[0:00:05] Good afternoon and welcome to the 2026 Per Jakobson lecture. Per Jakobson was an internationalist in every sense of the word. At the League of Nations and the Bank for International Settlements, he saw the collapse of international trade and cooperation during the Great Depression. But he also played a role in reviving it.
[0:00:38] After World War II, again at the Bank for International Settlements and then as Managing Director of the International Monetary Fund. He helped embed the rules-based system that encourages international cooperation and globalization at a time when, very much like today, international cooperation was at a great discount.
[0:01:05] The Per Jacobson Foundation was set up in his honor and is supported both by the BIS and the IMF. The Per Jacobson Lecture has always been about an issue of tremendous importance to the world. And today's lecture is no exception. We will talk about AI, a technology which promises a lot, but also necessitates global cooperation if we have to spread the benefits widely.
[0:01:36] While minimizing the risks. Let me turn the podium over to the IMF's deputy managing director and president of the PGAF, Nigel Clark, who will introduce the talk and the lead speaker. Thank you, Raghu. Members of the platform party, ladies and gentlemen, good afternoon and welcome to this year's Per Jakobsen Foundation lecture.
[0:02:10] As Raghu mentioned, for more than six decades, this lecture has been a fixture of the IMF's annual meetings, and it carries the name of the IMF's third managing director. And its purpose has always been straightforward, to put the most consequential questions facing the global economy in front of the people responsible for answering them, and to do so in a spirit of open and rigorous debate.
[0:02:39] When we sat down in January of this year to think about who we should invite to deliver the 2026 lecture, and I see members of the Perry Jacobson Board here in the front row, one thing was clear to all of us. Artificial intelligence is a force that will change the world, and its implications for economic policymaking are profound. You only had to listen. And even at these annual meetings this week in Bangkok,
[0:03:08] Every conversation in bilateral meetings, in multilateral meetings, in seminars, on corridor talks, and back at the hotels, AI has loomed large. If it wasn't the subject, it was mentioned. If it wasn't mentioned, it was the subtext. And so long before Mythos and Fable and Astra, and long before Hugging Face and AI Swarms became a part of our everyday lexicon,
[0:03:38] The Per Jacobson board had concluded that artificial intelligence had to be at the center of this year's lecture. But I'll be candid. This raised a little bit of a question for us. You look at the roster of Per Jacobson speakers, central bank governors, finance ministers, economists, presidents, over the 62 years of the Per Jacobson lecture,
[0:04:05] A co-founder of a company that is not even five years old has never given the Per Jacobson lecture. And so we ask ourselves, honestly, how exactly is this going to work? So yes, from a historical perspective, a backward-looking perspective, this year's choice is unusual. But frankly, so is this moment. And furthermore, the Per Jacobson
[0:04:33] Lecture is not meant to be looking backwards. It must always be looking forward. The issues are many. Capability, alignment and safety, interpretability, access, adoption, diffusion, productivity, distributional gains, how you tax, how you spend. And technologists and entrepreneurs probably can't answer the questions on their own, but neither can policymakers.
[0:05:04] We need both in the same room at the same time. And that is the spirit of today's lecture, an opportunity to hear an unvarnished view from the frontier and an opportunity to test it against the perspectives of economic policymakers and scholars. When we reached out to Jack,
[0:05:29] We suspected that it might have been a little strange hearing from us, but to their credit, Jack and his team, who are all sitting here, immediately embraced the opportunity, recognizing the value and the relevance. And on a personal note, I had the great pleasure of meeting Jack very briefly a few weeks ago and hearing him speak in California.
[0:05:57] By chance, and a very different setting to the one here today, no podium, no title cards, just a knapsack and a pair of jeans. And it gave me a glimpse of the man himself, a self-deprecating sense of humor, a forceful intellect, a warm and engaging personality. And as I suspect, you'll see all three of those this afternoon, though perhaps not the knapsack nor the jeans.
[0:06:27] So it's a real honor to welcome Jack Clark, co-founder of Anthropic, as this year's Perry Jacobson Lecture. Please give him a warm applause. So Jack's lecture, Inside the Event Horizon, Why Everyone Should Take AI Seriously, he will be looking beyond the current generation of AI models to the next phase of development and the profound questions it may raise.
[0:07:01] And to bring the economic and public policy perspective to the discussion, he'll be joined by Dr. Nagesawaran, the chief economic advisor to the government of India. Please give him a round of applause. And Professor Yan Shen of Peking University. Please give her a round of applause. And the discussion will be moderated by Professor Raghu, we call him, Rajan of the University of Chicago and formerly chief economist of the IMF.
[0:07:38] Please give him a round of applause. And without further ado, a big round of applause to welcome Jack to the podium. Ladies and gentlemen, Jack Clark. Thank you very much for that no pressure introduction. Thanks for having me. It's a great privilege to be here. And to start, I think I should list my relevant credentials. I am a human being. There's one of me, and it's taken me almost
[0:08:18] 40 years to build the experience to give this speech. And I'll come back to those credentials later. My whole career has been defined by finding information that seemed important to the future of the world and telling the world about that information. Two decades ago, I started as a journalist focusing on data centers. And through that work, I became so obsessed with artificial intelligence that in 2016, I left journalism to join OpenAI.
[0:08:47] And then I co-founded Anthropic a few years later. And ever since then, I've been trying to tell the world that one day AI systems would become so powerful they could alter national security, change the economic trajectory of the planet, and perhaps eventually gain the ability to build themselves. Well, that moment has arrived. AI is now changing national security. AI...
[0:09:16] is changing the economic trajectory of the world, as we've heard in so many sessions here. And it is contributing to its own development. And many of you know this. You are having to track AI as one of a multitude of destabilizing forces that is shaping your world and your work. And what I'm here to explain today is that AI is on a course to bring about even more change than what we are currently expecting, and we should prepare accordingly.
[0:09:46] And I'll explain why. For many years, researchers and science fiction authors have forecast a future in which AI systems would drastically increase the pace of overall scientific and technical progress, making things move so quickly that it could be hard for humans to keep up. They call this the singularity. And I believe that this year we've gained convincing evidence that it has now begun.
