Mallaby: OK. Natasha, I think last time we had a meeting a few months ago you had a pretty firm view on the lack of strong evidence for any labor market disruption from AI. Do you still feel the same way?
Sarin: Yes. I do. In that, if you look at the moment at basically any indicators of AI exposure in different types of occupations, and a lot of these indicators are coming from the AI labs themselves. And you can say that there are problems with the data and we’re not doing a good job of capturing it. But, like, the best that we have at present, you do not see any evidence of differential trends with respect to hiring, with respect to layoffs, with respect to overall composition about the types of workers, dislikelihood to be hiring new entrants into the labor force, any of that. You see none of the difference between AI-exposed occupations and non-AI-exposed occupations. And so in the data presently we do not see anything that indicates that there is widespread displacement that is happening at the moment as a result of the technology.
And in fact, part of what is giving me a little bit of pause as I try to evaluate this set of facts about the labor market, relative to the types of things that you’ve heard from a lot of these technology executives—like Dario Amodei saying, you know, we’re on the precipice of, like, you know, 50 percent of white-collar labor being displaced—is that it is actually the—it has for a long time been the case that sort of the pessimistic take about AI was coming from Luddites who didn’t really know much about the technology and were fundamentally kind of suspicious of it. It is now the case that what is happening is a lot of companies, who themselves are trying to deploy AI in ways that are productivity enhancing and ultimately revenue enhancing for their firms, are finding that they haven’t quite figured out how to do that yet.
So you’re hearing stories about how, you know, Uber went through its entire compute budget in the first three months of this year, or at Google they were trying to get everyone to do—to use a bunch of these tools in useful ways, and what they found was their employees were tokenmaxxing, so were using chatbots to, like, ask them what time it is rather than looking on their computers, in ways that were taking up a lot of compute and ultimately not super productive. And so all that is to say that I think that we are—I suspect that we are in relatively early innings with respect to seeing massive impacts on the labor market as a result of AI.
That said, I have to think that it is coming. In that my—and I’m sure all of you are having this experience as well—where a lot of the types of work that I do—I do a lot of research—a lot of the types of work that I do, where traditionally I would spend time with research assistants and, you know, go back and forth many times on, you know, trying to learn enough to be able to come and do this briefing this afternoon, like, basically AI is better than the best research assistants that I could marshal. And so you have to think that that means that there is going to be less of work for—particularly for early career-stage types of people in the economy. But we are just not seeing that yet.
Mallaby: Doug, let’s take this to Europe. There’s a paper floating around the internet. This is my last question before we open it up to members, so get your questions ready. And this paper, called Europe 2031, sort of tries to paint the picture of what’s going to happen in AI geopolitics, particularly how it impacts Europe. And it’s sort of one of these, you know, incredibly depressing narratives that essentially, you know, describes Europe. Every time AI accelerates a bit more they think, oh, we really understand we’re going to have to do something about this. And they do kind of 20 percent of what they should do. And then it accelerates some more, and they do another 20 percent. And they’re so far behind the curve that at the end they’ve got no significant compute, no bargaining power as to access to AI.
And the U.S., which is short of tokens, because that’s where we are, says, well, we need the tokens for our own economy. You can’t have any, or we’re going to throttle you back. And, by the way, if you’d like some, you need to hand over ownership of ASML, your lithography company. I mean, this is just a cue for you to talk about Europe in any way you’d like. But, I mean, there does seem to be—you know, we go through these phases of somewhat optimism—oh, Germany has changed its constitution to allow for more government borrowing, you know, there’s the Draghi Report. And then there are kind of periods of downgrading of expectations on Europe. And I’m kind of on a downgrade cycle right now, but where are you?
Rediker: Well, it’s funny, because we’ve had this conversation on the various panels I’ve done with you over many years. And whether it was AI specific, as your question was this time, or any other catalyst, it’s really the same story. It’s just accelerated. But, I mean, Europe famously acts only in a time of crisis. And yet, it fails to recognize a crisis when it is in front of their face. So even before AI, Europe has all of the makings watching this Trump administration actually accelerate the need for Europe to say, well, whether it’s what Macron calls strategic autonomy, whether it’s defense specific, whether it’s on a variety of other areas, you know, Europe really needs to step up and do stuff—Draghi Report, Letta Report.
