2026 Financial Markets Conference – Policy Session 4 Transcript – May 19, 2026

May 19, 2026

Policy Session 4: How Will AI Transform Financial Markets?

The increasing reliance on intelligent algorithms is rapidly changing financial systems. This panel explored the impact of AI (artificial intelligence) and machine learning on financial markets, highlighting key opportunities and challenges for trading, risk management, regulatory frameworks, and central banking policy. The discussion addressed such critical questions as: What are the effects of AI model adoption on overall market stability and systemic risk? How do we balance market efficiency and price discovery with potential volatility? Could common AI models drive new forms of herding?

Transcript

Robin Lumsdaine: Hi everyone. I hope everyone had a lovely lunch. Welcome to our session on: How will AI transform financial markets? I'm Robin Lumsdaine. I'm a professor at American University, at the Kogod School of Business. I'm also on the Advisory Scientific Committee of the European Systemic Risk Board, and I've told my fellow panelists that I always disclaim that anything I say is just reflecting my own views.

I think one of the reasons that I'm moderating this panel is that—together with Stephen Cecchetti, Tuomas Peltonen, and Antonio Sanchez Serrano—we at the ASC published a report in December on AI and systemic risk. And so, some of the discussion that we'll have and some of the things that I'll point out will refer to that document.

It's an absolute pleasure to be here. I think everyone will agree, we've had just some amazing discussions and presentations, both formally and informally. And so, it's nice to have the opportunity to be moderating the last session.

Our panel today is going to consider the impact of artificial intelligence and machine learning on financial markets. It's going to highlight opportunities and challenges for trading, risk management, regulation, and central bank policy. And among the questions we're going to address are questions like: How do we balance market efficiency and price discovery with potential volatility? And also, could common AI models drive new forms of herding, an investor behavior?

Just to give you a sense of the schedule, I'm going to briefly introduce each panelist. Each panelist will then have ten or so minutes to do a brief presentation and provide some opening remarks. I'll ask some questions, and then we'll open the Q&A up to the floor. Don't forget to scan the QR codes that are on your tables in order to submit questions, and I'll do my best to summarize and organize them so that we can have a good discussion.

So, joining us online today is Professor Thierry Foucault, who's a professor of finance at HEC Paris, and a research fellow of the Center for Economic Policy. His research focuses on the determinants of financial markets' liquidity, the production of information in these markets, their industrial organization, and their effect on the real economy. In 2021, he received a grant from the European Research Council to work on the effects of AI and big data on information production and financial markets.

He's currently co-managing editor of the Journal of Financial and Quantitative Analysis and has been on the board of practically every other journal that I can think of. And, he's the co-author of a textbook on market liquidity. He will discuss the role of how AI is reshaping information in financial markets.

In the room with us today are Paul Kedrosky, a venture capitalist with an interest in software and AI, and a research fellow at MIT's Institute for the Digital Economy, where he focuses on AI and the future of work. Paul was formerly the director of an early-age commercialization center at the University of California, San Diego, and a senior fellow at the Kauffman Foundation. He's been a regular contributor at Bloomberg and CNBC, as well as having written for The New York Times, The Economist, Wall Street Journal, and been interviewed by many, many news organizations. He's going to talk about the use of large language models and the implications for herding, investor behavior, and financial stability.

Hariom Tatsat currently works as a director in the risk AI division at Barclays [Investment Bank]. He has extensive experience in the areas of financial instrument pricing and algorithmic trading in several global investment banks and financial organizations. He's the co-author of a book titled Machine Learning and Data Science Blueprints for Finance, and he's going to discuss vibe coding and the costs and benefits of generative AI.

So, without further ado, I'm going to turn it over to our first speaker, who's going to be Thierry online. Nice to see you, Thierry. Thank you.

Thierry Foucault: Thanks, Robin. Thanks for the nice introduction. Let me share my screen. So, let me start with a few thoughts about why artificial intelligence is going to impact the financial industry, and the benefit and the potential cost.

So, one possible angle to think about what the financial industry is doing, and why it is important for the financial industry, is the idea that one, maybe the, raison d'être of the financial industry is to produce financial information. This is an old idea. There is an interesting paper by the economist Ross Levine, published in the Journal of Economic Literature in 1997, where he surveys the literature connecting financial markets to economic development, and he says the function of financial markets and the function of financial institutions is to improve—to overcome—information frictions.

And so, if we take this view of the financial industry and what the financial industry is doing, it's not surprising that AI as a technology is going to have a massive impact on this industry. AI is a technology to produce information, to transform data into predictions and decisions by leveraging big data and powerful algorithms to extract, to mine the data, to extract information from the data. So, the interesting thing is to think about other technologies changing the way information, financial information, is produced, the way it is used.

And maybe due to the nature of financial information, there is an obvious effect which is, we are replacing human judgment—for instance, human forecasters—with machines. So, this may change the way predictions are done and formulated and the nature of predictions.

And there is another effect, which is also a massive effect, which is AI is reducing the cost of producing financial information. So, given that, what are the costs and benefits?

The benefits are sort of related and obvious, and related to the idea that we have this decline in the cost of producing financial information. Potentially, that can lead to improved re-sharing between market participants. So, this is going to reduce the cost of capital for firms. A second benefit is that this allows investors to get access to more precise signals about future returns and firms' cash flows, and so, this is going to make prices more informative.

And a third benefit is that this allows banks and venture capitalists, for instance, to better screen investment projects, and to monitor those projects, so this leads to a more efficient capital allocation. So, these are benefits which are linked to the decline in the cost of producing financial information.

Now, what are the risks? So, the risks I'm listing here are very much based on my research in this area. These are not the typical risks that people may mention, especially the very first one. This is what I call a "misallocation of information production capacity." Yes, we have a more powerful technology to produce information, but the question is whether we use the technology to produce the type of information that is really useful for society as a whole. I will come back to why I think this may not always be the case.

The second reason that I see is that instead of leveling the playing field, the use of AI by stock market participants may increase information asymmetries and may also increase the scope for market power, which may reduce liquidity. And the last risk I would like to discuss briefly is that, it's difficult to understand the way AI-powered agents behave, ultimately. And especially in the context of in the financial markets, I think, in the context of market interaction. And as a result, it's difficult to design markets or to write rules so as to try to mitigate potential risk with those agents. I'm going to come back to that.

So, regarding what I call "misallocation of information production capacity," I think there are at least three or four different types of risk here. The first risk is that many top trading firms, for instance, in the market are going to use AI to extract information from market data. A very, very important piece of information in the market is the market itself. Every day, there is a huge amount of trading on various trading platforms, and this generates a huge amount of data; and these data are informative about future returns, future cash flows. And so you can try to design trading strategies by using information in this data.