[0:10:17] Verna Vinge forecast the singularity in 1993. He said, when greater than human intelligences drive progress, that progress will be much more rapid. In fact, there seems no reason why progress itself would not involve the creation of still more intelligent entities on a still shorter time scale. And he described the singularity as a point where our models must be discarded and a new reality rules. The singularity is here.
[0:10:48] The implications will reach across every aspect of life. But the most important implication is a very simple one. We should expect that in the coming years, the whole pace of progress of not just AI, but science as a whole will speed up. And this will feel confusing and disorienting and will bring with it great wealth and potentially great peril.
[0:11:15] So today I'm going to concentrate on the first point without discounting the latter. I'm going to talk about the benefits of rapid AI progress and what we need to do to share in those benefits together. So many of the ways that humanity will benefit from AI come from science. I believe that AI-driven scientific invention will change the course of the global economy, alter the ways countries plan and build their own infrastructure,
[0:11:43] Make us all healthier and able to spend more time with our loved ones. And at the same time, cause us to ask deep philosophical questions about choices we should make about the technology's continued development. So let me develop my claim. For the last half century, most countries have dealt with falling researcher productivity, while researcher headcount grows rapidly. Ideas and inventions appear to be getting harder to find.
[0:12:13] But something important has changed. We've now built AI systems which have the ability to augment, extend, and automate aspects of human labor. This has most obviously happened in software development, where AI systems are now more capable and efficient than the vast majority of the world's software engineers. And as a consequence, the field of software development is changing. A single person is now able to write software for anything they can imagine.
[0:12:42] Creating things that previously would have taken entire teams of humans months of work. We've seen concrete evidence of this inside Anthropic, as you can see here, where our researchers now produce almost 10 times the amount of code that they did for the first four years of the company's life. And that coding increases basically directly tracks advancing AI capabilities. These new AI capabilities, though, will bring new challenges.
[0:13:12] The coding improvements have meant AI systems have got much better at cybersecurity and at cyber attacks, as we've heard, leading the world to race to deploy these powerful capabilities to create cyber defenses, while bad actors race to use these systems to carry out hacks. And the important thing to notice is that these things are bound up inseparably. The advance of AI systems in a general area, like coding,
[0:13:38] creates positive and negative impacts in domains of international relevance like cybersecurity. This is a taste of things to come, of progress yielding new capabilities and new challenges. And these capabilities will not stop at coding. The same thing will happen to science and to robotics. It's already begun to happen. And it means some important properties of the global economy may change, like the rate of innovation,
[0:14:09] All the frequency of which we invent entirely new general purpose technologies that create new industries around themselves, like the internet, phones, and artificial intelligence. So what causes me to believe this? And what am I going to do to help you believe it too? In 2026, we've got the first signs that AI systems may be capable of doing science at the frontier. In mathematics, AI systems have been used to propose solutions
[0:14:39] to long-standing mathematical problems like the Navier-Stokes Millennium Prize, the Jacobian conjecture, and several open ERDOSH problems. And in many cases, AI systems were responsible for core insights that made the solution possible. In biology, researchers have used AI to design viruses from scratch and to identify new drug candidates for testing. And there are already drugs in trials today which have been in large part designed by AI systems.
[0:15:08] And in robotics, AI systems are now increasingly able to autonomously operate laboratory equipment, making it possible for us to further automate experimentation in domains from biology to material science, making progress faster in areas like batteries or fuel cell development. Taken individually, each of these examples is exciting, but the profound fact is that they all relate to the same underlying technology.
[0:15:37] And for most of these use cases, the same single underlying frontier model. We now have singular entities, AI systems like Claude or ChatGPT or Gemini, which can carry out work at the frontier of multiple scientific disciplines. We have virtual polymaths. There's more quantitative evidence that we can look at as well. One such metric is the score of AI systems at a test called Humanity's Last Exam. I did not choose the name.
[0:16:09] So this ominously named test was developed in response to the profound advances seen in prior years by AI systems at a variety of difficult and varied scientific challenges. And it covers topics as varied as chemistry, computer science, mathematics, classical literature, and ecology. And when it launched in January 2025, top-scoring AI systems obtained a score of 8%.
[0:16:35] As of this week, the highest scoring system is Claude Opus 5.5 with 67.7%. Another example is Frontier Math, a test released in November 2024 that sought to assess how well AI systems could solve hundreds of challenging math problems, curated and evaluated by professional mathematicians. When it launched, the best scoring systems got 2%. Now they get 94%.
[0:17:05] And more broadly, the point I want to make is it looks like this everywhere. I have a file on my computer where I've collected 100 graphs that I've been building just in the last week. I pulled these down this morning before giving the speech. If you pick different parts of science, you see the same pattern show up. Progress is accelerating, and it has started to accelerate in the last two years, sometimes just one. So I'll close this evidence section with a story.
[0:17:36] The latest example comes to us from a paper published last week. This is what I do for fun. I read papers, well, between giving conference talks. So this paper is called Theophys Eval, and it's a paper that tests out how well AI systems can solve unpublished problems in theoretical physics. The 50 questions in the paper were compiled by physicists around the world, unpublished answers, their own research.
[0:18:04] And unsurprisingly, AI systems got high scores. But what is surprising is how they got the high scores. So now I'm going to quote from one of the authors of this test, Indian physicist Suvrat Raju, who wrote a post commenting on the results on his Facebook. I don't know of any individual who can solve any significant fraction of our final collection of problems without AI assistance, he said.
[0:18:29] Our initial evaluation suggested that GPT-6 Astra's score was in the late 80s, followed closely by Opus 5.5. But as we started to investigate the mistakes that the models might have made, the models, especially Astra, insisted that they were right and that the experts' supplied answers were erroneous. When we went back to the experts, it turned out in multiple instances that the experts found subtle errors.
[0:18:58] Buried deep inside their calculations. And as they revised their answers, Astra's score started to inch up and it finally ended up with 98%. 49 out of 50 answers. So think about this. Some of the world's scientific experts built a test out of problems an AI could never have read the answer to. And when they tested the AI, not only did it score well,
[0:19:25] But it got a higher score by finding errors in the answers that the humans had worked on. There's a scene in films, usually two-thirds of the way through, when the young protagonist finally proves to their mentor that they have what it takes to go out into the world and succeed independently. It's a narrative trope when the teacher becomes the student, or the student becomes the teacher. That's what happened last week.
