Andrey Fradkin: What Happens to Markets When AI Agents Do the Negotiating

Ben Zweig: What once felt far-fetched are now potential realities that we need to grapple with now so that we are ready in case they come to pass. My guest today is Andrey Fradkin, Professor at Boston University, co-host of Justified Posteriors Podcast, and a wonderfully clear thinker on so many concepts surrounding artificial general intelligence, especially what AI does to the structure of markets, firms, and work. Here's our talk. Okay, welcome to the Economics of Work. I'm excited to welcome Andrey Fradkin, Professor at Boston University, and co-host of my favorite podcast outside of this one, Justified Posteriors, and one of the best ambassadors of modern economics. That's kind of how I think about you. And most recently, that means the economics of AI. So, Andrey, welcome to the show.

Andrey Fradkin: Thanks for having me on, Ben. Also a huge fan, and of course, forever grateful for Revelio being the first sponsor of Justified Posteriors.

00:51 How Andrey updates his priors

Ben Zweig: Yeah, no, of course. Huge fan. I mean, yeah, so Justified Posteriors, I love the framing of it, that you come in with your priors, and then you see how much those have shifted. So, just based on that premise, I want to just hear your take on the nature of shifting priors for you. So, I guess if I think about everything you believed X number of years ago, and everything you believe today, like the way I kind of think about it, let me know if this makes sense, is that like, there's kind of maybe like four dynamics at play. Like one is that maybe you have like really weakly held priors, and they can just like shift because they're just weakly held. Or number two is like maybe the evidence is really good, maybe you're like swayed by evidence, and that could shift your priors. Or number three, maybe you're just like getting older, and have different sensibilities, and just like different outlooks of the world. And number four, maybe there's like, maybe the world is just like kind of obviously changing. And it's just like noticeably changing the way you understand how things work, that really has nothing to do with evidence. But like, the world is the evidence. So, I don't know, I'm kind of curious to hear your take. I'm like, what are the dynamics at play in how your beliefs change?

Andrey Fradkin: So that's a great question. I generally have strong priors, and I think the reason for that is twofold. One, I've read a lot about how societies work, and I'm talking about specifically priors about economics related topics. I've read a lot of economics in my life, I've obviously received a lot of training, talked to a lot of people, and so I have a worldview. And those priors, by the way, include things that are really hard to demonstrate, like I know what things are hard to show, what things are easy to show. And so when I hear a new claim, I kind of slot it in to my worldview. And this makes me not update a lot from individual studies most of the time. Or alternatively, if you ask me a question, a specific empirical question, I can probably reason about it based on what I already know, which includes both kind of evidence that is in research papers, but also just evidence about what I know about how the world works. Now, AI has been really fascinating. I think when I first started hearing about grand claims about the singularity, and that was a long, long time ago, it's not that I didn't think they were possible, it's just kind of I assumed that it wouldn't happen really in my lifetime. I guess I had longer timelines.

04:05 What 'the singularity' means here

Andrey Fradkin: And then it's...

Ben Zweig: Well, wait, sorry. What do you really mean by the singularity? I feel like I kind of have a concept of what that means, but I feel like it gets used in different it can be interpreted in different ways. What do you really mean by the singularities? That machines just make more machines to infinity and we're all just irrelevant in the world? Or is it a more narrow view of self-reinforcing growth or something?

Andrey Fradkin: Yeah, I don't have a strong definition of singularity. It's just more like how it's being colloquially used. Let me reframe it as artificial intelligences that are as good as humans at most things. And can be replicated at a fraction of the cost of humans. So intelligence is abundant. I think that's kind of what I have in mind. I don't want to make any claims about self-replicating factories expanding through universe sort of stuff. But I just assumed we were a bit further away, we're quite a bit further away than it ended up being. And I think my mind started changing on these things around 2021, I guess, 2022, where I started seeing the language models, learned about the scaling laws there. And it kind of then became pretty clear that we had a formula and the formula was going to get us more intelligence. And then obviously there are a lot of details to figure out. On the technical side, I don't have any expertise in that, but more on the economic side of how is this industry going to be organized, how is diffusion going to happen throughout the economy? What are really the useful capabilities? Where does value come from? From human labor, that's something that AI has clearly forced us to think a lot harder about.

06:14 How AI is already changing work

Ben Zweig: Yeah, so fine. So that's mostly just seeing the world change, that has shifted your major prior, unlike the timelines for transformation. In that big question, okay, makes sense. So what do you think, I mean, let's say aside from timelines, so what do you think is happening that is very impactful already? Do you think, I mean, we kind of see AI everywhere, everyone's talking about it and using it and thinking about it. So do you think there are... Yeah, what do you think the major effects on the world are today, now that intelligence is more abundant?

