AI and Economics: Twins Separated at Birth? with Kevin Leyton-Brown (Live from Upper Bound 2026)

Speaker A: AI is not rocket science because instead of making rockets, we make absolutely everything. We make these incredibly general purpose technologies that touch agriculture and mining, rockets and rockets. It's not only rocket science.

Speaker B: Yeah, Right?

Speaker C: Yeah.

Speaker B: Ready?

Speaker D: Yeah.

Speaker B: Okay. Hey, welcome back to Approximately Correct. I'm Alana Fish.

Speaker C: And I'm Scott Lilwall.

Speaker B: And this session we have Kevin Layton Brown live with us at Upper Bound.

Speaker C: Yeah, really exciting. So we only have a limited amount of time and Kevin has so many titles, so I only go with a few. But he is a Canada Sefar AI Chair with Amy. He's a professor at the University of British Columbia. He's an associate member at the Vancouver School of Economics and a member a fellow of the Royal Society. He works on really interesting stuff. A lot about economics, multi agent systems and everything I read about what you do is really, really cool. So we had to narrow down what we wanted to talk to you about.

Speaker D: Well, thanks so much for having me. It's very, very cool to be here, uh, in front of all of these friendly people wearing blue headphones.

Speaker C: Yes. So I guess to start us off, um, you know, it's been said that a lot of the work that you do is at that intersection of AI and economics. And I think that's something that a lot of people don't initially think of right away when they think of AI. So like, can you give us a run through like first all for someone like me, what is microeconomics and how does it relate to AI?

Speaker D: That's a great launching off point.

Speaker A: I'm really happy you asked the question

Speaker D: that way because when I was a grad student and went to start my PhD program, I knew absolutely nothing about economics. I thought economics was basically being an accountant. I thought you worked in Excel or something. Sorry, economists. I've learned the error of my ways now. We just lost our entire much credibility with them to begin with.

Speaker A: So it's fine.

Speaker D: But uh, but really once you learn even a little bit about economics as an AI researcher, you realize that economics and AI are sort of two twins, separated at birth with completely different upbringings, but exactly the same genetic code.

Speaker C: Oh, wow.

Speaker D: Uh, in fact, they both were heavily influenced by John von Neumann, one of the progenitors of both disciplines. And they have exactly the same decision theoretic foundations. So they both rest on this question about decision making based on utility functions in the face of uncertainty, which is usually modeled in a Bayesian way. This should all sound pretty familiar to anybody who thinks about, uh, decision making, planning, reasoning under uncertainty and artificial intelligence. And not just the ideas, but even the language and the mathematical formalisms are exactly the same in economics. And yet they hardly talk to computer scientists at all. I mean basically when I was doing my PhD, the fields were just starting to talk to each other for almost the first time really. And the crucial difference between economics and computer science is economics is interested in the question of what happens when many of these agents interact with each other and what we should say about what will happen when everybody's kind of jointly optimizing against each other. And they're also interested in the question of how you should design systems that, that would do sensible things in the presence of that kind of mutual optimization. Whereas AI historically, maybe until just about now, has been thinking about just building one agent. We didn't have any agents at all. And so we were thinking about just how can I make any one thing that will act sensibly in this kind of economic way? So we were kind of this bottom up thing. They presumed the existence of agents that could do all of the stuff we were trying to build. And they thought about what happens when they come together. And so there's a really fruitful basis for interaction there.

Speaker B: That's a very good reason, ah, for the two fields to come back together again. It sounds like you've been working separately and you've come back.

Speaker D: Yeah. So it was the dot com bubble

Speaker A: that brought um, economists together with computer

Speaker D: scientists the first time. Because it used to be the distributed systems in computer science were like large networks on a corporate campus or something. And if somebody was misbehaving on the network, you'd walk three doors down and knock on their door and say, hey, kick it out. You're flooding the network or something. But when it's a troll farm in Russia or something, you can't do the same thing. And so you start to have to think about how to incentivize good behavior, how to make systems robust to misbehavior. And so just uh, the viewpoint of everybody's inherently cooperating went out the window once the Internet became a real thing.