I always say this, I feel it’s incumbent upon me to do it every time I’m up here, to say the IMF did a paper in 2023 which showed that the euro area can issue 15 percent of euro area GDP through a common borrowing instrument with absolutely no incremental cost to Europeans whatsoever, just because of the added liquidity. That was before Trump made the U.S. dollar and the financial system even more of a there is no alternative but please give us an alternative scenario. Europe does not actually step up and do it. Whether it is pre-AI or post-AI. They seem incapable of stepping up and declaring a crisis that catalyzes the political necessity to act when, in fact, those of us taking a step back would say, this is a crisis.
On AI specific, it’s actually worse than your question suggests, although I can’t say that the paper is right or wrong. I haven’t read that specific paper. But if you look at where AI is going, it is really now, on the back of the Trump administration’s export controls on the Anthropic models this past week, this is a real crisis, because now the U.S. has said, you know, we have the ability—by the way, even if they were to remove the export controls today, the crisis is already there. Because what they’ve said is, we have the ability, at a whim, to turn off your access to one of the most fundamental building blocks of the twenty-first century economy, right? So some of the G-7 leaders actually acknowledged that this week, and they said we have to step back and realize what this means. Well, that’s what it means.
It means you have access to a U.S. set of models, which can be turned on and off at will. Again, there’s always been the risk of a kill switch in some of our defense provision to Europe. Well, this is a blatant kill switch, without even, you know, any ambiguity whatsoever. Then you’ve got the Chinese models which, again, depending on who you talk to, are six months behind, eighteen months behind, whatever it is. But they are the alternatives. And nobody really wants to go to the Chinese model, except there is an appeal to an open-source model, which some of the Chinese models are. Not all Chinese, and not all open source are Chinese. But an open-source model means you can download it on your, you know, local computer. And then if the U.S. decides to execute the kill switch, via export controls or otherwise, then, you know, you have something you can still work with.
That makes these alternatives more palatable, even if on their face they aren’t. But note, I said there’s the U.S. models, there’s the open-source, primarily Chinese models. What I didn’t say is there are the European homegrown models. So the Europeans have Mistral, the French model. And I was mentioning this to you earlier in the lunchroom, Sebastian, that I learned this week Mistral, the French European great hope, is, you know, much smaller than its American or Chinese competitors. But Mistral trains its model in—dramatic pause—the U.S., because European copyright and data protection laws are so overbearing, from their perspective, that they can’t take the risk of actually training their own competitive model in their own EU regulatory environment without running the risk of legal jeopardy, which would undermine the entire exercise in the first place.
So I am hoping that Europe wakes up. I’m hoping that the Trump administration’s policies have spurred them to recognize a crisis when it’s in front of their face. But I’m skeptical that that is going to actually be the moment that we’re watching right now.
Mallaby: OK. Who has the question? And please remember, this is on the record. Yes.
Rediker: It’s on the record? I didn’t—(laughter)—
Mallaby: I didn’t want to tell you, Doug, but I’m telling—I’m telling the members.
Question: First, I so applaud your Knicks colors. Very well done.
Sarin: Ah, Knicks in five.
Question: Yeah. (Laughter.) You guys are all in the—if I go back up to 30,000 feet, we’re all in the business of forecasting. And it struck me, as you were talking, that, you know, all the sort of big risk that people used to look at came from outside. But with the apparent end of rules-based order, et cetera, you know, the risks are actually quite home grown. Like, a lot of them around turmoil in the world came straight from Washington. And even some on the positive side, the AI changes are coming from the U.S. We’re used to looking at the black swan coming from outside, and now it seems to be coming from our home. Can you talk about how or whether you’ve changed the way you actually forecast and think about it as a result?
Mallaby: Hmm. Who wants a crack at that?
