B (5:14)

Thank you, Jasza. It's really a pleasure to be here.

A (5:16)

There's many things that I would love to talk to you about, but the topic that I've been thinking about a lot is artificial intelligence. And I've had conversations in this podcast about just the technology of it with people like Geoffrey Hinton. I've talked about the dimension of existential risk with people like one of the co authors of if Anybody Builds It, Everyone Dies. You know, I thought about some of the broader public policy angles, but I haven't so far had a conversation really about the economics of artificial Intelligence. And I think it would be really interesting to try and get a handle on those questions. I think we'll talk particularly on questions about the labor market. But before we get there, sort of what in general do you expect the impact of AI to be? Is it going to be major or middling or minor? Is it going to lead to just vast economic growth as some people are predicting, or is it going to really decimate the number of jobs that are out there for humans? Is this going to be an economically revolutionary time? Or is it going to be one of many developments that are kind of in interesting and we're following, but ultimately not that consequential?

B (6:27)

So I think that I don't have a crystal ball first of all. So anticipating things is always hard. But let me give you my best take. Based on what we see, I think it's clear that a lot of knowledge work, even if the technology stop tomorrow, could be automated a lot of knowledge tasks. It's already very clear that for example, tasks that are routine, tasks that have to do with diagnosis, writing, crafting, printing documents, writing, doing research, for example, the task that the AI is already doing perfect coding work is really spectacular. So in terms of is it going to be big? I think it's going to be huge. It's probably as big a revolution as the industrial revolution. That's a very likely thing, except that instead of for automated work is for cognitive work. So I think that everything points out to a large impact and also accelerating. So. And there were people who were doubting, there were people who were wondering if AI would be a big deal or not. I don't think any of those people could be doubting given what we have observed, let's say in the last six or eight weeks, the explosion of new models, the way they work, cloud code is really kind of taking the world by storm. Everybody has noticed the software firms value is plummeting in the stock market, showing that people believe that many of the functions, that many verticals, many softwares that were actually accommodating for one particular usage can be replaced by AI. So yes, a big deal in many segments. Second, growth. Yes. I mean if this is as big a deal, we'll see big productivity growth and an acceleration, I mean not the kind of growth, and we'll talk about it for sure later, but not the kind of growth that many people in Silicon Valley predict. Because I think most economists think in terms of O rings and bottlenecks and weak links, meaning you can invent as many compounds to solve cancers as you want if you need to go through years of clinical trials and regulatory approvals, that's not going to suddenly accelerate massively. So those weaklings will const growth everywhere. We, we'll talk about those for sure. Labor, your third question. I would say the evidence so far, I mean maybe it's like the person falling through the window, but the evidence so far so good. It's more complementing than replacing. In the the three areas where we expect largest impacts, translators haven't dropped. I mean everybody thought translators are going to be decimated. Translation seems like a solved problem and yet demand for translator hasn't dropped. According to world labor statistics customer service agents, some people get fired, some people get rehired to do different jobs. Again, the BLS doesn't see much and even computer programmers, we are not seeing big drops. There was a couple of papers earlier in the year. Eric Brynjolsson has a paper with co authors called Canaries in the Coal Mine, which was starting to see drops in more exposed segments for more junior employees. And yeah, we see a bit of that, but there is a lot of discussion on whether that has to do with COVID and so on. So for employment at the moment, it looks like it assists more than replaces. I mean it's clear that it can do many tasks. My main quarrel with the Silicon Valley, the interpretation of things, is that they believe that if we can replace the task that can be more easily done by the computer, then the job is gone. And jobs are even a radiologist only spending 30% of his time looking at scans. The job of a radiologist is much more than just diagnosing scans. So all of those things happy to discuss.

A (10:38)

Yeah. And I was in a meeting with Sam Oatman in I believe 2018. I barely knew who he was at that time and I remember him pointing outside the window of this hotel in Silicon Valley saying, you know, in three or five years, you know, there's going to be robots building homes here. And none of that has materialized. So there's certainly a very real tendency of people in Silicon Valley not just sometimes to over promise on the technology, but to underestimate the obstacles to real world adoption of technology, which are particularly evident, as you've recently pointed out in something like house construction, where the constraints just aren't actually the inability to build homes. We know how to build homes. It's you know, all of regulatory approval and the zoning laws and the, you know, is the nature of my neighborhood going to change and all of those kinds of things.

B (11:27)

Two points about, about, about the comments you made one about the Silicon Valley position, I, I, I am very surprised that they are not just hyping the technology which I understand you want to sell, you want to sell enterprise subscriptions, but they are also hyping the risks of the technology and all the time threatening people with extinction will take all these jobs. I don't see the point of this tactic. I can see that if you want to justify their valuations, they need to say that all these things are incredibly transformative and they are transformative. But I don't see why this insistence, like the other day, Mustafa Solimandi, the Microsoft, the ex DeepMind co founder and Microsoft AI head, he was saying to the FT that they're going to automate all white collar works jobs in 18 months. And I was joking. Like does anyone really believe that Microsoft will actually get Outlook or Word to work properly in 18 months? I mean, I don't think, I don't think they can fix their two pieces of terrible software in 18 months. We all hate Outlook. We've hated it for 15 years. I bet you will hate it in 18 months. So they're talking about auto automating complex jobs that they cannot automate their own software. I mean that's just completely ridiculous.