Speaker B: So what is the micro of microeconomics?

Speaker D: So the micro means that you're um, studying um, at the level of individual agents. So you're thinking sort of your unit of analysis is a single actor in the economy. So if I want to think about how do people react to changing prices,

Speaker A: or if I want to sell a

Speaker D: Picasso to one of five people, what's the best way to sell it to the person who likes it the most, or to make the most money or

Speaker A: to Be as fair as possible.

Speaker D: These kinds of questions. If I want to think about something like if I raise interest rates, what's that going to do to unemployment there? My unit of analysis is not individual people. So that's macro.

Speaker A: I gotcha.

Speaker C: Okay. You talked about, well like why the two fields work well together. But why was it for you that you were kind of drawn to this combination of computer science and economics?

Speaker D: I think I'm really interested in social outcomes. I'm interested in trying to make good things happen with groups of people. And I think uh, another misunderstanding about uh, economics is that it's about money.

Speaker C: Yeah, I think that's the first thing people think of.

Speaker B: Yeah.

Speaker D: And they sure talk about money a lot. So you can see why people would think that. But ah, at their core economists are not so much interested in money, they're really interested in utility. The core concept that I think economists care the most about is welfare, which is like the sum of everybody's utility

Speaker A: for a certain outcome.

Speaker D: And the thing that makes economists sort of professionally the saddest is when welfare is left on the table when an outcome happens that collectively people like less well than some other outcome that could have happened instead.

Speaker A: And so a lot of so called

Speaker D: welfare economics is about trying to find ways of making collectively better things happen by eliminating barriers or eliminating kind of bad patterns that we're in that are preventing sort of collectively better outcomes. And I think there's something very appealing about uh, that sort of philosophy to me that there are not everything in life is a zero sum game. There are ways of making things collectively better by being smart about design.

Speaker C: And I mean that's something that I've kind of really respected. Kind of looking through your work and stuff is you do seem to have this really strong belief in kind of the social good or the public good. However you put like that this kind of technology and these approaches should be used for that.

Speaker D: Yeah, I think I have a real

Speaker A: desire to see that happen.

Speaker D: I think I also have humility about the ease with which that can happen. I think often um, the social good

Speaker A: is easier to talk about than it

Speaker D: is to precisely define. And I think a lot of harm and misjudgment, um, can happen by aiming narrowly at something that we think is the social good that doesn't take all um, considerations into account. Um, so I'm kind of cautious about it. But at the same time if we in academia don't use the freedom that we have to think about problems that matter rather than problems that are the bottom line for an enormous company, then I don't know who else in our field will. So I think it's something that we have the luxury of being able to do, and I think we ought to do.

Speaker B: And so AI has had this insane inflection point and has taken off in crazy amazing ways in the last five years. What has the knock on effect been for microeconomics? How have things been changing?

Speaker D: The short answer is that, uh, microeconomics is a field that is strikingly resistant to change.

Speaker C: Why is that?

Speaker D: It's a funny thing. I mean, I talk about it with

Speaker A: my friends in economics a lot. Uh, uh, part of it is just culture. Like it really has, it has a vastly different culture from computer science. They publish maybe one paper every few years in incredibly high quality journals after years of revisions. And so they craft these beautiful objects of scholarship that have every problem fixed in them. And we craft these things like 36 hours before the deadline, if you're lucky.

Speaker D: But as a result, the pace is

Speaker A: enormous and we just were constantly moving. And so economics is very careful about gatekeeping what counts as quality scholarship in their discipline. And they're not just ready to jump on a bandwagon. I think that's sort of the cultural reason, uh, the technical reason is that it's really not that obvious how you would use machine learning to solve the problems that economics cares about. So superficially, you might think economics must be about predicting things. You know, economists must care about what's going to happen in the economy. This sounds like a pretty good match for machine learning. Why are we not using it absolutely everywhere? The core reason is that you don't see IID samples from one distribution for a couple of different ways.

