About this transcript: This is a full AI-generated transcript of FULL PANEL: Federal Reserve Chair Kevin Warsh and Tech CEOs Debate AI's Future — AF15 from DWS News and DRM News, published November 17, 2026. The transcript contains 9,809 words with timestamps and was generated using Whisper AI.
"Even a few years ago, the idea that we could find in our pockets instant access to genius-level intelligence on almost any subject would have seemed far-fetched. Today, with the advances of SI, it is happening. The marginal cost of satisfying limitless curiosity on questions large and small is..."
[00:00:00] Speaker 1: Even a few years ago, the idea that we could find in our pockets instant access to genius-level intelligence on almost any subject would have seemed far-fetched. Today, with the advances of SI, it is happening. The marginal cost of satisfying limitless curiosity on questions large and small is falling fast, perhaps even trending towards zero, a subject that financial markets are trying to figure out every day. So now we're going to turn to the wonkier part of the discussion, but frankly, the more consequential. In the tradition of endogenous growth theory and economics, ideas are the ultimate resource. If living standards are to advance, conditions must be favorable to the generation of new ideas. It should be no surprise, then, that the United States is leading. Societies that bring forth the best ideas, well, they tend to prosper the most. The implications for economic policy can hardly be overstated. While an idea can be costly to discover, it's free to use. And because an idea is non-rivaled, that is, one person's use does nothing to diminish that of anyone else. It's value grows with the number of people who use it. The Nobel laureate, Paul Romer, among others, including my friend Chad, has shown that a doubling of ideas and other key inputs can more than double output. The key word there is more, allowing increasing returns to scale. Even so, in the U.S., over the last several decades, breakthrough ideas had been harder to find. You recall, at conferences like this, this is when there would be discussions about secular stagnation. By some efforts, research efforts have increased more than 20 times since the 1930s. Yet, Nick Bloom, another friend who's here, Chad Jones and other co-authors, have pointed out that research productivity has fallen by an even larger amount. The result? Productivity growth has fluctuated, and while in recent quarters it has been reasonably strong, it hasn't yet broken out definitively, structurally, reliably to the upside, at least not yet. Long-run growth, I should note, is also a function of the rate of increase in the number of people producing these ideas, setting them in motion. The conventional view holds that as population growth slows, there may be fewer new ideas to drive prosperity. The increasing challenge of generating breakthrough ideas, combined with slower population growth, helps explain why many in the economics profession, especially in the post-crisis environment, return to the old theme from the 1930s of secular stagnation. So, the question for all of us, the one that I want to explore with our colleagues today, is does superintelligence dramatically change this picture? Well, there's certainly some indications that it does. SI might be somewhat different from general purpose technologies, the kind that Mary referenced in the outset, the internal combustion engine, or electricity, or the transistor. Those inventions made for decades a follow-on innovation, so I don't mean to dismiss them, but they did not themselves automate the creation of new ideas. What might set SI apart as an economic influence is that it could supercharge idea generation. It appears to be changing the method and, at least as important, the speed of innovation. If idea generation broadens, growth could be sustained at a level substantially higher than the long-run average. But for all of that, however, let me make one essential point. Humans should not be underestimated. There is no replacing the minds and hands of men and women. Machines that can amass or even exceed human knowledge, well, they're still machines. They do not take the place of human judgment, morality, dignity, or responsibility. Humans are the designers and the builders, and humans must be the deciders. So, for our country in this era, consider how much is on the line. Consider the reason why, five and a half months ago, we put this task force in the front of the line. I remember Bob Mendel, another Nobel recipient, saying years ago that the most important fact of the 20th century was the rise of the United States as a superpower. For us today, that's more than a fact of history. It's our charge to keep. And so we must ask, can you be a superpower without superintelligence? My Fed colleagues and I are not at this conference as futurists to speculate on each and every turn in technology, and we're certainly not here to make predictions. We can't know the future, and as you've heard from me before in other settings, we ought not to pretend otherwise. Our job is to prepare for whatever comes next. That means contingency planning, a subject that I will return to in other task forces. Contingency planning becomes only more important when the world is changing at this fast speed. So our aim today is a humble one to ask the right questions. As policymakers, we have to be mindful that great opportunities rarely come risk-free. And the risks in this case, if overlooked, can be highly disruptive and economically quite costly. SI empire empowers all, including bad actors. So we're thinking especially of critical infrastructures and their vulnerabilities. And of course, I think of economic policy in this new era of SI. At the Fed, our role is neither as advocates or opponents. Our responsibility is to recognize that the widespread adoption of this technology will have a very significant effect on the economy, with hugely important implications for our day jobs on monetary policy. So let me say a few words about some of the big questions. They're mostly about magnitude, scope, and timing of this development. First, will this new technology deliver a rise in total factor productivity? Will the democratization of specialized knowledge significantly raise standards of living? When will this surge start, if it hasn't already? And when will it taper? Will it happen sooner or faster here in the U.S. than in other countries? And will adversaries race to deploy SI in whatever its form happens to be, intending to do harm to our economy? These answers will matter greatly, and that is why we're thinking about them today in a sober, serious way. And I expect more clarity with the passage of time. Still, we've