A (12:46)

Before we dive into all of the substance of this, I would just love for you to help us establish the premise that you're operating on, because I think a lot of my listeners are tech forward, a lot of my listeners are not tech forward. And I still find many people in conversation who've experimented with ChatGPT when it came out three or so years ago and have gone back to use it every now and again. Perhaps they use it instead of Google to search for certain things, or if they have a translation need, perhaps they use it, they use it for very specific things. You know, they're still convinced that it hallucinates a lot. They feel like the limitations of what it can do are very strong. Part of that, I think, is that the Most commercially used ChatGPT products are not very good compared to some of the competitors now, in part because they route your request sometimes to a really powerful model and sometimes to a really not very powerful model at all. I think part of it is that a lot of people use free versions of these AI tools which are much less powerful than the ones for which you need to pay at least $20 a month. And part of it is that probably only a fraction of people who listen to this podcast have used tools like Claude Code. So just to motivate what we're talking about when you're saying that there's been this tremendous progress over the last few months and more broadly over the last years, what are we talking about these tools being able to do today? How is it that people are using them that is so different from what you might think if you're just using the free tier of ChatGPT, let's say.

B (14:13)

Okay, so let me give you a cloud code example and a deep research example. So a cloud code example would be the following. What is interesting is that the machine can talk to you and it can send tools, it can put Python tools to work. So what do I mean? So let me, let me explain to your listeners in a very clear way. So I did a paper. I was a member of the European Parliament, and after coming back to academia, to do. After I returned to academia, I wanted to do some research on how the narratives work in the European Parliament. I wanted to show there are no tradeoffs in the narratives. So basically what I did is I collected 46,000 speeches, all the speeches, downloaded them, putting in spreadsheet. Each speech goes to, to the ChatGPT through API, which means it goes through a special byte, it gets processed, it comes back into a spreadsheet, it gets classified in certain ways, and then we analyze that classification with statistical tools. That took six months. It's a lot of work getting each speech, sending it, bringing back, et cetera, et cetera. I had done this for climate. Now I decided with cloud code tools to do all the work, six months of work to do all of these for AI. How is, how is the discourse in the Parliament evolving on AI? Okay, so I told cloud code, I wrote, this is with text, okay, it's no programming. I told cloud code, here is my directory, I put it in my directory where I had all these files, figure out these files. You don't have to tell it that, but that's essentially what it does. Write everything, the same pipeline to get a speech, to send it there, to classify, to analyze it, but instead of for climate, like in the original files, do it for AI. So there's a lot of programs, Python programs, multiple Python programs. I have to run six months of work, six, 10 hours later, there was a proper analysis by cloud code on all of this, with all in its directories, all the tables, every single one of the figures from start to end. So basically the difference is you talk to it, but it can deploy all these tools, it can do all these things, go over the web, it can run code. A second thing that I would tell you is that this is something that if they are not using, they would enjoy very much using is the deep research tools. Deep research tools are fantastic. So on the highest end research tier of these models, you ask it to research a question and you say, you know, populism has been growing. There are two explanations. There's cultural explanations, there are economic explanations. I want you to give me an in depth literature review of all the evidence comparing and you can ask it to it in your specific way the economic and the cultural theories and what's the evidence. And it's going to spend a lot of time, collect hundred references, classify them, tell you what's the problems of the evidence and write a very good research report. This is now better than what an RA could do over several months. So these are two examples of things that on the higher end, these richest tools, why are they useful for the world? So think of a lawyer, a transactionist lawyer. It's basically what they are doing is they are comparing a situation to existing presence and existing case law drafting, for example, the intellectual property lawyer or let's say a contract to buy your house. They're going to go and figure out similar contracts and it's going to be able to upload the knowledge that it has and convert it into a contract. Now if a law firm has this cloud code and incorporates all its contracts and it asks the system to use this knowledge to automate contract drafting and compliance and verification, etc. It's definitely able right now without any question to do the complete task.

A (18:29)

Yeah, and I have talked about this a little bit. I mean I started to use Claude code I guess about a month and a half ago. Now I have basically no coding background. You know, I had a couple of group lessons of C when I was in middle school and I, you know, did a little bit of programming and R statistical software in graduate school, but very limited. I was mostly a political theorist. And then I took a few weeks off CS50, the famous online course in computer science, just on edX. It was a very good course. But that was 10 years ago. You set me a kind of entry level coding task, like probing a number guessing game. I would not have been able to do it right. And now with this tool I've been able to program five different things that are actually of use to me in a very concrete way. And so it's just astonishing what it can do. I think more broadly though, some of the just pitfalls that AI used to have until a few years ago aren't really there anymore, right like when ChatGPT 3.5 launched, it didn't have an extended thinking modality. So basically the system was forced to think through the answer as it was talking. It's as though I ask you a challenging question and it's part of a game show where if you don't start answering within one second, and if you hesitate more than one second between any two words, you lose, that answer is not going to be very coherent. Now, these systems, if you offer free tier, talk to themselves and they talk you through the process by which they try to do the answer, and they try one answer and then they check whether that makes sense, and then they're like, no, actually I made a mistake, I should do this. So by the time they give you output, they're just thinking through it in a much, much bigger way. The problem of hallucinations. I wrote a post on Substack a few days ago about asking Claude to write a publishable paper of political theory. And a number of senior colleagues in the field wrote to me after I published this saying, absolutely, this would have been published in a top journal if it had been submitted. And I looked through some of the references, not for every single one. And it wasn't hallucinating. It now knows, by and large how to ensure that something actually exists and it flags if it's uncertain. It told me, look, I put in the page numbers for the canonical translations of Tocqueville. I'm not sure about those. Please go and double check them. I don't have access to that full page. If I upload the PDF of that book, it'll do it for me. But without it, it can't know. So it knows what it knows, it knows what it doesn't know. A lot of those problems have been fixed. All right, so now we go into the realm of economics. I don't know whether we have reached superintelligence as defined by Dario Amadei, where we suddenly have the appearance of a whole country of geniuses, but we certainly have the appearance of a whole country of middle class professionals, right? Suddenly the number of people who can competently draft a legal contract and do so in 10 seconds for very little money is vastly larger than it used to be in the past. So what does that do? What does that do? First for growth. If our economic growth was in some ways constrained by human capital, some ways constrained by the number of well trained people with access to a lot of knowledge able to carry out that work, well, that should mean that we're going to really increase economic growth. Shouldn't it? Or is it more complicated than that?

