Capitalisn’t: The One Thing That Could Decide If AI Takes Your Job

Episode Transcript

Luis Garicano: What we've learned with chess and with math and with coding is very interesting—that chess is the hardest thing, and you can spend all your life and still be pretty bad at it. It turns out that's the first thing the computers know how to do. If you're now in the job market and you're thinking of your future, you've got to think, "Which are the tasks in which the messiness will protect me from the reinforcement learning taking over?"

Bethany McLean: It seems like every other day we get a new headline: how many jobs is AI going to wipe out? Just this morning, I was reading that Apollo has produced a new white paper saying that it's not going to be full-scale job losses. Instead, AI is simply slowing wage growth, especially among lower-paying occupations. This was a Wall Street Journal headline from the other day: Big Companies Are Starting to Hire Again, Defying Predictions of AI Wipeout.

Luigi Zingales: What if all jobs aren't created equal? What if AI will decimate some jobs, yes, but augment others? How do we think that through? That is, in essence, the argument that Luis Garicano, who is a professor at London School of Economics, and his co-authors, Jin Li and Yanhui Wu, made in their book Messy Jobs: The Work That AI Cannot Reach. The book's governing insight is all knowledge work varies along one dimension: messiness.

Bethany: To decide if a job is messy or not, one first has to look at the bundle of tasks it comprises, if those tasks can be automated, and how interconnected they are with tasks that cannot be automated.

Luigi: Luis is a PhD from the University of Chicago, taught many years at Booth, and so he has a very Coasean way of thinking about this.

Bethany: Who is Coase? Remember, I'm an ignorant journalist. [chuckles]

Luigi: You are a journalist, but not ignorant. Anyway, Ronald Coase was a British economist. He won the Nobel Prize in Economics in 1991. For more than two centuries, we believed that prosperity comes from dividing work into smaller and smaller pieces. Then Ronald Coase came about and asked, if that's the case, why firms exist at all? Firms are islands of central planning in an ocean of market economy. If markets are so efficient, why doesn't every individual task get contracted separately to the person who can perform it more cheaply, and why don't we have the organization done 100% by markets with no firms?

His answer, that is both very insightful and very frustrating, is transaction costs. The cost of searching, negotiating, monitoring, coordinating—all this stuff is expensive. Sometimes it's more efficient to bundle a set of activities inside a firm, even inside one-person job. Sometimes it's not. A perfectly autonomous AI could, in theory, reduce the transaction cost of using the market. It could locate contractors, negotiate terms, monitor works, and coordinate countless separate transaction. That might make firms less necessary.

Messy works involve tacit knowledge, judgment, authority, trust, and unpredictable coordination, the very thing that make our lives, when contracting, difficult. In Coasean terms, messy jobs are where transaction costs remain stubbornly high. In Garicano's view, AI forces us to ask a different version of the question asked by Ronald Coase almost 100 years ago. It isn't just why firm exist, but why job exist. What kind of work become more valuable precisely because they resist being divided?

Bethany: This matter is not just as a philosophical discussion about long-dead economists and aggregate employment. It is economics, but it's also practical advice. If you're deciding what to study, what skills to cultivate, how to organize a company, or whether or not your own job is vulnerable, the answer is not simply learn to use AI. It may be figure out what jobs are messy and what jobs are strong bundles and will remain strong bundles.

For a few years, you stepped out of academia and became a member of the European Parliament. How did that experience shape your approach to academia or your thinking?

Luis: It's been an interesting trip. Luigi has done something somewhat similar. We are among the few economists who have not tried just the policy.

Luigi: On a much, much lower scale.

Luis: Yes. The difference is that most economists, or many economists, have done some policy. A few of them have actually done politics, which is a very, very, very different thing. Go actually to the voters and try to get the votes and try to persuade people to do stuff and so on. Actually, even more importantly and more surprisingly for an academic, all the internal party politics. I think that was the part that I found most new and most illuminating. You know this saying in Washington: if you want to have a friend, get a dog. That was more or less what I learned.

Luigi: The title of your book is Messy Jobs. Clearly, politics is a messy job. You're saying that this doesn't apply only to politics, apply to every job, except maybe coding. Can you explain to our listeners why?

Luis: A clean job, what we call Tier 1 in the book, is a job that is well-defined, well-specified task and that is suitable to reinforcement learning. It's verifiable. You can check the task, and you can make sure that it's working as expected. Coding would be one. A lot of transactional law would be like that. You draft a contract; the machine can verify it. There are many other jobs that have a bundle of tasks that are messy that need to be connected and performed together.

I think that the jobs that most clearly our listeners can identify are sales jobs. A relationship manager in a bank. This is a person who does the loans for the SMEs and for the small businesses. You visit a client. You need to be informed. The part that is not messy, the part that is clean, which is getting the information about the client and figuring out who the client is, what is his balance sheet, what kind of products your bank can sell to them, is intimately connected to the performance part, the relational part where you're going to see the client.