got a job to do in monetary policy, so it's worth considering the range of economic outcomes we could see over the next several years, depending on how SI progresses. Board staff has begun to evaluate various contingencies, there's that word again, applying stylized model-based frameworks to generate scenarios. They're illustrative scenarios. They're not forecast, they're not predictions. Moreover, many other outcomes, probably including the one that will actually happen, could be drawn. Consider them a starting point for the very useful discussions we've got lined up today. First, at one end, SI leads to an increase in productivity and economic gains that are shared broadly across labor and the owners of capital. In that event, real GDP grows persistently above its long-term average, and the U.S. stays close to full employment. Other outcomes, however, could be less positive. An SI-driven expansion could bring substantial labor market disruption, as some jobs are displaced rather than complemented by technology. GDP growth would still rise above its long-term average in this contingency. But widespread automation, combined with labor market frictions, could push the jobless rate above levels consistent with full employment. Let me reference a third. Or, SI could disappoint. After significant capital investment and stock market gains, expected returns could turn down. Equity valuations could fall, leading to an economic slope. Employment. In time, the economy would recover, but without the productivity gains. I stress that, while I'm not here to offer predictions, I'm not exactly impartial about American leadership, in a time so filled with both challenge and possibility. I'll end where I began, thinking back to my first visits out to this part of the world, to my days at the Hoover Institution, and to a collaboration with my fine mentor and colleague, John Cogan. As we wrote some years ago, without the knowledge that SI might be upon us, but I must say as optimists and believers in the US, and I'll quote a paper that he and I wrote together. If we meet the challenges before us, we said, America has a bright future, with strong, durable economic growth, abundant economic opportunities, vibrant and secure communities, and a healthy environment. Future generations will inherit a dynamic economy capable of delivering technological advances that will extend lives and promote human flourishing. That's the goal. That's the ultimate scenario. That's what all of us here hope to see. I can just end with a final note. It is a privilege, privilege of a lifetime, to be a policy maker here at a time of such consequence. And it's the most encouraging sign to be joined by some of the most serious, sober, responsible thinkers. Great men and women who are willing to join the Fed, even as an aside, to try to put our SI goals on the right path. Thank you all very much. Let me turn now to our first panel, and I'll ask my panelists to join me up here on stage. up here on stage.
[00:10:53] Speaker ?: Thank you.
[00:10:53] Speaker 1: Thank you.
[00:10:54] Speaker ?: Good to see you guys. Looks good to see you. So let's turn to this first panel.
[00:10:56] Speaker 1: It's an honor to be joined by these three gentlemen. I've known them all for decades. Let me do, let me introduce this like we would a proper beautiful old Russian novel. Introduce the characters, and then we'll see where the plot goes. I have the wrong one in front of me. Hold on a second. That's the wrong characters I was going to introduce. So this panel is called At the Frontier. The idea here is to ask questions from people that are users of the technology, who are some of the most innovative thinkers that I know. We didn't call these guys here because they're at the labs and they're designing this great stuff. We called them here because we wanted to know what's really happening inside of companies. That they happen to run technology companies might be an accident of history. Is that because these technology companies themselves are more innovative, are willing to cast aside how they've done business, or is it just three guys that happen to work down the street? Let me do the brief introductions. Nikesh is immediately to my left. He's the chairman and CEO of Palo Alto Networks, which he's led since 2018. Before that, he was the president at SoftBank, where I came to know him. And even before that, he was at Google, where he was the chief business officer. Next to him, Patrick Collison. I've known him and his brothers since they came to these shores as young men from overseas, poor immigrants, and they've done pretty well in the United States. Millions of businesses use Stripe, the business that they started and founded, to accept payments, send payouts, and manage other complex online enterprises. Max Levchin, who's at the end, has been a friend also for, I don't know, 20 years. He also came to the United States and is one of the smartest people that I know. He's the founder and CEO of a firm, the co-founder and chairman of other companies. And he created an innovation lab called HVF, which stands for Hard, Valuable, and Fun. I don't know if I was supposed to disclose that. Max was the chief technology officer at PayPal back when we were all much younger than this, and before at least some of us had gotten mixed up in public policy. And he was the chairman of Yelp. Not bad for someone new to our shores. So those are the three people I want to sort of probe on a few areas of questions. One is adoption. What's really happening? And I want to try to understand if what's happening is so special and different in their companies, what about the rest? What about the other 450 companies in the S&P? The other 2,000 companies in the Russell 2000. So I want to know whether this is just a technology that's at the frontier and going to stay there, or is going to find its way more broadly. When I start at the end, Max, with you, tell us about how your business is different, how you're using this new technology, and exactly what you think that means, not just for your business, but for the broader economy.
[00:14:23] Speaker 2: Sure. Thank you. Can everyone hear me? Yeah. Thank you, chairman, for inviting me to participate. So a firm is a lender, a fairly sizable one, and everything we do is done by machines. So our underwriting models are entirely ML, super intelligence based, et cetera, et cetera. And so we've been an early adopter of the earliest models. We've been using them primarily to increase our engineering productivity, which for scale is about five times more productive. So the cost per feature is one-fifth or what it was. We expect it to be one-tenth. So these are pretty extraordinary gains.
[00:15:07] Speaker ?: It's a great story.