Andrey Fradkin: I mean, it's changed how we work, how white collar work gets done in particular. And forget about whether it's affecting measured productivity or not, but just that it's hard for me to imagine a work day today without using AI. And that involves doing research for a project or responding to an email even oftentimes. How do you phrase this email in a polite way? I could have done it before the AI, but it would have taken me a while to craft it. And now the AI can just do it for me if I give it some bullet points. And just personally, I'm experimenting a lot with just how far can you push it? How much can you really do with AI in the workflow? But I think this for people who are curious about trying new things, once you try it, you can see how powerful it is and you're using it all the time. So I think that's kind of on a practical level. And then for consumers as well, they're using AI just for personal reasons, I don't know, relationship advice or whatever, shopping advice, tax advice, et cetera. So I think that's so far the biggest impact. That's kind of one. And then two, the other big impact is just the investment boom that it's generated. And it kind of seems that the entire economy is a bet on AI yielding huge productivity gains. And so all the investments in the LLM companies and the compute data center build out, that sort of stuff, it's just really important. So those are, I think, the biggest changes so far.

08:57 Will the AI investment boom pay off?

Ben Zweig: Yeah. Do you think that those will pay off? I mean, obviously very hard to say, but like gut feel, do you think that huge investment from the economy at large will be a lucrative investment? I guess I mean, on the one hand, for the investors of it, for the AI companies, but also just like, even if they don't have a good return on that investment, for the economy at large, maybe the returns accrue to the users who get a bunch of cheap intelligence. Do you think it will pay off for the economy?

Andrey Fradkin: Yeah, I think it will pay off for the economy. Maybe with a little hiccup, I mean, I'm not going to rule out internet bubble-type dynamic, but it's pretty clear that the internet bubble paid off for the economy, even with a lot of sloppy investment underneath it. And it's hard to imagine that this doesn't as well. If you look around, even in a way that's even much more visceral than the internet to me, it's already so powerful. It can allow, it gives all of us so many tools to use to accomplish what we want to accomplish. And we're just getting started. It's just so far from diffusing throughout the economy. And many have made this claim, but if you were to freeze the technology at today's level? We could spend a decade figuring out how best to use it, if not more.

10:30 Who wins from AI?

Ben Zweig: Yeah. Who do you think are the major winners? Are there some groups of people or countries or industries or whatever that you think will benefit the most?

Andrey Fradkin: I mean, it's hard to imagine the US and China not benefiting a lot from this as being the key developers of these technologies. But I tend to be an optimist. I think there's a lot of concern, let's say, about developing countries being really hurt by AI. I think that developing countries are going to be greatly helped by AI, actually, because everything we see is intelligence is getting cheaper at a rate. And we have research on this that is unprecedented for technological change. And even with today's level of intelligence, this gives people in a developing world access to world-class doctors, world-class professionals in every single domain. And that should, in principle, be very agency enhancing for them. Now, obviously, I'm not going to say everything's going to be perfect for developing countries or anything like that. But there are other, obviously, concerns economically. But I think some of the people who are worried are a little underrating this, right?

Ben Zweig: Yeah, I think that makes sense. I mean, I don't want to say that intelligence is scarce, in different contexts. But there's some version of that where that has to be true, where education levels are lower, where formal training is worse. There must be higher returns to intelligence where that is more scarce.

Andrey Fradkin: Yeah. Because intelligence plus expertise is what I would say.

Ben Zweig: Yeah. Yeah, it's funny. I don't know if they still make these, but I think in the Super Bowl, there was this OpenAI commercial. There were these strings of OpenAI commercials where they were framing the benefit of AI as helping blue collar entrepreneurs, like someone opening a restaurant or inheriting a farm or something. And it was very targeted to this, yeah, aiding blue collar work, which I thought was so funny because that's so not how it's used in the US. But that seems to be the vision for prosperity.

Andrey Fradkin: Yeah. Yeah. I wouldn't say opening a restaurant is blue collar work necessarily, but yeah, it is a small business person. Yeah. I do think that there's a point I like to make is that most things have diminishing returns, right? Most goods. So the first slice of pizza is really amazing. The tenth slice of pizza, not so amazing, perhaps. Quite negative. Yeah. And so with AI, it's similar in that actually for that person who's considering opening a restaurant, talking to a frontier model about how to do that can be really, really valuable. But as a share of total usage of AI, it's tiny and people are using AI for very marginal things. So things that are not that valuable, maybe even just like a role playing conversation, or I don't know, explain this podcast to me, something that you would have never even thought to do in the first place beforehand. And so we have to be careful that there are some uses that are very, very valuable and then other uses that are not, and the AI is the bundle of all of them. Yeah.

