AI Governance Initiative Book Talk: 'Technology and the Rise of Great Powers' with Jeffrey Ding

All right, welcome everybody. Uh, thanks so much for coming to this talk. This is sort of a homecoming. Uh, I'm Robert Trager, Director of the AI Governance Initiative at the Oxford Martin School, and we're really happy to have Jeff Ding here with us today. I think, looking at the audience, I would say about 80% of you know Jeff, uh, which is really good to see. Jeff did his DPhil here at Oxford and was, um, affiliated then with the Centre for the Governance of AI when it was, uh, a small place, uh, back in the early days when it was Allan and Jade and Marcus and Miles, uh, probably a couple of other folks who are here. Um, but, uh, we're really glad that he's back. We're really glad that he's written this book, *Technology and the Rise of Great Powers*. I suppose I should say also that he is now a Professor of Political Science at George Washington University and holds research affiliations at the Foreign Policy Research Institute, the Elliot School of International Affairs, and indeed still at the Centre for the Governance of AI. So, without further ado, please welcome Jeff Ding.

Yeah, thanks, uh, everyone for being here. It's really good to be back. I was just thinking about, yeah, starting out at GOVI, uh, at Oxford, researching China AI development, and I don't even know why this came to mind, but we had like a meditation session, and, um, that was when I first realized, oh wow, my neck kind of gets really sore, and I need to get better posture and figure out how to, uh, uh, be more mindful of not just sitting all day. So it was like a very healthy experience, not just for research, but also for, uh, being a more sustainable researcher. Um, and then maybe after the meditation session, I was just talking to Miles. We went to the Five Guys right by Tesco. So I don't know how that squares with the whole being healthy thing. Um, but let's get into it.

Um, I think often times when we think about the US-China AI competition, um, this image comes to mind. It's illustrative for two reasons, I think. One is two countries locked in some geopolitical competition, sort of the resurgence of great power competition centered on emerging technologies. Uh, but I think it's also illustrative for another reason, which is, uh, during the pandemic, I got really obsessed with chess again because of *Queen's Gambit*, the show. And anyone who plays chess will know this board, this position on the board is not possible. Right? And so I think this image is illustrative in that we don't, we have a pretty confused and muddled sense of what great power competition in AI actually looks like.

So the book asks that simple question: How does, how do technological revolutions affect the rise and fall of great powers? Maybe one starting point for what great power competition in emerging technologies looks like is this speech by Chinese leader Xi Jinping back in July 2018, uh, at the BRICS Summit in South Africa. So these are emerging economies that are making up more and more of the gross global economic output and population. And at this summit, around the theme of the Fourth Industrial Revolution, he gives this speech that talks about how these cutting-edge technologies are changing the trajectory of human history, specifically calling back to all of these past industrial revolutions. What happens in the Chinese system is often times the leader gives a speech, and then analysts and commentators go back and interpret that speech. So a few months later, this is published in *Qiushi*, which is an authoritative Chinese Communist Party publication, and it repeats some of the themes. But what I want to highlight for you here is the emphasis on taking advantage of industrial revolutions for productivity leadership. Okay, who, who becomes the number one economic power in terms of productivity? That becomes the basis for hegemony. Okay? That view of power transition by way of technological revolution also resonates with the classic literature in international relations scholarship. Here's Paul Kennedy's seminal work, uh, this pattern of the rise and fall of great powers, where you have differentials in technological change leading to economic growth differentials, right? And that leads to a shift in the global economic balance of power. And then gradually, that's going to shape the geopolitical and military balance of power. So today, I'm focusing on that first step in the causal chain: technological change to shifts in the global economic balance of power.

The standard story is what I call the leading sector account. It, it might sound pretty familiar to you. You have these new technologies, revolutions. One country, the, the new advances happen to cluster in one country. That country dominates new-to-the-world breakthroughs. Uh, they get to take advantage of these new fast-growing industries. They get monopoly profits. They get all the export advantages. They reinvest those to become the world's most productive economy. Uh, those familiar with international relations scholarship will recognize some of these names, the Robert Gilpins of the world, the Paul Kennedys of the world. Dan Drezner summarizes it by saying, "Hegemon status through near monopoly on innovation in leading sectors." Where I live and work in DC, it's a very influential perspective. I, I think it's pretty influential in the UK, Europe, and around the world as well. Just to flesh it out more, uh, a little bit more, development economist Walt Rostow developed this classic sequence of these great leading sectors, right? Some of these industries that sprout off the back of new innovations: the spinning jenny, mechanized cotton textile production, the steel industry, chemical industry, automobile industry. Some of these industries become the largest sector in the economy. And so political scientists have largely adopted this approach.