There is a messy job because, in order to perform the task, the cognitive part cannot just be peeled off and done by somebody else. You need to know what you're doing when you're talking to your client and trying to sell the job. Politics is pure messiness. There is very little analytical in politics, and it's all relational. I think all sales jobs, when you talk to executives of tech firms, tech sales is basically booming. They don't know how to replace this with robots, and I think for a long while they won't know.

Bethany: Another way, I think, of getting at the same question, in your book, you use the famous example of the guy who predicted that-- I think it was Marshall Hinton, who predicted that-

Luis: Geoffrey Hinton.

Bethany: -all radiology jobs were going to be-- Geoffrey Hinton, who predicted that all radiology jobs were going to be gone due to the advent of AI. You explain that through the framework of messy jobs, why it hasn't worked out that way. Maybe share that with our listeners. Why does what has actually happened in radiology, where they're more in demand than ever before, yet AI is very valuable? Explain this idea of a bundle.

Luis: Yes. We talk about strong bundles and weak bundles. Strong bundles are these cases where splitting part of the task—think of the relationship manager in the bank—splitting the cognitive task makes the rest of the bundle not work because there's synergies, knowledge synergies between the tasks. I think, for our listeners who might be economists, think of the theory of the firm. That's the theory of the bundle of tasks. It's kind of the analogous.

What we argue when we start talking about bundles is, look, a radiologist does a lot of things—does a diagnosis, signs the diagnosis, which is critical part of the job. Somebody has to actually sign that and be responsible, reliable. They train other patients. They decide on the treatment. They work together with the patient and with the other doctors to decide on treatments. There is a lot of tasks that are involved that are not just looking at the screen and decide that yes, no, yes, no, being a classifier. It's much more than being that classifier. That is a strong bundle because you cannot just split this out and have the task be done by a machine only.

Luigi: I love this analogy of the bundle. I think it's a very useful way to understand the jobs as they are, but part of the book is about the fact that if we want to take advantage fully of the AI innovation, we need to restructure, rethink. You have some very interesting chapters on that. To what extent this rethinking is basically a complete re-bundling? Today's bundles are not going to resist very long because new organizations will come along that have completely different bundles, taking advantage of the strength of AI. What we need to do is actually spend our time thinking about how the task would be re-bundled, not what the bundles are today.

Luis: The first thing I want to emphasize is that I don't think the messiness is a provisional thing. I think the reality will remain messy. I think there is a lot of discussion over the 20th century of these [unintelligible 00:09:58] movements that try to plan all our lives from fascism to communism and all that. We know that knowledge is tacit and is lodged in the minds of people, and it's local, and that organizations are trying to get that knowledge.

Largely, I think that the reality will remain messy. The part that I agree with you, Luigi, is that the actual composition of the job will change. They will be re-bundled, and they will be un-bundled. The way these bundles will form is to be seen. We have robots below configuration where what happens is the person who has the advanced knowledge has robots below, AIs, that they can use to leverage their knowledge. Then the way the task bundles is by giving them the say in many other areas where they wouldn't have had a say. Let me just be concrete to not lose people.

For example, if you think of an entrepreneur that would have somebody doing marketing, somebody doing finance, et cetera, with robots below configuration, you have an agentic AI that can do all those things. You, as the entrepreneur, are having to have the judgment on marketing, on finance, on many of these other things, apart from your deep knowledge of whatever it is that makes you a successful entrepreneur, maybe computer programming, maybe chemistry.

That's one way that jobs might re-bundle. The other one, which is the expertise on the box or the robots above configuration, is this idea of the new middle class, which is the idea that you continue doing the human and relational components and you get more of the cognition that is now cheaper in a normal middle-class job. A nurse, for example, becomes a practitioner, makes a diagnosis because the AI is able to help them with this diagnosis.

We still have to think of how the supply and the demand changes, and we can discuss that if you want in this middle-class example, but no, I don't think the messiness will disappear. I don't think organizations suddenly become these perfect agents, which the agent goes around looking for the solution, and with the two agents paired together, everything gets solved. We start the book, as you guys know, talking about the family.

The family is getting ready in the morning, and suddenly, all your cognitive tasks are solved. You know what you have to do. You know where you have to drive, but suddenly your kid says, "I don't want to go for a swim." You can have the best AI in the world, and there is a management task there. You have to figure out how to convince, persuade your kid to go swimming or change your whole plans for the day. You can have the best AI in the world. If your kid doesn't want to go swimming, you are in a mess, and the situation is pretty messy.

Luigi: One of the great things about great work is that it can be used in ways very different than the person who invented it designed for. Your book fits in this category because I know that you wrote this book to help people cope with AI. I think that this book can be taken and inverted and be used as the perfect manual for a Luddite. You explain in excruciating details how you can send gears in the mechanism that will replace a job. You need to maintain a human relationship and maintain messiness, put some licenses here and there along the way so that you create scarcity for your job. I think that this is going to be used as the perfect manual for the Luddite.