[00:15:08] Speaker 2: That's sort of what everybody will tell you. So none of it is new. So to contextualize what that means. As a lender, our job is to sort risk. So people come in, they ask to borrow money because they're trying to buy something. We need to pick the ones that are most likely to pay us back on time every time versus the ones that don't. And it's all done by machines. But humans are the ones deciding what makes an interesting predictor. So the fancy term is feature. The kind of colloquial term is a signal. As you try to decide what might be a great signal that predicts your ability to repay. The ideas greatly outstrip the available time to train and the engineers to put those things into motion. So there's lots and lots of really good ideas that end up on a cutting room floor. So what does this efficiency or this productivity gain really do for us? We're able to pick up a tremendous number of these ideas off the floor and put them into production. So in a kind of almost too meta a story, we used LLMs to design our most recent model, which in itself is actually built on top of a transformer model. So not exactly an LLM, but a token representation of repayment events that allowed us to do almost twice as well as the next non LLM model would have done. So we ran the two side by side and we saw an 80% increase over the baseline, which again for context translates to about two more billion dollars per year of loans for us. And so this is a direct consequence of just being able to free up engineering talent, all human zero layoffs due to AI at a firm. And just deploy them for what you might think of mass customization, which I think is what's going to happen. Maybe I'm too Pollyannish, but as a techno utopian, I'm prone to love our machine overlords. What I think will happen is we will see more and more of these long tail ideas that no one would ever have the time or the productivity or the staff to lift off the floor come to fore. And some of them will fail, but that's what we're supposed to see. But we get to find out if the thing that you need and only you, which is now can be produced at a much lower cost, is going to be finally created. And you might in fact find that there is the intersection between supply and demand that makes it work.
[00:17:25] Speaker 1: So Patrick, let me turn to you. I presume you're going to also tell us a story that Stripe is at this frontier and you're making great progress. We want to hear about that. We want to hear also from you about how long will the delay be between those at the frontier and the rest? What do you think is holding down? What's the bottleneck? When will the rest see it and what happens if they don't?
[00:17:51] Speaker 3: Well, again, great to be here. When you said it's like a Russian novel or something, I thought it was because we're all unhappy in our own ways as CEOs. But so, yeah, like Max, we've seen a very significant acceleration. And it's been changing quarter by quarter for, you know, 12 months ago maybe the, you know, the narrow work of engineers was being somewhat accelerated in code production. Whereas today at Stripe and I'm sure at both Palatine Networks and at Affirm, it's more about these longer horizon loops where the agents are identifying tasks autonomously themselves. They're going and actually implementing the feature or fixing the bug or whatever end to end and deploying it to production with substantially reduced human, I mean, human oversight nonetheless, but reduced, significantly reduced human oversight as compared to say six months ago. And in terms of what that means for Stripe the organization, I mean, it's interesting. There's always kind of this question, you know, to what extent AI is kind of a substitute versus a complement. At least in the near term, we're seeing substantially reduced hiring for our GNA functions as compared to some, you know, prior counterfactual. And, you know, GNA as a share of our revenue is definitely diminishing quickly. But for engineering, that is kind of R&D, you know, our kind of product production, and for sales, thus far we are seeing acceleration. Because, you know, effectively, as Mac said, the cost of feature production is diminishing. So, you know, in the face of that, there's a sale on for feature production. And in the, you know, in the face of the sale, we want to buy much more than we did previously. And similarly, you know, our sellers are much more productive than they were previously. And so, therefore, on an, you know, ROI math is quite straightforward, right? It's mathematically, arithmetic makes sense to hire more of them. In terms of what we're seeing across the economy generally, you know, I remember early in the year, I remember early in Stripe's history, the CEO of a very large mega cap company came to visit. And, you know, we were all, we were using all these new SaaS collaboration tools. And everyone in the company could see what was going on and all the things. And, you know, he was shocked. I mean, Stripe was a 30-person company at this juncture. He was shocked that we were, you know, that we would operate this way. And, you know, where is the governance and the SOC 2 and, you know, all the kind of enterprise considerations that, you know, were, that loomed large for him. And I'm now sensing some of this same disparity and kind of mental model impedance mismatch between incumbent organizations and new companies. Where large companies, and I put Stripe to some extent on the incumbent side of this. We have all the same questions that you have in terms of, well, where is the human supervision and where is the governance and what's the certification and the compliance and, you know, all the things that a sort of prudentially oriented infrastructure company, I think, is wise to consider. New startups are new startups. And they're happy, you know, we would not autonomously plug the agents into sort of just, you know, feature development loops and, you know, dust off our hands and walk away and say, hey, you guys do whatever you want to do. The new startups are happy to do this. And the 2026 cohort of companies started on Stripe is massive, so new companies, is massively outperforming any prior year's cohort by, you know, a dramatic magnitude. The number of new businesses being launched on Stripe every week is up approximately twofold year over year. And, you know, we've had more than one decade long time series here. You know, Stripe has been in operation for around 15 years. And so a 2x year over year increase is not something with historical precedent, even, for example, during the pandemic. And so I guess I'm reminded of, you know, the-- well, many people here in the room can correct me. It's the Paul David 1990 paper, I think, about how with electricity-- rather than existing incumbent firms adapting. And, you know, obviously, existing firms are trying to adapt and, you know, there's certain-- you know, AI is incredibly salient for CEOs across the industry. But I will say that new firms that we observe and whose business performance we directly see on Stripe are proceeding at an unprecedented pace. And I'm very curious he works.