Speaker D: So machine learning really rests on the

Speaker A: IID assumption very fundamentally, as I think everyone knows, if you're building a face recognition system, people's faces don't change in response to your system. But if you're building an economic system, people's behavior in that system absolutely does change in response to the system. Anytime you as, uh, an economist or a policymaker intervene in the system, you change the behavior that you elicit from the participants in that system. If you don't want to be just constantly chasing your tail, you want to anticipate the kinds of ways people would change counterfactually to all of the different things you might do and then find the thing to do that is going to have the response to it that you most like. And that's just a super different problem from the kinds of things that AI knows how to do.

Speaker B: But to me that sounds like reinforcement learning.

Speaker A: It's a bit like reinforcement learning. Um, but reinforcement learning mostly thinks about finding one stable outcome. It doesn't think about sort of every counterfactual that you could have for every way that the system could change. I mean also, I guess as you know, multi agent reinforcement learning is a, uh, much less well developed and maybe even less conceptually well founded field than single agent. The objective function you're optimizing isn't even that clear.

Speaker B: So you don't think of yourself as working in reinforcement learning.

Speaker A: I think reinforcement learning has some great tools and particularly when I'm in Alberta, I'll claim it if you guys will have me.

Speaker B: Yeah, you can come into our secret clubhouse if you say reinforcement learning.

Speaker E: That's the only reason I came.

Speaker A: I was just listening for those words.

Speaker C: Well, you do have to stay for the whole interview even though you got

Speaker A: whatever I need to do.

Speaker C: I mean on that kind of point about the multi agent, uh, system, because I know that's something that you've worked on a lot. I mean, I think you wrote one of the essential textbooks for it. Um, that's also seen a real rise in recent times. So how has that changed the kind of work that you're doing in kind of the economic area?

Speaker A: Yeah, it's funny to live in an age when multi agent systems is actually a thing that people claim to be building that has actual stock market value. That has not been true throughout my career. Um, I'm sort of ambivalent about the extent to which multi agent agentic programming or something amounts to an agent. M. I wonder how you feel about this. You might have similar reservations. Like five or ten years ago when we thought about building agents, we thought that specifying the utility functions was going to be the easy part and doing the planning and world modeling stuff was going to be the hard part. And now we have systems that are really good at world modeling and planning, but their preferences and intentions are specified almost implicitly. You never really know what they want and whether they want the same thing even from one moment to the next as the context shifts. And the way the tools we use to think about multi agent systems really are built on these foundations that there are going to be preferences that you can sort of point to and examine. Um, and I think the reason I mention those preferences is that in my mind there's an important difference between a multi aged system and a distributed system. So uh, in some sense you can call it whatever you want, but I think in a sense a multi Agent system is a bad thing to have. It's nicer to have a distributed system that does one thing but just has many nodes. It's easier to engineer than a system in which you actually have a bunch of independent agents that are going to do whatever they do and they need to coordinate in some way. You wouldn't build something that way unless something made you.

Speaker B: But it's sort of inescapable. I mean, that's where we're going, right?

Speaker A: Yeah, it does seem that way. But if these are all instances of the same model, if they were all fine tuned using RLHF or something by the same company, they're coordinated in a funny way. And so the degree to which they each have agency is different than a negotiation between your agent and your counterparty's agent when you're trying to buy a house or something. Those are also agents, but the degree of agency is a lot greater.

Speaker D: As we get a little bit more

Speaker A: familiar with these things, I think we're

Speaker D: probably going to wrestle with the degree of agency.

Speaker A: To what extent are these things subroutines and to what extent, uh, do they have their own preferences and desires and viewpoints or information sets? Um, and I think we're probably also going to make preferences more explicit at some point because it's hard to trust something that you don't understand the motivations of. Uh, I think it's a pretty good way to increase trust to be able to be a bit more explicit about what goals and desires ought to be.

Speaker B: Let's go too far off the whales. But do you think people know what their goals and desires? I mean, when we get, I don't know, it depends on, I suppose, the goal.

Speaker A: Well, so there's a branch of economics called agency problems, um, which talk about, uh, uh, having a principal and an agent. So the principal is like the boss and the agent is like the employee of the boss. And you try to think about ways to motivate the employee to actually do the thing that you want them to do, given that you can't watch them while they're doing it. And so they might kind of diverge

Speaker E: from what you want.