Luis: [laughs] That is interesting. It is true that their supply is crucial. In the case of the nurse that I was talking about, if nurses can do more work, but a lot of people can't get into nursing, then the nurses won't capture that advantage. If the nurses can do more work and they want to be the ones who are able to do the work, they have to fight, indeed, as you were arguing, for the regulation to make sure they are the only ones who can do the work.

I think that over the next years, economic growth, the changes in societies are going to be much more determined by this organization and political and economy discussions than by technological discussions. Think of New York. New York decides, "We don't want self-driving cars." Whatever productivity gains come from self-driving cars will not go to New York. Nuclear energy. Germany says, "We want to close all nuclear plants." Then the technology exists. Germany won't enjoy those gains.

I imagine that in some societies, some professional service categories will have a lot of say and will say, "We need to sign everything; nothing can change, all these steps have to be human." You'll not have the productivity gains, and you will have more of the technology being slowed down in the ways that you were suggesting. Given two things that happen in our societies, one is the fiscal situation we're in, and two is the aging, which means that the growth rates are going to be constrained.

AI comes at the right time. I wouldn't want people to take it in that way. I can see how, yes, the way we analyze the supply and the demand and the factors determining it can give people ideas in that direction. That's not what we want. [chuckles]

Bethany: Do you have a sense of what it is we do want? Where in that middle ground we should live? We've talked to one person who argued, "Slow it down, basically, put regulations in place that keep the human in the middle." We talked to someone else who said, "No, no, no, technological progress needs to go as quickly as it can because every time you try to slow it down, that's where stagnation comes about."

If you could decide how much of jobs we are going to keep messy and if some part of that is a decision to be made via regulation or via the political economy, what would your answer be?

Luis: Over the long run, what we have seen is that technology has enormously benefited humans. We are healthier, we are longer-lived, we are richer by orders of magnitude than our ancestors were. I like how Kevin Murphy likes to explain it. I don't know, Luigi, if you heard him tell it like this—our colleague at the University of Chicago. Kevin Murphy likes to make an inversion of the Malthus point. Malthus said land is fixed, population is flexible, so when any technology comes up, then more people get born. Eventually, because the land is fixed and agricultural produce is fixed, boom, people go down to the original level, and all the gains will go to land.

Kevin likes to say, "Forget about complementarity, sustainability, and all the rest. That's basically what happens with technology." Instead of making the land the fixed factor, it's the human which is the fixed factor. You get a new technological shock. The return to capital is much bigger, but capital is flexible, like humans in the Malthus story. As capital is more productive, people buy more and more and more machines until the competition between capital starts to drop down the interest rate to the original natural rate, R-star, let's say, and all the gains end up going to the humans, which is the fixed factor.

Of course, this is a simplified model because there are other fixed factors like land and electricity and so on, but the basic idea is, over the longer run, competition between capital owners pushes that down to the natural rate and the gains go to humans. My view is very shaped by that view as well. That doesn't mean that I think we should do nothing. You're right, Bethany, that there is a potential middle ground.

I think that we will need to make sure that there are compensation for losers. That is the one advantage that Europe has. Europe has a social system with social safety net. That means that maybe people will be less scared of technology than the US because they know that they are not going to be left to drop as much.

Luigi: I heard Kevin Murphy make the point; I think it's a beautiful point. However, I think it's missing an important element, in my view, which is knowledge or, if you want, data. As the agentic ability of AI will develop and as robot will develop, human can be substituted. What is really valuable is who has the knowledge. Knowledge these days is data, so who owns the data. My concern is that AI is becoming better and better at expropriating the data from all of us, even your messy jobs.

You know that in LA there are people that do a routine task with a camera on top of them. They see them fold their laundry and doing their stuff in order for this to be learned by a machine. All these nuances that before were difficult to be captured, now, with videos, they can be absorbed. Once they're absorbed, the person owning the data will get all the rent and everybody else is disposable. The real land, the new land, are the data. The owner of the data will become filthy rich, and everybody else will be a proletarian and earn the minimum wage, if any wage at all.

Luis: I agree with your inversion of Murphy's inversion partly, which is he says the fixed factor is the human. You're saying, "No, no, no, it's not fixed factor because the robots are kind of imitating humans, so then we can have a flexible supply of humans, and then the fixed factor is the data, and then the rents go to data." In a way, he's making the Kevin argument against the Kevin Murphy argument, against him. I think there is truth to that.

Let's say that I believe that the model layer is likely to be competitive. I think that recent events suggest that this is likely to be the case, with open models setting up a floor that we can always attain relatively cheaply, allowing companies to keep ownership of their own data, allowing individuals to keep ownership of their own data by using the open models. I don't think the model layer is going to capture a lot of the value. Maybe I'm wrong.