[00:22:15] Speaker 1: We're going to return to the Paul David-- Did I get that right? You did. 1989, not 1990. Productivity paradox paper with the economists. I wasn't born then. I'm going to keep you focused. So, Nikesh, you run a big company. You've got a history of big companies. I know you have one foot in the venture world and startup world. Tell me two things. One about how much is this new technology changing your own business at the frontier. But when I met you, what you were is the best salesman I ever knew. You would sell products to enterprises that weren't sure they needed us. See if you can tell us-- See if you can tell us-- Is that a compliment?
[00:22:59] Speaker 4: It is. I'm not quite sure where that was going.
[00:23:01] Speaker 1: They bought them, and Google seems like it's done okay. But what's the reason why there would be a big distinguishing feature between the frontier and the rest? What's the reason why, if we've got this hypothesis right, which is the Fortune 450 aren't coming along so quickly, what's the reason? Have they not changed their organization, their mental model, their way of thinking, their process, their governance? Help us understand on the diffusion question, not just your own internal workings. Got it.
[00:23:33] Speaker 4: Well, first of all, thank you for having me here. I think it's fascinating. We seem to be focused on talking about AI from an enterprise and enterprise productivity lens. If you look historically, since, you know, there was a time about 25 years ago, technology started getting built for consumers and not for enterprises. Historically, we built technology for enterprises, the mainframe, et cetera. When mobility came about, and when laptops came about, we started building for the consumer. And I want to go back to something you said. The idea that we will carry genius in our pocket and have access to intelligence at a marginal cost of zero is extremely compelling. And I think that's the optimistic view we have to start with. Because, you know, there's six billion people in the world who do not have access to expertise or people who are, they're physiotherapists and nutritionists. All that intelligence is going to be carried in the back pocket. So I think that's going to be hugely amazing for society and humanity at some level. Now, if you transpose that and say, why are we so obsessed with the enterprise version of this new technology? It's because it's expensive. We need to find a lot of money to go build these large data centers required to train these models to get really good intelligence. That's why the focus is inadvertently on how much money can we make out of this technology? And back to your point about how can we get enterprises to adopt it quickly and embrace it? I think there are certain things that this technology is really good at. Processing large amounts of data, retaining large amounts of information. If you have 3,000 people doing customer support, you can have one brain that understands all these interactions and it can be the complement of the human being that's sitting at the edge solving the problem of the customer. So, yes, it already has immense applicability in terms of how enterprises operate. But what you're noticing is, as Patrick said, it's really good for repetitive tasks where a lot of analytics are involved, a lot of understanding is involved, but it's still early to give it arms and legs. What I mean by that is giving it agency or giving it independent authority to go act on our behalf. I think that's where the fear is because that's where the job displacement begins to happen when you can say, oh, I don't need you. AI will take care of it. Now, as you mentioned, I worked at Google and not just in sales. I was there in 2009 when I sat in the first driverless car. It was a Lexus with many, and I still wonder what the hell was that doing, thinking of letting it drive me in 2009 or in 2026, which is now called the Waymo. Just think about it for a second. It's taken 17 years since then to take something and replace one human job. One human job has cost more than $30 billion. It's not the mainstay that we worry about. It's the edge cases. I work in cybersecurity. Our products are all about the edge cases. It's about that one time somebody will come, find a vulnerability, get into your organization, and exfiltrate data. It's not the 80% use case we worry about. It's the other 20% that we have to worry about. I don't think this technology is fully there yet in terms of being able to understand edge cases, being able to plan for those. That is why we still need that human judgment, the human in the loop. I know there are certain companies which are out there to the races because they found use cases where we need this superior intelligence. I think there are use cases and you'll be able to find distinctions like cybersecurity is one use case. Superior intelligence allows bad actors to go find vulnerabilities and come attack companies because that requires real smarts and lots of compute. I'm pretty sure it's going to work in cancer research. I'm pretty sure it's going to work in hard science problems which we need to sort of further humanity. But I suspect the benefit of this technology will be over time. It will become the new normal. We will also have access to amazing amounts of intelligence. I'll end with this. In 2004 when I was at Google, we used to say search democratized information. Everybody had access to the same information wherever they wanted. I think AI has the power of democratizing intelligence.
[00:27:29] Speaker 1: So let me start back with you. We're going to go down the line. I'm going to have a couple of individual questions. So, Nikesh, when I see you, I think cybersecurity, your business is focused on that. How worried should we all be? Well, we're relatively early in applying these technologies, both you on the business and revenue side for many of us for protection. How worried are you? What are the vulnerabilities to payment systems, infrastructures around the world, and what, if anything, can we do about it?