Speaker A: And thinking about desires has proven pretty important in that literature. If you want to find a way to motivate somebody that you can't fully oversee, then knowing something about what they want and being able to shape their wants through rewards, um, is a way to achieve that. So I think that's going to be

Speaker E: true for AI as well.

Speaker B: And I think like as the AI agents, I've seen this amongst my team. Like they learn how to describe the problems that they want the agent to solve after seeing how it has messed up previously and sort of like a back and forth. I think one of the things you're really interested in though is moving these kind of esoteric ideas into the real world. So do you want to talk a little bit about how you've sort of worked towards that?

Speaker A: I guess I think about that in a couple of different levels lately. I'm thinking a lot about a problem that Canada has of. I think Canada, this is not a problem that I'm unique in having recognized. I think this is a problem the Canadian government squarely understands and is putting a lot of attention into and has for many years. Um, but Canada is absolutely amazing at AI research. Depending on how you count. We're one of, we punch enormously above our weight. And even in raw terms, uh, we're a leader in academic research and AI. Canada, however, is not amazing at the adoption of AI by industry, kind of industry impacts from AI which, you know, if you're a resource centered economy trying to move into the 21st century, it seems like a problem. Canada's level of productivity growth is, is astoundingly bad. Over the last decade and a half, our rate of productivity growth, which is like the amount of economic value produced per hour of human labor on average has been 15 times less than the U.S. we have a relatively similar economy. So that's awful, right?

Speaker C: What's going to go on?

Speaker A: It's mostly technology adoption. So mostly the way you increase productivity growth is people invent things and businesses use them to allow their existing employees to do more stuff in the same amount of time. And Canada has almost flatlined for the last decade and a half. We're producing about as much stuff per unit time, as much economic value as we were 15 years ago. And AI right now seems like a pretty big part of the answer that you would guess at if you wanted to imagine how Canada was going to change that.

Speaker B: Do you have ideas of how we. Like, what do you think we're doing wrong? What is the hesitancy?

Speaker A: I mean, I think there are a

Speaker D: lot of pieces to it.

Speaker A: Uh, you know, I think in part we're a small economy. We're in the shadow of a much bigger economy. We've always had this branch plant problem that, you know, big companies, um, that have offices in Canada don't tend to have their main office in Canada. So let's just say DeepMind, you know, sets up an office in your city you know, at some point they might decide when they want to cut costs, they're going to get rid of that office. You know, they're going to do their most fundamental work somewhere else. So Canada always kind of suffers from this dynamic, which is not wholly our fault. It's hard to see how you would change that. Um, but in part, uh, I think the problem that's really interesting to me and maybe to you guys because you have a similar vantage point to mine as representatives of a university kind of research system, how do you leverage this enormous research strength to be an advantage in industry? How do we not make that be more of a boost for ourselves? And I was part of a group of people, uh, at UBC who wrote a policy piece for CD how last year, um, that we called AI is not rocket science. And um, here's the observation. I wonder what you guys think of it. Um, the tech transfer kind of incentive system that Canada has sort of exemplified by like NSERC alliance grants or my tax grants. So how does this granting system work?

Speaker D: Basically you say we've got a lot

Speaker A: of assistant professors who want to have a good career and industry is not that interested in going and talking to universities. Evidently, like the revealed preference is not enough of that is happening. And so let's instead incentivize academics to go talk to industry. Let's tell them that if you want to fund more than two grad students and get tenure, you're going to have to go knock on doors and find some industry people to talk to and try to find ways of building collaborations and ultimately getting your stuff out into the world. And so we've been doing this for decades in Canada and in particular the

Speaker B: decade that you just said there was,

Speaker A: we had been flatlined also throughout that decade. Yeah. Uh, and I mean it's again, in economics you always want to think about evaluating the counterfactual. So like it might have been, you know, yet worse if we hadn't been doing this. Right. This might be great.

Speaker B: Right.

Speaker A: But, but it's not. It's still leaving something on the table, evidently, because it's not looking as good as it might. And you know, what we realized is, let's say you're a rocket scientist. Um, it would be nice.