Hatzius: I can take a take a stab at it. I would say—well, if I go back to the kind of pre-2008 period—I basically started doing this in the late 1990s—then I actually thought a lot of the risks were coming from inside the financial system and inside, kind of, the—you know, the endogenous feedback in the economy, or between the economy and the financial system, because of the weakness of private sector balance sheets and the private sector financial deficits, and the over-leverage. So that’s—you know, that’s less—obviously been much less true since the financial crisis, or maybe since the aftershocks of the financial crisis.
Right now I would say, yeah, the U.S.—so if you define inside and outside as domestic and foreign, from a U.S. perspective, I’d say you’re right. A lot of the kind of shocks in one direction or the other—and, you know, could be negative shocks for activity, could be positive shocks for AI—does come from U.S. policy. So figuring out what’s happening with U.S. policy, and what happens in the U.S., and what the impact is on other economies along the lines of what Doug was talking about in terms of European access to AI, is more important.
I mean, one thing that we’ve definitely beefed up is, you know, our Washington Economist Alec Phillips has a much more prominent role in the team than, you know, would have been the case ten years ago. I mean, he’s always been great, but he now is, I think, much more top of mind for people than was the case. And if I look at Goldman Sachs as a whole, not just investment research, which I oversee, but also the other areas where people think about macro outcomes, yeah, we’ve definitely beefed up the amount of resources that we’re devoting to figuring out what’s happening politically in the U.S.
Mallaby: Sure, did you want to go quickly? I want to get more.
Sarin: Oh, I was just going to say that I feel like there’s needed to be a certain nimbleness in forecasting. In that, like, the effective tariff rate has changed eighty-five times over the course of—on eighty-five different days in this administration. Flip side, I think some of the sort of, like, volatility and the threats to institutions that we’ve seen over the course of the first year and a half have longer-term consequences for things like dollar dominance, and the safety of the U.S., and our stability. And I worry that we are—we, at Budget Lab but more generally, are not doing the world’s best job at thinking about how to contextualize some of those consequences in the long term.
Rediker: Just you asked, how are things changing? I travel a lot less because I’m based in Washington. And so I used to be traveling the world to sort of assess the risks. And now I just stay home. (Laughter.) And my background, as Sebastian mentioned, was originally in emerging markets. So I have the curse of living in Washington and having an emerging markets prism through which to view the events of the day.
Mallaby: Yeah, I asked you about which emerging markets are in trouble, and you basically answered about the U.S. (Laughter.) Let’s go to a question here.
Question: Thank you.
Question to you. Sebastian mentioned that there were two issues a little bit in tension with each other, growth in CAPEX and productivity growth. My question to you is, markets. Some of us are old enough to remember 1998-1999, valuations going through the roof. There was no NPV that could justify these valuations. Are we in the same situation today with the valuations of companies going public? Especially since a week from today SpaceX will be integrated in the indices. And so this is going to proliferate all over the market and the markets.
Hatzius: I mean, I look at the macro side rather than, you know, individual companies. I mean, clearly valuations are very high. If you look at it on a backward-looking basis, the, you know, Shiller PE now is not too far from where it was in, you know, 1999 or early 2000. So, you know, it’s a very highly valued market. I’d say one big difference, relative to the late ’90s, is the strength of earnings. Because if you looked in in the late 1990s, certainly in the national income account measures, profits actually started to come down in 1998-1999.
So by the time the market peaked, profits were already on a downward trend. Margins were on a downward trend. It wasn’t quite as obvious in the S&P 500 numbers until later, and there were also some restatements. But right now, the—I mean, earnings growth has been—you know, it’s been quite dramatic. You know, first quarter we had S&P 500 earnings per share growth. You know, I think it ended up in the high twenties. For the median S&P company, it’s, like, in the mid-teens. And as far as you can—you can make out from the national income account measures, there’s also strong growth there. So I do think that’s one difference.
