The AI Economist: The Data on AI and Jobs Tells a Different Story

I want to get to the bottom of the question: are jobs going away or are they not? I believe that millions of white-collar workers are going to lose their jobs in the next 12 to 18 months due to AI. The AI jobpocalypse is real. There's lots that's probably right, and yet at the same time the overall narrative is probably wrong. A lot of the jobs he described are actually in the top 20 and not in the bottom 80. So if you see your job is 75% exposed to AI based on this latest study, that might not mean anything. You could actually be making more. Absolutely. If your job is 75% exposed, then you're going to be spending your time on the remaining 25%. It could mean higher wages. I want to say one more thing about jobs that's important. Jobs aren't good. Getting paid is good, and finding meaning is good. But a job per se is something we do, and it's part of our society, but it's not necessarily good. And I do think on professions and jobs the long run looks good. There's risk in the short run. Are there skills that become quite successful? Are there skills that are quite non-negotiable? Yes. I think there's four things. There's a growing argument that the way we are rolling out AI is economically self-destructive. The more people AI replaces, there's going to be fewer people who can actually afford anything in the economy. The economy self-destructs. And as an economist, do you agree with this thesis? No. Today I'm sitting down with my former professor Avi Goldfarb. He is an economist and the Rotman Chair in Artificial Intelligence and Healthcare. He is the chief data scientist at the Creative Destruction Lab and a professor of marketing at Rotman School of Management at the University of Toronto. And today I want to get to the bottom of the question, are jobs going away or are they not? I'm Shane Bovell, and this is I've Got Questions. Professor Goldfarb, do you mind if I show you a TikTok? I am terribly sad to record this. AI is now able to do the work of a very, very smart human in minutes or even seconds. This is going to displace marketers, coders, designers, lawyers, accountants, call center workers, you name it. The result of this is going to be an ever more extreme winner-take-all economy where the top 20% or so are able to thrive and then the bottom 80% are left behind. And when people lose their jobs, it affects them. It's dry cleaners, dog walkers, hair stylists, restaurants, all local businesses that see less people who are able to spend. Personal bankruptcies are going to surge. College graduates are going to have a very difficult time finding entry-level jobs. People are going to be very, very angry by the fact that this path to middle-class status and social mobility is going to be broken. We should be doing much, much more, including a universal basic income. The AI job apocalypse is real and it's underway right now. What are your thoughts on Andrew Yang's take? It's a popular one. What do you think? There's lots that's probably right. And yet at the same time, the overall narrative, I think, is probably wrong. What I mean by that is if AI does all the things that he says it will, it means it's going to lead to extraordinary abundance. It only makes sense for AI to be replacing accountants and lawyers and coders if it is that much better, which means that there's going to be abundance for the rest of us of the ability to access code, access the law and the regulatory state, and do accounting in the case of accounting, for everybody who's not in those particular categories, they're going to be able to produce much more and consume much more. So the worry that I think he gets right is there are going to be some jobs, and I actually don't know if they're going to be those ones, I don't know what they're going to be, but there are going to be some jobs that are at meaningful risk because of what AI can do, and maybe even in the short term. The things that I think he gets wrong, are if that happens, that means the people who aren't in those jobs are going to be much better off, because they're going to have access to those services at a much lower cost than they did before. Another part of that is the jobs he described, he said, you know, the top 20 and the bottom 80. And what we're seeing so far in the data is that the professions that, are not experiencing rapid employment growth in the US, are those that are not AI exposed, which means they don't use typically writing, drawing, and coding. And writing, drawing, and coding jobs tend to be high wage. So what we're seeing so far, there's a recent paper called "Canaries in the Coal Mine". And in that paper, there's a meaningful worry, which is 22 to 25 year olds that are, focused on those AI exposed jobs, are doing worse in the labor market. But the hidden message in that paper is that everybody else is doing better. And there's been, in the US economy at least, rapid employment growth well beyond, you know, for every category except for that narrow one. And that narrow one represents a small fraction of overall work. Okay, so there's a bunch there. So when people hear, it is maybe possible, and let's say it is legal services that, randomly, let's just pretend that that's the job that AI ends up doing. We're all better off who are not in legal services, because those services become so cheap we can use more of them. Yes. And then what happens to everybody in legal services? So if that starts to happen for a few professions, what happens to everybody in those professions? So, a few things could happen. I can tell you what's happening so far and what might happen. So, one version of it is, lawyers simply need fewer lawyers. We don't need zero lawyers, okay, but we need fewer of them, particularly at the entry level if that's what the AI is doing. And the narrative, at least, he's saying is, well, AI can do a lot of entry-level legal work because it involves reading documents, redacting them, summarizing, etc., which is the kind of thing that AI does well. And if that happens, then legal work becomes more top-heavy. The folks with 10, 15, 20 years experience for now are probably fine. But the new entrants are going to be less productive than new people in that industry a while ago, which means we're going to need fewer of them, and/or the new lawyers that we get are going to be paid less than current new lawyers. Now, in terms of equalizing, unequalizing, junior lawyers getting paid less is bad for those junior lawyers. It is sort of in terms of thinking about the overall income distribution. If the rest of us get better services, and that effectively means we're better off, that would be equalizing. Yeah, so from a net macro standpoint, we are all better off when law becomes better and it's cheaper, even though at the entry level, there will be some people that may be out of a job. Yes, so it's absolutely not all rosy. So, as in, there's reasons to worry about the folks who are trained for years to do something, and then AI comes along and can do it as well or better, and so it's going to be hard for them to find jobs in that area. What we're seeing at least so far is that is in a handful of professions at the junior level, and it's an open question whether that's going to affect the overall economy. So, number one is we might need fewer lawyers. But there are other possibilities. Another possibility is we actually just need, you know, as lawyers get, as the AI takes some of that, some of the tasks that lawyers do, lawyers will spend more time on other tasks that relate to the law. And if they get good enough at those tasks, it's possible that there's no change in jobs at all. They just start doing different things. And if as law then becomes much more efficient, we realize that there's lots of opportunities to, you know, to use lawyers, maybe lawyers aren't the best example, but if we can use lawyers than we had before, that would be great. So, if you imagine if right now a simple contract still costs $500, $1,000, $5,000, if AI makes the law much more accessible and you can now write a contract for $3, we're going to have a lot more contracts. And if we have that many more contracts, then maybe we could need a lot more lawyers to help with whatever the last things are in those contracts that make it so that we need a human in the loop. So, then it seems like that thesis is kind of wrong from Andrew Yang or from anyone who kind of proposes that idea that we're headed for a dropocalypse, that they're not looking at the part two to that story. For it to be right, two very specific things have to happen. One is you have to, one possibility is you have to be able to automate an entire job, not just most of it. As long as there's a bottleneck on the job, then you still need somebody, people to do something. And second is as the price of things fall, do we actually want more? want more of it. And there, it's going to differ industry by industry. So as the price of food has fallen over the past 100 years, yes, we eat more food, but we actually, you know, we don't eat 100 times more food or 1,000 times more food. People, there's only, there's a satiation there. For other types of services, maybe there won't be a satiation, like in healthcare. It's possible we'll just keep wanting more and more and more and more. And so