there’s just kind of this implicit assumption here that you need like one person in each slot and so then there’s no reason to hire two people with the same skill set if that makes sense. Is that one defense of this?
Seth: I think that’s part of it. But even then, suppose I only have one slot for CEO. so I’m the CEO who’s not good enough to work at your company under normal conditions, but I’m willing to take a negative wage, right? Because I just want things.
Andrey: And actually that never happens. That’s amazing. That’s a great example, right? I think actually lots of people would be happy to take a negative wage to be CEO for a lot of firms and yet they don’t know unless
Seth: I’m trying to think but like maybe we could find like a sports example. I don’t know. Okay, fair enough All right, maybe that one does hold All right next one income distribution is skewed to the right and so here the argument is Because every firm is the product of everybody’s like positive value input qualities well Anything squared is going to be more skewed to the right than the thing itself, right? Yeah
Right-Skewed Income and Superstar Effects [43:13]
Andrey: Maybe this is the IO for me, is I think like people who get really high wages, although income can be thought of as broader, but let’s think about the wage distribution, have some sort of differentiated ability. And I guess this model does deliver that in that there’s just a single dimension of quality. And then your differentiation is that you just have high quality and those at the very, very top of the quality, aren’t very good substitutes for them. And therefore that explains their skewed income. I guess I’d want a richer, a bit of a richer model to truly explain distributional skews, like give something, anything about like superstar athletes or superstar media creators. It’s not like a single dimensional quality thing that’s going on. It’s actually just something about them. That means that there are no close substitutes in or it depends on the I guess in sports there are close substitutes. Let me take that back. Sports are close substitutes maybe and media and other and other places they’re not. Now that I’m thinking about it, maybe it’s just different cues for different industries, but then cues of fine abstraction.
Seth: board with this, I think you can tell a complementarity story at a couple of different levels that’s important for thinking about skew right production functions. If we think at the firm level, right, I go back to kind of the Rosen superstar story. So we’ve already, let’s just take as a brute fact that the firm distribution is right skewed, right, which it is, it’s kind of log normal, even power law at the top. Given that, right, if you get a sortative matching, course, incomes are going to be skewed because I’m a compliment to Walmart and you’re a compliment to the 500th biggest firm. So of course, our incomes are skewed. How do I think about complementarity as leading to skewed income distributions? The way I would think about it is just accept as a brute fact that the firm size distribution is skewed right. So the firm size distribution is log normal or even power law distributed. So very extremely distributed. And then, of course, it makes sense. that the CEO of Walmart is going to be paid a lot, a lot more than the CEO of the 500th largest firm. In other words, complementarity plus the firm size distribution is skewed gets you the income distribution is skewed. You could think about that same story, superstar story for superstar surgeons and rich people, superstar personal services and rich people. way I think you can get to the skewed right income distribution as arriving out of O-ring complementarity is within the individual. So we talked about this a little bit earlier, right, which is you can also think about individual ability as being combined. You know, my strength plus my speed plus my hard work determines, sorry, times determines how well I’m doing. If you imagine each of those individual components are random draws, Well, Andrey, what happens when I take the limit of the product of many positive random variables? What distribution do I get?
Andrey: PowerLog, your favorite.
Seth: Yeah, not quite. You get a log normal, right? No, no, I’m almost at power law. The product of many positive random variables is a log normal distribution. That’s the multiplicative version of the central limit theorem. Central limit theorem is some of a of stuff is normal. The product of a lot of stuff is a log normal. Now, if you keep, keep multiplying, that log normal gets more and more stretched out into a diagonal power law. So, Andrey was one step ahead of me. But why do I bring this up? I mean, I think about, have you heard of these things, Lotka curves, Andrey? So a Lotka curve is...
Andrey: It’s like Laka Volterra model.
Seth: No. I don’t think so. No, guy’s like a 1920s chemist. This is back when you just have curves named after you. What is the Lotka curve? It’s when you draw a distribution of the rank of someone in terms of individual success against some numerical value for their success. So you can do this for like number of scientific publications.
Andrey: Lotka.