14:33 Frontier models, open models, and integration

Ben Zweig: Right. Right. Right. I mean, yeah, I would think if we were thinking about creating the maximum value for society and we have these basic models that are intelligent enough, that don't need to be frontier models, being able to get very, very far for quite inexpensively, there's so much value there. And then now with more of the frontier agent based tools, it seems like they're much more focused on getting those marginal improvements. So do you think the benefit for society might come from open source and the benefits on a margin come from the frontier models? Do you think that's a way to think about it?

Andrey Fradkin: I'm not so sure. I think the way things are going, this is a weak prior, by the way. Is that a lot of the hard part is integrating the AI into the economy, into production processes that might involve training bigger models. It might also involve co-training models and scaffoldings in a way that's really useful. We're designing various verification systems that allow you to more confidently use the technology and how that work is going to be split up between frontier labs and open source labs. I'm less sure about... Like I just say, I'd be surprised if... Or not as surprised. We have plenty of evidence that both Anthropic and OpenAI are pursuing many activities that don't involve training models. So I wouldn't be surprised if they're the ones that are leading a lot of that integration and everything in technology and business is either bundling or unbundling. So right now we're used to this per token pricing situation. I wouldn't be surprised if we have more bundling in the future where you have per outcome pricing or some other pricing structure. Because ultimately the end user, Joe Schmo, or even just a manager at a big company, they don't want to be thinking about tokens and which model those tokens are coming from. They just want to know whether the thing they wanted to accomplish is done and how much that costs.

Ben Zweig: Yeah. I think it sounds right that the value might just be in these integrations. It kind of reminds me of this task chaining paper that the value that you get relies heavily on where in a process this can fit in. It can be chained to other things. And that glue is the difference between something being very valuable and not valuable at all. Yeah, I think that sounds right.

17:50 Is AI jamming the job market?

Ben Zweig: Well, let me hit you with a theory about what harm might be being caused by AI. And also a question about your thoughts on it because you seem like the person to think about this problem. So I suspect that the most negative consequence of AI today are on congestion in labor market search and match. That you just have a lot of people applying for jobs using AI. Employers are getting signal jammed. It's very hard to figure out who's the real deal, who's really interested. And you've done research on AI. You've done research on matching market congestion. What do you think is going on? Do you think this is a real thing? Do you think it's solvable? Do you think this is a policy choice? What's your take on this phenomenon if it exists?

Andrey Fradkin: Yeah, I mean, so I think it exists. I wouldn't put it as at the top of my current issues with AI, to be honest.

Ben Zweig: Really? Okay. Yeah. What would you put right at the top? Why do you think it doesn't quite rank?

Andrey Fradkin: Well, even before AI, it was trivial to send applications. The problem you're talking about already existed. And so AI has made it worse, but I don't know on a quantitative scale how much worse it's made it. That's kind of one way to think about it. That's not to say it's not important. It's just for me, it's not... There's so many hires already happening through referrals or other channels. We still have campus recruitment. We still have these kind of for new workers, GPA, and other signals like that. So while it's an issue, I wouldn't put it at the top. But now let's kind of talk through this. I think economists have a solution for various parts of this problem. I want to say just economists, but theoretically, we have a solution and there's a practical solution, which is we introduce costly signaling. If a job seeker has, let's say, two signals per week or per month saying, "I'm really interested in this job,", then the fact that they've sent them to a particular job and not another job is telling the employer something about the private information of that applicant, either that applicant really wants the job or they think that they're really well suited for the job either or both. That helps the employer sift through candidates. So that's kind of one thing that we have. I think one of the reasons that this doesn't quite work is that the labor matching market is very fragmented. There are lots of labor market platforms and you can also apply just directly on a website. And therefore, even if one company, say Indeed (and I should disclose, I've consulted for Indeed in the past), implements this feature. I don't know the current state, but they have been trying to think about such features since those are signals only on Indeed. Now, if you're using also LinkedIn and directly applying and also maybe sending an email to someone, the scarcity of the signals maybe becomes less important. So I think that's one aspect of it. A second aspect is, can we verify more things about an applicant, ex ante, so more objective measures like grades in a particular major at a university? These are things that you can learn what those signals mean. If you have a 4.0 in math at a top university, you're probably pretty smart. Maybe if you have a good GPA in another major, you might be smart, you might not be smart. And then maybe if you have a bad GPA, you're not that smart or you're not that motivated. So we have hard signals like that for entry level college labor, but we have less of that for later on in the funnel. And I think a particularly interesting thing to be doing is to be doing automated interviews and other things with some detection for fraud that evaluate what people are doing. But I do think that the labor market has to change. Things like a cover letter become kind of meaningless to some extent now that anyone can generate a plausible personalized cover letter for every job.