Your world will be one world where we get some really strongly monopolistic positions in LLMs, and there is just one guy who basically rules the world. I don't think that's what's going to happen. What happens is you get that data. Very expensively, you process that data to get the frontier model. Everybody else use the answer of your frontier model to shape the way their model works, which is what's called distillation, something which the frontier models really hate. As you're saying, the frontier models are stealing our data to start with. I don't know why they hate other people stealing their data. I'm so sorry.

My sense is that if this is the market structure, then the model layer is competitive. We have many people who are able to produce almost frontier models. For most applications, that's going to be enough, and that's going to allow the rents not to be captured by these data types. I think the message of this is it will be captured at the implementation stage. The part that will be valuable will be being the one who is able to use the AI to produce and to pair with labor to produce goods, et cetera, rather than be the one just overlord sitting in San Francisco and using this monopolistic power of the data.

The jury's still out. It could be that the frontier is all that counts. Many people in Silicon Valley are making this bet that it's all about the frontier and these guys will be the frontier. I just don't think the evidence so far suggests that.

Luigi: I do believe that there will be some competition at the model level, but let me make an example with a law firm. You have studied law firms, you have written about law firms, so you are an expert about law firms. My father-in-law works at a small law firm, and he specializes in car dealers. He has a lot of knowledge, messy knowledge, personal relationship. It's very difficult for him to monetize them by selling this knowledge to the next guy. He needs to hire a young associate, train this young associate, and he needs to give a lot of rent to this young associate because they are not going to work hard and do the job unless there are some rents.

Now, think about a world in which he can transfer all this knowledge to a specialized LLM model. Then he's able to basically hire or whoever owns this model, maybe he passes away, and you own this model, and you can hire poorly trained legal clerk to do most of the job. The difference between a legal clerk job and what he charges for his services would be fully appropriated by who owns the data. It's not only a few people in the Silicon Valley, but it's basically a few incumbents are getting all the rents now from the future, and the new generations are left behind.

Luis: A couple of objections. One objection is the relational part of the job, in your father-in-law's case, is, in my view, not something that somebody who doesn't have the knowledge can hold. If you're going to have to have the relationship and you're a paralegal, you're probably not going to be able to do that.

Luigi: Sorry, can I interrupt you a second because I think this is crucial.

Luis: Yes, it is.

Luigi: There is a human component, of course, but a lot of the relationships are based on the fact that I know what the guy likes, how to talk, how to approach. This stuff is transferable. It's not that it's not transferable. It's that people don't want to transfer most of the time because they cannot monetize. They are resisting like Luddite trying to prevent. If you have the right incentive-- I remember that once I was in an organization, I don't want to go into the details, but there was some person that was training me how to approach various people.

I was shocked because this never happened to me: that somebody voluntarily trains how to approach. This is incredibly valuable. Now, if you have a machine to say, "I know that Luis has a passion for everything that comes from the University of Chicago; I need to cite Kevin Murphy because he's going to go-- I know that he loves the example of the lawyers because he has written a paper about the lawyers." Once you know all this stuff, I think it's pretty easy for me to establish a relationship with you.

Luis: I still think there are pieces of knowledge that are very difficult to put into the central machine you're designing. The first is strategically withheld knowledge. You're saying, "Well, I give the right incentives." It's not clear. If I'm the buyer and you're the seller, I want to over-represent. You know this as well as I do. There is a situation with asymmetric information that we might not transact, and I might not ever tell you the truth about how much I value this or how much I value that.

The last thing is a lot of the knowledge doesn't exist in a closed, simple form, but comes from the interaction. That interaction is only fruitful. The things that I'm telling you now, some of the things I'm telling you now are not in the book and I haven't thought. They come from you asking me a question, me thinking through the answer, you asking the question. A lot of the knowledge gets generated in the interaction. It's not just this mechanistic construction that I just stick the knowledge in a box, and it's there.

While I agree with you that some of the tacit and some of the dispersed knowledge is going to be possible to put in the machine, a lot of the other bits are not. Let me still push on the competition bit. I don't remember the law that your in-law does. Let's say it's intellectual property. If intellectual property law knowledge is in a box, my sense is that that box is going to be zero price, meaning all that knowledge is going to be competed away. People are going to all know all this IP and be able to get that box eventually through competition.

That's the middle class example. The argument here David Autor has been making, and I like it, is look, the distribution of skills in the population, there are physical skills, there are emotional skills. It's much more like a normal distribution and less spread out than this weird distribution of math or chess because we're all humans, and every human has to hold a baby. There is no human in life who's ever been born without having a mother. Every human knows how to be a mother to some extent.