[0:19:54] So, you know, why am I telling you this? Why does it matter that AI systems are getting better at science? I believe it's because doing true frontier science, being capable of making breakthroughs that extend the frontier of knowledge, implicitly tests for one of the last frontiers of AI capability, creativity. My view is that if AI systems can creatively problem solve to succeed at science, we should expect this to be a sign of their ability to generalize into many domains.
[0:20:24] Including certainly AI research itself, which could have an explosive effect on innovation. And so if AI systems get better at science, this will strongly drive the growth and shape of the future economy. Which means that you can make decisions today to set up your nations and people to benefit from this trend in the future. So let's talk about growth. One of the main sources of growth is research yielding new ideas.
[0:20:54] And one of the things that drives growth is our stock of human capital, the people all over the world who ask questions and conduct scientific research and help the world move forward. But we also know that, as some economists have said, good ideas are getting harder to find. The cost of breakthroughs is going up, and we've had to pour an ever-rising amount of research into efforts to keep going at the same pace. So my question to you is, what if this changed?
[0:21:25] This technology allows us to change the number of researchers on the planet and how long it takes to train them. As I said at the beginning when I listed my credentials, there is one of me, and it took me about 40 years to get here. What would it mean to be able to run hundreds or thousands of copies of me simultaneously? What would it mean if new versions of me could be created instantly rather than waiting 40 years?
[0:21:52] And how would my work change if I could make copies of myself and order them to do work under my supervision? That is what the next few years will be like for large swaths of the economy. We'll be able to augment human labor in frontier disciplines like science in a way that we've simply never been able to do before. And this will likely change the rate at which we find good ideas, because we'll be able to add hundreds or thousands of times as many researchers to any problem.
[0:22:20] Force multiplying our human scientists. We'll also be able to do this quickly. If you wanted to train a new generation of researchers in a certain area, it might take you 10 or 15 years of targeted education and millions of dollars for a small cohort. And on the other hand, these AI systems can be created immediately in arbitrary numbers as long as there is enough computing capacity. And they can be used across domains to benefit from new scientific insights.
[0:22:50] Near instantaneously. And they can combine existing knowledge in new and surprising ways to make discoveries. So this means we can rapidly augment all of us on this planet with virtual scientists that can work alongside us. So in the coming years, every human scientist on this planet will become more important and powerful as they gain the ability to supervise hundreds or thousands of like-minded experts on any problem they find interesting.
[0:23:20] So the team of economists we have at Anthropic have done modeling here where we've tried to simulate different versions of the economy according to different trajectories of AI technology. And if you assume, as people at the labs do, that AI models will be able to autonomously complete a broad range of tasks and that their diffusion into the economy will increase in coming years, then under these singularity-esque conditions, one model
[0:23:48] predicts a vast economic transformation with US GDP potentially reaching 44 trillion by 2030, which sounds very outlandish, I know. But it's implicated by the technology continuing to get better. And if GDP were to grow this much, then economic doctrine suggests that new industries and inventions would need to have been created to power such a major economic expansion. And that's exactly what is implied.
[0:24:18] By this notion of having AI systems be able to perform autonomous scientific inquiry. So what do we do about this? How do we set up the world broadly to benefit from these advancing capabilities? And how do we balance this opportunity for progress with appropriate caution? We need to both plan for success so that humanity can take advantage of this technology and use it to create abundance.
[0:24:48] And we need to be prepared to make wise choices about how to use the technology itself. Just because AI will be a capable scientist, that does not mean we should use it for absolutely everything. On planning for success, I'll start with a story from my own experience. As I said, for the last 10 years, I've gone to governments around the world and I've tried to talk to them about AI progress. And I told them that because AI systems were getting so much more powerful and useful,
[0:25:18] The single best thing they could do to position themselves to take advantage of this coming revolution was to build enough electricity capacity and data center space to be able to host the kinds of powerful AI systems discussed here and to invest in the AI supply chain. Not many policymakers listened. I think now we deal with the consequences of that. In today's world, data center capacity is concentrated in the United States.
[0:25:47] China and increasingly the Middle East. And the supply chains are concentrated and are under such contention that prices are going up. Now that build-out remains important, but it was just the first phase of the economic and scientific revolution implied by AI's continued advance. Now there is more to be built, and I believe more countries can participate in this if they choose to. Because the next phase of the AI revolution is about the expansion into the physical world.
[0:26:17] Specifically by having AI systems conduct scientific experimentation and operate robots, sometimes together. So therefore, just as we did with compute and power, we will want to use much more scientific equipment and robotic systems in the world, or we'll run very quickly out of capacity and we'll see history repeat itself. So we need to build capacity now to let us take advantage of what's coming.
[0:26:46] According to UNESCO estimates, there are approximately 10 million scientists in the world. And even with only 10 million scientists, we have too few humans to oversee the work of today, too few laboratories to do material science analysis, too few biological labs to develop and test new drugs, and too few large pieces of scientific equipment in space or on the Earth for the experiments we want to run.
[0:27:15] Tell me what happens when instead of 10 million scientists, we have a billion virtual scientists directed by our own 10 million or 10 billion. That's the world predictably coming right for us. And we have to get ready. If we do get ready, we are going to be in a better position to translate the intelligence of AI systems into gains for us all. Longer, healthier lives, more energy, more food, more wealth.
[0:27:44] And just as with the data centers, the nations which build and develop this supply chain will be the ones that benefit most. That's a choice. It's an opportunity for everyone. If you're a developed country with a rich scientific ecosystem, you have an opportunity to make all of your scientists more productive. And if you're a developing country that supplies the basic ingredients into the AI and scientific enterprise, like raw materials, then you have an opportunity to leapfrog.
[0:28:15] Scientific testing facilities that, paired with human scientists and digital scientists, will let you grow a scientific community. And you can continue to develop your contribution to the supply chain, which will become more valuable. But infrastructure will not be enough. If discovery becomes abundant, we need to avoid bottlenecking ourselves downstream. An avalanche of proposed new drugs is of little use if only a handful can ever be tested.
[0:28:42] And the policymakers and their colleagues in this room will need to help us find ways to accelerate our ability to move discoveries out of the lab and into the clinic, the factory, and the field, and across and between nations. And the better we prepare for this world, the better able we will be to translate the gains of AI to gains for the world. But finally, we need to prepare ourselves as a global community to make choices about the future.