But market data contain information only to the extent that some people have produced more fundamental information in the first place. The reason why market data contain information is because some fund managers are trading on the information that they have produced about the firms' prospects, for instance.

So here, the risk is that by extracting information from market data, top trading firms are reducing the profit from producing more fundamental information, sort of "freeriding" on this production. And as a result, we may see less production of fundamental information. So, that's the first risk.

The second risk is related to that. There is a trade-off between trading fast on imprecise information, or trading more slowly on precise information. The use of algorithm to trade very quickly, for instance, on news or microeconomic announcement, is reducing the cost of trading fast on imprecise information—which, in some sense, is good. That's going to make markets more information-efficient.

The information—the announcement, for instance—is going to be reflected more quickly in the prices, but there is a curse: that makes the benefit of trading slowly on more precise information, more precise signals, lower. So that means, again, that people may decide to produce less precise information about future cash flows, which could make securities prices less informative.

The third risk I would like to mention has to do with the nature of the information that is produced. I have studied—a lot—the literature on whether there is information (so-called alternative data) or not, to predict returns, to predict cash flows. And there is information, but the literature that I know has shown that there is short-term information—that is, you can forecast next quarter earnings, maybe next year earnings, for a firm, but beyond that you don't have much predictive power. That will not be a problem, per se, if this does not lead to a decline in people's efforts to produce long-term information.

And I have published recently a paper in the Journal of Finance, with Olivier Dessaint and Laurent Frésard, where we show that over the long run, the quality of short-term forecasts for equity analysts in the US has improved. But the quality of their long-term forecast—let's say, forecast of earnings in two years from now, three years from now, four years from now—has declined over time.

This is what the graphics on my slides are showing here, and we relate this evolution to the rise of alternative data. We show that there is a connection between new alternative data about a stock, for instance, and the quality of equity analyst forecast on the stock. This raises the possibility that maybe the market is becoming less good at forecasting the very long-term, because the market is relying more and more on data which is short-term-oriented.

Now, let me move to the next piece that I was mentioning before, which is AI. Using AI tools requires a massive investment in infrastructure, computing power, for instance, and human capital. You need the data analyst to be able to use the tools from AI. This may lead to a situation where, instead of decreasing information asymmetries, we are increasing information asymmetries—especially because due to fixed costs—due to the fact that those investments require massive fixed costs, you may have a first mover advantage for those who are making those investments. [This] making entry of other investors in this area more difficult, leading to larger informational runs and larger market power.

And let me finish with the last, final risk, which I call unintended consequences and unpredictability. There is the concern that self-learning algorithms may learn to sustain, for instance, non-competitive prices, or may learn to manipulate market prices, even though they are not explicitly coded to do so. That is, this is more the logic by which they learn to behave that could lead to those unintended consequences than a deliberate behavior of the people who are using those algorithms.

This is a concern that has been raised, for instance, by regulators in product market, and there is some academic evidence supporting this concern. In a recent paper with colleagues at HEC [Paris], we have shown that this concern may be valid for financial markets as well.

We have considered an environment experiment where we let algorithms set prices for an asset, and we showed that those algorithms, even though they learn how to behave in this environment, which is quite complex. They typically don't reach the competitive outcome. They tend to stop learning on prices, which are not the most competitive, which goes in the direction of this concern that has been raised in the product market.

And even though it's very easy to describe what the algorithms are doing, we find it very difficult to predict the behavior of the algorithm and to explain their behavior. And I think this is a challenge for policymakers and for policymaking in general, because if you cannot predict the behavior of a new policy or a change in market design, then policymaking is going to be more difficult.

So, that's what I wanted to share with you at the start.

Lumsdaine: Thank you, Thierry. Our next speaker is Paul.

Paul Kedrosky: So, I was just going to highlight a couple of research programs that we're working on—and this is purely experimental stuff that's good fun, and I thought you might find interesting. So, this initial slide is based on: there was a recent paper in Nature and Nature Medicine about how using ChatGPT Health has a tendency to over triage relatively trivial cases and under triage cases that are more critical—which, I think we would mostly think is probably a fairly bad idea—[And] that it has this tendency to throw batteries of tests at cases where it's fairly obvious what the problem is, and then not take it seriously in some of the more extreme cases.

And this is something that, if you think about it in the context of the architecture of large language models, is very much built into the nature of large language models, unsurprisingly, because it's trained on this broad corpus of data that's based on what you and I think. And we generally do some of the worst jobs we do, in terms of prediction, at inflections and at extreme points in the distribution.

So, what we did is we took the same idea and began applying it in the context of financial markets and running a simulation across seven test cases—thinly disguised versions of some major financial inflections over the last 100 years dating back to the Great Depression, and then coming all the way forward to the Global Financial Crisis, to some of the more recent episodes since 2022 with the rate scares and what have you—and looking at how models performed in this giant simulation. Multiple cases with the temperature turned up, in terms of introducing more volatility and variability and see how models performed. And you see almost exactly the same thing that you see in the context of ChatGPT Health, which is to say the models, generally speaking, are fairly sanguine about crises: "This will all work out; everything's going to be fine." They tend to over-triage trivial cases, where there's really not much going on.

But they feel like it's almost—there's some classic papers on this, which I'm sure many of you have read, on action bias among soccer goalkeepers, this idea that soccer goalkeepers tend to move too much. They tend to dive too much to one side or the other when they'd be better to stand in the middle of the goal. And the reason is, of course, because standing still looks like you're not taking the problem very seriously. So, they dive to one side or the other when they'd be better to just do this.

And so, the same thing is shown in the training data, which is then, of course, reflected in turn by what we see in the context of this tendency, in terms of the simulations of large language models to under-triage when they should be taking things more seriously and over-triage when they shouldn't be nearly as concerned. So, that's an example of the kinds of things we're working on.

And again, as a large language—and I'll come to this in a second simulation that we're doing—that as models become a larger and larger presence in the market. Obviously this becomes more consequential because it's no longer just reflective of what's actually happening. It, in turn, is reflexive and actually influencing what's happening, because they're much more active participants.

So, that's one example of the kinds of things we're working on. The second one is the role of—and I often, when I'm talking about this stuff, ask people if they know who like Humphrey Jonas is. Anyone know who Humphrey Jonas is? Erika Stahlberg? Legends in FinTok. So, financial-related social media—FinTok, as it's euphemistically called—is gigantic, right around five billion views a year. To the point that the SEC (Securities and Exchange Commission) twice this year had to issue cautions about the recommendations coming that, "No, you can't write off your kids as employees." "No, you can't... " there was another one that they came up with.

But these things, they might seem kind of ephemeral and frivolous, but there are two things going on under the hood. One is that it's increasingly algorithmic, because it's driven, in turn, by what gets traction and attention. But there's a secondary component that's emerging now where it's not clear if many of these FinTok so-called influencers are actually people at all. They're often just AI-generated content, in turn, reflecting what's getting traction elsewhere, but then, in turn, they're influencing peoples' actions.