part of this is what we call elasticity of demand for the services as the price falls. Depending on what that elasticity looks like, we can end up with more people or less people as these things get cheaper. Andrew Yang talks about the top 20% are going to kind of have runaway wealth, and then it's the rest of us in the 80% that are kind of left to the ditches. But you pointed out that it's actually the top 20% that AI, if anything, may be coming for? So far, that's what's happening. Okay. Can't promise that's what's happening. But it will continue to happen. So you want to separate out skilled versus unskilled labor, like labor income relative to capital income, which is if you own stocks or if you're owners of these companies and these companies start to do, AI companies start to do much better and make a lot of profits, then the owners of those companies do better. On labor income, what we've seen so far. What we've seen so far is the things that AI tends to do well are things that relatively high-skilled workers do, think coding or aspects of the law, like junior people in those professions, but they tend to be, at least within their age cohort, near the top of the income distribution in the 80th to 90th percentile. And the research we've seen so far is that those are the jobs that are most exposed to AI. Okay. So there's a paper called GPTs are GPTs. It came out in Science. It came out in Science a couple of years ago that identifies those jobs as most exposed. Now, the reason I'm being cautious is exposed could be good or bad. So, yes, they're more exposed. And the narrative that Andrew Yang had there is exposed means is bad. But exposed could actually be good, which is that some parts of your job get automated and you become super productive. And in that case, you do better. So if we take his narrative as given, which is if your job is exposed to AI, that's bad for you. Then the second part doesn't follow, because the second part being that then rich people are going to do even better and poor people aren't, it's the rich people whose jobs, the relatively rich people whose jobs are exposed. Now, if we accept the second part of the argument and say, well, exposed is actually good, and that means you're going to do better and better and better as you can take advantage of what AI can do, then the first part of the argument doesn't hold up, because there it's, well, the narrative he had was AI is going to actually take jobs by doing things that humans do. The opening was AI can do these things that humans do, and over time, as they do that, those humans are going to be out of a job. So there's reasons to think high-wage people, that some people will be exposed to AI, and those exposures are going to be bad for them. It's going to be substitution. And in that case, some people will lose their jobs. And that's a real concern. And there's reasons to think that AI might increase inequality, as it superpowers people near the top of the income distribution. But those two things don't really work together. And so what, I'll give you an example from history. What we've seen with computers and the internet is, with computers and the internet, we didn't have the same automation anxiety that we have today. But computers. Computers and the internet very clearly increased inequality. And they did it not by replacing humans, but by empowering skilled humans to do even more. And so what computers and the internet did is they helped people at the top of the income distribution be much more productive. And while people who historically had been at the bottom of the income distribution didn't lose their jobs, they only did a little bit better, while everybody else did, while those near the top did a lot better. With AI and the early data, we're seeing some kind of change. It's something that looks like the opposite, which is it's those most at risk are near the top of the income distribution, and those who seem to be helped are in the middle or toward the bottom. Now, that may continue, but it might not. So there, AI can potentially become a great equalizer, depending on which parts of the chain it automates. Absolutely. Absolutely. And so the key thing to remember is that one person's automation ends up being someone else's augmentation. So if we automate, for example, what physicians do, what doctors do in diagnosis, that's going to be bad for doctors, but it probably will superpower nurses and pharmacists, because now they have access to these tools that others didn't. And so that would be an equalizing example, because you're automating what somebody at the top does and empowering what people at the bottom do. Now, there's other worries. On the other end, if we automate what taxi drivers do or Uber drivers do, then they're in the middle or the lower end of the distribution, income distribution, they're going to be worse off, and folks who now don't need to buy a car, buy car insurance, and all this might be better off, but those people might be in the higher end of the income distribution. So really sort of depending on what gets automated and what gets augmented, or what gets automated, depending on what gets automated, the people who will benefit and the people who will be hurt will change. So to summarize that, so if someone who says, okay, AI is coming for most jobs, and especially in the knowledge economy, there's probably going to be a job apocalypse, well, you have to actually first start to think, and if you're worried about income inequality, it's actually already higher wage earners that would be exposed to artificial intelligence. So there's a world in which we see a bit of equalizing, but there's also a world in which if AI doesn't end up automating everything within that role, we could see a bit more of a wealth divide because people, let's say the lawyers, the doctors become more productive, but if it does end up automating some of the lawyers and the doctors, those tasks could be paired with somebody who theoretically doesn't make as much, we'll say the nurse compared to the doctor, and then that nurse becomes supercharged, and we see more equality. Yes, and I'm not sure how much everyone can see your hands, but you had your top hand going down and your bottom hand going up. I think in many cases, the more likely scenario in the nurse-doctor example is the nurse's wages go up, the doctors figure out other things to do with all that training. And so it's equalizing where the folks in that case at the top are probably going to be stagnant, not doing worse, where folks who are relatively lower income, or aren't low income, but relatively lower income do much better. So then how do you think about a junior in this? In that world, because then how does somebody even step into the world of law, or is it that the field itself also starts to change? So right now, maybe a lawyer practices the way they've practiced for the last four years, but then we're going to have an agentic workflow in law, and maybe it looks more like simulations and agents and all of these things, practicing in trials, but it's agents doing it. So then the junior doesn't jump on the ladder the way the senior would have, but they onboard in a different way, and law also, because isn't the field also going to change? Absolutely. And I think we were just talking about medicine and law, and I hypothesize there's an interesting contrast there. So right now, to become a doctor, you do a lot of post-secondary schooling. And to become a surgeon, you may not really be a fully-fledged surgeon until well into your 30s, if not your 40s. It takes a long, it's a lot of training. Now, with an AI tool, potentially, the juniors who used to need to be in training for, you know, five, 10, even 15 years post-undergraduate, now will have some skills right at graduation. And so we might have a lot more juniors. They'll look different. We may or may not decide to call them doctors. But as medical professionals who can take advantage of what AI diagnosis and AI recommended treatments can do. And so in medicine, depending on what, you know, if you focus on junior doctor, maybe, but if you focus on junior medicine. If you're a medical professional, there's reason to think we're going to have a lot more juniors because just training to get up to the frontier is going to take less time. In contrast, law, the way we historically train lawyers, is by having them do a lot of reading and summarizing and redacting to just have enough of a sense of what documents look like in a law so that they eventually get experience and become partners and can identify new ways to think about the law. And ways to help their clients creatively. Now, in an AI world where, you know, as AI gets better at those tasks, the productivity of those junior lawyers goes down. They're just not as useful relative to AI. There might be some things for them to do, but they're almost surely going to be different. And so we may end up in a world where law ends up looking more like surgery, where to get the skills to truly be a lawyer. that can help your clients, you need to be apprentice for a very long time. And so junior lawyers' wages could go down with the promise of once they get the experience, it's going to be worth it, just like the way we currently treat our sergeants. Okay. So this means we have to stop thinking so