[0:29:13] Because as I've said, there is a risk here. Being able to use AI for science opens us up to the misuses of science as well, biological and cyber weapons and things we cannot yet imagine. And we have many choices to make, and it's only by recognizing this technology for what it is, for how powerful it is and how powerful it will become, that we will see the urgency behind making these choices. The first choice is about where we direct it.
[0:29:44] Just because we have an AI system that can do science, that doesn't mean that it should be left to do science unattended. One such use case is biological research, where we want the strictest human oversight and transparency to maintain safety. Another is AI research itself, where labs like Anthropic will need to make decisions about how much oversight we apply to AI-driven research.
[0:30:09] This could let us make more progress on the technology, but also introduces risks. And we will need to move extremely carefully here, because otherwise we could lose control of the technology. We will also likely need rules to govern this, to govern us. I was in Tokyo earlier this week, and a student at the university there said they worked in the physics department. And they said AI was already beginning to be able to do many experiments that they worked on.
[0:30:41] What experiments should they spend time on as a person? And what experiments should they let the AI system do itself? This is a subtle question to which I don't have the answer, but it is a question each discipline of science will need to answer quickly. The second choice is about how much we need to understand about what these systems do. As we've seen this year with the rise of AI and mathematics, we are already dealing with a world
[0:31:10] Where the advances made by AI systems are sufficiently creative that we as humans are now spending time trying to decode them. Trying to understand what they mean and how to apply these insights for ourselves. How much should we understand? Again, the answer here is going to be subtle and it relates to fundamental questions of meaning in the age of the singularity. For what parts of scientific progress do we want true understanding? We're only a...
[0:31:40] Few short years from AI systems not only doing science, but inventing their own scientific equipment built for the ways AI systems work rather than human minds. Do we need to understand that equipment? Should we? This is another choice. And finally, taking AI's progress seriously means we should give ourselves license to imagine a fundamentally different world. When I was a teenager,
[0:32:10] I read a lot about the world, and I was afraid, as I suppose lots of teenagers are. I believed that the 21st century was going to be difficult, that population growth, climate change, resource extraction, and other crises would add up to a Malfusian collapse of the human species. But after rising to invent computers and the internet, we might consume ourselves into the ground.
[0:32:37] And then over the subsequent years, I saw time and time again how technology allowed us to alter these odds. How solar panels came along and became much cheaper, much faster than people expected. How our growing understanding of the human body allowed us to invent cures for things that had killed many people, including some of my own family. How we could do things that I had never even imagined but find utterly amazing.
[0:33:06] Like how we are using AI systems to decode the speech of animals or how companies have built rockets that take off and then land back on this planet. I work on AI because it can help us do more of that. And I work on AI because whether it does that is about choices we make. Choices people here in this room will make about how we direct it and govern it. Choices I need to make.
[0:33:36] I work on AI because it's the most important technology our species has built. And it's a technology that, if we are able to use correctly, will let us deal with problems that might otherwise doom us as a species. But it's also a technology that has danger woven into it. For us to get through this, to chart a course through the singularity, we're going to need to work together to create unusual and serious rules for this technology.
[0:34:06] And the companies building it. And to exercise the hardest and I think most noble quality of people, which is restraint. We will need to build infrastructure to take advantage of this technology, but we will also need to go slowly and deliberately on developing the technology itself to pace or otherwise control its development. And we need to constantly differentiate between the things we can do and the things we should do.
[0:34:35] And we'll need to invest today in the scientific capabilities to make that progress possible. So I'm reminded in all of this by the great Carl Sagan, who said in The Pale Blue Dot that the humans who one day may reach the stars will be a species very much like us, but with more of our strengths and fewer of our weaknesses, more confident, far-seeing, capable, and prudent. That's what...
[0:35:07] AI demands from us. It's what AI demands of me and companies that I represent. And where Sagan thought we had hundreds of thousands of years to achieve this, to become a better and wiser people, we instead have a handful of years, less than a decade easily. I believe we can do it, but it's only by treating AI with the seriousness it deserves that we will succeed. Thank you very much.
[0:35:45] Thanks very much for that wonderful speech. Let me call on Professor Yan Shen for her discussion. She's professor of economics and chair professor at the National School of Development, Peking University. And she also serves as the executive director of the Institute of Digital Finance at Peking University. Please. OK.
[0:36:21] So it's my great pleasure to be here and hearing the keynote talk coming from the country that with the largest population on earth. I cannot stop thinking about the impact of AI on human. And so I'm sharing with you a little bit of our study, bringing you the perspectives of China's AI development. I call it a paradox.
[0:36:48] Because there are good signs, opportunities, a lot of them. And then there are also a lot of challenges that we deeply observe within China. To speed up the talk, I'm building up a little bit of my framework in thinking about the situation. So from AI investment, we just hear that from AI investment,
[0:37:15] To sustainable growth, there has to be a long road. So we just heard that AI investment indeed has brought so many groundbreaking technology advances, and we really need to think hard about the future. But from AI investment to sustainable growth to the human, there are several other links that we need to further dig deep into. Number one,
[0:37:43] If we skip the first one, is AI a bubble? Let's assume that that is not a problem. So AI is going to be really a real technology. But then the next step is, will AI turn into adoption, wide adoption across countries, across globe? And the next step is, if there is wide adoption, will this adoption turn into greater productivity?
[0:38:12] And then the next step will be will this productivity bring real gain and then how this gain is being allocated and diffused across the globe. And the next step will be if there's indeed gains diffused, distributed wisely, then can it really stimulate demand and then it finally turns into sustainable growth. That's something that we need to think about.
[0:38:41] So here we go from the AI investment. So the first step is we're comparing the AI investment in US and China. So on the left hand side, we look at up to January 2026.
[0:39:00] The comparison of investment of U.S. and China. And then up to this year, half year of this year, China also cynically reports that private investment is about 250 billion RMB. But compared with the U.S., the right-hand side picture is by Wall Street Journal of the already spent 600 billion and the committed 3 trillion U.S. dollars. The China's investment gap.