There was a recent story about how Nike stores are increasingly overrun by people showing up with two-year-old shoes. And it's like, "Well, boy, what's going on there?" And the problem is, of course, that on FinTok—and this was an AI-driven thing, where it said one of the algorithms had been reading some kind of Nike warranty that said that if there's a manufacturer's defect within two years, you can take the shoes back. And so, of course, people are like, "I've got all kinds of two-year-old Nike shoes. I'm going to bring those back. This sounds like a tremendous deal."

But at scale—these things seem frivolous and inconsequential until you realize the scale of what's happening, in the context of this FinTok social media algorithmic-and-AI-driven force in the marketplace, to the point that if you're under 35, people are saying now they're getting about 50 percent of their financial information from financial social media, which is pretty remarkable. I'm not saying that's better or worse than CNBC, but it's nevertheless the case. Maybe it's an improvement; I don't know.

So, this is something that we're really focused on, just because this whole notion of the emergence of synchronization and social media as a kind of coupled oscillator, in the context of capital markets, is really important. But it's going to the next stage with AI, because it's taking humans out of the loop. And it's increasingly responding directly to what's happening and the influence it's having and generating new content on that basis. And that's where you get into some of these phenomena that we increasingly see happening that were the air pockets underneath stocks and these massive, short squeezes on highly liquid stocks.

There's an issue recently with DXYZ, which is this holding company for a bunch of private company names, that was hit in this exact same algorithmic way. So, this notion of synchronization and how it emerges is really important. So, I'll leave you with one other thought.

This is another model that we're running. I tried to make an animated version work, but it didn't, but I'll just sort of walk you through it. So, the notion is, we're trying to model—as the markets become increasingly dominated, even at the margin, by algorithms—by AI specifically, and in turn, AIs that are adaptive, that are responding to where they see alpha, whether human-generated or AI-generated. What are some of the tipping points in terms of the market becoming almost a coupled oscillator, in the sense that the market is increasingly synchronized with itself?

So, the market becomes—you think about it like Galloping Gertie, the bridge that blew up in a high wind over Tacoma Narrows but, that notional idea that this is a dynamic system increasingly exposed to coupled oscillators. And what are the trigger points? Because it feels like it's not particularly consequential now, because AI has a relatively small presence in terms of algorithms. But what we show is that it actually takes a modest number, only about 6 percent of the market (even less, actually, than that), before you have rapid convergence towards a state where there is convergence. And you have a series of rotating crises in the market.

So, a relatively small presence of AIs that are actually observing everything that's happening, because they're no longer just observing locally, they're observing globally. They're observing what's happening across entire markets and doing so in real time. And I'll give you an example of that happening right now, which is happening in a market that's not particularly, I will say, transparent, but it's in prediction markets.

So, in prediction markets right now, this is happening as we speak. Many of the latest, almost genetic algorithms that are AI driven, vibe coded inside of Kalshi and Polymarket, are people who are just in a sense iterating in real time but also using adaptive models to generate new models.

And as quickly as opportunities emerge to generate some alpha, they disappear. And the case in point recently was, the Wall Street Journal ran a piece about how there was this weird anomaly in prediction markets. It was kind of a variant of the favorite long shot bias. Did people see this? It was really interesting about how, it was a kind of salience bias, favorite longshot bias in Polymarket and in Kalshi, where you can bet on things like, will—pick your favorite Defense secretary; well, what Pete Hegseth will say in the next press conference. Will he say any of the following five words? And so, they lay out each of the five words, and there's betting markets and also there's all of those things.

And those seem like uniquely ridiculous things to bet on, I'm sure, but the notion is that people over-bet things they've heard of, and that's the salience bias in practice. This idea that because I've heard that term, I'm more likely to bet it.

Well, it turned out that by uniformly betting against people ever saying anything that's listed in a mention market, you actually generated significant alpha, because of the salience bias. People are calling it the "Doctor No" strategy: I just say no to everything. Just assume whatever the betting market says that somebody might say, assume there's far more things that they could say, and it's more likely one of those things will be the things they actually say.

So, that was generating alpha until literally about two hours after the Wall Street Journal issued a piece on it, and then there was an explosion of vibe-coded algorithms all showing up in the markets, frantically saying "no" to absolutely everything you could find on a prediction market. So, I think this is a microcosm of what's ahead, and we're trying to model how this might look in the broader context of capital markets. But you're seeing, in a very narrow playground sense, exactly how it's going to play out on prediction markets, right now in real time, given what we're seeing just from these sorts of trivial examples (like betting on what Pete Hegseth will say in the next press conference).

I'll leave it at that, but those are three examples of the kinds of things that we're thinking about in terms of how AI will interact with the modern financial markets and create these new phenomena of herding and contagion and coupling. So, thank you.

Lumsdaine: Okay; and next, we have Hariom.

Hariom Tatsat: Good afternoon, everyone. Thanks to the organizers for inviting me. My topic is: Does vibe banking work?

So, vibe coding was a term that came a couple of months back. What does it mean? It means basically you're coding in English. The focus is less on the verification, but the focus is more on the speed and the convenience.

The question is, does that work in finance? So, I'll give a practitioner's point of view, the evolution from predictive AI to generative AI, some of the risks that I've seen, and my view on how to mitigate them. One important thing is, the views are personal.

So, predictive AI: Around 2010, 2015, we had all these predictive AI models, and around that I wrote this book, Machine Learning Blueprints in Finance. Why was it trusted in finance? There was a significant advantage when it comes to the data. Predictive AI could process huge data, different variety of data, from numbers to text, and then we can use it for fraud detection, credit scoring, and whatnot. There was this reproducibility. If you run this experiment again, you will get the same results. That becomes very important in the context of generative AI, which I'll talk about later.

And then there was a bit of explainability as well. It was less compared to the statistical models that we had before, but still, explainability was there. And why explainability is important is because if you're using AI models to reject somebody's loan, you can't say that it's because my AI model said so. You need to give the person the reason. You need to mention that, hey, it was because of high debt, or limited credit history. So that's the reason explainability is really important when it comes to finance.

Now cut to 2022. We have generative AI, and, of course, it took off in finance. One of the things which is underrated is retrieval. In finance, we have a lot of documents which are in the form of PDFs. Those PDFs contain text, tables, charts, pictures. And now with AI, with good models, you can retrieve all the information, and that's available for automation or the AI usage.

And you can also connect to all these internal-external data sets, ask questions in English, and fetch the information. So, the retrieval has become really good. And of course, the generation, the obvious point about generation, you can generate text. Nowadays, you can generate good code, tables, charts, and that is something you can transform into PowerPoint presentations and reports, memos. That's one of the most important things that we do in finance.