B (22:06)

No, I think the first order approximation is that you have an increase in productivity and that you have an increase in growth. I mean, I think that's a reasonable thing to start with. There are a couple or three kind of caveats that I think are important in trying to figure out how big that is. The first is, of course, organizations. The organization of work is intensely human. And as you were hinting from my recent post on the London housing, the reason 23 out of 25 boroughs of London are building zero housing this year. In 2025, they built zero housing. There were zero housing starts. Wasn't at all technological. And giving them better technology is not going to solve the problems with the neighbors, with the NIMBYs, with the greens, with all the things that stop construction, with the land, the regulation, the lawyers, all the things that stop construction that we already know. So first are organizations and all too human obstacles that mean that even when the technology is there, there are many other aspects that have to collabor. There are entire sectors which have these Beaumont characteristics. Right. So William Beaumont, I don't know if your listeners had somebody discuss this and maybe you discuss it, but Beaumont had this observation in the 60s that he's an economist who observed that string, famously, that string quartet would still take one hour to play a Mozart piece, the same exact hour as they would take 200 years ago. Four people, one hour.

A (23:59)

You know that this very old point, because nowadays no economist would talk about string quartets.

B (24:04)

That's right. So these people, this observation holds for a very large share of the economy, that in a large share of the economy, hairdressers, cooks, technology doesn't play any role. It's not just that there are bottlenecks, but it is that productive growth is very small because there's really no actual technology and no actual AI involved. Now, what is interesting is that in the sector of the economy that enjoys the technological change as the prices drop, it's perfectly possible, and we'll talk about demand elasticity in a second, it's perfectly possible that people reach association and that sector becomes smaller. For example, think of agriculture. It became technologically fantastic, but it became smaller and smaller and smaller as it was more productive because people's stomachs didn't grow. So the amount of workers employed went down. Now what that means is that the sector that has the technological expansion lowers, reduces its size. And the other sector, the one with the violinists, expands its size and its size. And as a result the weighted average not Just how much does this growth, but how much does this grow? But average. But maybe the sector is growing, is getting smaller. The other thing that I was referring at the start is the one way

A (25:26)

of thinking about this is just that everything that can be automated does suddenly become plentiful. And so that might not fully show up in GDP figures, but it does fundamentally remake the world. When I think about the agriculture case, as a result of our successful mechanization of agriculture, that's become a much smaller part of the economy. We're paying vastly less for food than we used to. Perhaps you'll tell me you'll understand the technical details better. That sort of underplays the degree of that change in the way that we track gdp. But it does mean that whereas for most of human history, even people in affluent countries, if you weren't at the very top of the hierarchy, were deeply constrained in how much food they could consume and were malnourished as a result, and died earlier as a result. Nowadays, if you are anywhere outside the bottom 20% of affluent to medium affluent country, food is not your primary expense. It's a significant expense if you like nice food and you go shopping for nice things. But if all you want is to be able to feed yourself on ramen and a few supplements in such a way that you don't get malnutrition, that is going to be a tiny part of your budget. And that is a fundamental positive transformation of human life, even if it doesn't fully show up in GDP figures.

B (26:48)

