Daron Acemoglu on pro-worker AI and vibes based capital
0:00Daron: Whenever I talk about pro-worker AI, which I’ve been doing quite often for about ten years, when I talk to CEOs or managers of companies, they are on board. Whenever I talk to technologists, I get three reactions. One is, “We love it. We’re doing it already.” Second, “It’s impossible.” Third, “Yes, just wait for it. We’ll get to AGI first.”
Seth: Welcome to Justified Posteriors. This week, we’re honored to host Nobel laureate Daron Acemoglu of MIT. He’ll answer the question “What happened to liberal democracy?” — the title of his new book. But first, we’re excited to learn about his evolving views on AI, and especially to pin him down on what pro-worker technology actually means. I’m Seth Benzell, always racing with the machine — and I don’t mean using a treadmill — at Chapman University in sunny Southern California.
Andrey: And I’m Andrey Fradkin. Very excited to have Daron on the show. He was an inspiration for me going to economics grad school, so this is truly an honor. Coming to you from San Francisco. Welcome to the show.
Daron: Thank you, Andrey. Thank you, Seth. This is my true pleasure and honor.
Seth: Thank you, Daron. Okay, so we have limited time, so we’re going to jump right into it. In a recent series of essays, including a paper with David Autor and Simon Johnson called “Building Pro-Worker AI,” and a new piece out in Nature about why we shouldn’t be talking about AGI, you’ve been calling for this thing called pro-worker AI. And of course, we love workers here on our podcast.
Daron: We are all workers.
Andrey: Yeah, it’s true.
Seth: We love the workers here. But as we’ve been engaging with this, we want to help you steelman it for us, and help us understand some bits about it where we still think there may be some tensions.
Daron: Help me or destroy me. Let’s go for it.
What Is Pro-Worker Technology? 2:00
Seth: All right, let’s do it. So the first question is: what is pro-worker technology?
Daron: I would like to step back and say that, in general, in economics, sometimes we black-box what technology does. That’s not helpful, broadly speaking, but it’s particularly unhelpful when it comes to AI or other digital technologies.
Seth: Right. AI is not a TFP term.
Daron: AI is not a TFP term, and it’s not a plain vanilla, garden-variety, labor-augmenting technology consistent with balanced growth — because we’ve assumed it has to be consistent with balanced growth. Most technologies make you better at one specific thing. They may make you less relevant by automating the specific things that you were doing, going back to the spinning and weaving machines, all the way to robots and other advanced machinery on the manufacturing floor, and now AI. They may also increase the new things that workers do, either because the machines that are being introduced require human input, or because the new technologies — directly or via organizational changes — create opportunities for workers to perform new tasks or gain new expertise. So my point, in general, is that these technologies have very different implications for worker wages, employment, the share of capital and labor in national product or sectoral value added, and, via their social effects, they have very distinct implications for liberal democracy, for social peace, and things like that. That’s why we should care about all of these things.
Seth: Great. One way you put it in your recent paper is that you define pro-worker technologies, including AI, as technologies that make human skills and expertise more valuable by expanding worker capabilities, which is aligned with what you just said.
Daron: Exactly. So pro-worker AI is that last category, Seth. Exactly.
Seth: All right. So staying there were a couple of different concepts that you brought up. You brought up the labor share being high as something that’s desirable. Making skills and expertise more valuable in these new tasks is something that’s desirable. And also this idea that when you create new tasks, it’s going to — maybe inevitably, maybe most of the time — get you these other things. But I’m sure you’ve thought this through. You can think of scenarios where you invent a new task for a worker in industry A, and then through some sort of general equilibrium effect, that’s going to reduce wages, or the wage bill, by more than that in industry B.
General Equilibrium, Sectoral Reallocation, and the Labor Share 5:01
Daron: Wow, you really want to go to the advanced material.
Seth: Well, this is the general equilibrium econ of AI podcast.
Daron: Boy, that’s the topic that I discuss only under very, very protected circumstances, where non-economists are not going to be bothered by it.
Seth: Only economists in the room. I know you don’t get to put this in your Nature essay. Okay.
Daron: But let me again take a step back before answering your question, which is an excellent question, and you’re right. Let’s first say that everything I’m going to say is dependent on a particular framework. No conceptual framework, no coherent discussion. And I’m using the task-based model.
Seth: You’re thinking about “The Race Between Man and Machine,” maybe.