Seth: You can do this for like points scored in basketball or lifetime RBI’s. So you get this distribution of kind of individual successes. And what you find when you draw these curves is that they are power law or log normal, right? And so the way that I think about why does that happen is, that individual success you can be thought of as O-ring in all of these different components. In order to be Novak Djokovic, you’ve got to be amazingly innately talented and get training when you’re a young kid and be innately dedicated and that way you become a superstar and you have 24, you know, tennis championships. Whereas if you miss just one of those weak links, you’re going to be in the mass of mediocrity. So I guess income distribution being skewed to the right, the story we get in Kremmer’s paper is that this has to do at the firm level with complementarity. But I think even individual level complementarity with inability still gets you this skew distribution.
Andrey: Yes. Now I will like add like a wrinkle to some of these explanations. So like the way the paper is written is really about like true productivity, like production. And I think about like something like sports from a societal point of view, it truly doesn’t matter whether LeBron is the star or the like fifth best player is the star. Right? Like these are like not like I think demand for basketball would be exactly the same. Some stars are obviously more photogenic, but those are not aspects of the on-field play, if that makes sense. I think it’s actually quite similar with the superstar doctor and the rich guy. It’s not like the superstar doctor. could be that they’re truly a much better doctor, but actually, I doubt that’s true. could just be one. Yeah, exactly. And so then both of those phenomena are in some sense, like not very important for economic productivity or growth or GDP, right? Whereas what we really care about is something like science, having O-Ring in science means that we get much better science. And that’s not just like a relative contest, but an absolute contest, because it’s an objective.
Seth: Citations might be a relative contest. don’t know. That’s the only issue there is how do you measure absolute progress in science?
Andrey: No, no, sure. like, I read citations as a metric, just to be clear, especially in our field. mean, economics, fine, but like business school, academia essentially has produced epsilon of value to society and has a lot of citations. So incredible. Hi, business school professors and students. Thanks for listening. No, but seriously, right? Like you have, you have these people on the top of citation lists. you know, you look at all their papers and they’re like, if they didn’t exist, would anything change in the world? And the answer is no. in contrast to O-rings, know, which has changed how a of people think about things. So yeah.
Seth: It really changed. If the O-ring isn’t there, the space shuttle blows up.
Andrey: Yes. Do you want to talk a little bit about the imperfect information version of the model? Because that’s kind of what he then devotes most of rest of the paper to.
Imperfect Information, Education, and the Big Sort [51:13]
Seth: Yes, that’s right. So I already started introducing it just a little bit, right, which is if you think about if a firm has many uncertain components, or even if within an individual, you can’t select how good you are at these different things. So you have kind of uncertainty in all of your individual abilities. I’ve already said what happens when you multiply a lot of positive random variables, you get a log normal distribution. Interestingly, that does not come up in this paper. I control F. for log normal, I couldn’t find it, but that is what you would get. And so he gets kind of three results that kind of come out of this sort of uncertainty. Kind of the big takeaway I took about from the uncertainty is when there’s uncertainty, you want to use the most valuable, the best workers for those very last steps, right? You know what? When you’re doing primary production, sure, any, any schmuck can work on that. And if they screw up, fine, you just dig another hole. But when we’re doing the ninth layer of painting on the Rembrandt painting, you got to make sure someone’s not going to screw up all of the painting that went under that. And so the two kind of development results that fall out of that are poor countries have higher shares of primary production in GNP, right? So there’s less to screw up at earlier stages. So you’re going to specialize in that if you have low quality workers. And then secondly, go ahead.
Andrey: Seth, I’m confused. I thought we were talking about uncertainty and now you’re telling me about average low quality workers.
Seth: The connection I see is you don’t know the quality you’re going to get at every stage of production. So I’m going to do a first stage of production.
Andrey: You can’t sort. You can’t sort. That’s the key intuition. You can’t sort people because you don’t know who is who. And therefore, you switch to production processes that are less, you know, the chains are less long, the production chains are simpler. And, you know, those are typically less valuable.
Seth: here to Exactly. And I guess we can think about sort of if there’s imperfect observation, right, then you’ll still see a sort of matching of the stronger workers in expectation to the later stages of production. Good point.