That is exactly right. Economists like to talk about welfare as the sum of the consumer and produce surplus. In this case, the consumer is really enjoying the biggest gain. So a lot of what happens with AI that we are seeing, at least already, is that a lot of the gains are going to consumers and not showing up in this GDP figure. So let me tell you an example. We have a dishwasher that is broken. We take a picture, we upload it to ChatGPT and say what the hell is going on? And says, oh, this thing is stuck, you should just remove this. And we remove it. Now our welfare has gone up, we are happier. We just solved the problem with dishwasher. Now there is a transaction that would have been some person comes to our house to fix the dishwasher that didn't take place. So the GDP would have been higher if this person had come and we would have paid them. But our welfare is increasing and if we can diagnose our own illnesses and we can know if our diet is good or bad, without going to the dietitian, if we can do our own contracts, all those things are increasing our welfare, but they are not indeed showing up in gdp. In fact, some of them could reduce the gdp. So it is true that a lot of the gains. I was talking to a CEO from China who was telling me, I think a lot of the gains are being smoked in the corridor. And I say, what do you mean? Well, I observe all these IT people and they are all more productive. I'm like, oh great, it's going to show up in better numbers at the end of the month and I don't get better numbers. So each person, it is more productive, they can solve problems faster, but then they are going home earlier or playing video games. So all those are gains that could happen and that could definitely not increase gdp. I think the other thing that I would mention is the difference between short and the long run. Imagine there are two sectors. One sector, sector A gets fully automated. So imagine we don't need, let's say, lawyers, to put one example, which is not exact, because lawyers have a lot of regulatory power. A lot of things they have. You need a lawyer, for example, for the court. But imagine these lawyers all, let me do something that has capital. So let's say, let's do lawyers. It doesn't matter. All the lawyers have to imagine we don't need any lawyers. We solve our own legal problems. All the lawyers have to move to sector B or all the people in sector A that gets automated needs to move to sector B. All the demand that is now like consumer surplus. We are not needing to spend money in legal problems. We can go and spend it on the other sector. At the moment, we don't do that and all the capital has to be moved to the other sector. All of these things take time. There is a moment when the GDP could be dropping because we don't consume legal or we don't consume dishwasher repairs. So in the meantime, when some tasks are automated, it could happen that in the transition, the capital needs to be reallocated, is written down, the labor needs to be reallocated, the demand needs to be reallocated. There is not sufficient demand also. So all of that transition could definitely not be like, oh, well, we're just growing and growing.

A (30:12)

I'm trying to figure out sort of what the aggregate effect of these changes might be. Right. Like on the one hand, you have traditionally, agriculture is a huge part of human activity. It mostly gets automated. The number of people working in agriculture is now astonishingly low. Output goes Up a lot. As a result, food prices go down a lot. So most of the consumer surplus is captured by humans and by consumers. And so it's a very good thing. I guess one thing that I don't fully understand is what actually kind of provides the basis of a negotiating power, the bargaining power of ordinary people in the agriculture world. The answer is that the production of agricultural products is now very cheap. But it turns out that humans are necessary for running all kinds of other elements of the economy.

B (31:13)

Exactly. And the agriculture, the machines.

A (31:17)

Yeah. And so there is a strong demand for human neighbor, and that's what allows them to continue to consume a lot of things. Now if we get. And this still sounds a little bit like science fiction, but I'm just trying to imagine this scenario. Let's say that we get to a world where AI can fully run agriculture, we don't need any humans in agriculture anymore, and it can fully run the system that is needed in order to manage the agricultural system. And it can fully run a couple of law firms that are needed to efficiently allocate capital to agriculture and make sure the most efficient agricultural firm is tilling the most land and so on and so forth. It may be that there's still all kinds of elements of a human economy where human work is needed. It may be that humans still prefer to have humans as teachers, and that humans are continuing to be required in medical decisions, perhaps because we don't trust the AI systems to do it, or perhaps just because there's regulatory obstacles to fully automating those. But if all of the underlying productive processes that actually produce material wealth are no longer in human hands, or at least they no longer require humans, is there a kind of perpetual mobile where the sort of circular economy of humans is enough to sustain affluence? Or does there need to be some kind of relation back to material production for this whole construct to sustain itself? If all of the need for human labor is produced by the fact that it's extremely expensive to look after old people, and the fact that, you know, stupid regulations won't let us buy houses, and you know, anybody who has capital is willing to pay a lot of money for a house because they need to live somewhere, and you know, some people who aren't really needed in the economy continue to have to be employed in human shape because of regulation, like, is that actually enough to sustain affluence within human workers? If all of the actually productive processes can be done by non human workers?

B (33:28)

Let me break it down into a few parts. So the first the demand, the kind of satiation case that we're discussing, where the sector gets smaller, doesn't necessarily have to be the case. In fact, in many sectors, as technology gets better and they get more efficient, in fact the sector grows in size. This is called the Jevons effect, after William Stanley Jeavons, an English economist who observed that coal was getting machines using coal were getting more and more efficient and they were consuming more coal rather than less. Why? Because as they were getting better, they were using so many more things that the coal consumption was going up. So in many sectors, think about health, think about energy. As things get more and more efficient, it's unlikely that the sector as a whole will shrink. In fact, it's more likely that it could demand more humans. It would grow in size and demand more humans. I would think the more elastic demand, the sectors that are likely to grow when the prices go down would be things like health and energy, for example, just to give two simple examples. Now second thing that is really important is the idea of complements, and you were hinting at it clearly in your question. There are many situations where a human is needed in a bottleneck. So Even if the 99 first tasks can be automated, if the hundredth task needs a human, the 99 are abundant, but the scarcity is still the human. And the human is going to get the rent and is going to get

A (35:03)

the labor, but that depends on the human being scarce. Right? Which is to say that if that task requires a very high level of qualification and you either need millions of humans to do that task because they are so productive, in which case a lot of people are going to be in relatively decent employment, or you need seven people to do it, but they have to be excellent, in which case those seven people are going to get huge rents. Some of that economic gain is going to go to those people. But that doesn't mean that, let's say you have, what is it, 5% of the male workforce in the world being employed as drivers, some percentage like that, that means that each of these drivers necessary for each of those rights. And so. But the rent from the need for human drivers is very broadly distributed. Each of these drivers is probably not very affluent, but the rent for that activity is very. Now we say 10 people have to supervise all of the self driving cars and let's say they have to be incredibly qualified. And there's very few people who are able to do that well, perhaps they're able to capture a lot of that rent, but there's only going to be 10 people who get that money, Right. Or let's say that it needs 1,000 people, but a million people are able to do that job. Right. Well, in that case, the wage for those thousand people is going to be really low because any one of them can be fired and there's 9,999 waiting outside the door willing to replace their position. So it depends a lot on those kind of details, right?