23:16 Signals, references, and reputation

Ben Zweig: Yeah, it's interesting. I feel like I'm sympathetic to the idea that we can do this well for entry level candidates. There are signals like GPA and university prestige.

Andrey Fradkin: They're highly imperfect, but they've always been imperfect.

Ben Zweig: Sure. Yeah. Maybe there's another issue to wring our hands about that those signals may be weakening with grade inflation and all that, but fine. So we have sets of signals. And they're good signals because they're differentially costly between the presumed top performers and less top performers. The limited apply thing, I don't know, it's costly, but it's not really differentially costly. It could be, I guess, with the right design, but sounds harder. And like you said, we have this multi-platform issue. But let's say we just take the experienced hires. I mean, who don't have those verifiable signals? One thing experienced hires accumulate is reputation and a network. So in the world of economists, we pretty much know who's really good. We just know because it's a small world and everyone knows each other.

Andrey Fradkin: Well, I don't know. We also know who's published where and what citation counts are. So part of that is through the network, but I actually think economics is pretty unique in that we have this publicly observable signal, which has its problems to be clear, but that we can all reference. Right.

Ben Zweig: I mean, yeah, that's totally true. But I also think if I think about, I'm just thinking of this other community that I participate in, which is alternative data in hedge funds. And it's like, as a company, we sell alternative data to hedge funds. That world of people that buy data and sell data and analyze data at hedge funds, everyone knows each other. There's maybe a thousand people. Everyone's got a reputation. And if someone's looking for someone to be their alternative data analyst, you pretty much know who's the real deal and who's not. So that's a market where there is no externally verifiable signal, but the network carries reputations pretty effectively. You've done research on reputation versus licensing. So I wonder, for things that do you think there's a way to extend reputation as a signal and in that way, not have to rely on licensing? Or yeah, what's the role of reputation in establishing credibility for workers that are not straight out of school or something?

Andrey Fradkin: I mean, I think reputation, it depends on which part of the labor market we're talking about, but I think reputation is obviously important. And for more senior roles, you'll have reference checks, of course, which are highly imperfect signals.

Ben Zweig: Do you think they should be more common?

Andrey Fradkin: What?

Ben Zweig: Oh, I was thinking references, employee references, feel like they're not very common. Do you get that same sense? Or do you think they should be more common?

Andrey Fradkin: It's tricky how common. I think for senior roles, they're extremely common, to be clear. I think there's a cost to soliciting them. And it's even sometimes tricky to get the person on the phone, if they're senior enough, obviously. Imagine someone was reporting to the CEO, getting that CEO on the phone for a reference check. It might be tricky, especially if this person is maybe potentially leaving. So there are incentive issues.

27:37 The AI panopticon

Andrey Fradkin: Let me flip this. Here's a world that we can live in the future. So the AI, so you have to assume that there's little things that humans can do that the AI can't. The AI can definitely watch us as we work, and it can form a very unbiased opinion about what the hell we're doing. And then you can have a mechanism. And so I think there's going to be a role for mechanism design here, for transferring that information to potential employers for deciding who to hire and how much to pay them and so on. So there's a world where AI is a panopticon. There's no private information anymore at all. So there's no point in reputation doesn't even matter. To the extent that reputation matters, it matters as an incentive mechanism. We've built up this reputation. You don't want to screw it up by doing a bad job. So humans do need motivation. But of course, if you live in the panopticon, you also don't want to screw it up, because the AI will notice it right away and will say, "Well, so-and-so is no longer trying." And therefore, you shouldn't hire them.

Ben Zweig: It's a very interesting world. I think it could... If we did live in that world, it would solve a bunch of problems. To me, it sounds hard to imagine. Because I don't know. Yeah. Why am I having a hard time imagining? I guess because of what I assume the backlash would be for privacy considerations. So it's a little hard to imagine that people are going to get there, where they just are always observed. And then their performance is directly observed and shared publicly. That sounds like it's got hints of dystopia that would freak people out.

Andrey Fradkin: Yeah. But I think that's the same time that could be inefficient, in some very narrow sense, inefficient things to do. And by the way, I think a lot of parts of life have this tension already pre-AI. Ambiguity is a feature in so many parts of life. Let me give you an example. The job description and what someone actually does on the job. Those could be very, very different things. And in fact, sometimes the best workers are purposefully not doing the things they're supposed to be doing because that's their judgment to do so. Right. And so...