linearly about jobs and professions, because if we look at the role as you're either a doctor or you're not, well, then it seems like, okay, depending on which part of that AI automates, we may not see as many people in the field there. But there's also a world in which we have a doctor, we have a nurse, and then we have something over here called medical A. And medical A is somebody that goes to a health school, they're AI powered, and they do something maybe in between, and it's something different, and more people can reach for that job, because it's not the 12-year having to get into med school. And so it actually brings the average wage up. And it may be harder to become a physician, because you're also doing something different. So there's fewer people doing the higher earning occupation and more people that can do something in the middle. Absolutely. Technological change changes, you know, affects the system of professions. There's a book by, a sociology book by Abbott, where he talked about how over the course of the 20th century, we saw these battles for the definition of the boundaries between doctors and nurses, and psychiatrists and psychologists and social workers. And here, we're, because of AI, in many. Professions, we're about to have new battles over the boundaries of what makes a doctor and what makes a nurse, what makes a lawyer versus what makes a clerk, and, you know, across the board. Now, there's this. Can you say more about that battle? Like, could you give an example of why you think that there'll be soon a battle over territory of who does what and why? Well, so think about medicine. And just to reiterate, so AI is. Depending on who you talk to, AI may already be outperforming many physicians on diagnosis. Okay. I think everyone agrees on many, but people disagree on whether it's the 10th percentile, the 50th percentile, the 90th. Okay. As AI gets better at diagnosis and conditional on diagnosis can figure out and recommend a treatment for most diagnoses, then so much of what the physician is trained for, memorizing information and being able to summarize it and getting a quick sense of how they can help a patient, might be done by machine. That is going to mean other medical professionals, like nurses and pharmacists, or the soon-to-be-named medical professional we just talked about, are going to start encroaching on doctor's territory, physician's territory. Now, one thing that physicians could do is, say, is fight back and say, no, you can't diagnose those things. By law, we physicians have a monopoly on most diagnosis, and that'll continue. And there will be some regulatory battles and how that all plays out. There's small versions of that in pharmacists diagnosing certain simple things. But with an AI, maybe you don't have to limit them to the simple things. But the physicians may also then, in turn, go into other things and focus on the exceptions, or focus much more on how to help patients navigate the stress of their health. In which case, then this physician might be encroaching on another profession, like the social worker. And that would be a different skill set. So you can imagine that, and I'm using the word imagine on purpose, because it hasn't happened yet, that as a technology gets better at one part of an occupation, it's going to be a different skill set. And that's going to be a different occupation. That occupation is going to adapt. And the other occupations that could take advantage of it will also. And we're going to end up with a different rule set and a different skill set. So what do you say to people then who look at AI and look at robotics and say, well, it's only a matter of time before AI does all of this. So there's, sure, for the next little bit, AI makes mistakes, we need humans in the loop. But after five, six, seven years from now, and then we can bring in robots into that mix, there's not going to be anything left for anyone. AI will just, if it can be learned, AI can do it. Or are they not understanding how technology and economies reshape? Yeah. So I think that's unlikely, but it's not necessarily wrong. Okay. And why do I say unlikely? So the first version of this is we've seen technological change before, maybe not quite at the same point, but electrification happened fast. Steam power went fast. The productivity transformation in agriculture was quick. And what tends to happen is people find other things to do. Okay. So, and the reason, there's an economic model behind the reason, economic understanding behind that reason, which I don't think is going away with AI. Okay. So there's historical data, but also this time doesn't seem different to me. And what that is, is as long as there is something that you need the human for, where the human is more productive than the machine, then more and more of the human work will go into that remaining thing. So I should actually caveat all this under, I'm assuming a world where the machines are still tools and they haven't taken over the world in some meaningful way. So it's, we have a bunch of tools and the worry is that these tools are going to replace human work as opposed to new consciousness that, uh, tries to take over. Sure. Okay. So as long as we're in that world, non-conscious things. Okay. So as long as we're in a world of these being tools, as tools get better, uh, as long as there's something that humans do, this is called Bommel's cost disease, which is as you get more efficient in some part of the economy, it's the part of the economy that's inefficient actually becomes much more important. And so what happened with agriculture is agriculture got more efficient. It moved from being what most people did to what almost nobody does. Like there's still, yes, there's still farmers, but farmers are, you know, well under 10% of the workforce. They used to be much closer to 50 or more. It's because as agriculture got more efficient, uh, people moved into the professions that actually didn't have the productivity improvement and still had demand like manufacturing. And then manufacturing started to get more efficient. And as manufacturing got more efficient, people started to move into services, which didn't have that increase in efficiency. And then in the last 20 to 30 years, as services have gotten more efficient, as many services got more efficient, jobs have moved into the services that had less productivity improvement, or at least measurable like healthcare and education. Now, so if we make those more efficient, as long as there's some parts of the economy or more precisely, some things that people want where, uh, the productivity of, of the human worker, um, isn't super powered by the AI, then that's what more and more of us will do. Now that sounds optimistic. Okay. As in it's not going to take all the jobs. And I think that's right. There's two caveats to that. Caveat number one is that's a long run story. I'm very confident about the long run. Okay. But the short run can be incredibly disruptive as in, if there are certain professions that suddenly we realize, um, uh, AI can do, and importantly, we figure out how to make that happen. Like it's not just that AI can do, but we figure out what that new system looks like. Um, then there could be meaningful large scale job loss in certain areas that will take time to recover from. So the long run is optimistic there, but the short run. So that's caveat number one. Um, caveat number two is that I said, as long as there's things that humans can do, then that will be what jobs are. Okay. Um, I can't, I have no idea of those things. I should say, I'm pretty confident there'll be something. I don't know if those will be things that, um, people like Andrew Yang think are meaningful jobs or good jobs. And, uh, to, you know, it might be that they're wonderful. They're about the caring professions and taking care of other humans, uh, or sort of, uh, some kind of intellectual pursuits or contests like sports. And, and, and, and, and, and, and, and, and maybe that's wonderful. Or it might be that there is just something that the robot can't do that involves picking something up and putting something down, um, that you need the human for, in which case we could end up being something that seems a little more dystopian. Okay. So it seems like what you're saying is we have to, when we think about a technological wave, we have to also ask what becomes scarce in that. So with the agriculture example, and again, with manufacturing, it wasn't that these sectors disappeared, they became much smaller share of the economy. And so in a world where AI can do what we would consider cognitive labor today, isn't it that humans themselves technically become scarce or jobs where humans are the value proposition becomes something that's scarce. So you had mentioned the care economy or nursing or teaching, um, and those could be some of the occupations. And it, of course, it seems like, well, not every job is going to be scarce. And so I think that's everybody would want to go into that. And of course, amazing professions and, and my mom's a nurse and, but maybe it's not for everyone. But isn't it also true that we can't predict? I mean, that's not the full menu of what we end up doing. Those are the care-focused jobs that we can see today. But isn't it also true that strange occupations and