[00:28:01] Speaker 4: I tend not to worry, but it could be hard otherwise. But this is a harder time than normal because what we've done is this capability that AI has, has the ability to find vulnerabilities and bad code and bad things that were written by humans for the last 50 years. And the problem is we, cybersecurity is an asymmetric battle. The bad actors would be right once, we have to buy it 100% of the time. It's very hard when you bring this technology to bear where things which used to take days now can be done with pretty much, you know, what we've noticed in the last six months is there are no people available in the industry who you can call and say, can you help me? I'm having a problem, literally. Everybody's backed up right now because every company is going through some version of bad actors trying to get into their infrastructure, understanding all vulnerabilities. So the technology which came out six months ago called Mythos has now become more commonplace and is available to most bad actors in open source models. Unfortunately, it's not that the hackers have become smarter. The technology has become so good, even bad hackers have become okay. So that's what we're dealing with. So right now, I think we're at a bit of an asymmetric situation. We have a disbalance in terms of what we need to get done. There are things which are getting exposed. We will see more cyber events. I think you already see there are more cyber events happening every week than we've seen in the past. I don't think this is going to stop. I think it's going to persist. I think it's particularly important to make sure that interconnected systems understand who the weakest link is. And I think the banking system is one such place where we have to be very careful because lots of people are connected to a lot of different people. And the weakest link is going to determine how safe our system is over time. I think the next two or three years is going to be a bit of a technology rebuild moment. We're going to have to go back and see every mistake, every thing we kind of did not pay attention to, how to rebuild. Now, the good news is AI or SI has a possibility of not just finding vulnerabilities. It also has a possibility of allowing us to refactor code and rewrite code and do a whole bunch of that stuff. So I think we take the positive outlook. We have to go spend more time investing in upgrading our technical infrastructure. We saw, for a totally different reason, the airlines were down this past week. We saw Microsoft had an outreach. So you are seeing examples of how, if you don't pay attention to upgrading old technology, we're going to end up in a slightly worse position than we are.
[00:30:25] Speaker 1: Patrick, since I met you and your brother, you always majored in technology, but you minored in economics. So let me give you a question that straddles in between. The question for financial markets and maybe for some people in this room is, what's the market structure on the other side of this? You were once a little startup company and by your own description, you sort of feel like a legacy incumbent after a decade or more. Who gets to make the money? Will the money be made by the big labs? Will the money be ultimately made by the users of these technologies, which are those things that could drive that productivity, which would be the biggest fundamental change in economics in generations? So what's your sense of what the world looks like in market structure terms?
[00:31:09] Speaker 3: I mean, I think that, well actually first let me just echo the last part of what you said and maybe not the first part, which is you said that the next couple of years is going to be a technology rebuild moment. I think that's exactly right. It's what we're seeing across the industry and even the most scrutinized parts of technology. Like the Linux TCP stack. There was, you know, tremendous attention devoted to making sure that it was as secure and as bug free as possible. A huge incentive for people all around the world to find, you know, every little minute error or exploitable sort of codicil. We thought it was pretty good. Like, you know, the Linux TCP stack is about as scrutinized as software gets. And, you know, turns out it is riven, absolutely festooned with bugs and problems and exploitable issues. And so I agree with that part of what you said. And you're, I mean, this is maybe why you're more successful and just, but I'm worried. I'm worried. I think it's going to be -- We make more money when people are worried.
[00:32:14] Speaker 4: So -- Yeah. So -- I could be thinking --
[00:32:18] Speaker 3: I think it's going to be a man's scramble. And so I just -- that's my Irish sensibility here. But on your question about the market structure. Look, I think this -- I think this reduces to the question of what does the -- as in the market structure question, it reduces the question of just, well, what's the world configuration, you know, 10 years hence. Maybe we can make near-term projections about, you know, what -- to whom different surpluses might accrue and so forth, you know, a year out. But, you know, the -- many of you have -- have -- have, I assume, read the essay "Machines of Loving Grace." No? Yes? Okay. Seeing some nods. Right. So this is an essay published by Dario Amode from Anthropic. And kind of laying out the positive vision for where AI goes. And it is a reference to a poem from Richard Braudigan in 1967. And it's a very short poem. You can read it. And the last verse of the poem says -- you know, he says that he dreams of a cybernetic ecology where we are freed of our labors. So, at least in Braudigan's conception, you know, this is not just a compliment to help us, you know, be more productive. And his vision we're freed of our labors and returned to nature with our mammal brothers and sisters. And all watched over by machines of living grace. And this is a vision -- Braudigan's vision -- for where things could go. Elon Musk has said that, you know, humans are a biological bootloader for superintelligence. And I guess I bring these up to say that, you know, there's some pretty far-out stuff being posited as, you know, potential destinations or at least, you know, intermediate points in terms of where these trajectories point. And so, given all of that, I found it pretty hard to, you know, then, you know, work backwards. So, well, what's the market structure of the cybernetic ecology? So, look, I think today and for the foreseeable future, the vast majority of the surplus is being captured and will continue to be captured by the users of these technologies. And there is immense productivity gain and new creative possibilities being unfurled. But, you know, what the general equilibrium looks like, I will defer to people here.
[00:34:50] Speaker 1: Max, I won't ask you to define what festoon means. But maybe in the old country, it's --
[00:34:57] Speaker 2: My socks are festooned with my logos.
[00:35:00] Speaker 1: Max, let me try this with you. I've been trying to frame this enterprises at the frontier versus the rest. I might have gotten this from you some time ago. Is SI really driven by individuals and not businesses? Are there people resident inside of your company where there doesn't have to be a corporate strategy on SI? They're just playing and excited by these new tools. So, is the transformation less about is the company small or large? Who are the individuals that are practicing these new ideas?