B (36:42)

Yes, absolutely. So I'm writing a book on this point. It's called Messy Jobs. And it's going to be going to be out well, it's going to be submitted in a couple of weeks, I hope, and out by May, let's hope. And the argument of Messy jobs is that there is a big difference between task and the task and a job. So Geoffrey Hinton, who you have in your podcast in the past, is famous for having said in 2016 that nobody should have studied radiology because radiology was just an expert system that could scan photos. And of course, any expert system was going to be better. It was going to be trained on hundreds of thousands, not hundreds of millions, millions of breast cancer scans. And it was going to be perfect at detecting those cancers. Now, the truth of the matter is demand for radiologists has never been higher. Their salary is growing. The numbers are growing. It's the third highest salary of any medical profession in the U.S. why? Because the task is very different from the job. The technologist imagines the project manager is a guy looking at Gantt charts in the computer and imagines radiologist just looking at scans. And only 30% of the time of a radiologist is used looking at scans. They have to do the diagnosis plan, they have to talk to their colleagues, they have to talk to the patients, they have to do many other things. So I think that the first crucial obstacle to your, to your dystopia is that automating parts of jobs tasks is not automating the job. And I invite all your listeners to think of what they did today and think of which of the things they did today. I went to a workshop. We had a job market seminar. I had a meeting with my colleagues. I had students walking in, I worked on a paper all of those tasks and think about how many of these tasks you could replace with a machine and you will discover that many of them can't. Definitely the task we are doing, which is having a human conversation about something, can't. So the second idea is it's really very different. A job and a task and many aspects of the jobs can go without the bundle, it will get rebundled, it will look different, but it will not go away. There are reasons for that. Part is the technology you need to direct the AI. You cannot just let it do its thing. Meaning the AI is psychopathic. It tends to agree with what you say. So if you want to direct it in direction left, it says, yeah, left is great. Yeah, let's do left. If you wanted to direct it in direction right, it would say, yeah, right is the best. You're saying smart, you're the smartest. Right was the good way to go. So what you tell it is going to matter. And it means that. That somebody's going to have to be exercising judgment. But it's crucial to realize that this is not solved by AI being smarter and smarter and smarter. Think of managing a family. Managing a family, which all our audience is familiar with. Everything that you do in the morning with the kids and moving around and deciding. And a lot of that is not automatable because a lot of the knowledge of what's going on is tacit. It's in your head, you know what's going on. And no machine can tell you whether the kid has to wear these boots or not, or whether the school is a day that they need to do this or that. You're not going to have to be deciding all these things. So authority is inherently human. Making the difficult decisions is inherently human. Being the consultant who does PowerPoints, yes, it can be automated, but does the consultant only do PowerPoints or does he go to the. Or she go to the company, listen to the work workers figure out where the problems are, how to automate this bit, how to do this other bit better. A lot of that is tacit. So I would push back against the idea that entire jobs are going to be done autonomously. Yes, you're right. Autonomy. The cars pass the autonomy threshold and basically the cars can self drive. And that means suddenly the supply is infinite. You can, you can have a lot of drivers like all the machines. And that means eventually the voyages collapse. So that's a good example that you picked. Is that a normal example? Is that an example where the task is very clearly defined? It's very repetitive every day going back and forth. It's all in the computer. Is that the majority of jobs? My claim, main claim of this book of messy jobs is if you think of the elasticity, many things will grow in demand. If you think of the complementarities, there is going to be scarcities, crucial, as you say, Yasha. But there are going to be Many scarcities that human can exploit. And this is without getting to the point that you were starting the question with, which is demand for human services, which actually I'm not sure is this large. I'm not sure that people necessarily, when they're old, they necessarily will want a person bossing them around saying, are we? Well, today, I mean, I might prefer a robot who's like, taking care of

A (41:38)

me, including a lot of the more intimate tasks that are involved in elder care and so on. Right. Like, would you rather have another human wipe your ass or would you rather have a machine wipe your ass? Like, you certainly want some human company. Right. I mean, once your ass is wiped, you'd love to be able to have a conversation with a human.

B (41:52)

Agreed. But you definitely don't need a person for that. You would rather not have a person for that.

A (41:58)

So I think I have a middle position in these debates insofar as I have a position at this point. But we'll come to that. Just to push back on a couple of the things you said, fau and again, I'm not coming from a kind of maximus position. I agree with you that a lot of the predictions that all the jobs are going to be gone in two years and just as testament of people who haven't thought about politics and who haven't thought about the real world in many ways. But some of the examples you gave, I'm a little bit less convinced of. To give one example, can AI outsource the managing of a family? I mean, part of a family is just that you're negotiating between human beings. You're trying to come up with a plan together. Even the AI can make a plan that is pareto superior to whatever plan you'd have. Part of what it is to be a family is to make those plans together and say, what should we do today? And so. So I agree that sort of on an emotional level, you might not be able to outsource those things. On a purely planning level, I think absolutely the things you talked about, AI could outsource. And in fact, I think many feminists would say that is what we've been arguing for for a very long time. Because it's often women who do the kind of emotional labor and the second shift and so on of, you know, keeping track of the fact that Timmy has to go to the dentist tomorrow and Tammy has to go to ballet the day after. And have we made sure that the dress that she'll need to wear to her ballet class has already been washed, et cetera? It would take an invasion of privacy like an AI that's part of all these conversations and that immediately notes down when Tammy says, oh, don't forget, for my ballet practice next week, I need X or Y. But can AI do all of those things? Absolutely. And could it in fact save some marriages in the process of doing that? Probably, yes.