Speaker B: I like that. Counterfactual. Yeah, exactly.

Speaker A: So if you're a rocket scientist and you want to go get a, you know, alliance grant, how many companies do you have to talk to? There's only small, like some single digit number of companies, you know, probably in all of Canada that do like aerospace research that would interact with yours. So over time you're going to know these people, you're going to have joint seminars with them, you're going to pass students who went to those companies. So those relationships are going to be strong and stable and few. And so maybe this alliance mechanism makes a different kind of sense for rocket science, for AI. AI is not rocket science because instead of making rockets, we make absolutely everything. We make these incredibly general purpose technologies that touch agriculture and mining and financial services and rockets. And rockets, space science. Uh, it's not only rocket science. And so this means that assistant professor who's trying to figure out like, I've got some new idea about lifelong learning for vision or something, which company in Edmonton cares about my thing. It could be anybody. And so this mechanism that, uh, maybe was working for rocket science is just so ill suited to us. And so that's what we've really been trying to think about. Uh, given the incredible dual use nature of AI, how can you scaffold this thing to make it easier? Because I think just stressing out assistant professors is not enough.

Speaker B: So it sounds like you're saying that the professors, I don't know what, the system should be different, that businesses should be coming to professors.

Speaker A: Yeah, I mean, I think this is the economics perspective. I think the system should be different. You know, we should structure the incentives for everybody in a different way to get the outcomes we want. Like I think we've got, you know, to come back to what we were talking about before. I think we've got a welfare loss. We've got an assistant professor who would be willing to do applied research. We've got a company that would love to, to work with them and they're not finding each other. To an economist, that's a market failure. And we would like to link them up together because they'd both be happier if they could find each other and the government would be happier too. And so I think part of the problem is matchmaking. I think we should invest in, um, actually people with PhDs who deeply understand the research that's happening at universities to actually do this linking up. I think we should have things like your podcast to help get the word out. Good for you. Um, I think we should maybe think about grants that involve not just kind of AI experts in industry, but maybe bring a third person into those grants who's an academic expert in the discipline. So if I'm going to work with a forestry company, maybe we should explicitly require that a forestry faculty member be part of that Grant and maybe even be the point person on the grant because they're going to understand the industry. Forestry is rocket science. Right. They have a limited number of forests. Um, maybe also we should think about incentivizing, um, low bandwidth conversations. So we also have an idea around, um, doing free consulting for industry where you would sort of have office hours. Industry could come to your office hours and you could recommend off the shelf AI technologies that would help them and you would get paid in student scholarships. So we would keep track of your hours.

Speaker B: And like, what you're describing is kind of what Amy does, I feel like.

Speaker A: So maybe Amy's just the solution to

Speaker B: all these problems anyway.

Speaker A: But we don't have one, um, in, uh, Vancouver. So at least for us, uh, we have a lot of assistant professors who are sitting on their hands.

Speaker B: Right? Yeah.

Speaker A: And they would like to be working with industry.

Speaker B: Yeah, we should, I mean.

Speaker A: Well, so tell me what Amy's doing that is in this vein.

Speaker B: I mean, we have like, there are professors who work on staff at AMY and talk to industry about how they could be. I mean, it's not exactly what you're

Speaker A: describing, but I mean, do you feel like the matchmaking problem is being solved?

Speaker B: There's probably stuff left on the table, as you said previously, but I do want to focus on you and not. We could pivot to Amy, but it's.

Speaker A: I feel like we should plug Amy for a minute. We all feel pretty happy about it.

Speaker B: Let's pat ourselves on the back. But I mean, you are really interested in getting out there and doing this matchmaking in a different way though, right?

Speaker A: Yeah, I think I'm interested in thinking about. I'm not interested in doing it.

Speaker D: It sounds like a lot of work,

Speaker A: but I'm interested in thinking about the mechanism that might do it because, uh, I think that's where thinking computationally and thinking about incentives kind of comes together and maybe informs that perspective. And I think it's something our country really needs. So, you know, right now there's a lot of national conversation about AI policy and it seems clear that something different needs to happen than what we've been

Speaker D: doing in the last little while.