Daron: Yeah, that’s right. That’s one example of that framework. Somewhat more general versions of it have been developed since that paper. So in that framework, the very unambiguous prediction is that whenever you do automation in a sector, that reduces the labor share in that sector.
Seth: In that sector.
Daron: In that sector. And whenever you introduce new tasks, that increases the labor share in that sector.
Seth: Very well put.
Daron: And those actually receive ample empirical support as well.
Seth: Yes. For example, your papers with Pascual about automation in the auto industry.
Daron: Well, David Autor’s papers, several other papers on robots, et cetera. The overwhelming number of them — and especially if you zero in on those that are well done — find these labor share results. Now, the general equilibrium effects of any technology, not just new tasks or automation, are different, and that’s something Pascual and I also address in some of our papers. It’s a very simple formula that you can have at the back of your mind if you want: any sectoral reallocation is going to change the aggregate labor share via the difference between the labor share of the expanding and the contracting sectors.
Seth: Makes perfect sense.
Daron: So in other words, if a labor-intensive sector contracts and a low-labor-share sector expands, that’s going to reduce the aggregate labor share. Hence, the scenario that you point out is entirely possible if, associated with new tasks, the sectors that have low labor share expand and the sectors that have high labor share contract.
Seth: So maybe one idea I have in mind here: you might imagine the movie business is invented. This creates all sorts of new tasks. I’m in LA these days. You can create all of these new tasks in the movie industry, which has a labor share that’s maybe eighty or seventy percent, and you’re crowding out the local theater, which had a labor share of ninety percent.
Daron: One hundred percent. That’s possible. But again, that’s only relevant if there are really big sectoral reallocations —
Seth: And really big differences.
Daron: And really big differences. In one of our papers, Pascual and I did that decomposition for the aggregate US economy from the 1950s to the 2010s, and we found that those sectoral reallocations were on average neutral, meaning that the expanding and contracting sectors roughly had similar labor shares. That doesn’t mean that moving forward with AI it’s going to be the case. But from a recent economic history point of view — Baumol effects notwithstanding, which are very interesting and specifically relevant for some types of skills — for labor overall, that didn’t seem like a huge deal. Moving forward, that might be different.
Why Care About the Labor Share? AK Growth, Dignity, and Liberal Democracy 8:56
Andrey: I was going to push back a little on the labor share. I understand that there are reasons to care about it in and of itself, but presumably if productivity is increasing a lot, then the labor share can go down and people are a lot better off.
Seth: Yeah — maybe we should be talking about reallocation to industries with better or worse median wages, right?
Daron: One hundred percent. I also agree with that. There are two different ways of thinking about the labor share, and then there’s a third reason to care, which I’ll come back to. One is that if we go from a sixty-six percent labor share to a sixty-two percent labor share and stay there, and in the process productivity is much higher — who cares? I’m willing to go with that. However, if there is a steady decline in the labor share, asymptoting, ultimately going down towards very, very low values, I think that is a cause for concern even if wages increase, because it means that labor is becoming more and more sidelined in the economy.
Seth: This is interesting. I had this as a topic we would talk about later — thinking about automation and capital accumulation dynamics in the long run. But I can tell a really positive story about the labor share going to zero, right? You think about moving from a Solow universe, where we have a cap on output because eventually growth asymptotes, versus an AK universe where you can get growth forever. So why not AK-universe growth forever as the thing we should be aiming for?
Daron: That’s a much broader topic.
Seth: Right. And so that’s going to be a political economy answer. Okay.
Daron: Before I come to that, I’ll just give a third reason, which I said to Andrey. From a practical point of view, if you want to think about what is going on with wages — which, Andrey, you said very much we should care about — then the labor share is a driver of that. If you want to see what’s going on with the wage, that can always be decomposed into something like a productivity effect, a displacement effect, and a reinstatement effect, and the labor share tells you about one of those components. So it’s also a diagnostic tool to think through, even if we don’t care about it ultimately. All right. Now, coming to your question, Seth. Since I’ve just published a book about liberal democracy —
Seth: Plug the book. Do it. It’s good.
Daron: Thank you. That’s great. All right, we can finish our interview now. I got what I wanted. So I think that a world in which labor becomes dispensable — where the AK model would be one version of that, where you don’t need labor for anything — would not be consistent with liberal democracy, or any type of society in which people are viewed as equals of those who control the machines, who control the AI models. It would be a very hierarchical, almost two-tier world. So I think there are a lot of problems with a world like that.