B (43:49)

So here's why I disagree. So there are information processing tasks, and you're right about information processing tasks, and we need to synthesize all this information. We need to put it in a form that can be processed, and we need to make a decision. There are other tasks that have nothing to do with information. I mean, your wife or your kid is upset and why is he upset? And somebody needs to talk to the kid and someone needs to decide, yes, I had told you the optimal plan from the pursuit of the family was that tomorrow you couldn't stay home and you were going with your friends. But I've listened to you and I decided that you stay. A lot of it is not information processing, is you understand the kids, you understand what a look means. You understand when a look means from your wife or from somebody else, when a look means yes, I will do it, or when they say yes, but in fact they mean no. And there is a lot of tacit local knowledge that goes in management, in the family and in a business. And we're talking politics, but we're not just talking politics, we're talking emotions. But not just talking emotions, we're talking local knowledge, interpersonal knowledge. You know your wife for many years and you know when you can push, and she knows when she can push all of those things, you're saying, well, the machine could know. I don't think it could know. I honestly don't think it could know. You are the contractor. You know, let's now go to the management. You're the contractor. You know which electrician is the one who's reliable and the one that, that played tricks on you last time. And yeah, Ken da now know whether you can use something to get this electrician to be on time or not. I mean, we are talking about, I mean, just calling it intrusion. We're talking about a level of interpersonal and tacit knowledge that is unreal because also think about this. A lot of the tacit knowledge on the jobs is knowledge the employees have that gives them power they're not going to be happily just sharing with AI. Oh, AI. You know, you should know that my colleagues and so has this problem with the boss and that he never wants to Work with that boss. No, I mean this kind of stuff is going to remain on the heads of the humans. So I believe that yes, the information processing task can and will be automated, but a lot of it has to do with not just the emotional and the social skills, but the tacit knowledge and interpersonal knowledge that probably the machine will never gain because it cannot capture it.

A (46:19)

I have two kinds of different lines of questioning about this. The first is just about if we go away from the extreme predictions, if we recognize that clearly at this point advanced AI tools are capable of doing a lot of the tasks involved in knowledge production, that presumably means that some jobs are going to go away, right? The idea that that AI is incompetent, that it can't do all of those things, that it's all hype, that it's all a bubble, we agree as suddenly, right? I think we also agree on the other end that a lot of those real world frictions are very real, that jobs are messy because, well, the world is messy, right? And that therefore the idea that the moment that Claude beats doctors in a bunch of medical, in a bunch of stylized medical questions, which it more or less does now to expect that therefore tomorrow there's no longer going to be any doctors, is really naive and doesn't understand the real world. But what happens in that middle space? What happens if suddenly the demand for white collar works is reduced by 25%, perhaps by 30% and it doesn't have to happen between today and tomorrow, it happens over the course of 10 or 20 years, right? You just see a continuing gradual reduction in the demand for that kind of high skilled work as existing firms automate out work, as firms that are too stubborn to do that, that are not able to do that, are out competed by new entrants to the market which are AI native in the same way that in many areas of the economy it took Internet native companies to out compete old ones, until you really saw some of those productivity gains come online. That's going to be a significant process. It's not going to happen from one moment to the next. But in a way that raises an even deeper and equally troubling possibility, namely that A the job market is going to slowly slump for an extended period of time and B, that there's the famous somewhat apocryphal boiling frog, right? That rather than, you know, if everybody lost the job in the course of two months, well, perhaps we would all organize and you know, demand. I don't know what some way of being made whole, right? If this is just going to show up, up as decades in which the bargaining power of ordinary people diminishes and diminishes and diminishes because the demand for human labor just continues to fall in a messy, gradual, haphazard way. That could still be an incredibly painful period ahead for ordinary people.

B (49:20)