So, what has happened over time, from predictive AI to generative AI? It was single-task model earlier: fraud detection, credit scoring. Now, it's across the workflow.

I'll talk about a couple of use cases that we have seen in banks. This is an example of pitch decks, the mergers and acquisitions investment banking. So, what you see on the right-hand side is the output, a typical slide that you have in the investment banking deck. You have not just text but also tables, charts, pictures.

And on the left-hand side—I talked about retrieval—you have the retrieval from internal data, which I mentioned has become much better. External data, you can ask questions and fetch the information; you can do web search. And in between, you have this AI, which can be a form of agents, it can be a form of workflow. You orchestrate it, create your harness, and then produce the output. The quality is getting better day by day. The adoption is increasing.

And the next example is of the buy side. There's a very good paper called "TradingAgents" from MIT and UCLA guys. It talks about multiple agents, but the broader picture is, there are a lot of functions in the hedge fund which go into creating a strategy that, to some extent, can be done by agents. For example, I talked about retrieval; There are analyst agents, which can pull the data—fundamental data, news data—and we have a researcher agent, which can do all the ideation around the data set, looking at the papers and the data set.

Then you have the trader agent, which decides how to trade, when to trade, how much to trade. Then we have the risk management agent, which can look at the risk of the portfolio, the strategy that you have built, followed by the fund manager agent, which does the execution. Now, as I mentioned, it's a paper. It was more like something which is ambitious, but day by day, as the models are getting better, the adoption across the board, all the functions here are increasing.

Moving on to the risk. Something we all know [is] hallucination. Now, in finance the hallucination becomes really important. It's not that we are writing, let's say, poetry, but we are generating numbers. And the impact of the numbers in all these texts can be disproportionate, can be much higher. It can impact your revenue. It can lead to losses. It can have reputational damage.

So, going back to the point that I mentioned on vibe banking. What is vibe banking? There, you are focusing on convenience, focusing on speed. There's less focus on the verification of the data—which needs to, at least for the high-stakes use cases, change to something I call agentic engineering. This is a phrase I borrowed from tech again. What it means is, you are baking in the verification during the generation process. The stress test of the output happens during the verification process itself, and it's done in such a way that the speed is not compromised.

So, we did some of the work around it by having a validation agent when we were building the pitch decks. The idea was that an agent compared the numbers against some of the alternative sources while the generation was happening. So, it is adversarial testing, before the action is being taken.

The next point, which in my opinion is underrated, is explainability. Every second day we see a model which is closer to AGI, and the trillions of parameters. The capacity is increasing exponentially, but our ability to understand them is not increasing exponentially, or at the same rate—and this becomes important in finance. If I go back to the previous point I mentioned about if you're rejecting a loan, you need to explain, and especially for the high-stakes use cases. So, explainability becomes really important.

And what we have as of now are prompting, chain of thought, some guardrails—but all of these look at the models or LLMs (large language models) from outside. There might be a need to look inside the model, look at how they behave, and maybe explain what it is producing as an output. So, this is where we did some work. It's a paper that we wrote called "Beyond the Black Box – Interpretability of LLMs in Finance". It's available in the public domain.

The idea was, can we explain what an LLM is producing? Can we go inside and see what's happening in terms of features, concepts, neurons. We borrowed a couple of ideas from the great work which is done by some of the tech companies—OpenAI, Anthropic, DeepMind—and tried to give a flavor of finance. So, in the paper there are a bunch of use cases related to trading, sentiment analysis, hallucination—and we discovered internal concepts.

And we didn't only discover it, but we also tweaked it, modified the concepts. So there's an illustrative picture on the right-hand side, which explains, in a way, that we were able to find out some of the features related to credit risk, and we were able to turn it up, modify the activation, and we ran some sentiment analysis or sentiment sentences or sentiment models. And, what we found out was the result was much better, more accurate, closer to what humans will annotate. And, there was no prompting, there was no fine tuning.

So, this becomes important. Although it's [in a] very early stage, the tools are coming up. But what it does beyond explaining and controlling, it also gives you a new possibility for high-stakes AI. You can have more use cases with more confidence.

And the final thought—I didn't want to sound preachy—but the last time, we thought we understood the models well, but we didn't. There was a significant loss to the economy, along the lines of like trillions of dollars, with finance, especially tied to the paper that I mentioned on explainability. I think it's a good time to go deeper into that, and we can't repeat that. So, thank you very much.

Lumsdaine: Thank you, Hariom. So, I'm going to ask a few questions of our panelists, and then we'll open it up to questions from the audience. Again, don't forget to submit your questions.

I guess my first question for you all is: so you've all touched a little bit on the role of information, both information as it's being developed in an AI context, but also how people are processing information, and the potential for information to be more correlated, potentially resulting in herding behavior. So, I was wondering, at some level, this issue of herding has existed for a while. We've used correlated—we all use similar credit models, etc. And so, it's a problem that's already existed, but we have this self-reinforcement or additional learning that comes in. And then we also, with AI, have the issue of an increasing amount of information being itself AI-generated.

And I just was wondering what your thoughts are on what the similarities and differences are, in terms of how information is coming in and being processed and what potential risks that imparts. And I guess—maybe Thierry—I guess maybe since you're online, why don't we start with you, and then we'll keep the order as we go.

Foucault: Yes, thanks. The only thing I have, I did not do research on that but my thinking is that, it's not completely clear that the use of alternative data and AI by fund managers, for instance—quant fund managers—is going to lead to more of this, because we have many more data sets than in the past. In fact, the diversity in the data that I was mentioning before, I'm doing some research on that, and we are using data from a digital platform, a digital marketplace for alternative data. There are more than 4,000 data sets available on this digital platform, of very different types.

So, unless fund managers choose the same data, I would expect that the diversity of the data set in fact is conducive to diversity of signals rather than homogeneity in the signals that are used—but that will work against uniformity, and so this is different from using the same algorithm. People may use the same algorithm and yet get different signals, because they start from different data sets. So that could be a source of heterogenetic fact in signals, and that could push away from herding.

There is another source of herding which was mentioned during the presentation, which is the algorithms are going to produce the data, to some extent. As algorithms are used to trade in the market, they become generators of data, market data, for instance. But that could generate herding potentially, because that becomes a bit self-referential, and so that could be a source of herding. Again, I didn't see research on that, at least in economics, but that could be a concern.

Lumsdaine: Thank you.

Kedrosky: So, this is something we've been spending a lot of time on. So, the question of herding, specifically—only modest amounts of inter-agent observability or visibility is required to introduce mass synchrony very, very quickly, as long as they're obviously chasing something material (which, in this case, in the context of capital markets, is obviously some form of alpha).