[00:35:33] Speaker 2: So, I think like most profound innovations, it begins with individuals. And it's up to us as leaders and, frankly, all of you as leaders to institutionalize the gains. I think the -- in the industry, the -- for a little while, there's been this narrative of, well, there's about 500 people in the world that are good enough to think up the next advance in LLM structures and sort of the math that comes with it. And I haven't heard that one in a little while because I think that number is expanding. And I think today I can name the handful of geniuses on my MLAI team. And I know what they're working on. I'm excited to sit with them because they teach me things that I just had no idea were even possible. I think this is true for most companies and all of us are hanging on to our researchers, lest the frontier labs come and get them and overpay them. But over time, like everything else, our job is to take kind of the patterns of genius that we're seeing and institutionalize them. And I think that's a -- that's reasonable and a good thing. And there's some degree in some reaching into places like national security and kind of a short-term concerns around leakage of IP where we must know who they are and really care about where they go and what they work on. I think over time this becomes diffused and maybe we care a little bit less. And since we're all on the topic of security predictions, I'll stick with my -- I am the optimist apparently on this panel. Or maybe the most optimist anyway. I actually think that we're going to go through a period of upheaval for sure. There's quite a lot of discoveries of these latent bugs that are being unearthed very quickly by sort of the likes of mythos. But I think that motion has always been there. The entire security industry is all about lots of very brainy people going, wait a second, that's a buffer overflow. If you don't know what that is, I encourage you to look it up. It's one of the more fun parts of computer science. And then you're like, oh, here's -- there's an if statement missing. We'll fix that right now. And what we're going through is an extremely compressed period in time where we're finding all the missing if statements and we're fixing them with SI as our propellant both in the finding and the fixing. The fact that the bad guys have access to open-weight models and they're chasing the one thing that they need to be right and then all of us have to be right most of the time or hopefully all the time, that's a temporary effect. The long-term effect is that we'll build more robust software. The thing that we're getting from this revolution, which is probably the best word for it, there's a tremendous amount of creativity that's being unleashed in all of us. The thing that LMS have not gotten to, and I will take the bare case, they have yet to come up with a single brilliant idea that was not originated with a human. The job changes that we're going to witness is going to be all about pivoting all of us towards more creativity. The real question I think all of us have to ask is how creative am I really and what am I doing to unleash all of that? If you're finding yourself devoid of good ideas, that's the moment to worry. But retailers will become better at marketing, more creative at marketing, because the fulfillment and price comparisons are all going to accrete towards just efficiency, because this productivity gains will apply to all those areas. And so creativity is going to be the currency of the moment for security researchers too.
[00:39:11] Speaker 1: Max, let me follow up on an important thing. We'll come to this with C.C. Rouse and other economists in that panel. What's happening to your workers? What do you expect there to be? You use the word upheaval. Algorithms that follow the Fed just traded several basis points on upheaval coming up at the Fed conference. What does upheaval mean to your company and to the enterprises you know in terms of your own labor force?
[00:39:38] Speaker 2: So I think in companies, I suspect, like all of ours to a varying degree, but probably not too different, upheaval is mostly a positive motion. We are primarily trying to figure out where to apply these gushers of productivity. You have some degree of budgeting, some degree of capital allocations we have to engage in, even if productivity is just running off the charts because tokens are free. We're not, you know, electricity is free, et cetera. And so in my world, prioritization has become ever more important. The sort of what is the appropriate risk to take? You know, Black Friday, Cyber Monday is upon us. That's the sort of a hunker down part of the year for the two of us. You have to have your five nines or six nines or seven nines because retail is expected. And there's a small army of people saying, I have this great idea. Let me ship it. And it's a human. It's not AI that says, I thought it through and that's the right thing. This is the risk I want to take. And the risk is particularly high in the moment before Black Friday. So on the flip side, if you ship it and it works, that's another five basis points of growth that you wouldn't have had. So I think upheaval is mostly judgment and attempts to decide what is the right bet to place right now. And that's always been our job. That's, you know, the thing that brings us to work every day. But the pace of decision making is increasing. And I think it's being propagated down where every single human that's sitting in front of this rapidly increasing capability doesn't decide, well, do I write five more lines of code or is this time to call it, you know, Friday night for, you know. It's more like, well, do I put out this feature? It'll take me and my agents several hours and then we'll ship it and hope it doesn't break anything. And so I think the tension we're all feeling is this productivity needs to be directed and well managed lest it causes actual upheaval in our code bases and introduces more bugs. The one thing that AI unfortunately does right now, it writes code that is not entirely perfect yet. And therefore, all the bad guys probing for holes or finding holes that were introduced, of course, by the very same AI they're using.
[00:41:44] Speaker 1: So, Nikesh, I'm going to start with you with a question, go down the line. So we've limited most of this discussion to thinking about our day job at the Fed. So we're thinking about implications for the U.S. economy. You three are leading important U.S. companies. What about the diffusion of the rest? What about the U.S. versus the rest of the world? How should we think, economics isn't zero sum, but how should we think about what our enterprises are doing, what our individuals are doing in the U.S., and whether that's being replicated among big trading partners around the world, and what this does for U.S. economic growth relative to our trading partners?