we find weird things to do that we call valuable? I mean, you were also a former professor of mine of marketing. And when you think about the origins of marketing, it came off the back of the second industrial revolution because we had way too many goods now that we could produce because of industrial capitalism. And so we had to – local demand couldn't absorb it. And then we had the railroads. Now your market may be somewhere else. So we had to figure out ways. How do we coordinate between finance and production? And maybe we shouldn't just make goods and try to get salespeople to sell them. We should think about things like consumer demand and consumer needs. So that entire part of the knowledge economy we call marketing, and we know so many people in it, came from a technology. So isn't it possible that AI leads to strange things that at the time were unforeseeable? And I don't know. I don't even mean that as an optimistic kind of view. I don't consider myself an optimist. I'm not a pessimist. I'm just – I don't just subscribe to the title and then hope for the best. But that has technically been what we've done. And we invented this – the podcast, the whole thing that's now so important that cable news is going to be the beginning of the end of that. So it's possible we do strange things that are just as strange as marketing was. And now it becomes something that we think about. But that may also not happen. So I think we should expect exactly what you just described. Which is? Which is that we should expect that there's going to be new jobs, new professions, new industries that we can't even imagine. And the problem is it's really hard to – it's much easier to figure out what's going than what's coming. And so Ajay Agrawal, Joshua Ganz, and I have this paper that came out in 2019 where we were asked about the labor market impact of AI. And the dominant model at the time said, which jobs are going to come from the arrival of new tasks? So you worry about substitution. Okay, you can identify which jobs are at risk from AI. And then you can anticipate the new tasks. And so our original assignment as we were thinking about that paper was to first identify where substitution was possible and people's jobs would be at risk, which we've started to talk about. And then to say, where were these new tasks? And what was both enlightening and frustrating about writing that paper is we realized we don't know where the new tasks are. We can't anticipate them. They're going to happen, we think, but that's essentially – that's a trust in the model and faith in human ingenuity as opposed to empirically we can say this job is meaningfully at risk. So where would I expect to see things? A lot of the places we just talked about. So if AI continues to get much, much better at intellectual work, then maybe there's going to be certain caring professions. Whether the same ones we have today or different ones where there is just real value to the human consumers and interacting with other humans. There's going to be competitions, right? We still watch foot races even though cars can go faster. But it's interesting. We enjoy watching that. And so a world of competition and a world of the arts where there's something about the human creation that we enjoy consuming. There's likely a whole bunch of opportunities in that. Alex Imast, who's an economist at the University of Chicago, has talked about status-related work. So that if in a world of abundance, we humans are still going to want to compete with each other and we do compete for status. And if the AI is making everything abundant, then the remaining things will be how we get status. Can we do something that the AI can't do? Can we interact with other humans or the world in ways that the AI doesn't? And there's some things that are always going to be rare. Like coastline in California, even in a world of abundance, is still scarce. And so different versions of that are exactly how that plays out is unforeseeable, I think. But to anticipate that there's going to be something left for us, I think that's exactly what we should expect. And I do think on professions and jobs, the long run looks good. The transition. Yeah. Jobs aren't good. Okay. Getting paid is good. And finding meaning is good. But a job per se is something we do and it's part of our society, but it's not necessarily good. And to the extreme. In Econ 101, when we teach labor economics, we say people have a trade-off between work and -- or between leisure and consumption. So if you have more time off, then you get less income and you can consume less. And so that means that in the standard economics framework, jobs are bad. We'd rather get paid a lot and not have to work. And we can see that in the world, right? So. So the victories of the 20th century labor movement weren't that people worked more. They were that we got weekends and 40-hour weeks and we got to retire at 65 and all these sorts of things. It was more leisure time was the win. And at the other extreme, in the movie The Matrix, see The Matrix? That's a dystopian movie, as I understand it. Yeah. And every single human works from the day they're born to the day they die. They're batteries, but they work. And no one, to my knowledge, works from the day they're born to the day they die. And no one, to my knowledge, thinks that that kind of work is what we want. And so to the extent that there are folks who are thinking about what does the optimistic, exciting future with AI look like, the goal of jobs per se only makes sense if we think that's the best way to distribute the wealth that AI is going to create, or that's the best way for at least some of us to make a living. If we think about the worry that jobs are going away, in the short run, if somebody is working and their job goes away, or someone expects to be working and their job goes away, that is really hard. And that is something that is absolutely of first-order importance. In the longer run, if we get to retire at 50, or we get to only work a few days a week, there are other opportunities to spend our time that are even more meaningful than work, and we have enough wealth as a society and distribute it fairly that we can benefit from it. That sounds wonderful. So there's an economist at Michigan, Betsy Stevenson, who's talked about these ideas. I just want to give credit to her. So she's framed it as if AI gives us the abundance as claimed, so that it creates extraordinary productivity gains and extraordinary economic growth, so that it creates extraordinary productivity gains and extraordinary economic growth. Then the worry isn't jobs per se. It's whether we can find a fair distribution of income that our society accepts. And two, whether for those people who get meaning from their lives through jobs, can we find other ways for them to find meaning? And if the answer to those questions is yes, then the job issue per se isn't the right focus. Right. So there's jobs. There's work. There's income and purpose. Sometimes those things are bundled, and often they're actually not. And so if you were to ask someone, if you were to design the workforce from scratch, would you keep it this way? Would you fight for it to be this way? Or if you had an opportunity, if this was a design space, what would you do differently? And would you do what you're doing even if you didn't get paid? Is that the thing that you would want to be doing? For some people, the answer is still yes. And then for some people, it's no, I would rather not be contributing to somebody else's dream. I'd rather be pursuing my own. When you ask somebody, is work the most important thing in your life? They think, oh, no, and I never want to answer that. But then when you think, okay, we might take that part away, that also feels very violating and scary. So we know it's not the thing that really brings us meaning per se. We also don't want to lose it. So in a world where you get income and you can pursue different things, that's actually potentially on the table if we design it right. But then I guess there's also the fear of what the heck do we do with our time? Because when you look at what we've done with the free time, we have more free time today than we probably did 200 years ago. And we're scrolling, we are yelling in the comment section, we are buying nonsense, some nonsense, or being sold nonsense. But we are also living longer and better educated. Yes. The structure we all exist within right now, including our education system, came out of the Industrial Revolution. The factory model, the idea of the job became because of specialization. And in the factory, it wasn't some long lost quest that we pondered upon and thought, okay, let's all run to the factory, invent the firm, and then go work in it nine to five. I don't think that's how most people would have designed society. And it actually goes further than that. How would you design society if you could redo it? And we might be at that design moment. Yes, we might be. And what I'm good at as an economist is talking about the trade-offs. Okay. And if we move to fewer hours of work. then to what extent are we still motivating people to produce in a way that allows society to keep giving us more of what we want, whether it's healthcare or education or food or clothing or other things. If the AI's