Speaker A: And so I think this is an opportunity to think about new kinds of design.

Speaker C: So I mean, uh, it's obviously something you're really passionate about. Do you think the desire is there in, uh, the larger research community as well? Have you heard from others?

Speaker D: I mean, I guess you guys can speak to that at least as well as I can, but my sense is, at least in my university There are a lot of professors who would happily take on another grant to work with industry if they could find something that was sufficiently aligned with the kind of work that they do. So I think we have residual capacity in the system. I think we have people with. Another thing about computer science generally is it's pretty excited about applications. We feel like real data, real applications keep us honest, they let us prove out our ideas. So I think even people on the pretty theoretical end of the spectrum are often happy to plug in one new applications project if it's sufficiently aligned with the thing they do. And I don't think we're harnessing that capacity in the ways that we might. And I don't mean to suggest that it's easy. Industry doesn't necessarily want your regret. Minimizing offline learning algorithm Finding the way to connect those things together can be tricky.

Speaker C: It can be hard to. Because it's not always speaking the same language a lot of the time.

Speaker B: Yeah.

Speaker D: And it's not always working on the same time horizon and it's not even always solving the same kind of problem.

Speaker E: Problem.

Speaker D: But I still think we could be doing a lot better than we are now and I think that's a problem worth attacking. And I think. I'm not sure that we know exactly what the solutions should be but I think the solutions look like new incentives, like new mechanisms doing things differently. Not just saying encouraging things and holding some m. Networking events. I think we really have to structure things differently.

Speaker B: I think what I was trying to lead you to was the kinds of projects you've been doing in uh, like for farmers in Uganda.

Speaker E: Yeah.

Speaker D: At the beginning said I feel like I could go two ways and I went the Canadian way because we're in Canada.

Speaker C: But I mean I'm glad you did. It was really interesting.

Speaker D: But yeah, that was sort of the other thing that I was thinking to say. I mean I've also been really interested in ways that uh, economic reasoning can help underserved communities where I guess by underserved I just mean problems that aren't getting a lot of attention in the academic literature that sort of a Google or OpenAI or something isn't going to see as a revenue source for them. There are a lot of papers written every year on how to better optimize um, click based advertising. Um, because there's a, we've all seen

Speaker A: these at conferences and you can see

Speaker D: why that's on the critical path for companies. But there's a lot fewer papers written about food insecurity in sub Saharan Africa. Even though it's an enormous problem. And I think many of the same kind of core technologies of market design and machine learning and reinforcement learning can be, uh, employed to move the needle on some of these problems. So I've been excited to think about ways of doing that, make it concrete.

Speaker B: What is the problem?

Speaker E: Yeah, I think in a sense that's

Speaker D: really the hard part. Right.

Speaker E: I think we as computer scientists want to say, tell me the problem so I can say, solve it. I think what I've had to sort of painfully learn over a long period of time is I really need to resist that impulse in myself because I think we often, our students feel like time I spend identifying a problem is wasted time. Time I spent solving a problem. That's where the paper comes from. And usually you end up solving the wrong problem. Anyone who started a startup knows this. The first pitch you have to investors is never what you're doing like two years later and basically doing development economics, doing like, trying to solve some hard problem that hasn't been solved in an underserved community in Canada or elsewhere in the world. It's basically all of the difficulties of starting a company of an underspecified problem and trying to get people to adopt it without the revenue stream. So it's kind of even worse as a kind of problem space. So I think you really have to be willing to iterate on what the problem is. The issue that we went after in this problem in Uganda was you have

Speaker D: another kind of market failure.

Speaker E: You've got a lot of. So 80% of, uh, the economy in Uganda, at least at the time that I was working there, like a decade ago, um, was subsistence agriculture. So people living maybe a little bit above the poverty line, they've got a little piece of land, they produce crops mostly for themselves to eat. But then a couple times a year they'll sell their crops to a market and they'll get the money they need to buy things they can't grow themselves, like salt.