Ben Zweig: Job crafting.

Andrey Fradkin: Yeah, exactly. And I think an intuition for why we like that ambiguity is we have the sense that with the right incentives, this person will have the opportunity to do the most good for themselves. And also part of the job is consumption value. You pick one job versus another because you enjoy doing that job versus another job. Now, if you go to the Panopticon, you would have to be very enlightened Panopticon where like, "Oh, this person is going on a podcast. I don't know how to trade that off against this person making a sale to a client." Yeah.

Ben Zweig: Ouch.

Andrey Fradkin: Yeah. Right. You see what I'm saying? Yeah, yeah, exactly.

Ben Zweig: Yeah.

Andrey Fradkin: It would have to be very enlightened to actually work, but I am afraid that market forces might push us there regardless. There's certainly Panopticons and lots of workplaces already, like call centers and professional sports is a Panopticon. Everything about your performance is observed. And people are kind of okay with that. So it's... Yeah.

Ben Zweig: Yeah. No, it's an interesting thought. I mean, yeah, to the extent that we have it already, it's very imaginable that we can shift more toward that world on a margin. Yeah. And to the extent that we're incentivized to do so, it's hard to imagine that we wouldn't, unless there's very extreme backlash.

32:29 The Coasean Singularity

Ben Zweig: Okay. And yeah, while we're actually talking about things that kind of exist on the margin, that kind of exist in the world today, but we can also imagine a very extreme world where we go full tilt on all of them. I want to talk a little bit about your idea about Coasean Singularity agents as market participants or agents interacting on our behalf. What is the vision for what that is, what that looks like, and why do you think it's kind of plausible?

Andrey Fradkin: Yeah. So we wrote this paper called the Coasean Singularity, and we didn't define what the Coasean Singularity was in that paper.

Ben Zweig: That's excellent. So yeah, here's the opportunity.

Andrey Fradkin: It's very evocative, right? I think at a high level, the idea that the Coasean Singularity is a thought experiment of a world where essentially a lot of the things that prevent what economists would call perfect markets, so like perfect information and incentive alignment and that stuff are kind of solved. And as a result, we get more efficient allocations of all the goods and labor, prices are better, and so on. And that has also implications for the structure of the firm. So one of the reasons, this is kind of the Coase part, is that Coase wrote this very famous essay on the nature of the firm, why do firms exist? And they kind of exist because there are lots of transaction costs if you have to use the market for everything. You don't want to go out and for every single task you need done for your company, hire a separate contractor and negotiate with them and so on and so forth. And now why not have an employee? And then an employee can do those things. You can just tell them to do it, have them negotiate every time. There's kind of this advantage of being the firm. Now imagine a world where I have my agent, Ben, you have your agent, our agents talk to each other, that's not our cost, that's the agent's cost. And then they decide, hey, you guys are going to do this podcast today, just show up at this link and that's that. And we don't have to even think about it. So imagine all parts of the economy were like that. So I need to buy a lamp, my agent goes and talks to the agents of all the people that are manufacturing lamps, decides which one is the best for me, bargains for the right price, and then I just get it and I don't have to think about it. So that's kind of the thought experiment that we have. Why is it plausible? Well, it's pretty clear that agents can, their problems don't get me wrong, but it's very easy to imagine now how an agent might be talking to other agents and representing me. And there's of course this open set of questions of how do we ensure that the agents are representing us well, Let me restate this. So how do we know that the agents are representing us well? How do we know that they're not being tricked or manipulated by other people's agents? Do we ever need to put the human back in the loop? What are those things that the human needs to be put in the loop for? So these are very open questions, but I think we'll figure them out.

Ben Zweig: Well, do we have some version of this today?

Andrey Fradkin: In the sense that we can hire real estate agents who represent us?

Ben Zweig: Oh, we can have human agents represent us for sure. I mean, yeah, obviously the concept of agent is not new, but do we have automated agents working on our behalf? What is the closest analog to that you think exists today?

Andrey Fradkin: I mean, you have very simple, very dumb examples. You get your paycheck and then automatically there's an automation in the background that puts it into a savings account and then you can even automate it to automatically invest in something as you get it. You're going to have like bill pay, automatic bill pay and so on. It's a very rudimentary version of it, but any kind of automation like that is one. I think it's very limited in a lot of ways for obvious reasons. It's hard coded, it doesn't have flexibility and so on.