tools are providing extraordinary abundance, then that's wonderful. But there still needs to be some kind of incentive system to grab that abundance and distribute that abundance in a way that makes societies work. We're sitting at the University of Toronto. There's a lot of smart people here and around the world thinking deeply about what those new systems look like. And some of that involves actually thinking about policy changes, like does a universal income make sense? Should the retirement age be older or younger? Should we move to a three-day week versus a five-day week? Incremental, but with enough of them, it could be meaningful. But each of those has a handful of assumptions embedded in them that before you grab and say, this is what we want, you need to think really carefully about. As in, once that system comes into place, how will us humans, potentially superpowered by AI, be able to game it for, you know, some people's advantage and not others? Mm-hmm. I think, and what does agency even look like in that system? Absolutely. Because I think that that's the kind of the real question. It's people may be able to say, you know what, I like my job, or maybe I do, maybe I don't. I like having the agency to pursue something that I at least think maybe interests me. And then it can be rewarding to see growth in that field and to earn something from it. And yes, I don't know if it's human nature. I guess we'll have to talk about that. But I think it's a really important thing to talk to a biologist about that. But our tendency to game things, our tendency to seek, to be power-seeking, there probably will be winners and losers in any new system. Absolutely. And how, so would you say then that when you look at something like software engineering today, is software engineering the canary in the coal mine for the, what's about to happen to the knowledge economy as something that's going to become a smaller share? I mean, this is for the first time. I mean, this is for the first time in 15 years, after 15 years of strong growth in computer science, there's been a drop in enrollment. And we've seen, of course, we talked about the fewer, probably junior hires doing software engineering. So would this be a rational attempt by students to avoid the psychological and economic suffering of an AI-dominated field? Or are we, again, misreading what is happening here? I think software engineering is a super interesting area to talk about. So first, because there is a sense that there's fewer new job openings, at least fewer new job openings at the wage expectations of new graduates. My understanding is graduates of software engineering are still doing relatively well compared to other 22-year-olds, both in wages, certainly in wages conditional on getting a job. And then you've got to think through, because, you know, sort of wages adjust, and then you've got to think through, you know, is if employment, if unemployment is high, is that because wage expectations were sort of set earlier? So it's like, it's an interesting profession to think about, because they, in some objective ways, software engineers are still doing very well, even though I think there's evidence they're doing worse than they used to. Now, the second point is, if AI is enabling us to code at scale, what is, do we now have a system professions point about software engineering? Okay. So I suspect you don't remember having to code in HTML in the late 90s. No. Okay. So if you wanted to build a website in the late 90s, you had to do it yourself and code in a relatively easy to understand coding language, but it was still code. And so there were a number of people who developed that skill and were web developers, where web developer meant something like today's software engineer. You actually needed to know how to code. Now, you have a website. You may even be able to edit your website yourself. Okay. But you don't think of yourself as a software engineer. Not at all. Okay. And so what happened is, as no code web design came along, is a whole bunch of people who didn't used to be software engineers or web designers are now web designers. Some of them identify as web designers, who use no code and don't think of them as programmers, but they themselves as programmers, but they do think themselves as web designers. Some of them are more like you, who I don't think you'd put that in your. It's not in my bio. It's not in your bio at all, but it is a part of what you do. And so what I think is going to happen with software engineering, it's not that we're going to have less code. As I understand it, we're going to have more code and more of us, especially once tools are developed to help everyone take advantage of what. AI-based coding can do. Many more of us are really going to be coding, but we're not going to think of ourselves as software engineers. But the real consequence there was the whole system of what it meant to be a web developer changed. And I think we're in the first step where we can see the risk. And those of us who are going to benefit from being able to code haven't really experienced what that means yet. It seems like computer science, those who are going to be in the very top percentile are going to be kind of systems designers, socio-technical architects, software engineering, those who remain in it. It's actually going to become much more complex. If you think about what's coming at us, the robots that we're going to be syncing with the different applications people are building, connected to your driverless car, and maybe some AI wearable, and you create some sort of immersive movie that way. Picture what it would mean to be some form of a software engineer, whatever we call that, some system architect. In that world. So I think it's eventually software engineering, whatever we call it, will be even more complex. Those who remain in the top of it. And then there will be a cohort of people who kind of do the, I did for my company website, I hired somebody. And then for my personal one, I just did it. And that can be kind of the rest of software engineers. I think that's a really interesting vision. I have no idea if that vision is three years away, 10 years away, or 40 years away. And so I'd be hesitant to say to a recent grad or someone's starting college saying, okay, that's where things are going to land at the time you graduate. So you got to prepare for that now. Instead, maybe even to the extent that you have conviction on that is where the future is going at some point, then what's useful perhaps today is to find one piece of that. There's only so much you can learn in a year. So one piece of that, that you think you're excited about, and that you're relatively good at, and focus on learning that piece. And then next year, it'll be something else. And maybe one of the macro lessons of how we can anticipate the change with AI on all sorts of jobs is that whatever it means to be a software engineer in 2026, I'm confident that it will be different from 2031. And so what you want to be able to do is learn enough different pieces of it, and get exposed to enough different opportunities that as the world changes, you're capable of changing along with it. The adaptability piece as a skill, which is something that we all need to be doing. I was at an AI roundtable dinner a couple of weeks ago, and there was a woman there, she's in the CEO level or C-suite level at a financial services firm. And she was saying that their organization is designing AI systems that intentionally keep, humans in the loop in the workflow. And so the thesis for her was, we want to make sure employees stay engaged, and their skills don't atrophy. We also want to make sure we have a succession pipeline, so our juniors can become managers and then can become partners. So the lifespan of our firm isn't 20 years or something. So she's designing, they're designing AI systems where it checks in with the human, the human has to have this core part. Is that the right way for companies to be thinking about AI right now? Are they better off to intentionally only try to augment workflows? Or is there something that companies that are having or that are taking that lane are missing? If there's an opportunity to automate something that will make your company much more productive, it's hard for me to understand how that would not be a good idea. Okay. But that if at the beginning was very important, which is it turns out it's very hard to fully automate a job. There's a study by Dan Gross, looking at telephone operators. So in the 1920s and 1930s, perhaps the most common profession, certainly for young women in the United States was telephone operator. Okay. Today, there are pretty much none. But it took, the automated switch was invented in the 1890s. And it took until 1978 for, in the US, the last sort of traditional telephone operator to be out of work. It just took a long time to figure out how to automate that job fully. And so if companies do see an opportunity to automate, then that does meaningfully impact the bottom line. I don't see why they wouldn't. But there's a barrier because it's hard to automate. And there's another barrier, which is that even if you automate it, it has to be worth it for your bottom line. And a lot of what we talk about in our book, Power and Prediction, is this difference