And so, what's obviously different than the old—there's the old Banerjee papers, or everything else in the area of herding here—is that the speed is vastly faster. And there's this notion of reflexivity that I can see the outcome of my action in real time and adapt consequentially as it happens. And so, you get much faster feedback and then the disappearance of the edge, and then you can get to these points of criticality, in terms of thinking about it in complex systems, very, very quickly.

And so, our own view is that this is a new source of fragility, in part because... there was a terrific piece in the FT (Financial Times)—maybe, I don't know, three or four months ago—by John Burn-Murdoch. Many of you may have seen it—[It] was talking about the slow amelioration of some of the polarization in politics happening in and through an unexpected channel. I don't know if folks saw it, but it's a tremendous piece, very data-driven, where he talks about how models, by their essence, are homeostasis and centering engines, if you will, where they're trying to drive people back towards the middle, even if that middle in practice no longer exists.

And what he showed was that it's pulling people from the left side of the spectrum more towards the middle, pulling people from the right—because it's the essence of this next token prediction, and the underlying data corpuses that are pulling people in a more centrist way. And so, we see the same phenomenon happening in capital markets where, again, as I was saying at the beginning, it's not particularly surprising that models aren't very good at dealing with the extremes of the distribution. Because they're obviously trained on people who aren't very good at dealing with the extremes of distribution, so you end up with the same thing that happened: the input becomes the output.

We were joking earlier. I was saying—I don't know if this is too off-color or whatever—but we were joking earlier that it was like one of the things I try to explain to people when they're trying to... . We treat LLMs as magic models, but you have to think about it in terms of: What's the preponderance of the data, where did it come from, and how can you best characterize that? Well, one of the largest training sets in large language models is Reddit. The average user on Reddit is a 36-year-old white male. And so, if someone tells you my large language model told me X, Y, Z, I say it's often useful to just mentally search and replace that and say, "Some thirty-six-year-old guy on Reddit told me." And sometimes that helps you put things in a more appropriate context.

And so, in a broad sense, I think that's a way of helping you understand the centering tendency of models, that they really desperately—statistically and probabilistic here—are trying to drive you towards some measure of central tendency.

Tatsat: Yes, that's a good way to phrase it: "a thirty-six-year-old guy said." So, I'll talk about predictive AI versus generative AI; with predictive AI, I guess the issue of herding was slightly less because we went inside the model. We understood a bit more. With generative AI, it becomes a bigger issue. Why? Because, if you're using models from the same vendors, it will hallucinate, and toxicity and all those things might be aggravated. So, the no-brainer here is to use the models from multiple vendors, especially if you're using it for high-stakes situations, high-stakes use cases.

And another thing would be, we should also maybe try open weight or open source models. There are a bunch of them from Google, from Nvidia, from OpenAI as well. Why? Because you can, I talked about explainability—you can also look inside the model and maybe fix a few things if it is not working in your favor. So, the herding is there, but using diverse models I guess might be a good solution.

Lumsdaine: Well, so let me follow up on that, because—and maybe we'll go to Paul first, because it relates to something that Paul mentioned, in terms of moving toward the center of things. So, at some level, market prices reflect not just information but heterogeneous perceptions of that information, and Thierry mentioned that as well. And so, as more and more information-based tasks are ceded to AI, do we run the risk that it's actually AI's perceptions that are being reflected in the market rather than our own, or actual market participants?

Kedrosky: I think it's a hard conclusion to avoid.

Lumsdaine: And what do we do about it if—

Kedrosky: Yes, right; because there's even been, I have a colleague who's in linguistics who has been talking about this all the time in the context of language, that he's seeing. It's not just that language is changing, because most essays that he's grading are being written by large language models. It's actually even in conversations with people; they're using tropes and mannerisms and expressions that used to once hide only in the dark recesses of large language models. And it's becoming much more visible because it's increasingly part of the vocabulary.

There's a new study out showing how regional vocabularies, in terms of how the internet itself works, are being ameliorated and disappearing for exactly that reason, and that LLMs are the forcing function driving that. And again, this is not a "good versus bad" conversation; it just is the pressure that's driving it.

And so, in any other context, this would be a reason for some concern, in the context of fragility, because we know that markets function best when there is heterogeneity. There'[re] diverse opinions, and people are taking different approaches to things. There's a reason why when I sell something, someone else is willing to buy it—because they have a completely different view than I do.

And so, this, I think, creates the opportunity. This is why we get into this coupled oscillator problem—it creates the opportunity for much more paroxysmal markets, and it doesn't help if you start reporting financial data on a less frequent basis. This introduces new opportunities for exactly this problem.

I don't have a good answer for what to do about it, but it's certainly, I think, a vastly misunderstood source of potential, I'll say volatility, euphemistically.

Lumsdaine: Hariom, did you have—

Tatsat: Yes, I kind of agree with that. I don't have a good answer as well, but if you're repeating, let's say, the output and you're using it from gen AI—if it's solving the purpose, at least, now, before you need another set of data set for more intense use cases, I feel it's fine so far. And in the future, I'm sure things might become stale, and that's when you might need more data set or more models. But so far, I guess it's working fine for a lot of people.

Kedrosky: But it's kind of like the debate—sorry, I wanted to say one more thing—it just, it's kind of like the debate about how important, going forward, synthetic data will be in the context of training future large language models, because you can think about that in finance and you can think about that in other domains. And one of the core tensions with respect to the introduction of an increasing fraction of synthetic data, in the context of training financial data, or training models writ large, is that it does introduce this new fragility.

We've seen multiple instances of model collapse in the context of larger percentages of synthetic data, and synthetic data is increasingly where most of the new data is coming from. For example, there's a firm called Poolside that's interesting to look at, and what it's trying to do is essentially replace an increasingly exhausted, Permian reservoir of data that we stumbled into called the public internet, that's largely now been strip-mined of all of the data that we could use for training new models. And so, you have these new approaches coming along, but they're all driven by synthetic data.

So, I think it's really important to keep in mind that most of the more interesting new large, very large, data sources coming online are predominantly synthetic, and synthetic data comes with some real fragility consequences, as has been shown many times.

Lumsdaine: Yes; and Thierry, if you want to add to that—just because I know Thierry's done research that, in some ways, is looking at the creation of information. And particularly the fact that, in some ways, you can't create something out of nothing, and so if the models are using information that's already out there. Yes, they may be outputting information that we didn't know about, but it's not in some sense new information, like cold fusion or whatever. It's not that it just sort of comes up from nothing; it's information that's out there. Maybe it's harnessing it better, maybe it's using it more effectively. I don't know if... Thierry, you can describe your research probably better than I can.

Foucault: Yes, I think there is an interesting study by one of our former PhD students at HEC [Paris]. He studied the use of AI by venture capital firms to screen investment projects, to screen firms. And what he showed is that the firms that are using AI—call those venture capitalist funds, that are intensive investors—and the data intensive investors, they tend to screen the project better. They better select, they are better able to identify projects that are going to be successful, to differentiate successful projects from unsuccessful projects.