This means that if the knowledge component has a lower cost and a lower premium, the inequality is going to go down as everybody can do all these tasks that are non-cognitive, emotional, relational, plus that knowledge that gets more freely distributed. To the extent that, I guess, again, we are discussing whether the data is really a monopolistic, non-competitive situation or whether it's somewhat competitive, but to the extent that it's competitive, you could argue it's Bertrand because it's freely replicable. It's basically a public good.

If that's the case, then the other people, this paralegal that in a way you're putting down, "Well, the paralegal's great. He's like a lawyer. He can go around solving problems and filling up forms and gets a middle-class job." That's not necessarily a bad outcome. It's a bad outcome for many professionals, many high-skilled lawyers, et cetera, et cetera, who are going to lose that monopoly they right now have, but it's not necessarily a bad outcome for society.

Bethany: Yes. I think I'm more inclined to agree with Luis and maybe even a little further along that line because when I see people, very high-end lawyers, litigators functioning, you could pay them $10 million to put all their knowledge into a box and they wouldn't be able to do it because they can't even explain what the decades of standing in front of a courtroom and being able to react and what that skill is nor the personal skills that go with the years of having learned how to hold somebody's hand.

You could pay them any amount. You can pay me any amount—I'm not sure anybody would want my skills—but you could offer me $10 million to try to put what I do as a journalist into a box so that somebody else could come along and be me, and I wouldn't be able to do it. I wouldn't know what it is that makes me able to get people to talk to me. I wouldn't be able to explain that to you. [chuckles] I had a slightly different question that follows from this in some ways.

One of the things we talk about a lot on this podcast is wage growth and the importance of wage growth to the continuation and the belief in capitalism. In your book, productivity doesn't automatically become wage growth. Just because a job is a strong bundle, that doesn't mean that wage growth is going to occur. Can you talk about the conditions that are necessary for wage growth to occur even when a job is a strong bundle versus when it won't occur?

Luis: Basically, what happens is you're going to get some income. Let's assume that we're not in the automation replacement case, as you said. We're not in the Tier 1 jobs that get replaced by computers. We are in a Tier 2 job, and there's a certain set of things I can do. On those things, there are some things that the computer can't do, and they remain within my career. Now, that doesn't mean I have a high wage because there is more of us, in some sense. I now can do twice or three or four times as many things in my time as I could before.

The question is, is the demand for my services saturated or is there a very elastic demand? The first case, I love an example that comes from Nordhaus's work, which is the illumination paradox. Nordhaus says in 1996 NBER piece, "The cost of illumination of light decreased by a factor of 40,000 between the 19th century and the 20th century. Basically, it was a really significant cost for a poor person in 1830, 1840. They would go to bed when the sun set because they couldn't see anything, and maybe they had some form of really a little bit of wax or something they could use to illuminate themselves. It would cost hours of work to get that.

Now, we are bathed in artificial light wherever we are. He says, "And the result is the industry's importance disappeared." Basically, the demand was inelastic. We got [unintelligible 00:32:24]. We have definitely enough of the light to not want more. The result is, as a share of GDP, illumination is very, very, very small. The wages of the people or the value of the illumination industry goes down. It doesn't mean that the wages of the population as a whole go down. That's a separate question. The specific people in a particular industry could have so much more productivity if the demand is not sufficiently elastic that there is excessive supply of the hours.

If each lawyer counts for 10 and the demand by people or lawyers doesn't increase by 10 times, then that's bad news for the lawyers because they are competing more with each other. That elasticity, the question of how close to association we are in different sectors, is very different for different things. My sense is, for example, when you think of radiology, to go to one strong bundle we've discussed, I wouldn't be surprised if the price was sufficiently low if we all had a CAT scan every month or every two weeks. I could see a world where the radiologist is there, the machine is AI, most of the time he doesn't need to intervene. You walk in the booth, you press the button, the guy takes a quick look, you go out.

That would be a situation with very, very elastic demand where the cost goes down and people can demand much more. When you talk about law, for certain things like criminal law, probably there's not going to be-- if each lawyer gets 10 times more productive, there's not going to be 10 times more people to defend. There, we could have the opposite result. Now, as I said with the Kevin Murphy example, that doesn't talk about what happens with the average wage.

The average wage in the population, at the end of the day, over the medium-long run, I would expect it to go up because the return-to-capital wouldn't absorb that, and under the idea which I disagree with, which is that data is not a fixed resource, then it would be humans who capture that increase.

Bethany: One of the things you write about in the book is the signaling process by which people are hired and valued internally. Talk a little bit about how AI can change that signaling process, and I guess why I might be wrong in having an incredibly negative interpretation of this, which is that if AI makes human networks even more important, doesn't that make it even harder for people who don't have those networks in the first place to break in and stay in?

There's this recent study done that found that women who submitted resumes created with AI assistants were evaluated as less competent and less trustworthy than men who submit identical AI-assisted resumes. With AI further mixing the signals and requiring in-person networks for success more than ever, it seems like AI might further and perpetuate divides instead of fixing them.