You're more or less describing my scenario of transition between sector A and sector B. And we know that in the industrial revolution, what is called was the. What was called the Engels pause, which is something between 1790 and 1840 or between 1800 and 1850 where basically this was happening. Workers wages were stagnating or dropping and the workers were in trouble. And then GDP was multiplied by two over the following years, following 50 years, 20 to 1900. So yes, it could happen that over a period of time, the transition is hard. Now against that, I would say now human beings have been automating tasks for hundreds of thousands of years. I mean, I think all the human existence been automating things since the abacus and since counting sticks to count. So it is something that we're used to doing. I would also say it is enormously promising to people who are thinking of trying to stop this. I would say it's enormously promising to, for example, have, if we have this cancer genes in the data center, to have the cure of cancer or the cure of Alzheimer or the cure of many things. If these technologies can really advance science by decades at a time. So I would not think that that organizing to stop it is the right thing to do. I would think that the combination of bomball sectors, sectors where nothing is going to happen because they are outside of this. If you add up the sectors, the fact that you have bomball sectors, sectors where nothing is happening because there's no technology. And everything from public sector jobs to arts to I mean like the music, to the barbers and the hairdressers and the cooks and all of this you add, by the way, dog walkers in the United States is. Or pet care in General is like 1% of the population just doing all the pet care sector, which is a sector that we wouldn't imagine that. And of course it's not affected by AI. If you add sectors with very elastic demand, which are sectors which are going to grow like health and energy, and then you add the messy jobs, the jobs where even though they're part of the task getting automated, but there's many tasks that don't get automated and the jobs continue, from managers to entrepreneurs, Then you have and you add the complementarities we haven't talked about this but there is this idea from David Autor about the new middle class which is you think of a nurse who is empowered with this genius on a box, this nurse who can now diagnose really complicated illnesses and she can. Or he can hold the hand of the patient and do all the other parts and now solve more problems. Of course, as you said then maybe everybody wants to be a nurse and we would have to think about the supply of nursing, the supply stc. But you add up all these things. I think you go away from the feeling that there's a cataclysmic change and you more go to yes, it's automation. Yes, it's going to be a bigger revolution than what we've seen in the last 50 or 60 years in terms of automation. It's more similar maybe to industrial revolution. But no, I don't think it's going to cause widespread long term unemployment. We are going to be able to from tiktokers and Instagrammers to people walking dogs, there is so many new jobs that we wouldn't even think. I mean this podcaster who was going to tell you you would be a podcaster.

A (53:03)

I mean I don't know if the vision of the future is that humans are going to be fine because we're still going to be tiktokers and Instagrammers and dog walkers. I'm a little bit skeptical about how

B (53:11)

ringing in those the pet care sector is 1% of the population. That is not just dog walkers. It's nurses and people taking care of the pets and all these other things.

A (53:22)

Let me ask you about the dog walkers. I think one interesting thing that's happened over the last 10 years, which just shows you how epistemically modest we should be about all of this is that I remember all of the conversations about all the drivers in the world are going to lose their jobs and and somehow that was linked in the conversation about populism to that's why the Midwest went for Trump. I think there's never really a connection there and so they should all learn to code. So now it turns out that AI is really good at coding but because of a set of technical issues that ended up being more hard to solve for a while now this technical issues are mostly solved. Waymo is very efficient and much safer then human drivers are there still a lot of regulatory obstacles and so even for the number of rides that Waymo is offering is going up exponentially, it's still a very small share of the Market, most human drivers are still fine. Again, this is going to take longer to play out than a lot of people think. And we're now in a world where knowledge workers are seemingly about to lose their jobs, but all of the manual traits are safe. We're assuming a world in which the plumbers are still going to be fine. To revert to your example, the dog walkers are still going to be fine. Well, I watched, as many other people did on the Internet, the quite remarkable display by Chinese robots for the annual Chinese State Television gala. The progress in their dexterity from a year ago to today is just astonishing. We know that the ability to combine the manual dexterity of these machines to visual processing and understanding of a world is going up very quickly as well. So I am personally waiting for the ChatGPT 3.5 moment in robotics. I think that it won't take very long for there to be some consumer products that is actually usable. We're getting close to that. And the applications in the industrial sector are likely to increase as well. Again, I don't think it's going to happen tomorrow. I think it'll take time to fully be implemented in the economy. But when we're talking about a timescale over decades, when we're saying, well, in 20, 30 years, the fact that more and more of these knowledge work tasks are going to be automated because those skills can already be done by AI, and perhaps it'll take a long time for firms to reorganize and for new firms to enter and so on, you know, but it's okay because perhaps we'll all be, you know, in the pet care sector. Well, sure, but you know, that assumes that in 20 or 30 years we're still not going to have figured out household assistance, that if you're off at the office or doing whatever you are during the day, you can't have a little robot who walks your dog in your stead. And that seems to me, given the rate of progress of this technology, like a pretty significant background assumption, I think

B (56:31)

that is very, very correct. I think that robot's physical AI, let's say, is not that far. What we have seen in the past is that the capital is in what economies call inelastic supply. You can always invest more in capital, and that means the capital, eventually the rents on the capital get competed away and the robot gets sold at a competitive price. And that means that people can use robots for care. And remember, we have a lot of fertility problems and growth, population problems to pay our pensions. And having robots could be A solution to all that. It's like more population growth. Now in a world where this is competed away again, we're back to consumer gains. So the capital doesn't earn extraordinary returns because there is infinitely elastic supply of capital. More people can invest in making more robots. And what is the scarce resource? Well, the scarce resource obviously is going to be land, it's going to be energy, but it's going to be whatever human labor is needed still. And that human labor, it could be that we're working less hours, it could be that we are able to enjoy more leisure, and it could be that human labor is employed in a whole range of jobs, which indeed, you're right, we can't anticipate. What we shouldn't imagine is that somehow there's an economy that works without humans because all the value is in the humans. What does the economy generate value for if nobody's buying the products? The definition of value is something that is worth more to humans than it costs to make. That is what value means. If there is no human who can buy stuff because they are all poor, there's no value. So it is the way the economy works that the return to capital gets pushed back down to the normal return to the competitive return, and that the rents get captured by these cash resources, in this case the complementary labor that is needed still in those moments. You're right. Nobody can anticipate what happened in 30 years time. And yes, I agree that both physical robots and AI and cognitive AI are going to be a big revolution. I don't think we should be thinking of this as an apocalypse. I think that there is a lot of complementarities, there's a lot of scarcities that favor human labor still. And there are a lot of areas where this doesn't really bind at all.