Andrey: But then, and this is an interesting point that he makes, it’s almost an aside here, is well, then you have these potentially multiple equilibria for society. So if society invests in education and sorting, then it can sustain very complex production processes that increase GDP. On the other hand, if everyone thinks that no one else is going to get educated and else is to get sorted well, then why would they invest in their own education? Because being high skilled is not as valuable in that situation. Alternatively, they might just leave the country if they get educated. so, societies can fall into low-skilled traps versus high-skilled boons. And it relates to a concept in the internet discourse, which is the big sort. So, think Patrick McKenzie likes to use this metaphor where or like the sort where society has gotten a lot better due to information technology in sorting people by their abilities. And so, you know, back in the day, you might have someone very talented, you know, working at a very local, you know, small business. But these days, because it’s so much easier to find out, like how to break into more lucrative, higher impact industries. And at the same time, it’s much easier to signal objective quality, although maybe we’re ruining that ability with our current set of technologies. You get people sorted into very talented people end up being in the highest marginal productivity jobs, right?
Seth: Right, get the impoverishment of the periphery and the enhanced- everyone skilled goes to the Metropole. Yes. And... yeah.
Andrey: Yeah, or even like they go into tech, like they go into tech or even if you’re like, you know, like maybe let’s give an example. Like, let’s say you’re like an expert in governance or governments. You might start working for Anthropic all of sudden, you know, because they’re going to, you know, an equilibrium, they pay you a lot more for that set of skills. If you’re very good, then if you’re working at a think tank or, you know, a local government or something like that. That’s kind of most of the paper. yeah, where do you want to take it now?
Seth: I’m ready to go into posterior’s if you want.
Andrey: All right, let’s posterior.
Sponsor Break [56:02]
Seth: All right, and so now we’re going to take a little break for you viewers who are playing along at home to contemplate your posteriors and to see whether any of these ideas have changed your views on complementarity in the economy. This chance to contemplate your posteriors is sponsored by Revelio Labs. Revelio Labs is a leading provider of labor economics data and data services for companies, academics, and independent researchers. Andrey and I have been working in economics of AI for a long time, and we can confirm just how useful Revelio’s data is. Revelio’s team combines comprehensive micro-level data on employee professional profiles, job postings, and employee sentiment with standardizations, mappings, and enrichments available, all to make that data useful without making your modeling decisions for you. The data can be flexibly aggregated to company, market, or industry, and be used to study questions ranging from career trajectories to occupational transformation to the returns to skills and the impact of AI on labor demand for tasks. Can’t imagine anyone be interested in those. And Revelio data is available on WRDS. So if you’re an academic with a good library, you might already have access. And if you don’t, you can reach out to their excellent economics team and they’ll hook you up. So now we’re in our posteriors. So the first thing we thought about, Andrey, are do we think that O-Ring is the best explanation of the positive correlation between workers within firms and then the big gaps of incomes across countries? So big gaps across firms, gaps across countries. I think you started off 70 percent yes for firms, 30 percent for countries. Do you move having refreshed your memory about O-Ring?
Posteriors [57:54]
Andrey: 60
Seth: 60 and 30, excuse me.
Andrey: Yeah, I think I’m maybe a little more on the firm side, Yeah, 65. Yeah, pretty similar on the on the countryside. I think there’s an element of this paper that it’s very simple. And it seems like it’s micro founded in some weird way. Like, yeah, you have this production process, which, you know, they’re like end tasks and we just multiply them together. But it’s actually not that micro founded and kind of you could do a lot more like you can try to model the firm. hierarchy and how different types of firms compete with each other in a much richer way.
Seth: You could think, yeah, I mean, all the things I would want to endogenize. You’d want to endogenize investments in de-complexifying and changing how tasks are bundled. You’d want to endogenize the complexity. You can pay money to have a more complex production function that might have a larger TFP term, but would have more steps that could fail, right? There’s all these margins you’d want to let firms invest in.
Andrey: Yeah. And then at the macro level, like, it becomes harder to think about what are these cues, you know, how do different industries combine in a total production, if that makes sense. And so if you take the analogy of its broadest sense, we can fit a lot in here. If you take it at a very narrow sense, you can fit, you know, relatively little here. But I think just, there’s very clearly showing this force for complementarity. and some simple implications, I think it’s very neat paper.