Luis: Yes, you got exactly the gist of argument, and I won't belabor it too much. Basically, if everybody can do a good cover letter, if everybody can do a good report, write a good paper or a good-looking paper, all those signals get devalued. We think that the power of weak links that used to be high—or at least my grandfather argued that weak links were how people got around in life—is probably going to depreciate, and we're going to have, Italian society maybe, a strong linked society where people basically, as you argued, Bethany, are going to have to rely more on, "Well, what job did you do before? Why did you do that job?"

Because somebody hired you. "Why did that hire you in the previous job?" Because you knew somebody, or you got in the door initially. It's probably going to be much harder for people who cannot break those initial steps, those initial barriers, to signal that they're actually pretty competent. The central exam in China is going to become much more important. Maybe the SAT is going to become much more important. Objective signals are going to become much more important.

We have already evidence, and we mentioned in the book several studies, the ways that people filter resumes online are gone. They now no longer use those ways to try to find somebody who would have fallen through the cracks. That person is just now probably falling through the cracks because they don't have a good way to signal.

Luigi: Isn't this pretty dramatic? I would like to double down on what Bethany is saying. I think we're going to our new feudalism because I remember when I was coming from Bocconi trying to apply to the United States, I did not have the letter of recommendation of the most important person at Bocconi at the time. The reason is he was not my advisor. I asked him to be my advisor, and I had a very good GPA, but he was busy and other people with more connection got his attention. I didn't.

I had a stellar GRE, stellar GPA, and as a result, I got basically every university in the United States offer me an admission. If I were to apply today with a GRE, people do it with some cheating thing, et cetera, and the GPA you cannot trust, the only thing that matters is that I know the important guy. It becomes a feudal society which may talk us down the drain, and what you add is only contacts. The children of people who are influential add contacts. The rest is left behind.

Luis: I think that this is a scenario that could play out. This is playing simultaneously to this other move in which we have devaluated exams and objective criteria for admissions. I think Booth, the University of Chicago, has a fixed grade point average that forces people to have As and Bs and Cs. Harvard, I saw they had voted it. I don't know if they're actually going to put it in place next year, but they voted a fixed average so that there is clear information in the grades.

I think that the only way to counter this risk that you say, Luigi, is that we go back to really introducing all of these things that we have been devaluing—exams and blue books in class where people actually have to write, as opposed to projects where they are making them up, and SATs and GREs and all the rest, which right now, sadly, many universities don't even use to accept people.

We're going to have to reintroduce the meritocratic tools that our societies had in the '50s and '60s that were quite successful to pull people up from the bottom.

Bethany: I hope that can happen. I have a last question because I unfortunately have to run, and I'll let you guys continue if you want to, but I had a really practical question. How does your book make you think about what advice you'd give somebody who was looking for a job now, or somebody in her 50s who wanted to keep a job?

Luis: The big conclusion of the book is in a letter to a young person I wrote before we finished the book in January, which is, go for the messy job. Try to think of, "Is this a task that I'm going to be doing that is subject to reinforcement learning that is verifiable, that is simple and clean?" If that's the case, even if well-paid, you're going to be learning something that is going to get probably depreciated.

Instead, find a job where you're going to be growing in all these other dimensions that includes coordination and relational aspects, and embrace the mess if you want. It's not a difficult job, right? What we've learned with chess and with math and with coding is very interesting—that chess is the hardest thing, and you can spend all your life and still be pretty bad at it. It turns out that's the first thing the computers know how to do. If you're now in the job market and you're thinking of your future, you've got to think, "Which are the tasks in which the messiness will protect me from the reinforcement learning taking over?"

I think that we have experience on that from the last 40 years. Remember the Excel example, which I love. People were spending hours building up a model from scratch with pencil and paper. If one of the assumptions changed, hundreds of people had to redo all the accounts of the company, all the tables, tabulating it and summing it up. You put Excel, and people think Lotus 1-2-3, and people think, "Oh, accountants are going to go away."

It turns out that there are many more people doing this because they're doing the much more sophisticated, complex work of building the models and using these tools to complement them. It's not the case that the accountant, that the person that is doing numbers, went away. We know how it happened in that case. I don't think people should be depressed. I don't think people should think, "Oh, I can't do anything about it." There are going to be plenty of jobs. Just think relationship, relational work, social, emotional, integrated with cognitive.

I wouldn't say not think of the cognitive. I would think of the tech sales. Tech sales is the perfect job right now. It's getting the productive gains from the tech. You can use your AI to learn what the clients are and to get better, but you're doing the relational part. Those jobs are really doing very well.

Luigi: One of the issues, because you often mention, is that, "Oh, the prices will go down and so the purchasing power will increase, and so people will start buying other services," this really presume that the purchasing power is not equally distributed, but widely distributed. If you are in a world in which most of it comes to a small elite, then of course, you're going to add a lot of job, but a lot of jobs serving the small elite. It is really a feudal society in which you have the rich lord and then a lot of people doing messy job and very high-touch job for the feudal lord.