[0:39:29] Based on my study, it's actually widening. So in terms of adoption, of course, AI has been the fastest adopted technology compared with earlier revolutions. Both in US and in China, we have observed a fast adoption rate. So the most recent number by CINIC is that the generative AI users in China
[0:39:57] For the first half of 2026, it's already 700 million. And for large enterprises designated up to a threshold, that adoption rate is about 30%. And if we look at the adoption from the computing supply and demand side, we observe fast growth too, about 170% year on year in terms of computing power, about 1,400
[0:40:25] Times of daily token consumption compared March of 2026 with March of 2025. And the compute demand for the first quarter already increased over four times. But if we look at the impact on productivity, the story is mixed.
[0:40:51] So this is a study from NSC. We have three main message to deliver here. Number one, on average, we indeed observe that more AI usage adoption can
[0:41:05] have positive impact on productivity. But number two, there is a massive effect in the sense that those who already have higher productivity can gain more in terms of the impact on productivity. So there's a massive effect. The stronger will becoming stronger. On the other hand, we find that
[0:41:27] More AI usage, more increases in productivity is not necessarily linked with more profitability. So the right hand table shows you that the study about the impact of AI on ROA or ROE of China's listed firm.
[0:41:45] No matter we are looking at ROA or ROE, currently we don't observe positive impact yet. So we did do field study, and based on our field study, there are three main reasons that why it's not productive. Number one, firms may not exactly know what kind of AI technology is really suitable for them. Number two, there is so-called pseudo effect because of competition. Firms may...
[0:42:11] Adopt AI not because it's profitable, just because their opponent, their competitors are using it. They just cannot afford losing the game. And number three, AI development, it iterates so fast, so their earlier investment cannot, before they can make money, it's already obsolete.
[0:42:33] So that's the pain part of the media enterprises in China. They cannot afford losing the game, but it's not making money for them at this point of time. And if we look at international perspective, the picture here is also quite, oh yeah, here, sorry, this side. So this first step is about AI's impact on income, and here the effect also is mixed.
[0:43:01] So this is another study from NSD where we look at impact of AI from the household's perspective. So we measure, we evaluate for households if they are in the region with higher AI exposure and their career, interact with their career exposure to the AI. So what we observe is that for those who have
[0:43:28] Higher AI occupation exposure. The current empirical study shows that there is a decreasing impact on the labor income of the household head. And also an increase in their debt. The probability that they are having debt is increasing. So that shows that the impact of AI on net income
[0:43:52] From household level is something of a concern. And of course, on the other hand, there are positive side that we observe that AI has been widely used in China and it has a lot of impact on demand discovery. So here I'm showing you application.
[0:44:10] It's a health app. It's called in Chinese. So it's linked. It basically aims to solve the problem of China's physician shortage. So it links four players, patient, hospitals, doctors, and then farmer. So when AI is used, I'm not going to go through the details, but here we observe that it
[0:44:40] Number one, it increases the coverage of health care. And number two, it increases the efficiency and also it lowers the cost. So the positive impact is that it's going to have a promising future of helping to demand a better demand discovery and satisfy the demand.
[0:45:05] But the fifth part that I'm going to show is about the international diffusion of China's AI. Here, I'm showing a mixed pattern, too. On the left-hand side is China's AI. It's the token share of China-originated AI on one of the world's, it's called OpenRouter, one of the largest
[0:45:33] For these APP usage. So what we calculate is comparing the US share and the China share. So we can see that after March of 2026, it seems that the consumption of token of China originated APP outpaces those of the US. But on the other hand,
[0:45:57] This probably is not necessarily representative. So on the right hand side, I'm showing the APP downloaded worldwide, the AI related APP downloaded. And I'm calculating the share of China originated APP download with respect to the US.
[0:46:17] So the denominator is a total number of downloads from US and China. And the numerator is China from the whole world. What we observe is that there indeed has a spike after the appearance of DeepSeq. But it's gradually becoming slower down now. So it seems that US gains
[0:46:41] better margins. If we look into deeper about what are the APP has been used and we are comparing 2025 with 2026. So the main message here is that China's AI indeed expand over the years but
[0:47:01] The impact is that probably if we are comparing just as the keynote speech has already showing what Anthropic is doing, China's AI expansion, it seems to me that it's more or less on the lower end. We observe relative comparative advantage in areas like image and video. If you know China, there's a lot of short videos, AI generated, not using any real humans to do play anymore. And not only China.
[0:47:30] And also across overseas. Video and search assistant. But for areas like productivity and so on, for those front end, China's margin has actually decreased. So overall, I think I have 10 minutes. I'm using up my time. So the take home message.
[0:47:54] So the take-home message I have is like three points. Number one, China indeed invests firmly on AI, and we do observe a lot of fast adoption. But number two, domestically we observe that this development is something that's promising but together with pain. Because China is currently facing deflation pressure, but AI itself may exasperate.
[0:48:24] This pressure. And also, it may exasperate the situation called strong supply. It's making the supply become with a relatively lower cost, but the demand is not increasing given that the income is not necessarily increasing together with the technology. And also, from the international perspective, even though China's AI has been increased quite fast, but it's not at the frontier yet. Yeah. Thank you.
[0:48:53] Thank you. Thank you. And our last panelist is Prof. Dr. Anand Nageshwaran, who's the chief economic advisor to the Indian government. I thought it was a bit amusing that someone born in 1963, who just completed 63 years, asked to be a discussant at the 63rd Perry O'Curbson lecture. OK, let me get serious.
[0:49:31] Thank you, Raghu, and thank you to the Perry O'Copsen Foundation for inviting me. In August 2005 at Jackson Hole, Raghu asked whether financial innovation had made the world riskier. A distinguished member of the audience called the premise slightly luddite. Three years later, the answer arrived. I may own the same label today. I will accept it on one condition, though. I'm not against technology.
[0:49:59] I'm against technology that takes away human agency and against finance that hides its own risks. You might think that I'm here to rain on the AI parade. Rather, my remarks are about not letting AI rain on the parade of human agency. Jack has given us the builder's view. I offer the view from the receiving end and from an economy that is rich in labor.
[0:50:27] Finance ministries, central banks, and households will pay the bill if today's optimism on AI proves wrong. They may pay part of it, even if it proves right. I will set out five risks and then what governments could do about them later. The first risk is to the first rung of the career ladder. Firms are hiring fewer juniors. In a study of 41 countries, the junior share of the workforce fell in 23 of the 31 that could be measured.