between point solutions and system solutions, right? So a point solution is you look at your existing workflow, you take out some human process, drop in the machine, and keep the workflow the same. And in that scenario, typically, the best you can do is what you're already doing, but a little bit better. And that might not be worth it for the disruption to your workers, the disruption to your pipeline, to just do a little bit better. But if you can reimagine your system so that you're becoming much more productive, serving your customers much better than you did before, reducing costs in some significant way that allows you to do things differently, then for a company to say, I just don't want to do that for these other reasons is hard for me to understand. Another sort of economic concept of what you just described, I think, is really important, which is it's wonderful that that company is training their employees, and the way they are. And the question is, are they training their employees to have what we call general human capital or specific human capital? So general human capital means that the things they train their employees for are useful at every other company. And specific human capital mean they're only useful at that company. If they're training their employees for things that are only useful at that company, then the pipeline you just described sounds great. If they're training their junior employees with skills that are going to be useful everywhere, then if I were the leadership of that company, I would worry that as we train employees after three, five years, and they're ready to thrive in an AI world, that all of their competitors are going to poach their employees. And so what I imagine they're doing is thinking through not just training people, but training people to thrive in their company. And that makes a lot of sense, which is to figure out for your new employees, as they get tooled up, how do you train them to thrive in your company, in that environment, to be good at what you do, so that they can be the next generation? And don't you have to first ask, is this how our company is going to continue to compete in the future? Because if you're assuming that your business model can stay the same, and you just need to have more efficient, faster. Employees. And so, yes, they're going to be AI-powered, and we've kept them in the workflow, but we're not actually changing anything else. We're just kind of trimming a few things here and there. If you haven't first done the business model work to make sure this is the best way to compete, couldn't you be calling into the question the longevity of your firm in its entirety anyways? Yes. So that's a great point. So the senior leadership of every company should be doing the work to understand what the competitive. What the regrets are from AI, what the opportunities are, and what does the future of their firm look like. And the training processes that they're going to develop in-house, if that's what they want to do, have to be consistent with that future. There's another possibility, which I hope is already happening, which is, yes, companies need to train their own employees, absolutely. But we are in higher ed. We have an important role to play in getting people ready for the workforce where AI is ubiquitous. One of the reasons why in the first years after ChatGPT's launch, employment for some young workers was lower than they might have anticipated, employment rate than we might have anticipated, is because in the middle of their education, ChatGPT arrived. And we in higher ed weren't ready for that. And one, it made it easier to write essays in ways that we couldn't detect. But two, our curriculum wasn't designed to teach people how to use this tool and use it well and take advantage of it. I think now, three and a half years later, we've made real progress, both here at University of Toronto, but also at other universities around the world, in what does it mean to train people to work in a world where they have access? Well, I think that's a really good question. I think that's a really good question. And because of that, I'm hopeful that the work you just described the company having to do, they're going to have to do something. But the students will graduate and on day one, they'll be able to contribute in a way that students from three years ago couldn't, because new students will be used to working with AI and will know how to use that effectively. So you could actually theoretically be doing a disservice to your employees, because the reality of the world is AI will be ubiquitous, and soon AI agents, everything will be ubiquitous. And so if you were trying to engineer a workflow with the assumption that AI isn't, and a human has, and of course, this is within finance and certain, we're assuming the workflow doesn't have something dire where a human should oversee it, something kinetic or something physical. So in the world of, let's say, finance, if you're designing a workflow and you're saying, okay, every 15 minutes, we've designed in a section where the human comes in and they do some review, it goes on, you're actually designing for a world that no one's going to actually be living in, because AI will be ubiquitous. And they'll either be a new grad that has learned these tools from scratch, or a company that is built in a different way that assumes you're streaming AI the way you stream electricity, and you're potentially doing a disservice to the people in that company. If it's simply you have your existing workflow, and you've just randomly picked a point in that workflow to make sure the human stays there, then absolutely. There's another possibility. I have no idea what this company's doing. So there's another possibility, which is they've actually thought very carefully about each of their workflows and said, we can redesign the entire workflow as long as there's a human spending a lot of time focused on this step. And if the human is focused on that step, the entire workflow will be more efficient. And in the process, we get this bonus that we train that person better than we could before. Now, the, the easier process might be to just sort of mandate, okay, just find somewhere to put the human in. And then I agree with you, that's worrisome. If instead, it's, let's think through strategically, within each of our divisions, each job within those divisions, and each task within those jobs, when, if ever, the human is needed. And then can we get people, to really understand that well, and that's fantastic. And that would be great for, I imagine, I'm using the word imagine on purpose again, great for the workers there, and great for the firm. One thing that we're seeing AI can do is it helps you draft emails. Okay. And another thing we're seeing AI can do is it helps you summarize emails. Okay. And I'm often in executive classes, and I ask the class, okay, what are you, what are you using AI for? And someone says, I'm drafting emails, and someone else in the same organization says, I'm using it to summarize. And then the question is, well, why are you even bothering with the AI at all? Or why are you even bothering with the email itself? It should just be, you know, somehow, the bullet point to bullet point should make things much more efficient. But then once you're doing that, maybe you can question whether you even need that communication at all. And when I push them, they say, well, I still want to read the bullet points before I send them. Because there's, you know, there's, you know, there's, you know, there's, you know, there's a there's a, you know, there's a, you know, there's a, you know, there's a verification of is that what I want, that still matters. And so for tasks where it does matter that the message is what the human wants, then that ends up being, I think, in many cases, this verification step, something that humans are, like, are going to be central to, even in a world of AGI. There's a paper by. A couple of former students of mine called The Simple Economics of AGI. So Christian Catalini, Jane Wu, and in their paper, they say, like, as the AI gets better and better, even if we're in a world of artificial general intelligence, that we still need the humans to verify that it's what they wanted. If ultimately, the role of humans is judgment to figure out what we want, then that verification is going to be important. And so I think, you know, there's a couple of former students of mine called The Simple Economics of AGI. So Christian Catalini, Jane Wu, and in their paper, they say, that the role of humans is judgment to figure out what we want, then that artificial layer becomes, like, in their model, almost all of what we do. And you can see that with, you know, in the story you just told within a workflow, you can see that play out. And so what do you say, because there are some, even some economists that say, but to prevent that, the threat of the job apocalypse, it may happen even in the short run. So I think we're both aligned that in the long run, things may actually be okay. The next 15 years, that's what I'm really concerned about. So we could avoid all that if we just only augment. roles instead of automate them. Okay. So honestly, I don't think that makes any sense. But I know very smart people disagree with me. And the