If you are filthy rich, you value the best food prepared by the best person, so you're going to have a personal chef; you're going to have personal masseuse; you're going to have a personal this; you're going to have personal that. You're going to have the world divided into feudal lord, want data and get a lot of the return, and then everybody else employed, but employed doing very low-level jobs. How do we fight this?

Luis: This is the pure relational economy of Alex Imas, a colleague of yours at Chicago. By the way, let's do this one, the last question, because I promised my wife I would go to the barbecue. I don't have as narrow a view of the demand. I think that a lot of the jobs that are normal jobs in the economy, like a nurse or a lawyer's assistant or a paralegal, are jobs that, with more cognitive, with the expertise on a box, they can actually increase in productivity and in reach to the extent that that happens in many aspects of the economy.

I don't necessarily see this dystopian world where wealth is concentrated. I can see one scenario where that happens. I think we have to fight it, and the way to fight it is to keep the open models going and to keep the open weight and to make sure that the data is not concentrating on Elon Musk and his three friends. I strongly believe that cognition-- Think of the developing world; I think that cognition is a huge barrier to development, that a lot of what happens is you don't have medicine because you don't have doctors, and you don't have a repair in the electricity network that fell down because nobody has any idea how to do it.

If that village somewhere in a poor country has ChatGPT and they make a picture and they say, "This fell, how do we repair it?" A lot of those problems can be alleviated. I think that cognition is scarce, it's expensive, and it's a source of rents for people like you and me and for big lawyers and for big consultants and for big investment bankers, and that democratizing cognition is not making the society more elitist, but potentially and easily is making the society more democratic and more open.

What that requires—and we go back to your very first question—is that we don't go into a data monopoly, that we don't go into a world where only the frontier models rule and everybody has to go to the frontier model, that they have the data, they have the model, they have the fine tuning, and that's it, but that there is an open-weight ecosystem, an open world where everybody has access to this cognition.

Bethany: I was surprised by how optimistic he was. I liked it. He's a humanist and a believer in humans. I liked his broader, more philosophical argument of all of the good things that come out of messiness because I do think it is this fundamental tension in humans, not just with jobs, but with everything, where we're always trying to make things neater and put things into more of a box and optimize and optimize. I really liked his broader viewpoint on that. Did you agree?

Luigi: The comparison with Coase is perfect in two ways. Number 1, because it's true that he does with job what Coase does with firms. Coase insight was great, but from a practical point of view, it was very, very difficult to implement because these transaction costs are very difficult to measure. You can make a story for everything. There is an entire field of economics called transaction cost economics that goes around, look at firms, and then—I'm a little bit unfair—but rationalize, expose what you see as optimal.

If you spend enough time, you make up what are the important costs and not the important costs, and you can rationalize everything. The same is a bit with this messy jobs because unless you have a theory of messiness, which that's a very high bar, but unless you have a theory of messiness, then you don't really know what is messy and not messy. The other thing is you don't really have this notion, for example, of how elastic demand is because the big change is going to be how the economy will be reorganized and how jobs will be organized.

Unless you know how they will be organized, and if you do, I'm sorry to say, you should be doing the work of reorganizing because you're going to make bazillions. If you just pontificate, you might get it right, you might get it wrong, but it's very difficult to know what is the elasticity of demand, for example. If you have a reorganization, it's going to be very different. The same is very difficult to know whether you can separate a task or not.

If you keep doing it the way things have been done in the past, of course, those tasks are now separable, but the world is changing. There are different ways to organize. Honestly, the only thing I know is the past because, as you said, the future is very hard to predict. The secret about the Second Industrial Revolution was the introduction of replaceable parts. You standardize all the pieces to the point that you could produce a piece in one factory or another, and you put them together, and they work perfectly. This was an incredible difficult job to do.

It was started actually by a French general. In 1800, you were fighting, there were a lot of pieces of rifles left over, but they were produced in different ways, so you couldn't reconstruct. With two half rifles, you could not reconstruct one rifle. You say, "It will be incredibly useful for us that all the rifles are produced in the same way so that we can reassemble parts." The idea actually of what is called the American manufacturing system was invented by the French.

The French were never able to implement it. Why? Because you need to eradicate the artisans. The secret of an artisan is that my product is different from yours. The standardized department is exactly the opposite of what an artisan can do. In the United States, we were able to do it in part because they imported a lot of unskilled labor. They put them together, and they organized replacement part manufacturing, which is what led to Henry Ford and mass production.

Had you seen the production before, you say, "Oh, this is a very messy production; will never take place," because how do you do this? You need to fit every part. Imagine two mechanical parts. You need to arrive to the precision that they are exactly identical to the last millimeter, whatever it is that thing. It's very difficult. Eventually, they succeeded, and the artisan was wiped out. If history is any indication, it's going to happen the same.