[0:50:57] In the United States, employment of 22 to 25 year olds in the most exposed jobs has fallen about 19% behind that of the older workers in the same jobs. A firm once recovered the cost of training a junior from the routine work she did. AI now does that work cheaply, so her learning becomes a pure cost. Each firm, acting rationally, hires fewer juniors.
[0:51:25] Together, they dry up the pipeline of future experts. The people who will check AI's work tomorrow are made on that first rung today. For developing countries, therefore, the stakes are much higher. The second risk is to human capability. Researchers at the MIT Media Lab tracked people writing essays over four months with and without an AI frontier lab model and recorded that brain activity.
[0:51:53] Brain connectivity weakened as reliance on the tool grew. In the first session, 83% of ChatGPT users could not quote a single sentence from the essay they had written minutes earlier. The authors call this cognitive debt. The study is small, but it matches what teachers report. A lost job can be replaced, but a lost habit of thinking is much harder to rebuild.
[0:52:22] The damage does not stop with individuals. Woodrow Huttstock and Jessica Silby from the Boston University School of Law argue that AI also wears down institutions. When a machine makes the call, know-how is not handed down and nobody quite answers for the result. Central banks and finance ministries are such institutions. Our authority rests on judgments we can explain.
[0:52:51] The third risk is safety. OpenAI's chief scientist wrote last month that no lab has solved alignment well enough to keep scaling responsibly at maximum speed. The Bank of England reports that in tests, some models with weakened safeguards reach systems beyond their task. Calls for restraint are welcome, but a lab that slows down loses the race and its valuation.
[0:53:20] Nobody pays the lab for restraint and we have seen this structure before. It was called banking before 2008. The fourth risk lies in the financing. The US data center build out may need over 10 trillion dollars by 2032, about 3.6% of GDP a year. This year, for the first time, the big cloud firms will spend more on capital expenditures than they generate in cash.
[0:53:48] That fills the gap increasingly of balance sheet. Meta's Hyperion data center is about 90% debt financed through a joint venture whose bonds do not appear on its balance sheet. The money also moves in circles. Hyperscalers invest in AI labs. The labs sign large computing contracts with the same hyperscalers.
[0:54:13] In the first quarter, about a third of the big five's earnings came from marking up such stakes. On top sits the embedded leverage. Leveraged exchange-traded funds concentrated in AI stocks and hedge fund borrowing from prime brokers that has doubled since 2023. The fifth risk runs through public finances.
[0:54:37] Anthropic's own modeling has an extreme scenario in which the American economy is a third larger by 2030. But in that scenario, labor share of income falls from about 60 to 45%, and the wages of knowledge workers end up more than 10% lower than otherwise. The outcome that best justifies today's valuations of hyperscalers and labs
[0:55:05] is one that shrinks that income tax base. If AI disappoints instead, the Bank of England warns that fiscal forecasts assuming higher productivity will suffer and so will government debt. Either way, governments are exposed. So what should governments do? For economies rich in labor, the first answer lies in the very same endowment.
[0:55:29] As knowledge workers wages fall, wages in other occupations end up about a third higher. Work that needs human presence, hands, and judgment gains value as cognitive work gets cheaper. Think of nursing, elder care, care for children with special needs, mental health counseling, teaching, coaching, cooking, and hospitality. Think of electricians, welders, plumbers, and the technician who keeps
[0:55:59] Who keep machines and data centers running. AI can assist these workers. It struggles to replace them and that's the good news. The demand is there. Aging societies will need carers, nurses, and tradespeople in large numbers. India alone needs about 1.8 million more doctors, nurses, and midwives just to reach the WHO minimum. Skilled people are becoming as scarce for the world as some minerals.
[0:56:25] A country that trains its young well will engage the world as a partner and not as a supplicant. What holds the young back is standing. In a recent survey, 71% of young Americans said trade school carries more stigma than a degree. Only 7% thought that parents would prefer that. Indian parents feel much the same. These careers will attract young people only when pay
[0:56:53] Certification and respect rise together. That means upgrading training institutes first, giving vocational qualifications formal parity with degrees, as Germany does with its Meister, and starting apprenticeships while students are still in school. India is upgrading its industrial training institutes with this aim. The task deserves speed.
[0:57:18] And India stands for skills would be a worthy national mission and I'm sure it applies to many other emerging and developing economies as well. Three other steps would help. Supervisors could treat signed leases, residual value guarantees and joint venture debt as the debt they are with the fund and the financial stability board mapping these exposures across borders.
[0:57:45] Finance ministries could keep AI's productivity dividend out of their forecasts until it shows up in the data. Large listed firms could be asked to disclose their entry-level hiring and training. On safety, testing cannot be left to the labs alone and the developing world belongs in the room where the standards are set. Jax spoke about Carl Sagan in his closing remarks.
[0:58:16] One of India's ancient spiritual texts, the Kathopanishad, compares a human being to a chariot. The senses are the horses. The mind holds the reins. The intellect is the charioteer. AI adds a very powerful horse. It is welcome. But horses without a charioteer do not take us where we want to go. The real risk is not that machines will start to think.
[0:58:45] It is that we will stop. I hope it is not too late to prevent that. Thank you. Okay, that was an excellent set of presentations. I think there are questions embedded in those presentations for all the presenters. So let me start by asking, Jack, you said very clearly you were talking about the upside, and the others talked about some of the downsides. So let me point to...
[0:59:23] A potential downside that so many people are worried about, especially young people, will they find a job and will they be able to keep them as AI gets better and better? What's your response? So maybe three things here. One, we agree that early graduate hiring is a thing where you need support, you need to try novel things. Anthropic has funded an initiative called...
[0:59:49] Claude Corps, where we're trying to create 1,000 jobs for early graduate students that we deploy into nonprofits and NGOs around America. We pay their salaries because the best way to get people onto the career ladder is to get them a first job. We need to run many experiments like this. And then obviously, the AI industry would need to show up with the wealth to pay for that at scale. That's one. Number two.
[1:00:14] We can study occupations by their exposure to AI and what happens to employment in those occupations. There are occupations where high AI exposure may lead to a job loss. I think call services and customer service agents are an example. There, that becomes a matter of policy, a matter of discussion. And there are jobs where it seems like the application of AI might have a null to positive impact on jobs. We see this in some of the scientific disciplines.