reason I don't think it makes sense is going back to this idea that augmenting some people automates others. And we saw this with the rise of the personal computer and the word processor. What did that do? It augmented what the knowledge worker could do because they could type directly and delete their mistakes. So they didn't have to be great at typing. But there was a profession called typist. And that profession ended up becoming irrelevant, effectively, as word processing became part of what knowledge workers did. And so word processing was very clearly an augmenting technology. What it did is it made knowledge workers able to think as they write and erase, and they became much more productive. But that doesn't mean it didn't hurt people or create job loss, because the typists were no longer relevant. So I don't really. The narrative that we should focus on augmentation and not automation, honestly, unless that's a statement about augmenting everybody without automating anybody, which I don't. I haven't seen a story where that's possible. It seems to me that a constructive alternative is to augment everybody without automating anybody. And I think that's a good alternative, is to say, if you're choosing what to augment and who to augment, and what to automate and who to automate, then think carefully about whether you're automating the roles of the vulnerable, or instead potentially augmenting the roles of vulnerable relative to less vulnerable people. And I want to ask you about your recent paper, O-Ring Automation and the Logic of Task Concentration. Because every few months, a paper drops about jobs or sectors that are exposed to artificial intelligence. And then we hear all of these viral headlines about if your job or your sector is exposed to this technology, it's the beginning of the end for you. Your paper shows that actually something else could be happening. If a job is exposed to artificial intelligence, or if a sector is exposed to artificial intelligence, it's not only that it might not be automated at all, that could actually mean your job is going to increase in value, and you could be making more money. So can you set the record straight, on what these AI automation indices are getting wrong? The paper's called O-Ring Automation. And the reason it's called O-Ring Automation is because it's about how all the components potentially of a workflow, of a job, can be essential to the final product. And the metaphor comes from a Michael Kramer paper on O-Ring production, where he says that the space shuttle Challenger, the disaster in 1986, it's a very complicated machine, a space shuttle. And the failure turned out to be because of a little tiny washer-sized piece, a seal that was broken, that allowed fuel to leak. And so a one tiny piece was essential, and so the whole thing ended up failing. It's a catastrophic failure. To the extent that in a particular workflow, a particular job, there is one piece that requires a human, or actually, let's be more precise, to the extent that each piece is essential to the overall workflow. So you can't just get rid of one. As we start automating some of those tasks, what's going to happen to the human workers? They're going to start spending more and more time in the other tasks. What that means is they're going to get really good at them, which is then going to make it even harder to automate those remaining tasks. And so as long as there is one task within that workflow that the human is more productive than the machine, then those humans are super productive, they get high wages, and they remain in the organization. Now, so that's the essence of O-Ring automation, which is that as long as there is one thing that the humans can do, if they spend all of their time on it more productively than the machine, then there's a real opportunity, and there'll be higher wages, and they'll keep their jobs. So they'll actually make more. Absolutely. If your job is 100% exposed, that's a different world. But if your job is 75% exposed, then you're going to be spending your time on the remaining 25%. And that could mean you're going to get better at that. It means you're going to be more productive. And that could mean more people like you or less people like you. It could mean higher wages and lower wages, depending on another thing, which is back to elasticity of demand, which is as your firm gets more productive, are there people who actually. Want that many more of what your company builds. And if so, then that combined with an O-ring model means that the task exposure models just miss something fundamentally important. Now, I want to be a little cautious, which is that a lot of the published research papers on task exposure are very careful. There's this Elan Duetal paper that OpenAI put out along with Dan Rock in Science two years ago that was, at the time and probably still, the best of the task exposure papers. And the newspaper headlines around that paper are, these jobs are exposed, the headlines, the TikToks, whatever else, are disaster. But if you read that paper at least once on every page, they say, exposure does not mean substitution. It could mean complements. So the measurement, often the people doing the measurement are being careful. Right. Interpreting the measurement. Interpreting. It's just so tempting to say, ah, it's automation and humans are out of luck. And I've been dying to ask you about this AI paradox, because there's a growing argument that the way we are rolling out AI is economically self-destructive. So the more firms race to try to replace workers, as long as that's the thesis, they're also actually automating their own consumer base theoretically. So the more people AI replaces, there's going to be fewer people who can actually afford anything in the economy. The economy self-destructs, and this becomes a tragedy of the commons. As an economist, do you agree with this thesis? No. Or I should be cautious, as I always am as an economist, which is there are a very, very narrow set of assumptions that on a knife edge that could possibly happen. But- But for that to happen, essentially, as AI gets better and better and makes us more productive, it creates stuff that nobody wants. Okay. And that there's nothing left for us humans to do. So there's no O-ring. There's no Bommel's cost disease, as in there's nothing left for humans. And not only is there nothing left for humans, but the AI is producing nothing more useful than we already have. So another version of that is like human wants are satiated. Perhaps. Under that scenario, then it's possible. I don't think that's likely. I don't think it's likely because first, I don't think all jobs are going to be automated. Second, I think the productivity benefits of AI will mean we get more of what we want. And so there will be demand. And as consumers are demanding stuff, then the machines are going to produce for that. And if people are don't want stuff, then why would the machines produce? And so the circularity of that argument, why would we be automating if there's no demand for the thing downstream? And so sort of a necessary step for that automation to happen, where again, I don't think it could happen in the way those people describe it. Even if we accept that, there has to be a reason why the owners of the machines would want that to happen. Yeah. And then there's another layer, which is that presumably the automation of certain workers in companies creates profits. And those profits are distributed throughout society, not just to the billionaire owners, but a large fraction of the population has exposure to the stock market. For example, if you're a billionaire, you're going to have to have a lot of money to for example, everybody who has a pension plan. And so, or not just the stock market, but to investments. And so again, this extraordinary wealth is going to be distributed to people who can then spend it. And so there have to be a whole set of assumptions on like where that wealth goes to and what people continue to want that goes away. And the final thing is this transition that I wanted to come back to. I think that we both share, I mean, you share it from the perspective of an economist. I share it from more of the perspective of the belief in the papers that I read that we, it is unlikely that all jobs are suddenly going to be automated by this technology, we're more likely to see people become more productive or augmented. It's much harder to automate a full workflow than many of the AI leaders would like us to believe. And it will probably take longer than anybody is forecasting, especially people who are quickly jumping into the stock market and hoping for the best next week. But it is probably true that we're going to transition to a different type of economy. And maybe we're in economy A right now, and we'll get to economy B, and maybe that's 10, 15, 20 years. But it could be very shaky over that time period. How do you think this transition could go? And what do you think could happen? And how do we think about preparing? Yeah, tough set of questions. So one scenario that I think is most likely, but I don't have any empirical reason to say that, is that a lot of the transition happens much more slowly than people anticipate. So in the long run, it may even be bigger than people anticipate, but