Seth: Right, agreed. And again, I would agree with you, Andrey. Maybe the reason is I read this paper just a year ago, so it’s a little bit fresh in my memory, so my priors weren’t able to move that much. But yeah, it’s still come away thinking that this is an important force for those issues that it brings up. Now for the spicier posterior, which is around how do we think about how AI interacts with this, right? Should we be thinking about the Acemoglu-Restrepo model or this model? How would they differ in terms of implication? Where do you come out from that?
AI, Programmers, and Increasing Complexity [1:00:07]
Andrey: Thinking through a little bit more about this, it really does show how in certain cases, having access to very powerful AIs, as long as not everything is automatable, can result in a lot of returns for people with high skill. And it kind of pushes against some of the narratives about, let’s say, programming jobs going away.
Seth: Anyone at the bottleneck?
Andrey: Not all programmers, but a lot of programmers are extremely skilled, have a high quality. They’re smart people. They have a lot of agency and capability. it’s not crazy to think that they’ll find ways to make use of all these AI capabilities in a way that actually increases their earnings. I don’t know what share of programmers that is, but I’d imagine it’s probably a higher share than some of the AI labor doomers suspect. There’s definitely a type of programmer, often one doing outsourced or commoditized programming, that I think is really highly at risk here because I just don’t, I think their q is too low to take advantage of this. But I think in the US, a lot of programmers have a high queue and I can imagine them doing a lot more stuff.
Seth: A lot of AI podcasts talk about agentic AI. Here we talk about agentic programmers. Let me tell you the way that I’m thinking about this question, which is, first, I want to make a caveat, right? And I talked to Andrey about this for a second before the show. It’s not even clear to me that there is a direct contradiction between this model and the Acemoglu-Restrepo model, right? If you think about a Cobb Douglas production function,
Andrey: Yes, yes.
Seth: That’s one where a bunch of terms raised to a power are multiplied by each other. So if you just raise that to another power, you’ve got a multiplicative production function, right? So there’s some sense in which this is a special case of CES with a scaling. Now, of course, lots of different things fall out of it because Kremer here is going to make different assumptions about what inputs do you get to pick and what inputs do you not get to pick. And I think that is important.
Andrey: To be clear, the CES function that’s used in the literature is very far from this one in the macro literature, in the calibration of it.
Seth: Yes, the calibration is going to end up differently because they’re going to end up thinking about different measurements, right? I’m just making the simple point that K to the alpha times L to the alpha. Yeah, I get it. All right. The next point I want to think about is what happens in a model like this when there’s a technological shock and we get access to a new production technology that either increases or decreases the amount of difficult steps?
Andrey: Yeah.
Seth: and also increases the kind of the TFP part, right? And so, you know, the TFP parts gone up. So now we want to switch to either the more complex or less complex production function, right? By complexity here, I mean the number of complementary steps that have to go right. And you’re to get very different answers, right? More steps is going to be more skewed, more extreme gaps between rich and poor, productive and unproductive, stronger sort of matching, right? if you get more steps. So then to me, the question becomes, is AI going to increase the amount of steps or decrease the amount of steps, right? That’s the $64,000 question. Are we mostly automating uncertainty or are we mostly creating huge new vistas of much more technologically advanced things than we could have ever contemplated before? Andrey, I have to be the two-handed economist here and say kind of like, both things seem to be going on. There’s definitely a case in which we are automating uncertain steps. And there’s definitely a case in which we are making more complexity and new steps possible. Who’s going to win that race? My guess would be on betting on things getting more complicated, right? I think that’s the general trend of society is towards higher steps on the value chain, increasing complexity. And therefore, reading a model like this leads me to think that AI is going to do things like increase income inequality, increase the inequality across firms in terms of firm size and productivity.
Andrey: Okay, well on that note, thanks for joining us for another episode. Yeah.
Closing [1:04:44]
Seth: Please join us on Discord.
Andrey: Discord Seth is trying to make it happen. Maybe it will
Seth: We’re going to make the discord happen. All right. Andrey, give us the outro one more time.
Andrey: Keep your posteriors justified.