Speaker A: So it's really hard for these people

Speaker E: to figure out, where should I sell my crops? Because they only do it a couple times a year. Prices fluctuate enormously. Their counterparties who buy crops have way more information about prevailing prices and try to rip them off, uh, or use their market power, as we say in economics. Also, like, the people buying crops don't really know. Is this person, do they have high quality crops? Can I rely on them? Are bandits going to rob me if I go down this road I've never driven down before? Um, is the Person going to change the price after I drive to this remote place. And I had to pay the fuel

Speaker A: cost of getting there.

Speaker E: So it's a really tricky thing to match people up. But there's tremendous value in doing it well, because right now the people in

Speaker A: the cities are saying, I can't find

Speaker E: the crops I need. And the people in the rural areas are saying, nobody wants to buy my crops.

Speaker A: And that's terrible. And that's the kind of problem that

Speaker E: economics exists to solve. Right.

Speaker A: That's a welfare loss. It's just about matching people up.

Speaker E: And it's the kind of problem that

Speaker A: information technology seems like it could change the game around.

Speaker E: So, um, at the time, like 10 years ago that we're working actively on this problem in Uganda, most people had a mobile phone, um, not a smartphone, but a phone that could send text messages. And so we set out to kind

Speaker A: of build the Chicago Mercantile Exchange, which

Speaker E: is to say like a commodities market, um, that would run sort of. So not like Craigslist, not like individual people posting individual, um, things. Because you can't browse really on a non smartphone. So you don't want to read like everybody who's got beans in your neighborhood, you want to sort of know, find me a good match for beans. That makes sense given, um, what I'm looking for and where I am. But you want to do it all over text messages, basically. So we built a system that did that. It transacted, um, tens of millions of dollars US in agricultural commodities in Uganda. And ultimately it shut down because we were paying for enough research grants. And at some point it was kind of unsustainable to keep it going.

Speaker B: Um, but it was like a proof of concept that, that.

Speaker E: So, yeah, I think it was us. We wrote some papers about the need to, uh, better use information technology in Africa, which have been highly cited. And I hope that means that people are building on it.

Speaker A: And I think we also learned a

Speaker E: lot about both the opportunities and, um, the places to be humble about, um, the ability for technology to solve these kinds of problems.

Speaker B: And also it's a setting where the outcome really matters to real people.

Speaker D: Absolutely. Yeah.

Speaker C: You're not talking just livelihoods in this example. You're talking about food security. Um, how does that factor knowing the importance of the outcome of the systems you're building? Like, how do you approach that?

Speaker D: It's a hard question. I guess at the end of the day we were feeling like you sort of think about the outside option. What are people going to do if they don't transact on our system and we're not taking away the outside option. If the system fails to do everything that people would want it to do, they're kind of made no worse off than they were if they didn't use our system. We didn't charge anybody anything. That's part of it. Wasn't, uh, uh, a masterminded business.

Speaker A: We had literally no revenue stream, which

Speaker D: is maybe why we had to shut

Speaker A: down, which was hard.

Speaker D: You might imagine we would take a fraction of transactions or something. But the transactions happen in person. Many of them fall apart. There's no way to monitor their existence. Um, interestingly we found out that we built the system, um, to facilitate transactions. Um, what we ended up finding out was that there was some value in that, but there was even more value in publicizing more broadly the prices at which these transactions happened. So the people who didn't use our system at all had better price data because the prevailing system for making price data available was they would go to the markets and interview people, would be hired by the World bank and stuff to, to go, um, interview people in markets and say, what are you, um, buying beans for today? And what do you know? They're not going to tell you the truth because they know that's going to set the price for beans. So they say something a little bit favorable for them. And when we're watching actual transactions happen in a market, we're seeing real, kind of unvarnished and geographically grounded prices. And so that kind of price advisory thing ended up being, I think what we ended up learning was that that's really what the market needs.

Speaker C: M I guess before we kind of wrap up, one thing that I know you've been looking a lot more into recently has been like the economic reasoning capabilities of large language models, things like that. So I just would love to know what direction do you think that's moving in? What are your thoughts on that area?