Seth: You basically think wealth dynamics in the long run are always more unequal than labor distribution dynamics, something like that.
Daron: Well, I think you need labor to be essential in some way, so that, A, the vast majority of people actually have some contribution to society, which is important for their self-worth, which is important for their dignity and all sorts of things. And it’s also critically important, as you anticipated, Seth, for the political economy. Because if they are dispensable, why would anybody listen to them?
Is Technology Directable? Three Reactions from Technologists 12:59
Seth: Okay, we’ll come back to political economy. Let’s stay on pro-worker AI. The second question I have here is one that, whenever I hear you talking about this, I want to jump into the podcast and say, “this is the question Andrey and I have.”
Daron: Now is your chance.
Seth: Here’s my chance. When you talk to technologists and you say something like, “We should direct AI to be more about healthcare AI instead of military AI,” the way they say it is, “AGI is on the golden path to getting whatever technology you want.” They’ll quote Sutton’s bitter lesson at you — this idea that in the long run, the generalized, heavy-compute technology will always beat the specialized technologies. Whatever individual expertise-enhancing, task-expanding technology you have in mind, the shortest path to that is AGI first and then that. So is technology directable in the sense that you have in mind, if that’s the case?
Daron: Whenever I talk about pro-worker AI, which I’ve been doing quite often for about ten years — when I talk to CEOs or managers of companies, they are on board. Whenever I talk to technologists, I get three reactions. One is, “We love it. We’re doing it already.” Second, “It’s impossible.” Third, “Yes, just wait for it. We’ll get to AGI first.” And all three of those, I cannot disprove. They might have some truth to them.
Seth: But you can’t disprove anything. That’s what your book’s about.
Daron: Right. Exactly. All right, you got that one. So it may well be that some of the things that some of the companies are doing are already pro-worker, and I’m not seeing it. It’s a possibility. Though I must say, when I see reports from companies like Anthropic, et cetera, claiming, quote-unquote, “augmentation,” typically it’s not what I would call pro-worker AI. It’s using AI to do some of the things you were doing before a little bit more efficiently, or something like that. That’s just the standard labor-augmenting technology, which has very, very different implications.
Seth: I think what it also suggests is that doing that taxonomy just from looking at the technology is extremely hard — that you really want to understand how wages and labor shares are changing in order to back out what kind of technology it was. But keep going.
Daron: Second, the statement that this is impossible — the version that I think is not disprovable is that we are ultimately, and in the relatively near future, going to get to AGI. Once you get to AGI, you’re going to get to ASI, artificial superintelligence, and then at that point, or somewhere along that spectrum, whatever workers could do in new tasks, whatever they can do in any place, can be done better by machines — robots with AI, or AI by itself. So in that case, nothing can be pro-worker. So again, I cannot disprove that.
Seth: Well, nothing can be pro-worker in terms of labor share?
Daron: Nothing can be pro-worker in terms of anything.
Andrey: Well, there are relational goods, right? There’s still going to be work to do —
Daron: If it’s superintelligent, then it can also replicate those relational things very well.
Andrey: So there’s that argument, but I actually think that’s a wedge point. Some people think that those jobs will remain even in a world of ASI, and others do not. I think there’s some debate there.
Daron: I don’t see how they could remain, truly. If you define AGI and ASI in a coherent way — that it does everything better than humans — then it must be that it can do the relational things better than humans as well.
17:01Seth: Can you explain “everything better”? This Ricardo guy told me once that we would always have a comparative advantage in something.
Daron: Not necessarily. You would have a comparative advantage, but if something else has absolute advantage in everything and is extremely cheap, why is your comparative advantage of any use?
Seth: Why did England trade with Portugal, even though Portugal was better at everything?
Daron: Because nothing was very, very cheap.
Seth: Right. It was just a little bit cheaper in Portugal.
Daron: A little cheaper. So if you literally have ASI, at least in the way that I’m defining it, and the marginal cost of producing these ASI services is very, very low, the only way human labor can compete against that is by becoming very, very cheap — which humans cannot survive on. It will fall below the subsistence wage.
Seth: Okay, fair enough. And so now, number three, the sexy one. Tell us about it.
Daron: Number three is essentially also a coherent argument: that ultimately, on the path to something like AGI, you can do things in different ways, but once you are at AGI, the best way of doing everything is using AGI. This does not say you cannot redirect AI. You can redirect AI, but it would be futile, because ultimately the best thing from a productive efficiency point of view would be to just let it rip with AGI and ASI, and then we’ll use them in whatever way we decide to use them.