Another difference is that the financial imbalances, which became even more severe once you got into the sort of 2006-2007 period, they were already pretty significant in 1999 and 2000. The private sector was running a financial deficit of 5 or 6 percent of GDP. The, you know, amount of—well, net borrowing in the corporate sector, a lot of this was concentrated in the corporate sector with the telecoms boom/ We’re not seeing anything like that at the moment. So I would say we have, you know, a very highly valued market, but at the moment still strong profit growth and, therefore, I think still some support for—you know, for where markets are going to be. But I would watch, you know, margins very closely. I’d watch things like, you know, aggregate financing gaps in the economy and private sector balances very closely, because this could change.
Mallaby: But I wonder whether, you know, some of the, you know, more comforting metrics on the public market reflect the fact that since 1999 the private markets have grown out of all recognition, both on the debt side and the equity side. And so a lot of what might be—what might have dragged down the numbers in the past, you don’t see it anymore because it’s all private. So if you look at, you know, OpenAI’s valuation in the secondary market, it’s way below its last valuation in the round that it did. If you look at the private credit runs that have been going on—I don’t know, I feel like there may be similar problems, but they’re just less visible because they’re not public.
Sarin: I similarly share that view. And I’ll say, like, if you look at these sort of three mega IPOs—SpaceX already, and Anthropic and OpenAI on the horizon—taken together they exceed by almost two times the size of the IPOs in the internet bubble from 1995 to around 2001. And when you look at analysts who are talking about these companies, what they say is that essentially the way that the valuations can be justified is as if you are pricing in something like a monopoly or an oligopoly outcome for these firms. That we’re in this sort of winner take all race, and at the end these are the firm standing such that everyone is paying for their models, and they’re the best models, and all of that.
Something that seems hard for me to grapple with is I haven’t quite understood that sort of revenue case for these firms, ultimately. In that it strikes me that these are pretty competitive markets at the moment. And it’s not obvious to me that there is going to necessarily be that—it seems likely to me that you’re going to be in a world in which there are going to be a few very dominant players. It seems likely to me that some of us are going to need access to the very best LLMs. But also, some of us and the way we use AI isn’t necessarily going to be willing to pay for those types of LLMs, and it’s going to be fine with a substitute that’s almost as good, but much cheaper. And so in that sense, like, how do we justify the valuations that seem monopolistic or oligopolistic in a world that—in a market that feels more competitive than that? It seems an open question to me that I don’t quite know the answer to.
Mallaby: Hmm. Let’s go over there.
Question: You shouldn’t need a microphone in this small a room?
But actually, the most interesting question that your fabulous session has come up is a military one, which is chokepoints and the future of the military. But I’m going to let that go, because that’s not your subject. I’m stunned by all the things that not only don’t we know, we’re not doing research on. That we have very unusual financing mechanisms between the users and the developers of these big models. How do you analyze those? Where are they going? You have CAPEX spending, which is doing what traditionally was an operations function. It’s ultimately going to train models, not so much to build capital stuff. A lot of talk about open source. We think open-source software is much more vulnerable to hacking and to being used in a hacking. But we don’t really know. I mean, why isn’t there—
Mallaby: So the question is?
Question: Why isn’t there more relevant research going on in this?
Mallaby: What’s—yeah.
Question: Even the definition of productivity?
Mallaby: I feel like Natasha might have some research on this.
Sarin: (Laughs.) But it’s hard. I think part of the challenge is that we’ve all been circling around the sort of, like, known unknowns, but the unknown unknowns are much more challenging. In that it becomes—one of the things that I’ve been grappling with is I’ve been thinking a lot about private credit. And I’ve been trying to understand the extent to which we—private credit, and a potential sort of bubble collapsing where you start to see a lot of the AI investments that these firms are funding through private credit vehicles start to go bad. And so what does that actually mean from a financial stability perspective?
And the answer to that question is, it is just genuinely really hard to know, because if you look about—if you look at public credit markets, things like bank loans, for example, we have, like, a lot of information about the extent to which there are interconnectivities between different types of financial institutions, about who ultimately holds the bag, about how to think about the ratings that are coming on these types of debt instruments. Whereas, for private credit we just know much less. In part because the word “private” is in the name, right?