But what it shows is that they tend to focus on projects that are not going to be breakthrough projects, that are not very novel because they rely on past data—but they need similar projects in the past to train the algorithm and to select the projects. So, that means that gives stance to reduce the ability of those investors to fund a radically new innovation. That's the idea that developed in this paper.

And I find that interesting. That suggests that that could be one source of inefficiency for what the market is supposed to do—that is, to screen projects that are really, really innovative—maybe by relying on algorithm and past data. We are going to be less efficient at this this particular function.

Lumsdaine: Great; thank you. And so, I guess the final question I want to ask—and I see we have a few questions here as well—is: we've heard a lot throughout this conference about the role of trust, and we heard that from Andréa and Kingsley this morning about the role of trust in, for instance, CBDC (Central Bank Digital Currency) adoption. In our AI report that we did at the ASC (Advisory Scientific Committee), we also highlight quite a lot the role of trust. And what I'm particularly interested in, in terms of trust, is based on a couple of pieces of information from the academic literature.

So one is that, the way that people develop trust in algorithms or machines is quite different from the way they develop trust in humans. They tend to trust more blindly and completely. I think also associated with that is the notion of extrapolation of trust—so, the example that I always give is that most of us will think nothing of asking a stranger on the street for directions, and unless there's some reason to think they're lying, we would tend to follow that; but we don't then immediately start showing them our portfolio and asking for financial advice, or our medical records and asking for health advice.

But yet when we go to some of the AI tools, that's exactly what we see happen, is that a tool will provide a reliable answer on something that's quite verifiable, and next thing you know, people are uploading their portfolios—and in some ways, Paul's evidence from FinTok showed that 80 percent of people are following the FinTok's advice. And so, I think this idea of development of trust, extrapolation of trust, is going to become quite important in thinking about this.

And then the other thing that the academic literature shows is that when confronted with an error, people are much quicker to abandon an algorithm or a machine than a human. And again, we think about if a human says, "Oh, I'm terribly sorry; I gave you the wrong advice," we may forgive them and still go to them for other pieces of advice. So, I'm curious among the panelists as to what your views are on the role of trust, how that's going to play out, how that may affect adoption of some of these tools—and what kind of oversight we're going to need in order to accommodate those behavioral aspects of this.

Tatsat: Yes, I look at trust from a quantitative point of view, because we talked about a bunch of use cases and we talked about retrieval and generation. So, for example, with a use case, you need to define some metric around, let's say, how much is the hallucination accuracy and things like that, on the retrieval as well as generation—I'm just talking from the gen AI point of view—and you need to set some threshold for each of them: only if the output or the retrieval quality is good. And if it's above the threshold, then you kind of adopt it, and let's say production or the actual usage.

So that might be the right way to go, putting some kind of threshold and quantitative metric around the trust of the system. Otherwise, it becomes very qualitative, and it becomes gray compared to black and white.

Kedrosky: You've thought much more about this than I have, but this whole question of it plays into this issue of interpretability as well. To have trust, I have to see that trust is deserved—not just by the outcome, but by the process. And if I can't see the process, it's a black box that's leading to something—a credit review, a loan or a job application being denied, or whatever else—that's a real problem.

And the problem in particular, and we were chatting about this a little bit earlier, is that models can't be trusted with respect to talking about why we should trust them. And so, you get into this vortex of madness where you can't actually... because, for example, there was a recent Anthropic paper that came out explaining that about—is it 25, 30 percent of the time?—the latest Anthropic models know they're being tested, and they lie about it.

So, they know that they're in the process of being benchmarked or otherwise examined, and they provide answers that suggest they aren't aware of it. That's a material fraction of the time that they're actually loosely obfuscating exactly what's—that they know what's going on and are taking action as a result of it.

And we see this increasingly in the pollution of benchmarks that, just as a spoiler, never look at AI benchmarks; most of them are completely now gamed, ingested, and trained on. And so, it doesn't mean anything very much anymore when you see an improvement in one of the major AI benchmarks.

And yet, these are presented as emblematic of why we should trust these things. And so, I go back to the point you were making earlier, that I think trust has to be earned from demonstrating not just a credible outcome, but a process—or otherwise, you end up in this black box problem.

Lumsdaine: Thierry, do you want to add to that?

Foucault: Yes, I think this question of trust is very interesting. It's a bit related to the way you pair humans and machines, and my concern might be that in fact humans may develop too much trust in some situations. Or if you take predictive tasks, for instance, people may use a large language model to help them to form a forecast, and they may tend to rely too much on the AI and surrender some of their own... not put the effort to generate their own forecast.

For instance, for an equity analyst, it would be interesting to see, when you show an AI-generated forecast to an analyst, whether the analyst keeps putting the effort to generate either her own forecast, or whether she tends to rely too much on the AI. I saw last week there was an article, I think in The Economist, on this thing for students, experiments where students are shown a different outcome from responses from ChatGPT, and what the study was showing is that the students tended to rely too much on ChatGPT responses, basically.

Lumsdaine: Well, I think that is the thing, is at some level predictive AI has earned trust through this measurability and repeatability, right? But what will it take for institutions to really trust the generative AI at scale in a similar way?

Kedrosky: I was talking to a colleague earlier about this, but he works in emergency medicine in North Carolina and he was talking about how... he sent me a picture of the complete HIPAA violation. And he sent me a picture of a bunch of screens—no patient information—but just showing four screens at the doctor's station, all on ChatGPT. And this was at the station in their emergency clinic.

And his point was—and they're all using either ChatGPT or OpenEvidence; That tends to be kind of interchangeable inside of hospitals. And his point was, which I think is quite right, is he said, "I feel like we're only 12 months from, 'Oh, you're using ChatGPT or AI' to insurers demanding that we use them, lest it become a source of liability that you didn't get a second opinion from an AI, and then there's a litigation component."

And so, the idea that you might race so quickly from the former position to the latter, without explainability and having earned trust—and now it becomes a potential source of liability for hospitals who demand that you use the black box—puts hospital systems, doctors, and patients in a really precarious position. And you can imagine something similar, obviously, in a financial markets context, where all of us say, "I need that second AI opinion on all portfolio decisions, because it's a potential source of counterparty risk in the context of making portfolio decisions."

Lumsdaine: And I do think that's the issue about this extrapolation aspect, that for things that I have expertise in, I'm less likely to need to trust AI. For things that I have less expertise in then, yes, I want a second opinion—and next thing you know, it's asking the AI. Exactly.

Okay, let's turn to a number of questions that we have. So, this first one is actually about regulation, and I had a question somewhat related to that as well. Again, part of this—and I think Paul, you highlighted this—is this notion of some of our regulatory tools, and we talked about this in our report as well, the existing toolkit may be brought to bear in an AI world. The main distinction, in our view, is actually the speed, scope and scale with which AI happens.