If You Were Dario: Frontier Models vs. Applications 18:29
Andrey: So I have a question for you. Let’s say you’re Dario, and you get to choose what Anthropic works on next. How would you go about making that choice pro-worker?
Seth: Dissolve Anthropic.
Daron: No, no, I don’t think so. Anthropic is a very, very creative company. The problem we have right now in terms of productivity is that we don’t have enough applications. We don’t have applications that are custom-designed for specific tasks, that are easy to use, that can be integrated into different organizations and enable those organizations to do what they were doing more efficiently, or do new things. I think the first thing I would want to do is create an effective production line for applications. And then, in the context of those applications, you can relitigate the question that Seth raised, which is: what’s the best way of making these applications better, more effective, more usable in the short and medium term — short being four or five years, medium being ten to fifteen years? Is it to put all of your energy and best talent into the underlying Claude model, or to actually use some of that to improve those applications? And if so, what are the places where we can make those applications more pro-worker?
Seth: And your hypothesis is that if you did that analysis — if right now their knob is at ninety-five percent frontier model and five percent applications — maybe that knob should be at eighty percent.
Daron: Eighty, seventy, sixty. We really want the applications. Right now, if you are very knowledgeable as a programmer, you can use Claude to do many, many things. But most companies aren’t.
Andrey: So I’m a little bit confused by this statement, in the following sense. I see tons of companies trying to build applications for AI. There is even an argument that they’re all using the Anthropic models. One very plausible thing is that Dario can make the model more capable, and then the application layer can develop it for specific applications. And now we can try to go through each company one by one and say, “Is this pro-worker or not?” I think that might be a little tricky, but certainly people are trying. I don’t have the same sense as you that people aren’t trying to make the technology pro-worker.
Daron: Look, things are changing fast, and you guys are in San Francisco. I’m in the less de—
Seth: Andrey’s in San Francisco. I’m in LA.
Daron: LA, okay, fine. The West Coast, let’s say. Sorry. I’m in the less developed part of the world, on the East Coast.
Seth: You’re off the technological frontier. “Diffusion’s more important than pushing the frontier,” someone not at the frontier would say.
21:36Daron: Exactly. So I’m not seeing all of the different places where creative energy is going. But if you look at the pre-agentic age, up to the beginning of 2026, say, it was remarkable how few usable applications we had. Perhaps that’s changing now, and next year we’re going to see a lot of these applications. We’re going to have companies properly start integrating some of those technologies much more into their production line. Then you may be right, Andrey. But again, I have not seen that yet. If you look at where a lot of the spending is going, in terms of both talent spending and compute spending, it seems to be the development of the general model and the AGI agenda.
Will the Market Sort It Out? Vibes-Based Capital and Two Market Failures 22:31
Andrey: Just a follow-up on that: how do you think about the market forces here? There are going to be a ton of startups. There are already a ton of startups. Some of them are very focused on augmentation, some of them are more focused on substitution. But isn’t it likely that the market will determine which of these technologies wins? I know you have policy proposals to try to address this, and I’m curious whether you want to speak about those.
Daron: Okay, that’s fantastic. No, I don’t want to talk about the proposals yet — — because I haven’t convinced you that there is a need for proposals. That’s what I want to do first. I’m an economist, as you might know.
Seth: I thought you were a political commentator.
Daron: Exactly. That’s it. Thank you. That’s below the belt, Seth, but it’s fine. I’ll get back to you.
Seth: No, you’re getting hated by all the right people these days.
Daron: Thank you. Thank you. By John Warner [?]. So I’m very keen on the market, but it’s also important to understand the limitations of markets. I think there are two limitations of markets as it comes to that. Let me start with the second one, which isn’t the classic set of limitations related to imperfect information, monopoly power, et cetera. We don’t have the standard market deciding where money goes at the moment. None of the companies are making money. They’re all losing money. They’re losing a huge amount of money. The way that we’re deciding where funding goes is based on future projections, vibes, and beliefs.
Seth: Hasn’t it always been the case? How is that different from how the market’s worked in the past?
Daron: It hasn’t always been like this. Most of the time, try to take a loan from a bank as a small entrepreneur. Describe your dreams. You know how much money they will give you? Zero dollars.