At the beginning, there are all this messiness, "You cannot do it, you cannot do it," until the moment you do it. When you do it, bam, the world is different. I love Hemingway say, "How do you go bankrupt? It's slowly and then all of a sudden." I think the innovation is the same. You hear that this might be happening, it might be happening, and by the time you start not believing, boom, it changes the world.

Bethany: It's interesting. I think my takeaway from what you just said is that the question you and I asked each other in our conversation before our talk with him was exactly the right one, which is how much of messiness is innate to a job and how much can be stripped away. All of this discussion that we had with him hinges on that answer because if the messiness can be stripped out, and it can be automated as it was in the Industrial Revolution, then nothing is safe.

Luigi: Actually, I learned a lot in the conversation and I refined a bit my thinking. The way to look at it is this. Of course, messiness is endogenous. Why? Because the moment you attribute task in a particular way, the soft information generated by the task becomes the livelihood of the person exercising and is going to fight tooth and nail to protect it. I have to give a lot of credit to a co-author I'm working on, Tano Santos, who educated me on all these things. When the philosopher of the Enlightenment started to create the Encyclopedia, one of the thing they wanted to create was precisely tables on how to produce what the artisans were producing.

They couldn't do it. Why they couldn't do it? Because the artisans, they don't want to share that information. It's as simple as that. Now, how did we get around this? By having the artisan confess under duress? Maybe in the Soviet Union they were able to do that or in other countries, but actually not, by bypassing them completely. What I describe you as the American manufacturing system does not ask the artisan how you do it. It completely bypass what the artisan do, with enormous benefits and some costs.

Standardized products are not as beautiful as the ones done by hand. Will the artisan survive? Of course. Tailor-made suits are much better than the one you buy at Filene Basements. However, how many tailor-made suits I have? At the moment, zero. Why? Because there are good enough at the price. I think the same thing will happen. We are going to dispose the messiness. We're going to have a moment in which the local guys are going to try to resist their job and they're going to hold on to all their messiness.

The solution is not going to be trying to extract their information, their soft information; it's going to be to completely bypass it. There will be a moment of adjustment in which the product that we produce with AI is not going to be as good, but then it's going to be so cheap that it's my choice of a suit. I go for the one at Filene Basement.

Bethany: [chuckles] I will not criticize your wardrobe, Luigi. Then do you think that Luis's optimism is misplaced?

Luis: Depends on what you mean. I think that if you look at the big picture and the pie, this is fantastic. If you think about the effect of the Second Industrial Revolution, they've been enormous. I think for our grandchildren, this is the greatest thing that will happen to them. I think that, for our children, now, I feel I'm old enough that probably I'm protected, but for our children, this is going to be a messy situation. That's really the messiness. There will be a lot of blood in the street.

The transition ain't easy, and some people are going to make a fortune, and a lot of people are going to be impoverished in a way that is completely unexpected. He keeps the eyes on the overall pie and that's what the optimism is. I might be more pessimist by nature, but I am more sympathetic to the transition cost. I don't want to, in any way, say the transition cost should make us stop the industrial revolution that is taking place now.

However, do we need to proceed at maximum speed? My favorite comparison is, think about the first industrial revolution. England and France. It's not France is much poorer than England these days. They might have adopted their first industrial revolution a little bit more slowly. They are right to the same point. This is relying more on my readings of literature than of history, but all the terrible novels of Deakins and child labor, et cetera, that we saw in England, we didn't see in the same proportion in France. If you arrive a little bit later, but with less body bags along the way, I vote for that.

Bethany: Then I guess it becomes the same question we've been asking, which is, is there the political will to do that, and how that will be implemented if there is. Speaking of messiness, that will be quite the fight.

Luigi: My concern is in the future. I'm taking very extreme positions. We're not going to get there immediately, but just for the sake of argument, imagine that all the knowledge is embedded into LLMs. The scarce factor is the knowledge embedded in LLM. Everything else is flexible because you can produce an infinite amount of drone, an infinite amount of robots, an infinite amount of whatever. What is scarce is the knowledge, and the knowledge is going to get all the rents. We're going to have people being poor, but data owner being very rich.

Bethany: I don't know if you saw; I'd actually included you on a tweet that I saw recently that augments your point of view because it argued that, increasingly, the platforms, OpenAI and Anthropic, are taking over all of their customers' business. They're developing drugs better than drug makers can. They're stepping into Microsoft's business and offering corporate customers Microsoft's products. That would seem to be evidence of all of the knowledge potentially getting embedded into a few very large and very powerful companies.

Luigi: Believe it or not, not only I saw the tweet, but I even like it. [laughs]

Bethany: Oh, my goodness. [chuckles] Luigi, your social media skills—you have to then say something and respond to it.

Luigi: No, I put a little heart, okay? [laughs]