A (59:10)

Tell me a little bit about the state of the empirical literature. I understand that there's a real distinction in micro and macro studies, a real distinction in studies that look at the extent to which particular tasks can be automated and the extent to which the overall job picture has changed. When I look at the fields that I know a little bit, I worry that that is an indication of what is yet to come, rather than an indication of the fact that AI won't have a big impact. You mentioned translations earlier. Another thing I've been thinking about is index making in the publishing. There's all of those things where basically there's been no change so far as I can tell. My next book is going to be translated by human translators. Perhaps they don't actually do it and they privately send it to Claude and capture the consumer surplus by going out and having a nice vacation while we pretend to be working on the book. But in terms of the actual kind of economic flows, nothing has really changed. I don't know how long that's going to continue to be the case. It is very sticky and very complicated to change those processes. Somebody needs to say, I'm going to be the asshole and fire over translators. And I'm going to deal with a backlash of the agent saying, my author doesn't like the idea that it's AI rather than a human who's doing this. And perhaps it's a newspaper story and your customers might be upset. There's all kinds of arguments to be risk averse from being the first mover to make that change. But I will tell you that one of the things I've created for myself with Claude Code is a personalized translation tool because I publish my articles, including some podcast transcripts, not just, just in English, but also in German and French. And that is not just better than the off the shelf tools. It is at this point better than all but the very best translators I've had, the very best translators I've had, who I'm deeply grateful for, particularly in France, I think are still better. But 90 plus percent of the translators I've dealt with, professionals who've translated famous books by famous people are significantly worse now. Right? For now, I agree. If economists tell me, hey, actually translators haven't lost their job and none of this has changed that much, I believe it. Right. I can see that. If economists tell me, and based on the fact that in the three years that AI has existed, in which two out of those three years it really wasn't at the level yet that it's coming to be, and people have not yet integrated those processes sufficiently, we can make predictions about the future. I'll say, come back to me in 15 years and let's see whether those translators still have jobs in the library.

B (62:02)

I don't think anybody's predicting that translators are still existing. I said, so far so good. Maybe the person falling through a window. So I do think jobs go away. Newspapers are digital and there were lots of people in printing presses and paper and all the industry and all the newspapers were automatized, including my grandfather, whose

A (62:21)

job it was as a young man to lay the newspaper letter by letter, and then later he helped to manage the printing side.

B (62:30)

But yes, yeah, that's gone and that's been human history all the time. So you said was the empirical evidence. The empirical evidence up to now is positive. So when they've done randomized controlled trials where they have given in a control setting the micro evidence evidence, they've given a control setting the AI to a It was a previous AI, but customer service, customer support agents, what has happened is the least advanced customer support, the most junior ones get a performance similar to the more senior ones when they've done it with writing tasks. The worst writers get a performance similar to the more better writers. So it helps the least advanced people when they've given it to software programmers in three different tasks they've seen the the software programmers were less good get to program closer to the skill of the better software programmers. So microstudies seem to be finding all the time complementarities rather than substitution at the aggregate level there is much more confusion and much less clarity. We don't see big drops in demand. There are some canaries, as I was telling you from that paper, canneries in the comment, there is some preliminary evidence that hey, maybe there is some drops in junior jobs. I think when we think that the research task, the PowerPoint task, the Excel task, those are the cases to automate. We have to imagine the junior lawyers, junior consultants, junior investment bankers will not be recruited as much because you can do a research task without the junior person. Now it turns out the McKinsey class of this year is bigger than before. They keep hiring people. It doesn't seem like they're hiring less. So far so good. I agree with you. That is not a forecast of the future. I don't mean to say a forecast of the future on the basis of the fact that we haven't seen much. That's not the point. I think the point however is there are indications that complementarities are important, that people who use this AI produce better and that substitution is still limited. But it's hard not to think that tasks that have to do with just basic PowerPoint or research etc. Is not going to be fully automated. No, no, I, I, I don't think we want to make this 15 year forecast. I agree.

A (65:12)

Thank you so much for listening to this episode of the good fight. In the rest of this conversation, Lewis and I talk about what young people should do to future proof their careers. This moment is disorienting for everybody. It's particularly disorienting for people who are just trying to figure out how they can have a meaningful and hopefully well remunerated job not just for the next five years, but for the next 50 years looking Lewis gives some really interesting advice for how you can prepare yourself for a future in which messy jobs are king. We also talk about the phenomenon of bullshit jobs. My dear producer Leo thinks that a lot of jobs in the economy are bullshit. I wonder whether that's true and if it is true, whether that indicates that a lot of jobs can easily be automated and are eventually going to go away, or whether that means that bullshit jobs somehow might persist even when more and more jobs in the world might turn out to be bullshit because you could just have AI do it instead of you to see what Louis thinks about this, how we puzzle through this question, Please become a paying subscriber. Please set up the premium feed of his podcast in your favorite podcasting app by going to writing.Damonk.com listen.

C (66:56)

Putting off replacing your window treatments because you think it's complicated? At blinds.com, we've spent 30 years proving it doesn't have to be, and today is your last chance to save big on Spring Black Friday deals. Whether you want to DIY it or have a pro to handle everything from measure to install, we've got you free samples, real design experts, and zero pressure. Just help when you need it. Shop up to 45% off with minimum purchase, plus get a free professional measure during the blinds.com spring Black Friday last Chance sale. Rules and restrictions apply.