Ben Zweig: Yeah, I like that as an example because when you say, "Oh, invest in Wealthfront or whatever," then you have some agent deciding the best investments for you. It's like, "Oh, well, here, I know about your preferences based on your retirement age and your risk profile and whatever." You put in your preferences and it goes and finds investments for you. Then I guess there's something, I mean, there's buyers and sellers on the other side, but like, yeah, I'm trying to think if there's a way to extend that where the buyers and sellers on the other side are automated. I guess that's not really the case in financial markets, but yeah, it's maybe just a two-sided analog of that phenomenon.

Andrey Fradkin: Yeah, I mean, the other thing is any of these centralized matching systems that economists have helped to design, like the residency matching system, I mean, it's not per se a full agent and you have to literally specify your preferences over the different residencies you want and then hospitals have their preferences over the candidates. But then once that's put in, then you get a match at the end and there's kind of a black box. As far as you're concerned, there's a black box and then you get your assignment, right? And so agentic systems, and John Horton and co-authors have done some work on this might instead of forcing you to be the one to manually set the preferences for every single option, say like, I prefer this hospital, this hospital and so on. The agent will just do that for you. And then, and so one of the things that we write about in the paper is maybe we have more situations where these mechanisms are involved. Right now, they're really annoying to do, they require really high stakes decisions because of how much effort it takes. But in the future, if you have agents participating, then you have them, we can use them in a lot more places.

40:00 Thinking about the messy middle

Ben Zweig: Yeah, yeah, it's a very cool thought. I mean, so one thing I struggle with sometimes in the economics of AI community, and I'm curious your take on it, is like, we can imagine scenarios that are maybe possible in the world of AI and then we could analyze the economics of those scenarios. And we obviously have scenarios today and we can analyze those scenarios. But I struggle with getting from A to B, how are we... Sometimes I feel like some subset of conversations in the economics of AI are insufficiently at the margin. You know what I mean? And as a economist, we should be thinking on the margin, can we get a little bit of the way toward this outcome where we have better facilitated coordination between preferences or something where we overcome some class of coordination failures and maybe on the margin have slightly smaller firms that are more decentralized or something. So I don't know, what do you make of this? I don't know, do you think we need to... I mean, I think there's a role to imagine some scenario and what are the economics of that scenario. But to me, it feels like there's still a gap in some economic analysis of the path to a scenario and what that kind of looks like. I mean, it's kind of like the messy middle idea from Martha Gimbel. What do you make of this middle lands where, to me, they're still kind of confusing?

Andrey Fradkin: Yeah, I mean, I think they're confusing. I think people work across that spectrum from studying the near past to trying to study the near future to studying kind of hypothetical scenarios. I don't want to speak for my co-authors, but the purpose of this paper is to also create a vision and a set of trade-offs, describe the set of trade-offs, describe kind of what things could look like, and then that can provide inspiration for other people to work towards that direction to the extent that they want to. I do think a big tension as a researcher, your instinct is, "Well, I'm going to write another paper." You're asking for us to predict the future, but the near future, but not the far future. The best way to predict the future is to invent it, is a quote that someone said at some point. And naturally, it kind of raises the question of, "Well, why aren't you an entrepreneur?" Or, "Why aren't you actually building these things rather than writing about them? Does that make sense?" So you're asking something of researchers that their instinct is not to do that. And then, at the same time, I kind of think it's hard to predict period, but it's really hard to predict also the specific ways in which technology evolves and diffusion happens. I'm not saying predicting the far future is easy either, but it's hard. Yeah.

43:48 The case for superintelligence scenarios

Ben Zweig: Yeah. I mean, there's one view of, I guess I think about it as there's some, within the economics of AI community, I guess, I feel like there's a contingent that's like the super intelligence crowd that wants to write and think about the future of super intelligence. And in that model, there's no diffusion. It's just a totally different paradigm. And I wonder if, one thought of the usefulness of that is that, I don't know, I'm kind of thinking of this trope in social media, not social media, in movies and TV and stuff, where you have this scientist who was the first one to do research on X, and then lo and behold, X happens, and the scientist becomes like plucked from obscurity to run the show. And that is a kind of familiar theme. And maybe that's a model for where these scenarios become really useful. But a counter argument to the usefulness of those scenarios is that if those things do come to pass, we'll have plenty of lead time before that. So we won't overnight shift to this new paradigm. So maybe there's just enough time to transition the way we think. But then again, as I'm saying, I'm just thinking, all right, but there's also a usefulness in just analyzing the economics of some toy scenario because it helps us understand all sorts of things and all sorts of mechanisms, because that's what models do. They clarify our thinking on some themes. So yeah, in your mind, what's the best kind of defense of this super intelligence analysis of a scenario that's not our current scenario? From a usefulness perspective.