Seth: You’ve got to be in the San Francisco ecosystem to get money that way.
Daron: But yeah, those previous venture capitalists, Seth — just to close the loop on this — would do that, but they would withdraw the money after one or two years if you couldn’t show returns. Here, we have created almost an ideology that the best thing is to lose billions of dollars, hundreds of billions of dollars, and then somehow, at some future date, everything’s going to be sorted out. So that is a very different way in which the market allocates capital than what we are used to. And it would not have been possible, for example, if we did not have a huge amount of private wealth from sovereign funds and very, very wealthy individuals — and some of it from venture capital — that was looking for a place to be parked. So I think that ecosystem we have to bear in mind.
Seth: Is AI a symptom of the global capital glut?
Daron: I would definitely not say that it’s because of it, but the two are synergistic.
25:32Andrey: The overabundance of capital suggests that if there was a profitable opportunity to make pro-worker AI, it could get funded. That’s one version of the argument. So you might be right that we’re over-investing in certain types of technologies, but there is a ton of capital to go into lots of technologies, right?
Daron: That’s true. There is an overabundance of capital. But two things. First of all, a lot of that is now going into data centers, et cetera. So before the generative AI boom, there was an overabundance. Now that overabundance is more tempered. Second, talent is scarce. The smaller application companies hire some very talented people, but the cream of the crop go to Anthropic, Google, OpenAI, Meta, et cetera.
Seth: Talent’s an interesting one too, because — and this is tied into the directability issue — people make a decision in their early twenties about the expertise they’re going to have for the rest of their lives. So there’s a sense in which you have to make your decision based on vibes about what this industry is going to look like for fifty years, and you might not be able to turn your expertise on a dime.
Daron: That’s right. I don’t know how important that is, though, because some of these AI skills are fairly portable. The very good talent that goes to Meta for hundreds of millions of dollars probably could have been employed in some applications doing a great job there as well. But just to close the circle: second, there are the more standard reasons why what the market decides may not be efficient, and there are two in particular that I would emphasize. One is that we have a tax code that creates a big bias in favor of capital instead of labor. That’s something I’ve written about with Andrea Manera and Pascual Restrepo. The market outcomes, conditional on that distortion, will not be optimal. Second — something I have written about with Pascual Restrepo — if there are labor market imperfections and workers are paid above their outside option, as in a search model, bargaining, efficiency wages, et cetera, then the equilibrium will have too much automation and too much demand for automation from companies, because companies automate according to the wage that workers receive. A social planner would use the opportunity cost of labor in that decision, and the wedge between the two can be quite large.
Seth: I was thinking about that paper. It’s an interesting paper. I’m sorry if you address it in there, but you could tell the same story about the other side: there’s a financially constrained company, and there’s a wedge between the interest rate and their opportunity cost of capital. Do you have an argument that that’s less important, that I missed?
Daron: No. If capital wedges take a similar form, that would also be relevant.
Capital vs. Labor Taxes, Pensioners, and Who Holds the Wealth 28:37
Seth: Okay, so let’s go back to this idea. One policy instrument it seems like you’re proposing is to rethink our capital-to-labor tax mix. This is my last question about pro-worker AI, and then we can get on to the next topic. A classic argument for wanting to favor capital over labor is that we think capital is supplied more elastically than labor, and in the long run you always want to tax the inelastic thing. So maybe you can engage with that. And then, relatedly: is a pro-worker policy an anti-pensioner policy? Professor Kotlikoff was my PhD advisor, and so I have had banged into my head that Social Security is a Ponzi scheme.
Daron: I thought Pascual was too.
Seth: Oh, I have too. That was my committee. It’s Kotlikoff and Pascual, dude. And so now you know exactly why I’m exactly this amount of insane. So is part of your argument that we’ve done too much in our political economy to favor the rentier elderly class, and we should be favoring labor incomes?
Daron: There are bigger beneficiaries of the favorable treatment of capital than the pensioners, last time I thought. When you look at who holds capital in this country —
Seth: Well, you tell me. In a life-cycle savings model, it’s the old.
Daron: Right. In a model with heterogeneous agents —
Seth: It’s the patient.
Daron: — it’s the people who get all of the capital at first, right.
Seth: So it’s a mix of that.
Daron: Yeah, it’s a mix of that.
Seth: So why are some people wealthier than others? It’s the people who’ve been accumulating their capital for the longest, and that could either be inherited wealth, which maybe we frown at, or some of that’s just middle-class people saving for retirement. You can tell me which is empirically more important.