That it’s sort of, like, in fact, there—I mean, there is some information, and Apollo is trying to be really transparent, and all this stuff, but, like, the extent of knowledge that we have about the actual sort of relationships, and the exposures, and where they lie in our financial system, are much weaker than more developed public credit markets. And so all that is, like, a long-winded way of saying, I feel like part of my challenge as an economist trying to analyze in real time a lot of these developments is I don’t necessarily have all the information or tools that I might want in order to be able to do that important work.
Mallaby: Perfect, and economists—
Sarin: And therefore, we need to make a lot of assumptions and do our best, which is what we are trying to do.
Question: (Off mic)—our research focus.
Sarin: And our research focus needs to be well-informed.
Mallaby: OK. Let’s go to another question. Let’s go over here to Earl Carr.
Question: Thank you. Earl Carr, CJPA Global Advisors. Thank you, Sebastian, for a fascinating discussion.
Given chokepoints, and given that countries will be focusing more on internal renewable energy-type sources, do you think that one of the byproducts of the Iraq war—the war in Iran will be that the U.S. petrodollar will decline and the internationalization of the renminbi will increase, given that global supply chains are so embedded with China.
Mallaby: Doug.
Rediker: I think the first part of your question, yes. The second, not so much. I think that the petrodollar foundation, from decades ago—which was U.S. provides security, the Gulf states agreed to denominate oil in dollars, and they recirculate those dollars into investments into the U.S.—I think that that is breaking, and going to continue to break. Whether the renminbi becomes the beneficiary of that is probably a vast overstatement. I’m not sure what comes next. I think the renminbi—certainly on the back of the chokepoints, the manifestation in the Iran war and elsewhere—clearly, the dollar’s use as a bilateral trading currency is being selectively undermined. It’s not being undermined overall as a reserve currency.
But if and when the Gulf states, at the end of this sixty-day period, and then thereafter, find that U.S. security ain’t what it was cracked up to be, and that the U.S. withdraws in a way that they are not comfortable with, they’re going to have to rethink because they’re going to have to redeploy their assets in a different way. It makes more sense for them to be actually denominating oil in a trade-weighted basket of currencies, just from an objective standpoint, if you didn’t have this implicit/explicit petrodollar deal, so if that if the security part of that becomes more questionable, then all of the other steps become, you know, a little more vulnerable to change. But I don’t think the renminbi becomes the de facto beneficiary of oil priced in renminbi as opposed to dollars. I think it’s going to be something much more multipolar.
Mallaby: I wonder whether either Jan or Natasha wants to come in on this, because I’m always a bit confused by this petrodollar argument. In the sense that if you ask the question, why do people hold dollars? Is it because defense comes to the United States was it because they want to hold dollar assets because the U.S. economy generates very attractive, very liquid, very high-growth, you know, technology-connected opportunities to invest? I would say it’s overwhelmingly the second. People hold dollars because they want to buy stuff with dollars, assets in dollars, because U.S. capital markets are so attractive. It’s not really about who provides the Patriot missiles.
Sarin: I very much agree with that. And, by the way, look at how foreign investment in China has, like, basically collapsed in recent years. No, that is another argument against this.
Mallaby: Yeah. I mean, China has great technology, but it also has overcapacity in all these technologies. So if you’re a shareholder, it’s not great to buy stocks in BYD. Yes, let’s go over there.
Question: I’d like to ask a naïve, hopefully not Luddite, question. About two and a half years ago, even the leaders, including Musk, of AI companies said they would entertain a six-month pause to allow the major economies, basically the U.S. and China, work out a regulatory regime. That seems like an awful long-ago world. But in your conversation today you touched on two things which remind us of the need, or perhaps the need, for state control over the new technology. One was the recent kill switch issue that just arose this week, and the other was back when Hegseth and Anthropic came to blows back then. You, Sebastian, mentioned a six month or so, or eighteen-month lead over China’s technology. Obviously restraint here—and there’s also the states in our federal system, which also threaten to regulate. The question then is, is there still an opportunity for the major powers to slow the development to establish a common regulatory regime? Or is that such an idealistic, fantastical vision at this point that none of you forecasters have to take it into account at all?