getting industry by industry to figure out what their business model looks like as AI takes off, what that new system looks like, could be slow. Electricity, it took 40 years. Computers, it took something like 30 to maybe even 40. Figuring out what the new business model is, in the AI world, will take time. They had to redesign the factory before they could take advantage of what electricity could offer. And even in things like digital payments, there were many retailers as of 2019 who were not taking credit cards. And so to say now suddenly something's happening, and those people who it took a pandemic for them to move to digital payments, are going to transition everything they do instantly because of AI. I find that hard to get my head around. So I think it's going to happen more slowly than many in Silicon Valley and elsewhere anticipate. But there's other scenarios that are more pessimistic or optimistic, depending how you think about it. One is that because this technology is an intelligence technology, we figure out stuff faster. And so the change does happen fast. And that could mean a bumpy couple years as people lose jobs. It could mean the opposite as we suddenly have cures to diseases we never imagined, as long as we can figure out what to do with ourselves when we have all this free time. And so I think it's going to free time. There's another more pessimistic version of this, which was that I once gave a talk to a group of executives and government people and said the optimistic scenario I just described or described earlier where AI so far is helping people in the middle and the bottom and hurting people in the 80th percentile of the income distribution. And I framed that as an optimistic because it's equalizing. And they said, I don't see why you think that's optimistic. Historically, revolutions come from people toward the top of the income distribution when their opportunities are lost. I don't know if that's historically true, but that was the argument. And that seems pretty compelling. So that's the, well, if the disruption happens and it happens to people who want to hold on to power, then things could go bad quickly. So that's- And I could keep telling you these scenarios. I could tell you optimistic, pessimistic. Empirically, what we saw with the internet, what we saw with computing, what we saw with electricity is in the long run, the impact was extraordinary and had it meaningfully helped society get more of what we want, while at the same time having a real negative impact on segments of that society that it will be important for us to think carefully about how to mitigate that impact. But the scenarios go from, we can tell stories that are super, super pessimistic or super, super optimistic. And so if you're, if you are a student in this moment, let's say you're in college right now, how should you be thinking? What are the concrete things that we could be doing? Are there skills that become quite non-negotiable that it's okay, wake up tomorrow and think about how you're going to build this scale? Yes. And I do this, this is part of my standard teaching now. So the first thing is whatever industry you're going to go into, there's still things you need to know about the world. You need to understand some facts. If you go into marketing and you don't know what CPM is, it doesn't matter that you can look it up on your AI tool. You're going to look like an idiot to the other people if you don't know the simple language of the industry. So you test that with pen and paper exams. Second, you need to know how to use AI. So at graduation, you should have experience with lots and lots of projects where you embrace everything AI can do and try to superpower yourself. So for example, I now teach a AI and marketing class. I used to have my students when it was called digital marketing, write a marketing plan. AI writes really good marketing plans. So I don't ask my students to do that anymore. But now I'm going to say write the marketing plan, create the ad copy, design the website, collect and analyze data, do so much more than any student project could have imagined over a few weeks or months before, and feel and understand what it means to work with an AI. So number one, know the basic facts. Two, be able to use AI well. Third, be able to work through really hard problems. So I got this idea. I give a talk at Columbia University on an AI and education conference. And one of the other panelists there was a math professor. And he said, in math, my undergrad classes, I'm not training future mathematicians. Like, it's wonderful when the hundred kids in a class, one of them becomes a mathematician, but most of them aren't going to become mathematicians. But I'm not going to be a mathematician. I'm going to be a mathematician. He believed it was still useful for them to be math majors, because in math, they teach people to solve really, really hard problems that require just a level of perseverance that is perhaps not true of many other subjects. But I think that skill is going to be even more important in a world where AI can do a lot of the easy stuff. And the perseverance? And perseverance and being able to figure out how to solve hard problems. So trying this, trying that, trying that. Figuring out that this didn't work, that wasn't good enough over and over again until you figure out something that you're excited about. And then the last piece is harder to teach in university, but it's knowing what matters, having judgment. And having the ability to recognize an opportunity for your organization, having the ability to understand what you want and what you can accomplish, you know, a key part of what anybody can do. And how are you going to learn that? You're just going to try lots and lots of different experiences and different things and figure out which ones you enjoy and which ones turned out to be valuable for the people you're interacting with. The learn by doing. Learning by doing. And you just have to immerse yourself in it. The many different experiences you can. And there's an example that everything that you encompassed. I was talking to a father who his son just went into private equity, and he's the junior in the firm. And there really isn't that much of a difference between the two. And he's like, I'm going to do this. I'm going to The things they were having him do were kind of antiquated. And he's like, OK, my job is probably going to be done by an AI really shortly. But he knew the vernacular of the field. He had learned enough in university that he went up to the partners and said, can I build you a bunch of agents to do what I'm basically what you've hired me to do? And in building that agentic workflow that took his own job, he is now doing things for the partners and helping them run different agentic simulations. And all because he's a part of the firm. He knew how to frame the problem. OK, this is what they're trying to achieve. I can technically automate my workflow, but I know the goal of what my team is trying to build. I can actually do that with AI. And now this is what he's doing for the summer. That's a great example of all four, right? So he understands. He knew the basic facts about the industry. He was able to use AI. That's almost surely required him to think deeply and solve hard problems. And he had the judgment to know that this mattered. So that's a fantastic example. I love it. And what would your advice be to policymakers for this transition? As we recognize all the legitimate things to worry about as AI gets better and better, to also understand it's our best chance at productivity growth, which means it's our best chance for us to get more of what we want. Our economies grow through productivity. And a lot of times, that means we get more of whatever we want, whether it's leisure or health care or education or other things. And that happens through productivity growth. And when you look at the suite of technologies available right now that you might think are going to lead to productivity growth, AI is by far the most likely to be a general purpose technology, which means it's by far the most likely to have a meaningful impact for the better on how we live and work. And so to protect people, so in the event that it might get bumpy, because there could be scenario A, scenario B, scenario C, a black swan for sure. How do they think about distribution, protecting people, training? What is the plan? How should they think about building a plan in this? So different governments are going to have different values, but regardless of the values, they need to see some trade-offs. When you ban something, it means that that thing will not be useful for what you want. When you price it, as in you put a tax on it, it means people will use it less. When you subsidize it, it means people will use it more. And so to the extent you said two policymakers, I don't know the politics of these policymakers, and they're going to have different preferences. But regardless of their preferences, they need to know that their decisions to to slow down or speed up, end up creating other trade-offs throughout the economy. And as long as they go in with that eyes wide open, then I feel like I've done my job as an economist. Avi, thank you so much. It has been a pleasure. Look forward to having you back.