[00:42:22] Speaker 4: Well, let's start with the comment you asked, Max, about upheaval or how we're deploying it and how other companies are going to be deploying it. I think what you're noticing and seeing is a gradual reskilling of our workforce. It's not explicit, but it's gradual. Every employee at Palo Alto is now experimenting with AI. We're watching them. This is a technology nobody knew. You can't go to school and learn how to use it. You literally have to be naturally curious. You have to be naturally creative and learn it. Now, you're seeing that across the board. It's not as uniform as you'd like it to be, but you're seeing that across the board. I think by virtue of the U.S. being the center of the AI storm in a positive way, you think there's a lot more ability to pivot, change, embrace. Every CEO is interested in seeing how their organization can embrace this new technology. So I think you're seeing a bit of a groundswell desire. I'm not suggesting that everybody will make it across the other side, but I think the desire to embrace, the desire to understand, the desire to absorb it and leverage it for economic gain or organizational efficiency is the highest in the U.S. I think your question around the U.S. versus the rest of the world, I think we have to separate two parts. I think we get over-indexed on who's producing the fastest, smartest model versus who's diffusing it most effectively and where are we seeing real economic gains in enterprises which have competitive advantage as a consequence of using it effectively. I think, yes, the U.S. is the center of where a lot of these models are being produced or where this intelligence is being produced, and the mass of that is still to be seen, whether that results will find out in anthropic. But it's still to be seen in terms of is $50 billion every time the right number to go build these models. Most countries in the world can't afford to create this intelligence of their own. So what you're going to discover, there will be a few centers around the world which will build the intelligence. And the question becomes, let's assume that, let's take it for the positive side and we can debate the, you know, how this intelligence is in the hands of people who could use it for bad things. But let's take the positive scenario, which is 98% of the scenario. There, I think we're more than likely to see the diffusion in the United States faster than anywhere in the world. I think that will result into efficiency gains. That will result into margin gains. That will result in global competitiveness. So I think, you know, this is ours to lose from that perspective. I just think we get too hung up on who's building the next fastest model, who has it. I think the diffusion here is the most in any world still.
[00:44:42] Speaker 1: Patrick, let me follow up with you on the same question. So we care in the United States about our absolute level of growth, our potential productivity, potential growth. Is Nikesh right that the delta between American growth and the rest of the world could gap out, unlike any period that we've seen? Or is that is that unlikely?
[00:45:05] Speaker 3: Not only do I think it's likely, but it's happening. And we see it directly in our data. AI usage intensity is higher in the US than elsewhere. Obviously, on the supply side, an overwhelming majority of the revenue and the firms and the creativity and the innovation is occurring in the US generally and indeed in the Bay Area specifically. And, you know, as you mentioned in the introduction, I grew up in Europe. And, you know, the European economy is-- when I left Europe, the European economy in aggregate was, you know, not wholly dissimilar in magnitude to that of the US. And, you know, we don't need to be mean here. But suffice it to say that there's been a considerable divergence since then. And again, every indicator and sort of leading datum that we have access to would suggest that that divergence is going to increase. And then, generally speaking, you know, I guess in econ, you know, you always want to avoid corner solutions. And, you know, on the one hand, there's sort of this sense with AI that maybe like the international, at least in its early days, it's sort of showing up everywhere. But in the productivity statistics, now you mentioned that, well, maybe we're seeing some kind of wiggles of hope and so forth. But it's certainly not a dramatic trend break or not, you know, the peaks that we saw, you know, in the aftermath of the Second World War and so forth. So, in that sense, some degree of, you know, humility or caution or something would seem to be warranted. And then, you know, at the other extreme, and I mean, something I would hope perhaps to convey to all of you is the magnitude of that which is contemplated in -- San Francisco doesn't really have, you know, smoky back rooms. It's more in kind of desultory and drab conference rooms. But that which is spoken of in these drab and desultory San Francisco conference rooms is two people I know have a -- and, you know, these are both people in positions of, I would say, considerable influence over how AI unfolds from here. They have a bet between them as to whether in some year of the next 10 years, global energy use doubles in a single month. Now, you know, you don't need me to point out that is not something with historical precedent or that, you know, has occurred hitherto. But that is at least what one of these individuals feels sufficiently confident in that this VAT has been placed. And so you sort of have these two extremes of, well, we can't exactly see it in the productivity stats and global energy use doubling in a single month -- some month of the next 10 years. And I would say Stripe sits somewhere in between where we do see it. Again, we are observing this dramatic, arguably exponential acceleration. And on the relative framing of your question, U.S. versus the rest of the world, I would say the U.S. is in a definitively preeminent position today.
[00:48:21] Speaker 1: Max, join in on that question, the U.S. versus the rest of the world. And give us your sense about whether that gap is growing.
[00:48:31] Speaker 2: It certainly feels that it does. I agree with you. I think that it's a little bit earlier than the numbers would allow us to sort of declare. But just looking at the sort of external factors, perhaps the ones we're very used to, I think European Union is still debating the correct framework for debating about AI. They're sort of a -- they're a couple of meta levels away from accepting that this is happening, whether they like it or not.
[00:49:02] Speaker 3: We're the -- excuse me, we're the regulatory superpower.
[00:49:05] Speaker 2: As someone who also grew up in Eastern Europe, you know, I take -- You've seen regulatory superpowers? I take a lot of pride in being an American by choice. And no -- maybe one of those strands is the -- my love of bureaucracy being cut. I think the -- present company excluded. I think -- I think we are seeing extraordinary gains that are going to add up to some pretty -- some pretty staggering -- I'm not sure about the doubling of energy use, but it's very clear that there's a tremendous number of tokens that get consumed if they were cheaper. And so it's incumbent on us to build access to those tokens, therefore access to energy, compute, and markets are really, really good at figuring out the intersection of supply and demand. So I'm not even a little bit worried that we're going to over-produce or under-produce the number of GPUs, TPUs, you know, your favorite chip. That will sort itself out in time. But I think the time to make those investments, however perfect, is right now. And the fact that U.S., generally speaking, independent of your political spectrum, is embracing that, and it doesn't seem to be a tremendous party divide around the question of should we be the world leader in this all-important technology, I think gives me a tremendous amount of hope. I think if you sort of look at Europe, the conversations about, you know, we have to stop this thing. This is -- it's going to take us 10 years to figure it out. Like, 10 years from now, maybe we'll be doubling energies. I think our obvious adversaries/trading partners in China have exactly the same idea, and I think there's an enormous amount of future tension that we're going to have to navigate, hopefully in peaceful and primarily trade ways. And I think that's sort of it, which maybe is the sort of -- the world that we were always going to have, but it's coming sooner than we think.