Andrey Fradkin: Well, I think one question is, are you determinist or not? So if you're a determinist, then maybe you're right. There's nothing we can do today to affect what happens five to 10 years from now or 20 years, whatever your timeline is. And it's inevitable. Artificial intelligence is going to spread through universe at the speed of light. And it'll also be smarter than us. So it'll tell us what the economics of AI is at that point once we get to it. But I do think a lot of people have a viewpoint, which is a natural human viewpoint, is that humans have some amount of control. Policy decisions made today, product decisions made today, might affect the shape of that super intelligence. One version of that is obviously people who are very interested in AI safety or think they can make progress on having these future AI systems be better aligned with humans, whatever that means. But there's a version of the economics version of that is, what is the way to organize the economy today so that when we get this capability, it will be beneficial in the greatest possible way. And that's a very, once you say that's a really ambitious thing. I mean, there are other viewpoints on it like, this is going to happen and we're going to need something off the shelf at that time to implement. It's kind of the people make the argument of pandemic preparedness, right? If you're prepared for the pandemic, once you get to it, you're going to have a lot easier time dealing with it and there are proposals going around there. I think some set of people just really love thinking about the science fiction features. So that's another reason. And the final reason it says, on a personal level, it affects what you do. I struggle with this. If you think that you're going to get artificial super intelligence in some amount of time, that's very soon, within five years or something. Well, what is the point of doing anything? You just wait and then the super intelligence will do it. I think about this in my research projects. It's like, well, I could really hurry up and finish it today or I could wait a couple of years and then one shot prompted. And then why suffer today? Or the alternative version is, well, my work is primarily intellectual, so I'm really going to have nothing to do in the future. And so I need to save up now so I don't become part of the poor class in the future. There are different versions of this. Some people, they're wealthy enough. They're thinking, well, if we get this version, we're going to have revolution. I'm going to build my bunker and so on. So there are these sorts of considerations for some people. Yeah.

Ben Zweig: Yeah. I'm very sympathetic to the... I think out of those reasons, the part that I'm really drawn to that I like is this idea that when you do analysis on some scenario, you are inviting entrepreneurs to build that if it's a good scenario. So this like, Coasean Singularity is a portrayal. Like, hey, if we had this world, it would solve a lot of problems. It would be like a pretty good thing. And then the people who build entrepreneurs, whoever might be so compelled by that idea that they might say, okay, I'm going to innovate in that direction. I think that's very compelling. The nihilism of what we do and just I want to just have fun writing science fiction. I just have a hard time being sympathetic to it. Not because it's unreasonable, but really just because I guess it just offends the American dream sensibilities of... If we all believed we had a limited time before we were all part of the permanent underclass, then isn't that depressing? I'm almost averse to thinking it.

Andrey Fradkin: It doesn't make it any less true.

Ben Zweig: No, it doesn't make it any less true, but it is just like a visceral... Yeah, yeah. I don't know. I guess if you... Just to think about that idea on the margin, it reminds me of this deflationary spiral in a way. If you keep thinking prices are going to fall, then just wait and wait and wait. I guess that's something that can be quantified on the margin. If the price of something keeps falling, there are some times when we consume less and that spirals. But I feel just empirically, it's not that meaningful of an effect. People still prioritize today and maybe discount the possibility of things being meaningfully different in the future. I don't know. Do you think that's a comforting thought or do you think that's true or do you think like...

Andrey Fradkin: I think that's true of people, but I'm not saying that's rational.

Ben Zweig: Right. So maybe true of people in aggregate, but not true of this person or that person.

51:37 AI and interest rates

Andrey Fradkin: Yeah. I think a related argument is Halperin and co-authors' argument about interest rates. So if you think artificial superintelligence is coming and it's going to yield this abundance in the future, why would you save? Because when you get to the future, we're going to have abundance and everyone's going to get it. And so you should just spend today. Alternatively, if you think that AI is going to kill us all, then why save? It's funny.

Ben Zweig: I think about that argument a lot. And just when long-term interest rates just ticked up, I was like, I wonder if this is about AI. Yeah. What do you think? Do you think that's we're seeing little bits of that or not so much?

Andrey Fradkin: Well, I'm very consciously not a macroeconomist. And I do believe in Tyler Cowen's, I don't know if it's second or third law, that all statements about interest rates are false. Yeah. I don't want to try to explain short-term interest rate movements.

Ben Zweig: It's a dangerous game.