Daron: I don’t know. I think probably in the fifties and the sixties, life-cycle savings was very, very important. But as wealth inequality has increased, I think the numbers I’ve seen suggest that more and more of it is in the hands of the very rich. There are many measurement challenges when it comes to wealth inequality — big debates between Emmanuel Saez, Gabriel Zucman, and others on the one side, and those who think wealth inequality hasn’t increased that much on the other. I find it very difficult, given the level of income inequality and capital income inequality, that wealth inequality hasn’t increased. But it’s not my area of expertise. But let me take the question that you asked, which is: is it efficient to tax capital so low? That would depend on many things. Supply elasticities is one of them. Labor market and capital market imperfections is another. And how this actually affects production decisions is a third. Our simplest models impose that the long-run supply elasticity of capital is infinite. The data doesn’t support that. In the data, the supply elasticity of capital seems to be moderate, not huge.
Seth: This is the elasticity with respect to the interest rate?
Daron: Yeah, elasticity to interest rates, yes.
Seth: Okay. Well, I just want to say that’s more consistent with an OLG model than a representative agent model.
Daron: Yeah, OLG models would have that. Exactly. That’s right. Your training again, Seth. When I said “our standard models” — sorry, I was not locating myself in BU. I was thinking of representative-agent standard models. I should be more careful.
32:25Seth: Okay, so anyway, that would be a policy that you’d be interested in, and you would be theoretically willing to trade a little bit of growth for a little bit of labor share. Is that putting words in your mouth?
Daron: Okay. So both on the issue of AGI versus pro-worker AI and capital taxes versus labor taxes, I have tried to have my cake and eat it too. I’ve argued, in fact, that the growth negatives from these policies wouldn’t be that major. In the Brookings paper that I wrote with Andrea Manera and Pascual Restrepo, for example, we make a bunch of assumptions. We use a specific model, but a relatively flexible task-based model. We calibrate the parameters of that model from existing estimates along the lines of what we were just talking about — supply of capital to interest rates, labor supply elasticities, quasi-labor supply elasticities. And actually, we find that if you went to a more equal treatment of capital and labor, you would increase GDP, not reduce GDP. But push me, and I would say yes, I’m willing to trade off a little bit of growth for a fairer distribution — especially one which would be more consistent with liberal democracy, social cohesion, the dignity of workers, et cetera.
Seth: Those things seem good.
How Close Is AGI? Pause vs. Slow Down, and Lessons from China 33:56
Daron: I am definitely not the person you should talk to if you want to understand how close we are to AGI. There are experts. But my assessment is that true AGI, along the lines of what we were talking about earlier, is not that near, and going after AGI right now is not necessarily the strategy that would maximize productivity either.
Andrey: That’s a very interesting claim, because I think there’s another policy proposal out there, which is to pause AI development. But the stated purpose of that policy is that we’re so close to AGI and we’re not ready for it yet. So would you, for instrumental purposes, support a pause on development?
Seth: Will you get in the broad tent?
Daron: Oh, man. There are two arguments out there among the slow-down crowd. One is that we are very close to AGI, and if we push a little bit more, we’re going to be at AGI and all bets are off. The second is that we are doing really badly on safety, even though we may or may not be that close to AGI. After all, Mythos was not AGI, but there was broad agreement that it should not be unleashed without any guardrails. So I think there is an argument on safety that does not necessarily rely on being epsilon away from AGI. But more directly to your question, Andrey: I have spent too much of my career studying political economy and the history of technology, where I see so many attempts at blocking technological change, that I put a high bar on saying you should block technology, stop innovation, et cetera. But slowing down temporarily in order to redirect is a different phenomenon. If you are driving two hundred miles an hour towards a cliff, you may want to first hit the brake temporarily to steer better.
But no, I’m not in favor of stopping progress, so long as we democratically agree — based on expertise, and a democratic compromise and communication — that the technological direction is socially beneficial. Right now, we have not had that democratic debate. And even if we had the democratic debate, this is an international technology that’s going to affect seven billion people. The remaining six billion, or close to six billion, who don’t live in the US and China haven’t had any say on this. So I think there are more complex issues here.
Andrey: A related question. Let’s abstain from thinking about whether the investment in the data centers is justified or not. But if we implemented a hard pause at this moment, wouldn’t that crash the economy?