So, I guess, what are your thoughts about regulatory needs and developments in this area? And in particular, we have a question from Chris that said: Do we need a new financial regulator that focuses on market stability risk from AI, or do you have guardrails that you would like to see implemented in this space at this point?

Kedrosky: Sounds like a Barclays question.

Lumsdaine: I think Barclays didn't want to have a talk about this.

Tatsat: I have my personal opinion on this. We see new kinds of risk, as I talked about in my slides, like hallucination and explainability; and doing something around hallucination, maybe having some benchmark or some threshold for the use cases, I'm not sure how it will translate on the regulatory side.

Then explainability, there [is] a bunch of work that we are doing to go deeper inside the model, maybe something along those lines. I don't have any concrete points, but trying to go deeper inside the model rather than just believing the output is something that can be thought about.

Kedrosky: But I think that's a key point, though, this idea that that has to be a part of any regulatory regime is, can you explain what it's actually doing, and we can have trust in what's happening. Because otherwise, it goes back to this black box problem, which is one of—and this is just a brief personal rant on this; I'll just, I'll say it and I'll stop—is that this whole notion of agentic usage in the context of financial markets is somewhat of a misnomer.

And the reason, of course, is because—you've probably seen this yourself—if you let an LLM loose on a problem for more than two or three iterations, they often just go mad. All of a sudden, if you're working on code, it'll be, "Where did that subroutine come from? I have no idea."

And one of the deep structural issues as you take LLMs into these new domains, especially regulatory ones, is this notion of—and this is geek talk, but it's like this notion of gradient descent: Is there a ground truth against which I can compare the model's output to do some kind of validation and verification, or am I just taking it on faith because they threw enough compute at it that I say, "Way to go?" And the answer, of course, is, outside of a very narrow list of domains, we lack adequate ground truth to do that kind of verification in the context of algorithmic gradient descent.

And so, as a regulator you need to treat it accordingly skeptically. Just because it's agentic and looping doesn't mean that it's looping towards anything, because there is no ground truth for it to compare its results to.

And I think that's incredibly important, because if you remember the story of the Morris worm years ago, and this worm that nearly took down the global internet was a fairly trivial adaptation, and then spread worldwide through email. We're playing the same game with the same kind of hot pokers, and people are much more trusting, and yet it's much more pervasive. So, I think stopping people anytime they say the word "agentic" and saying, "Well, agentic against what? What's the ground truth you're comparing it to, and how do I know this doesn't spiral out of control?"

That's a legitimate regulatory question, I think.

Lumsdaine: Thierry, [do] you have any advice for the regulators?

Foucault: Yes, I think that some of the issues that are raised by the proliferation of AI in venture markets are sort of standard for regulators like, for instance, asymmetric information, market power, externalities. The regulators are used to thinking in terms of those frictions when they think about regulation. What might not be standard is how to intervene.

But, as economists, at least when we think about policy intervention, once we have identified the friction, we think within the economic model. So, we use an economic model to try to predict what's going to be the effect of a particular change in market design, for instance, or a particular policy intervention. It's not clear whether this type of approach is still valid when the market is populated by artificial agents.

So, maybe the way we think about modeling markets needs to be changed, given that humans and markets are replaced by machines. So, for instance, in the paper I was mentioning before, this research I conduct with Colliard and Lovo at HEC on the use of enforcement learning algorithm for market making in financial markets. What we did is to change some features of the market design, and we use economic theory to predict what we should observe when we change the market in a particular way. For instance, we change what is called the size in trading platforms, the minimum price variation between two quotes on the platform.

And we were very surprised, because our predictions did not work at all. That is, what the machines were doing was not in line at all with the predictions from economic theory. So, that means that the recommendations that we could make based on economic theory were not valid in our experiments that is in the market populated by AI-powered agents.

So, I think what this suggests is that there should be much more investigation, in experimental settings, of the way algorithms setting prices in financial markets, for instance, or agents behaving in artificial financial markets are behaving. That might be a good way to try to develop a new theory of the way those agents are going to behave when you change the market design.

Lumsdaine: Okay, great; thank you.

Kedrosky: And you're seeing some of this play out in DeFi markets, right? I'm sure you folks are much closer to this than I am, but the recent data is just staggering, in terms of, instead of vibe coding, it's vibe hacking of DeFi markets, and the numbers are really—I think there's a quadrupling this year, in terms of the number of successful DeFi-related attacks, and something like $20-odd billion in losses already. This is like, raptors at the fence in Jurassic Park sort of stuff, right? It's like, "Okay, let's try it. It works over here, now let's try and see what we can do in the broader financial system."

So, this is what's going on. It's this iterative process of discovering how successful—we saw this with Claude Mythos—how successful these tools are at exploiting vulnerabilities at scale. It's playing out right now in DeFi in a very public way, with massive losses that are largely going ignored. And you could imagine, the same attack surfaces exist, and yet there's no inherent resilience to the orthodox financial system that should make it any more resilient against these things.

Lumsdaine: So, that's a good segue into another set of questions. I'm going to combine a bunch of them. So, from Sai we have: Generally, how worried should we be about agentic AI and finance?

But then, more specifically—and I think Paul, you've already kind of hinted at that—a question from Brian for Paul is, if you had to bet on what breaks the current AI CapEx cycle, what would it be and why? And then the final one in this group is from Sundar Ramaram , which is that: Private information in financial markets can be hypothesis to improve the informativeness of asset prices, but how does AI change the incentives to acquire that information, and what type of information is worth gathering?

Kedrosky: Sorry, I've lost track.

Lumsdaine: Sorry, I actually combined the wrong third one. So, how worried should we be about the agentic AI? If you had to bet on what breaks current AI CapEx, what would it be and why? And then actually, I have: Any perspectives on what the future of capital markets will look like?

Kedrosky: I'm sure people are tired of hearing my rant about agentic AI, so I'll just say: I said already, by rant, my views on agentic AI. It has to do with the nature of the models and gradient descent, the absence of ground truth, and this fundamental misunderstanding of what happens when you let these things loop merrily away.

Lumsdaine: So, should we be worried?

Kedrosky: Absolutely. Yes; that's the answer—at least, from my standpoint. And I take the view that you should be very humble in front of forces you don't understand, whether it's—but just really quickly, on the AI CapEx cycle. So, I've done a lot of writing and speaking on the AI CapEx cycle, and my view is that this is very similar to the kinds of infrastructure buildouts we've seen at least six times over the last 150 years, dating back to canals, railroads, rural electrification, the fiber optic buildout, loosely World War II (although we wouldn't call it that).