Andrey Fradkin: I mean, I think on a practical level, right, there's a ton of investment in AI right now that should increase the interest rate. So yeah, I don't know. That's kind of my sense. Okay. Yeah. The opportunity cost of capital is higher because so much of it is going to AI and there's so many people that are optimistic about the returns to these AI investments. The US government's fiscal position is bad, but it's not unpredictably bad. I don't think anyone would have thought that the Trump administration would improve the fiscal sustainability of the US. So I don't know. Right. It didn't just get bad overnight. That's what I'm trying to say.

Ben Zweig: Yeah. Yeah. Yeah. Interesting. Yeah. No, I think that sounds kind of plausible that we'd have this crowding out. I mean, we have to- I think even just like it's noticeable that the venture capital landscape has shifted dramatically. Now every VC investor is putting it all on AI and obviously that creates less capital for other parts of the economy. So it stands to reason that it would happen with sovereign debt also.

54:17 What the economics of AI community should do

Ben Zweig: Okay. Yeah. I mean, yeah, this is like, again, don't want to deviate too much into macroeconomics. But yeah, one kind of like final question for you is just about the economics of AI community. What is your kind of like hope for the orientation of that community? How do you think people in that sphere should be thinking if you kind of like could steer it?

Andrey Fradkin: I think the thing that I would love more people- so first of all, we have a great community. And people are really motivated for the right reasons. They think AI is important. They're trying to do the right work that they think will make a difference. Not everyone agrees on what that is, but I think it's really like a pleasure to be a part of such a community already. I think that's one thing that I would say. Not all parts of academia are like that. And obviously, a lot of members of the economics of AI community are not part of academia anymore. So yeah, something. So I think that's not what they should be doing. I'm just saying that it is what it is. Starting from a good place. And I think going forward, and this is mostly to people who are coming in to the community or people who are thinking about what they should be doing within the community. I think really taking the responsibility that we have seriously, right? The world is very confused about this important change. And we're confused. And earnestly trying to figure things out is very high value. And not only figuring them out, but also communicating them to other people, whether those are other experts in AI or other policymakers or just to the general public. It's hard. It's hard to know what the returns to that are. But that seems really important. And it seems important to also work to be a little bit open-minded about other types of work. I think a lot of us have been writing more blog posts, Substacks. We sometimes spend a lot of too much time on Twitter maybe, but there's a sense in which that sort... I always go back to like, you can single shot a paper these days. And if we just, as a community, only focused on paper writing in the traditional sense with doing the traditional circuit, waiting for the referee reports. It's just a waste of time. And to be clear, there's a reason why the system evolved that way. People need it to be valued, and so on. But the opportunity cost of time is so high. And spending that writing, slightly tweaking the written artifact of some paper that you started working on five years ago just to please referee three's sub-comment five. It's just not a good use of time. And so thinking really hard about what is a good use of time, I think that's really important and not just sleepwalking into the default. And I don't have the answer, by the way. It's not like I'm sitting here being like, "This is the optimal use of my time today." But the trap is for a lot of academics is just to continue doing the circuit, doing the same old stuff that they've been doing the entire time. And I think that's not going to be a fruitful way to do economics of AI.

Ben Zweig: Yeah. Yeah. No, I think you're right. It's one of these areas where I feel like it is kind of changing. And there are more Substacks, more Twitter stuff, more podcasts. And I also kind of feel like there's more private sector initiatives. You mentioned John Horton, he's starting a company and this stuff we do. Revelio is kind of trying to add to the conversation through a different channel.

Andrey Fradkin: Let me just say something. You as a data provider, you have some competitors and so on. When I think about all the papers that have been written with this type of data, it's pretty clear that compiling the data is the high value thing. Because writing the papers is trivial. Sometimes I see these papers like, "Oh yeah, yet another shift share paper using labor market policy. I don't know. Another diff-in-diff." And those papers are great, don't get me wrong. But actually creating the data set is way more high value than any of those individual papers.

Ben Zweig: Yeah. Yeah. No, I mean, thank you. I agree. And yeah, I think you're right to kind of invite the curiosity of the community to think about other ways to be productive because that's ultimately what we all presumably want. Yeah, fine. Thank you so much for this. This was so fun. And major shout out to your Substack, Twitter presence, but also Justified Posteriors, best podcast out there if you want to learn about the economics of AI. It is fantastic. Thank you so much for doing it and being part of this.

Andrey Fradkin: Yeah, this is so much fun. Thanks for inviting me, Ben.

Ben Zweig: The Economics of Work is brought to you by Revelio Labs, workforce data for research and benchmarking. If you enjoyed, please remember to rate, review, and subscribe. It'll help us reach other curious listeners. I'm your host, Ben Zweig. Our producer is Cole Wagner. Thanks for listening.