Daron: Yes. There’s a problem. So that’s why, again, slow down and pause are not the same thing. I think we really need applications. Look at China. Despite what we are sometimes told from Silicon Valley, there are things to learn from China. They are behind in large language models and some applications of generative AI, but they are far ahead of the US in integrating AI into the production process, and getting it into the hands of consumers in a manner that’s safe. So there are things that we can do on the applications front — hopefully in a more democratic and more human-centered way than China — that would actually be a contribution to GDP and consumer surplus.
Andrey: What do you mean by “they’re making it more safe for consumers” in China? I just want to try to understand where that’s coming from.
37:45Daron: Meaning that because there is more government regulation, I think some of the blowback from consumers using it in ways that may not be warranted — for example, in the classroom — is going to be easier to control in China. The evidence that is now coming out from a couple of papers is that the use of ChatGPT and other generative AI models in the classroom has been pretty bad for student learning. And how are you going to deal with that? I think we in the US have no way of dealing with that, because we haven’t set up a regulatory structure, we haven’t developed norms about how AI should be used, whereas in China there is more of that. And the consumers are also more savvy, because they’ve been using e-commerce more actively. So there may be more ways of adjusting to some of these things in China than in the US, because of that lack of regulatory infrastructure here. But secondly, I think partly because of labor shortages, Chinese companies are much more interested in applications of AI integrated into the production process.
Updating “The Simple Macroeconomics of AI”: Agentic AI and Three Bottlenecks 39:11
Andrey: How have you updated your estimates of the economic effects of AI on US GDP per capita since “The Simple Macroeconomics of AI”?
Seth: One of the first episodes of this pod. So we’re very much in your shadow.
Daron: Oh my God. Well, look, I have not gone back to the data and updated them. But if I were to do them today, they would be higher, no doubt, because that did not incorporate agentic AI. Agentic AI is not AGI, in my opinion, but it’s pretty impressive. On the other hand, I think the diffusion assumptions made in that paper are still pretty much in the money. The diffusion assumption was that about a quarter of what can theoretically be done will be done economy-wide. Now, we’re actually well below that, because the applications aren’t there, as I said. With agentic AI, there’s much more that can be done. But look at the production process: there are very few companies — large companies, or even medium-sized companies — that have integrated AI deeply into their production process. So it’s going to be a race between the model advances and the applications.
Andrey: What do you think is the biggest bottleneck? A lot of people have proposed many bottlenecks. There are organizational bottlenecks; there’s “we don’t have enough computer chips”; “we don’t have enough energy.” Do you have one that you prefer?
Daron: I think three. Three bottlenecks are really important. One is we don’t have enough engineers. Again, the Chinese experience is that any one of these technologies — robots or AI — integrated is so engineering-intensive. So many hours of actual hands-on engineering you need to do. We have a real engineering shortage in this country, and all of the people who actually understand generative AI are working in the generative AI companies. So there’s a real engineering shortage. Second, organizational constraints are real, absolutely one hundred percent. And they are different for large companies, which are bureaucratic, and small and medium-sized companies, which are stretched too thin. But a third — and this is where my uncertainty is, this is why I’ve always emphasized it’s so uncertain — is what a couple of people in industry and a couple of people in academia are now calling the jagged frontier of AI.
Seth: The Josh Gans term.
Daron: I thought it was Ethan Mollick, but it might be —
Seth: Oh, he stole it from Ethan? Of course.
Andrey: Yeah, he stole it from Ethan.
Daron: Okay. But the jagged frontier is also the reason why productivity improvements are going to be very difficult. So that’s why I wouldn’t say, “Oh, my earlier estimates are completely wrong.”
Seth: It’s because sectors are gross complements with each other, or businesses are gross complements to each other.
Daron: Each one of them has a different challenge. But the jagged frontier — especially if where the jaggedness is uncertain — means you still need the humans. So you need both the AI and the humans, so you’re not saving that much.
Seth: It also sounds like we’re complementing humans, then. The story you just told about all these bottlenecks seems very optimistic about wages and labor share, right?
Daron: Oh, absolutely. My estimates in that paper — again, all of it is uncertain — did not find huge negative effects at all on wages.
42:42Seth: One last comment about uncertainty and “The Simple Macroeconomics of AI.” You talked on Tyler’s show about having a methodological point, an empirical point, and a rhetorical point. The rhetorical point that I felt wasn’t made, and maybe should have been, is that you’re multiplying four numbers together, each of which is highly uncertain. Maybe the rhetorical point should have just been: when you multiply, you add the variances, so the confidence intervals should actually be pretty big.