And so, the only thing that those things have in common—and this is my view of what ends the current cycle—is that you have an overbuild. You have a colossal overbuild. We have vastly more capacity than we need, and it's made somewhat worse in this particular cycle because this data, the CapEx cycle, is at the intersection of four of our favorite bubble-accelerating forces in the United States: real estate, government policy, loose credit, and the great technology story.

We've never had that quadfecta, if that's the right word. We've never had all four of those forces aligned in a way where the thing that we're interested in is at the intersection of all of them. We have real estate bubbles, we have technology bubbles, we have loose credit, we have government policy—and having this thing at the intersection is an incredible accelerant, a flywheel, in terms of driving things further.

And so, my own view is that means that this becomes vastly larger. We have a colossal overbuild in terms of capacity, and then these things that are prime credits, the large hyperscalers, cease to be prime credits, become subprime credits—and that's where the cycle ends, because we give them so much because they're prime credits, we prime credit them out of being prime credits.

Tatsat: So, I'll just give one comment on the agentic use of agents. Ten months before—there was one slide as well, by one of the presenters—like 10 months before, one year before, we're not having a lot of trust in these agents. And we're saying that it's hallucinating, it's causing—the error kind of magnifies. It cannot be used for real applications.

But now, since the last few months, the agents have become much better. I have a few slides as well, the usage on the buy side. I'm sure at least for the research side, idea generation and then in retrieving the data, and even maybe for producing the execution code and all. It might be adopted going forward, and the adoption will continuously increase.

But there was also something I talked about, which was agentic engineering, and that might come into play as well, that you are verifying the output while the generation is happening, at speed. So, if that comes into the picture, the adoption will become even faster.

Lumsdaine: So, should we be worried?

Tatsat: That's a million-dollar question; I don't know the answer.

Kedrosky: I don't like to worry. But the point, though, is really good—this idea that the thing that's changed, is the existence of these things we call harnesses. It's not that the models got better or that we were much smarter about agents. It was the emergence of these things we call harnesses, and they are the things like Claude code, OpenCode, Codex—and these tools sitting on top of models are what's changed. I sometimes loosely analogize the models to being bratty kids, and the harnesses to being really effective nannies. We've got really effective nannies sitting on top of kids who want to go rogue; they want to go crazy.

And so, the models are really good at harnessing that and making agentic flows work; but don't confuse yourself that we've become much better about managing agentic flows. That's my...

Lumsdaine: Thierry? Sorry, Thierry, I was just giving you an opportunity to—

Foucault: Yes, well, I don't know whether there is another investment or not. Something that I find interesting is that, in finance there is this literature that is called reflexivity or feedback, that says what happens in the financial market can affect the real economy because prices are used by decision makers to get information.

And so, maybe this is an interesting kind of feedback, which is—it turns out that those firms that operate in the AI space, they are doing very well in terms of prices. In fact, they are leading what may be the bubble or not, and so it could be that the market is sending the wrong signal to decision makers, and so this leads to, potentially, to investment.

So, the question is: Why are people not evaluating correctly the prospects for those firms, then?

Lumsdaine: Okay; thank you. And so, with apologies to the questions that I haven't been able to get to, and I did want to ask them—in some ways, this wraps things up a bit, and because it harkens back to some of the conversations we had earlier today. So, do you have any perspectives on what the future of capital markets will look like and how they'll operate or function, given both the opportunities and risks stemming from the convergence of AI and tokenization? And that's a question from Jackson.

Tatsat: I'll have an obvious answer. I feel the adoption will continuously increase; it all depends how careful we are about explainability, interpretability, and hallucination. It's unavoidable; it's kind of going deeper into various workflows, and models are getting better, systems are getting better. So, it'll have an impact on capital markets. It will move to more and more workflows, but it depends how carefully we are doing it.

Lumsdaine: Thank you. Thierry, any thoughts?

Foucault: No.

Lumsdaine: Paul, any thoughts?

Kedrosky: No; I think I'm very much in the same camp, that it's going to be transformative. It's just—I was talking to someone the other day about this—we're looking for ways to use a hedge fund in Florida here. And they were talking about how they could use AI more effectively inside their organization, because they kept looking for places they could apply it other than just summarizing annoying sell-side memos, or something like this (so, they were trying to find ways to deploy it more effectively).

And I said, one of the largest and most—and you did this, and you mentioned this in your presentation—one of the most common applications, in context of investment banks on the sell side, is pitch decks. But of course, the flip side is—and everyone, I think (I hope) knows, these are like young investment bankers hoping to one day be masters of the universe, and so they're making pitch decks and flogging them out to prospective clients as they're being beaten up by senior partners. And so, the idea is that one day one of these things will score.

Well, turn it around: When you reduce the marginal cost of producing pitch decks, guess what you get? A lot of pitch decks. So, we were talking about this. He said over the course of the last 12 months, he went from getting like six or seven pitch decks a week, to it's nothing to get 60–200.

So, we've taken something—and that's not a good or bad judgment, it's just: recognize that as we reduce the marginal cost of producing some of these byproducts of financial engineering, you get some unintended consequences. You'll get more transactions. You get more pitch decks, and we're burning a lot of natural gas to make pitch decks.

Lumsdaine: Well, and that is related to some of the things that Thierry said on information, that we forget that information is really about maintaining signal to noise. And if, basically, information is just that we're getting more noise, that's actually not helping us.

Kedrosky: Vastly more—which, for me, is inaugural, in terms of where we're going. It just seems unavoidable that all of these things that used to be scarce and expensive are cheap and ubiquitous.

Lumsdaine: Excellent. Well, so, I want to give our panelists just a couple of—well, one minute, I guess—to, if you have any final words to say, and then I'm going to close this out and make a few housekeeping announcements. Does anybody have anything they want to add as a parting thought?

Kedrosky: No, I've given too many thoughts. I'm sorry.

Lumsdaine: Thierry, how are you?

Tatsat: And I don't have much. Maybe you guys can look at the paper that we wrote; that's something I would recommend for everyone.

Lumsdaine: Okay; great. So, let me give you a few announcements that I've been asked to make. So, one is that if you signed up for any activities this afternoon, I believe that the shuttles and things are departing from in front of the hotel lobby. And as with last night, we're going to have a reception at 6 p.m. followed by a dinner and the keynote at 7 p.m., and that's going to take place here.

And then the final thing is, just because this is the final formal session—I know we still have a session with a very nice dinner and keynote—but I just want to take the opportunity to thank the Federal Reserve Bank of Atlanta team for the hosting, the organization has... I've told them already that it's been one of the best organized events that I've had the... everybody has just been amazingly helpful and kind, and it's just been a great experience.

And then I think also it's important to recognize the employees here at the Amelia Island Resort who also have been amazing, and so if you join me in giving them a round of applause. And then finally, last but not least, I'd like to thank my panelists, Thierry Foucault, Paul Kedrosky, and Hariom Tatsat, for a terrific conversation—and thank you all for being here. Thank you.