[1:00:42] You add AI, you make them more productive, maybe you even raise employment in those areas. We need to share that data, Anthropic shares data from its own platform and works with governments to join that with local data to their country so we can know that. And I think finally, the bigger thing here is what I talked about is my model, it can't predict jobs that might be added from these new industries. It does predict potentially very disruptive employment impacts.
[1:01:12] And the purpose of my speech is actually to say just very frankly that that might happen. The implication of changing all of science and technology and how innovation happens is you may change like pretty fundamental things about how employment markets work. And I don't say that lively or like it's going to be easy, but rather that's a serious issue that we actually need to work on together. So I'd say that in response. Professor Shen talked about some ways that
[1:01:41] What you said could be implemented when she talked about the Chinese medical system and using AI to help expand medical access. I mean, that's clearly a place where the world needs far more. Absolutely. And so combining AI with people can actually do more. But, you know, turning to China itself. Can I add one tiny thing? Only just because you mentioned it, you said.
[1:02:07] Firms should disclose hiring and training. We happen to publish a post today about hiring and training practices at Anthropic, including employment trends about different employees and their seniority where, again, it comports with the data. I think that we absolutely should just mandate that firms share that stuff because that's the telemetry that lets you make decisions. Sorry. Absolutely.
[1:02:32] You know, a point about China. You're talking about it exacerbating some of the problems, for example, pricing and the fact that companies are not able to make profits as much. But has the technology implementation where China is ahead of many other countries, you know, you talked about some job creation. Are the youth in China also as worried as they are elsewhere in the world?
[1:03:02] And how do you see that playing out? Are there new initiatives to help? Yeah, true. So it's both sides. So I do hope that the positive side can dominate in the future. So we did a field trip in both this summer and last summer to look at the enterprises. So I do observe that it has been used widely, not only in traditional
[1:03:30] For example, we visited energy production firms. So what they do is that they used predictive AI widely in their system. So for forecasting, for risk controlling, and so on. So it's a listed firm, and they are profitable. That's number one, to make their existing business more standardized.
[1:03:56] More controlled. And number two is a new part that I just mentioned about health care, like making the health care a much better situation. And number three is we observe new platforms that can help to find new demand. For example,
[1:04:18] For young people, so one day there's a social post that a young lady complained that her boyfriend never separates the socks with other clothes. And then what happened is that that post then...
[1:04:35] was delivered to a machine producing factory. So they invented this new washing machine that you do one load, but this one load can do three separate.
[1:04:49] Washings with whatever you want to separate. So that means that you can cultivate. So we do observe it's not only that the trend is changing from production to demand, it's becoming so-called prosumer. We also observe that some of the production is done by voting. So the producer
[1:05:13] We'll put a note, and then the commentators will vote which one they want. And then it will then be produced. That's the number three. And also, we also observe it has been used in areas that we don't think it can be standardized. So I give one example about milk tea.
[1:05:34] So we visited a milk tea firm. It was created only less than three years. It already had 15,000 branches in China, 15,000. And their goal is to expand to 20,000 this year. So what they do is that they produce this tea fresh and standardized within two days.
[1:06:03] So that standardization, you know, milk and fruits and so on, that cannot be simply done without AI system. So that is, we observed that it has been used.
[1:06:16] Widely. And there, we observe something called a premium of engineer. So there's a tip-down effect. In one sense, because of AI, some of the engineers lose their job from IT industries. But then it creates this premium for service industries so that they can use these engineers with a relatively cheaper price. That's what the owner told us. They told me that they can hire an engineer
[1:06:46] With about 200,000 RMB. But compared with US, it's going to be much more expensive. So that will be something that probably can be used. Because China is graduating 5 million STEM graduates this year. If we think about this tip down to other for the service industry. So there is still great hope there. Absolutely. Dr. Nageshwaran.
[1:07:15] There are many countries that are not at the frontier, unlike the US and China. India is one of them. How do you see the role of AI in India, and how will that play out? You gave us a few examples, but is it a real problem not being at the frontier, or are there opportunities that actually emerge because of that?
[1:07:45] Very capital intensive nature of being at the frontier. Perhaps it is a blessing in disguise that I'm considering the way the leverage models are currently beginning to unravel. So probably it's okay to be a second or third mover in terms of being out at the frontier. So that's one. The second, I also noticed that in the Indian startups, the edge AI or the frugal AI computing models are at different stages of development.
[1:08:12] Which will probably make it more cost effective and will help in diffusion. Because even now, enterprise wide adaptation, even though per token cost is falling, the aggregate costs are still formidable for many enterprises. But these edge AI models may probably make it easier. So that's the second part. The third one I think says, as she mentioned in the case of the dishwasher example, which is very interesting.
[1:08:38] In healthcare, I also learned that diabetic-induced blindness early detection would help, and apparently AI and Aravind Eye Care is piloting it. Apparently it is doing a very good job of coming up with this detection much more quickly. And also both in remote learning and in education in remote parts of the country, I'm sure AI can...
[1:09:02] Substitute where humans are not willing to go. So these are all the developments that are happening from the Indian perspective. And last but not the least, as I mentioned, it is probably making it more urgent for a country like India, which is labor rich and has to create about 8 million livelihoods a year to focus much more.
[1:09:22] On labor intensive skills and labor intensive services. So in that sense, the fact that we are not at the frontier is probably concentrating minds on where we should be concentrating. Great. Jack, I'm going to come back to you, but I want the others to also answer this question, which is, Anthropic has been in the forefront of talking about the existential risks that AI development poses.
[1:09:52] At the same time, you have a reluctance of countries to regulate for fear of other countries getting ahead. Now, is humanity doomed because we fear other humans more than we fear extinction? What regulations can we cooperate on and what is needed to start this process? And I want you both also to come in on this. I'll just be brief. No, humanity is not doomed. Good to hear. Good, good.
[1:10:22] I think especially there's a pattern where people will try everything but the actual hard thing until the very end, and then they will do the hard pivotal acts like when the crisis arrives. Our view and the disposition of Anthropic is we would like to avoid the crisis. So we are trying to talk at length about what we see in front of us because we think these are choices that not only companies make,
[1:10:50] But are actually matters of policy and not national policy, but international policy. I think that the most confusing part of all of this is that any time we talk about regulating AI, it is as if we are talking about needing to invent general relativity or something. And yet, we talk about that in like
[1:11:10] Offices that are highly regulated. And we took a car there that was highly regulated. And then we go home and eat food that is highly regulated. And then we're like, it is a total mystery how to regulate technology. It's really confusing to me. And we're surrounded by regulation. And I think it's just simply having the recognition of the seriousness of this technology that we apply those basic safety testing standards, transparency standards, disclosure standards.