Daron: One hundred percent. I was just leaving you to make that point.
Turkey, Memes, and The Economist 43:25
Andrey: Do you have any thoughts on AI in Turkey? We hear less about how non-American companies are adopting AI, or even how the government is adopting it.
Daron: There are three layers there. One is that the public is using AI a lot. Because of press censorship, a lot of young people were on social media, on the internet, much more than you would have expected at that level of income, so that’s facilitated it. There’s a lot of use of AI. But second, with the exception of a few leading companies, most companies are behind the frontier, and they’re not really ready for integrating AI, or even for the AI age. Third, the government is using more and more data collection and AI — not frontier AI — and some of it is good and some of it is bad. The good stuff is that, while it’s not perfect, there are many more services that are now online. Turkey was famous for having to go from one office to another to get even a stamp on your passport. Some of these things have become much easier. On the other hand, Turkey has its own surveillance state.
Seth: All right, here’s a fun one. What do you think of Acemoglu memes, and what is your favorite meme? Do you get asked this on all the podcasts?
Daron: No, it’s the first time I’m asked, Seth. And I don’t follow them, so I don’t have a favorite one.
Seth: You don’t follow — oh, very respectable.
Daron: I am touched that somebody has spent their time on these things.
Seth: All right, I’m going to tell you my favorite three.
Daron: Okay. Tell me. Please tell me.
Seth: The law of large numbers has been renamed the law of Acemoglu citation count.
Daron: Wow, that’s a good one.
Seth: That was a good one. “Nations fail, Acemoglu does not.”
Daron: I had heard that one. The first one I had not heard, actually.
Seth: My actual favorite is “Prices take Daron Acemoglu for granted.”
Daron: Oh, yes. I heard that one. That’s right. Okay.
Seth: Okay, but now a follow-up question. So you’ve gotten a lot of love, but you’ve also attracted some recent hate, including this really interesting Economist article — I don’t know if “interesting” is the word. Here’s kind of a meta question for you. One of the themes in What Happened to Liberal Democracy?, and a theme that I just see ambiently right now, is people increasingly concerned about superstars and very-top-percent inequality in general. Do you think that this ressentiment that is coming towards you is just general anxiety against superstars, and you’re just the economic superstar?
Daron: No. Absolutely not. Look, my book sales fall way short of Yuval Noah Harari’s book sales. And the following week, they had a very, very long interview with Harari, eulogizing him. So it’s obviously not going after everybody who is doing a little bit of visible work.
Seth: Sure, but I’ve never accused you of being a superstar in writing political books. You’re the superstar of economics.
46:49Daron: I see. I don’t know. I don’t think so. But let me say the following. First of all, I see that as a sort of hit piece. But I also consider myself fortunate, because if you’re going to have a hit piece against you’d better pray that it’s an incompetent one. And that was a really incompetent one. They did not even go and interview the people they cite. What kind of journalism is that?
Seth: It was very insinuating. “I had people behind his back say mean things about him.”
Daron: And the criticisms are just laughable. So I’m baffled exactly what dynamics led to that, and what dynamics led to other frictions I’ve had with The Economist.
Seth: But just coming back to the theme of your book, which is resentment of elites — do you see no connection there?
Daron: No, I don’t. If “elite” is access, well, The Economist has a million times more access than I have. They’re the elite. I’m the working class.
Seth: There we go. Everybody’s always punching up, Andrey.
Closing 47:56
Andrey: Any final thoughts for our listeners?
Daron: Well, this was so much fun. I’ll come back again if I have anything else to say.
Seth: Dude, we didn’t get through half of the things we wanted to talk about. I apologize we didn’t get to talk about your book a lot.
Daron: That’s fine. I’ve had many podcasts about the book, so it’s okay. And my guess is that the topic of AI won’t quite go away, so we’ll have other opportunities.
Seth: Perennial. Daron, thank you so much for your time. As I said before we got on the line, it’s been a real honor. We’ve always been working in your shadow, and we look forward to what you come out with next.
Daron: Thank you, Seth. Thank you, Andrey. This was truly delightful. Thank you.
Andrey: Thank you for joining us for another episode of Justified Posteriors. As a reminder, we’ve registered all of our priors and posteriors on our website, justifiedposteriors.com, so please check it out.