[00:51:08] Speaker 1: So we've got two minutes. I'm going to ask one question. I'll just add a statistic, which is imperfect. I think you guys will concur. Token prices are falling about 50% per quarter and are falling at an accelerating rate. So on the question about who gets the gains, boy, it sure looks like markets might work there. Let me end with a final question for you. I began mistakenly by beginning with a great Russian novelist. So let's end with a depressing conclusion. This has all been way too optimistic. We central bankers are two-handed, and I think that I've lost control of the three of you because you think this is easy and we're going to win. So I'll start Nikesh with you. What could go wrong? I know, Kevin.
[00:51:54] Speaker 4: You're not going to drag me down that path. I know you want to be unhappy at the end of this, but we don't want to be unhappy. Look, we've powered through a pandemic in society. We all thought it was the end of the world as we knew it. And here we are sitting in the room and no masks and life is wonderful. So I don't think there's anything we'll not be able to get through. I think this is going to be a phenomenal technology. You have to look at the positive sides of it. All the different tasks. Do I want my 20-year-old kid to go become a customer service agent and say, how can I help you? Your internet is broken? That's not a job I want in the future in my society. So the fact that this technology will allow us to unleash the creativity at max set and get people to focus on more important things. We're seeing a gradual reskilling. This is not going to happen as a shock to the system. This will happen as every other revolution has happened, as over time people will reskill and you suddenly discover 10 years now when you look back and say, holy shit, that used to be a category of jobs that doesn't exist anymore and we seem to be fine. So it's going to happen. Are we going to have some little bumps in the world? For sure. Are we going to have a cyber event that's going to be there? That's going to impale something for a while? Possibly. Are we going to have some system that doesn't work for a weekend? For a week? Possibly. But they're not existential.
[00:53:05] Speaker 1: To you. When we've got to do quick answers to quick questions, what could go wrong?
[00:53:10] Speaker 3: Well, being an immigrant, I felt like I should read Democracy in America. So I just finished it. And at the very end, during the last chapter, De Tocqueville describes how the sort of failure mode that he can imagine for democracy is a kind of soft despotism where life becomes too comfortable, the state provides succor and comfort and a kind of suffocating tranquility that inhibits and enervates and does not support the cultivation of a great society and ambition and achievement and great works by its individuals. So many people in the AI sector expect drop-in remote workers within the next 12 months. So you can have a little critter. It's on Slack. Join the Zoom meetings. You can do anything that somebody with access to a monitor and keyboard can do. And I think there's a question for us as to what the -- I mean, there's all those sort of doom scenarios and cyber hacks and bio-risks and all the rest. But there's also the countervailing question or set of questions about the positive scenario. Let's just say the catastrophes do not unfold. We manage to tread the perilous path. I think there's a question about what the utopia looks like and how it is, you know, what the shape of the society in the future that we want to inhabit is, what the nature of work is and so forth, and how we avoid de Tocqueville's soft despotism.
[00:54:45] Speaker 1: With a final word with you, Max.
[00:54:48] Speaker 2: I'll try to be less positive than the -- although, for the record, I'm extremely positive on this topic and many others. So I think this -- the thing about AI revolution is, unlike most other, at least in my memory, revolutions, the incumbents, in particular incumbents flush with extreme access to capital, are very aware of it, aware of it from the very beginning. And so the disruption by the plucky start-up is not actually given enough time to go reconfigure the industry. And so maybe if there is a concern to be had, it's that the incumbents to correctly invest in the opportunities, further solidifying what sometimes can be a monopolistic or oligopolistic grip on the industry. And so what could go wrong? We could end up with a hundred greatest American corporations completely dominating our economy, and not a whole lot in the middle tier. I do believe that the long tail, the creativity being through the solopreneur, the bedroom creator of a unique product that one person in this room will buy, but that's okay because it's a profitable purchase, will endure and better than endure. And so there could be a weird economy where we have this incredible number of businesses starting on Stripe that never really get to Fortune 100. And Fortune 100 just obliterates the Fortune 10,000 because they can't, because that's a productive purchase for them. I'm not sure that's a terrible thing, but it's at least a weird thing that the market may or may not be able to quickly turn around. That said, I think every person in that infinitely long, long tail, the thing that makes us human is agency. We look at things and say, hey, let's fix that. That looks stupid. Like, I don't want to live in a world that looks like that. So the little guys will start repopulating the hollowed out middle, should my scenario unfold, I think. But in the meanwhile, we may see sort of a boom of M&A that may or may not be well advised for things like consumer pricing.
[00:56:45] Speaker 1: So on behalf of the Federal Reserve, to the three of you, we're grateful. Thank you. And we'll turn immediately to the next panel to figure out the implications. Thank you. Thank you.
[00:56:57] Speaker ?: Thank you.