Speaker D: Yeah, so sort of we, we started out with an idea that subsequently, uh, I think a lot of people have gotten excited about in social science, which is the idea of doing human subject experiments on LLMs. So some people are now calling this in silico experimentation. Right now if you want to do behavioral economics work, which is another area that I work on partly with my Amy colleague, uh, James Wright, um, is to do human lab experiments with mostly psychology undergraduates and you know, pay them to engage in little sort of pseudo economic mechanisms and you see what decisions they make. And you try to use this to build models of how people reason. Um, just to kind of better understand how people think. And um, the problem with this is

Speaker A: that these experiments are pretty expensive. You've got to pay people a decent amount to really motivate them.

Speaker D: You get small amounts of data and

Speaker A: it's not clear how representative like first year psychology undergraduates are of the population you really want to know something about. Anyway, so an idea that uh, occurred to us and has occurred to other people is LLMs are trained on kind of all of human knowledge. In some sense they might represent um, the way people make decisions a lot of the time. And so wouldn't it be cool if Instead of doing 100 experiments on psychology undergraduates, you could do 100,000 experiments on slight variations of the question on ChatGPT and Claude? Ah, the problem with this is you don't know the extent to which they are representative of people. And especially when you kind of maximize against them, you might just end up drawing out weird behavior where they diverge from people. Um, which sadly much, um, of the psychology literature that now is using LLMs is just blindly ignoring this. They're just studying it anyway.

Speaker D: Um, but I think if you want

Speaker A: to do this work responsibly, you need some kind of external validity, you need some sense that you believe these things, reason sensibly like people do. Uh, so that kind of sent us down the rabbit hole of trying to understand to what extent do LLMs, um, behave economically, rationally, um, in what corners of the space do they diverge from this? Um, what kind of data could we build that you could fine tune them on to make them more rational or more like people?

Speaker D: Um, and ultimately I think if you

Speaker A: want to trust an LLM or an agent built around LLMs to uh, act on your behalf, you would like to know that it's going to make rational decisions. So that's kind of what led us to uh, basically take economics textbooks, um, look sort of concept by concept at the kinds of reasoning that economics has surfaced to say this is an important element of rationality, and then kind of taxonomize all of these things and build giant libraries of um, of skill testing questions that you can give to LLMs that we can use to assess their degree of rationality sort of in different dimensions. And what we found is um, sort of what you would expect that LLMs are the frontier. LLMs are getting better and better at solving these kinds of questions, not always for the right reasons. Sometimes they're really good at recognizing, oh, this is an instance of that question, which means that if you embedded it actually in the world, it you can fool them actually more easily, like in ways that you wouldn't want.

Speaker C: So it's not really doing the reasoning.

Speaker A: When we do these experiments at scale, we see things like if you ask the question in first person or third person, they give really different answers. Uh, if you change the gender of one of the. You've changed the gender. If you change the subject matter, if you ask it about, um, like house hunting, it'll give you really different answers than if you ask it about medicine. Because, uh, all of this, um, alignment stuff has made it really scared to have an opinion about medicine. And so it'll just say kind of nothing. If you say, is it better to save five people or three people? It'll be like, no, you shouldn't ever hurt anybody.

Speaker D: So even really obvious choices, sometimes it won't do the right thing. We just tried to build essentially an experimental framework for thinking about when and how should we decide that we can trust LLMs to make decisions in a sensible way. And I think, um, there. We've essentially used economics as a raw material because economists have thought in a really nuanced way about, um, what individual ingredients of decision making look like.

Speaker C: Yeah, it sounds like you kind of gave chatgpt an, um, economics final exam.

Speaker D: We gave the economics final that you would wish that you could give a student if you could ask them unboundedly many questions. It is sort of funny, uh, how like, the LLM evaluation space has turned into the automatic test generation space. Like, you know, it's almost like circle back and become like education.

Speaker C: Yeah, well, I think that's a good way to place to leave with me, I think. And, um, yeah, there's a lot more that we would love to ask you, but I know we have a certain amount of time if you want to hear more from Kevin, uh, without us on the way, two o' clock here on this stage.

Speaker D: And if you want to hear more

Speaker A: from these guys, subscribe to their podcast. Very nice.

Speaker C: We didn't ask him to say that.

Speaker B: Thanks so much, guys.

Speaker C: Thank you very much.