Sarin: I worry that it is a little bit of an idealistic vision, in the following way. If you take seriously what some of these executives at these leading labs are saying—and they’re saying different things at different moments in different audiences at the moment. Where you had Sam Altman on CNBC recently say he doesn’t think AGI is all that useful of a term, and maybe that’s—sort of, there’s going to be this, like, stepwise accumulation of improvements with respect to super intelligence, but it’s not going to happen all at once. But I think that was sort of more trying to calm all of us, the public, down. I think if you take seriously what they’re saying, it’s possible that by, you know, 2028-2029, we are going to live in a world of recursive self-improvement, or a world of essentially AGI.
If that’s true, what that means, if you want some sort of coordinated global regulatory framework sort of policy perspective, is that in this administration—in this one, in the Trump administration—
Question: He’s already two years in. So let’s put that aside. But that’s—then we’re further behind the ball then, but go ahead.
Sarin: Yeah, but, I mean, what I’m saying a little bit is I see no capacity for the policy process to ultimately find its way to regulate successfully or coordinate across countries successfully, based on my belief about how policy ultimately tends to work and how far behind it is relative to understanding this technology and the way in which you might even want to regulate it.
And then, second fact is I don’t actually know what a pause means. Like, I actually—like, structurally, even if we want—I know they were all saying it, but, like, even if we wanted to do that, how would one effectuate it? And how would one effectuate it in a world, if what Sebastian is saying is true that China is sort of eighteen—we have an eighteen-month lead. Then you’re going to say we’re going to pause and then we’re going to lose some of that ground, and to do what exactly? Because we don’t have a cogent policy process for what it would mean, like, even to regulate successfully. And so all that makes me—like, I think you need it, but I don’t know how you get it. And I’m quite nervous.
Mallaby: Yeah, OK. Shall we take one last question? Right here in the front.
Question: Thank you. My name is Vanessa (sp). I’m a corporate member.
It’s about unemployment, and what are your views about that. But I just want to make a comment on that. Is it’s really hard to price. And, like you said, there’s not enough evidence in the data. But one thing we know is the technology capabilities is so much higher than the business is actually being able to implement, enable. The gap just keeps growing. And we know there’s a bunch of resistances that make that happen. Once that happened, that feeling you had about your research assistant can be a huge effect, because when you look at end-to-end processes in a corporation involves a gigantic number of people. (Laughs.) So having said that, what is your perspective in terms of the unemployment rate for the upcoming years?
Mallaby: OK. We’re going to answer this in a rapid-fire way, because we’re just about out of time. But so we’re going to frame this as unemployment, let’s say, three years from now, 2029. What’s your number, Natasha?
Sarin: Four, five—four and a half, five (percent).
Mallaby: Doug,
REDIKER: I’m not going to answer that question. I’m going to answer it differently. I’m just going to say, I think it’s going to be a big problem. And I think there’s going to be a political reaction.
Mallaby: You speak rapidly, even if it wasn’t rapid fire.
Sarin: Yeah. You made me give you a number. (Laughter.) He just goes, like, eh, on the one hand, on the other hand. (Laughter.)
REDIKER: I’ll just say, I think there’s going to be a political reaction to the populist need to address a rising unemployment that is currently taken with a bit of whistling through the graveyard.
Mallaby: OK. Four and a half, five (percent). Populism is a problem. (Laughter.) Jan.
Hatzius: I have a similar, Four and a half, four and three quarters (percent).
Mallaby: OK. All right. There we will end it. Thank you, everyone, for coming. Thank you to all the speakers. (Applause.) It was fun.
Sebastian Mallaby
Paul A. Volcker Senior Fellow for International Economics, Council on Foreign Relations; Co-host, The Spillover; Author, The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence
Jan Hatzius
Chief Economist and Head of Global Investment Research, Goldman Sachs Group, Inc.
Natasha Sarin
Professor of Law and Cofounder, Budget Lab, Yale Law School; CFR Term Member
Douglas Rediker
Managing Partner, International Capital Strategies; Nonresident Senior Fellow, Brookings Institution; CFR Member