[1:11:39] Be able to price the risk of this appropriately. So I think we can do it. Because I'm thinking more about the regulation from the finance side, because I'm also working on institute of digital finance. So the creation of agent, I think probably will need.
[1:12:04] to pay more attention, especially for the financial market about, you know, we could have more market risk, we can have more operational risk and so on. That will not be something that one country can do. So in China there's always this dilemma set that if you are tolerating, so there's always a balance between how we are advancing technology innovation.
[1:12:32] As well as controlling the risk. China's perspective often happens is that when we are encouraging technology, then there will be a lot of wild things and create risks. And for this AI technology nowadays, because of the black box in terms of its characteristics of calculation, I do call for more international collaboration in terms of financial systems. That's something probably least
[1:13:01] Political, but that may be something that we can work on. Other things, maybe it's difficult to discuss, but maybe we can work on something that's least political. And finance, please. Well, thankfully, finance is non-political, so it should be easy. I hope. We should. Well, to your point, in the last panel, we heard the head of the IMF, the FSB, talk about the need for cooperation on regulating finance. But, Anand.
[1:13:31] Thank you. So it'll be presumptuous on my part to say whether we are doomed or not. But I think there is potential for us to face some catastrophic setbacks given the fact that it is already diffused.
[1:13:44] I mean, when it came to nuclear technology, I think because it was in the hands of few states, we could put some safeguards in. But I think this particular horse may already have bolted. So my only hope is therefore that whatever accident, because I remember one of Jack's slides and he said, you know, in biological research, AI can create viruses and also
[1:14:07] The cure that will follow. But I hope the gap between the two is not too long. And therefore, I would only hope that at this point, if there is an accident that happens, that it is not catastrophic enough, but serious enough to wake all of us up to the extent that we may have to disable some of the features of the technology and put it behind some safeguards, if that were possible.
[1:14:35] And also take a step back and look at some of the technologies which we had mothballed for the reason that not that they didn't have benefits, but they had cost which exceeded the benefits. And I would like that kind of a conversation to happen with respect to AI as well, and it hasn't really happened yet. So if an accident that is not catastrophic but serious enough makes all of us think about having this conversation, that I would count as progress. Well, that's always the case.
[1:15:04] You know, with technology, if we get a small shock but not an existential shock. Let me end from the panel with the last question, which is you've talked about AI becoming much more capable. So isn't there a question, you know, Dr. Nageshwaran said this, which is we always need to have the reins in our hand. Do we?
[1:15:33] If they're much more capable than we are, at what point do we say, hey, take the reins? Yeah, I think this is a real choice that lies ahead of us. And in the conversation around pacing the frontier, our CEO, Dario Amadei, said in a public post that we may want to introduce limits on the ceiling of intelligence that you allow for these systems. I think that that is...
[1:15:57] A conversation that is maybe a little at the edge of the Overton window, but is exactly the thing implied by your question. My view is that we're in control of this technology now, and we are watching it get much more capable, much more quickly than all of our intuitions. That's what I've tried to explain in my talk. And the window for us to act is a small number of years. And we can act, but we have to recognize this as a
[1:16:25] As a real technology trend, or as David Chalmers has said about philosophical questions, he said, philosophy now has a deadline. Well, I think the policy conversation has a deadline now. Let's take two questions from the floor, and then we'll ask the panel to answer. Question, gentleman there. Yeah. Do we have a mic? We have a mic coming. Yeah. To the gentleman there.
[1:16:56] A lady somewhere. Lady there. Thank you for a wonderful session. This is Sandeep Mehta. First of all, you predicted 2008 crisis. And now with AI, climate change, so much.
[1:17:17] I saw a report by IIF Washington that we are now having $365 trillion, with a T, as a global debt, which is 311% of global GDP, with such a high risk. With AI and all this data forecast, what you did in 2005 and 2007, stating we are going for a financial crisis.
[1:17:45] And it actually hit. Do you foresee similar financial crisis coming ever soon? Okay, I'll ask the panel that. The question there, the lady there. Hi, I'm Ginny Yan from London, Chief Economist of ICBC Standard, Chinese African Bank. My question to the panel, actually I have lots of questions, but one question importantly. If you have kids, what degree would you ask them to do? And what would you like them to do?
[1:18:18] We could talk about labor markets, which is crucial at this IMF, and about resilience, particularly climate resilience. What is that one job if you had to choose for your kids, if you had a child to choose? Thank you. Okay, let's go down the panel, answer whichever question you want. I won't comment on the financial crisis. It's not my area of expertise, since it was addressed to another panelist. On degree, I...
[1:18:46] I think that lots of degrees which have a sort of rote knowledge learning capability are, by implication of what I've talked about, going to be under extreme pressure. But there's extreme value in creativity, as I've said, that I think remains one of the final kind of frontiers. I may be talking my own book here, given my background, but I find that the liberal arts and having a good sense of history and philosophy and basically the
[1:19:13] The question of what human values are and human ethics are feels like it's only going to become more relevant and will help people orient themselves during this time. My kids are three years old and one year old, though, and I'm not sure how old yours are. So I also want to respond to the education question. So I think we need to have a switch of paradigm. My kid is already in her PhD.
[1:19:43] Well settled. I think there needs to be a switch in education paradigm from the Confucian style for Chinese, the ways that the teacher teach and the students learn and write and memorize, to the Socrates style. That's a Q&A, Q&A, and Q&A. I think for parents, the biggest thing we need to do is to protect the curiosity.
[1:20:10] Thank you.
[1:20:19] So on the financial crisis question, I'm sure we will soon find out the answer. And on the question of learning, I would simply say that we need to, both educators have to reimagine education by teaching them how to ask the right questions, teaching students how to ask the right questions. And from the students' and parents' perspective, it is about going back to the basic skills of reading and writing with hand.
[1:20:48] Well, thank you very much, everyone. This was such a fantastic panel. Please join me in giving the panelists a hand.