My argument is, uh, actually, power transitions by way of technological revolution come through a different pathway, and that pathway is based around general purpose technologies, or GPTs. And these are foundational, uh, advances that have the ability to transform all different sectors of the economy. And so today, when we talk about US-China competition, think about all the different emerging technologies. Let's just pick one out of the hat. Electric vehicles is on everybody's mind right now. Um, AI is more general purpose than electric vehicles. Think about the variety and the range of uses of AI compared to a single use case for electric vehicles. Another test you can do to think about is this technology more general purpose than the other? Is, uh, how are advances in one technological domain transforming, uh, advances in another technological domain? Right? What's happening in the biotech industry is not, not reshaping what we do in AI, but what's happening in AI is completely transforming what's happening in biotechnology. One is more general purpose than the other. And so that, that pathway by which general purpose technologies shape economic power transitions, one that that's emphasis is on diffusion and adoption. Uh, that's going to inform the institutional adaptations that countries have to make. How do you have to adjust your country's political structure, your country's approach to science and technology planning, uh, your country's system of organizing businesses, industrial organization? For in the for the purpose of this book, I focus on a country's system of skill formation, and particularly for general purpose technologies, it's about this wide base of engineering skills and the ability to cultivate that.

So I gave you the argument in a nutshell. Uh, in the next 20 minutes or so, I'll keep this pretty brief so we have a lot of discussion, because I see a lot of faces in the room who, um, I think are going to disagree with the thesis I'm going to present to you. So I want a lot of good, open, uh, back and forth today. So I'm going to flesh out the argument in more detail. I'm going to give you evidence from the historical case studies, mostly from the Second Industrial Revolution, and then we're going to talk about what this means for US-China competition in AI today.

Returning to that original slide, so the basis of all, all of these theories about how technological revolutions affect the rise and fall of great powers, the framework is that there's this evolutionary relationship in which new technologies present some demands, and then whichever country is more successful in adapting to those demands is going to be the technological leader. So for the leading sector story, it's about adapting to the demands of these new fast-growing industries, where it's all about which country can monopolize foundational innovations. This is our obsession with the eureka moment. Uh, think about the front page of the Wall Street Journal, Bloomberg. We celebrate who do we celebrate? The technologists, the people who are pioneering new innovations. And this innovation centrism is going to shape the factors that we think are are really important for success in technological revolutions. And so it's about institutions to corner and monopolize innovation. So who's spending more on cutting-edge R&D? Do you have a really strict patent system that protects the rights of these heroic inventors? Are you training the best and brightest talent? This idea comes from basically treating the nation as a firm, and this firm has a new product innovation. So a nation developed this new smartphone, and so they spent all that R&D on there. You're the first innovator there. You have that brief window when you can capture all the monopoly profits from being the only one with this new smartphone. But eventually, new competitors come, and you lose that brief window. So, so you have to, you have to capture that brief window in the product cycle. Uh, this is just some quotes from representative international relations scholarship that shows, okay, it's about, um, the, the main innovations cluster in one country, right? One country monopolizes innovation. One, one, one firm has that product innovation. There's that brief window of time. So, so the impact of a technology comes really early on in its development and is really crucial. Leading sector theory, it's about who captures the monopoly profits, the pioneer, whether the pioneer can capture the monopoly profits before diffusion happens, right?

But for general purpose technologies, uh, so, so for leading sector theory, the story ends at that diffusion stage. For me, that's where the most interesting stuff is happening, and that's where my argument starts, because GPTs don't make a mark until they're diffused throughout the entire economy. So here's GPT diffusion theory. Why do I focus on GPTs? I think so many times, and maybe those of you who are really in the AI policy space might feel some similar frustration. So many times, it's like we have these policy papers and these research papers about emerging technologies, and it's almost like you could find and replace "electric vehicles" or "quantum" or "AI," and it would be the, the same paper. Um, but for me, not all technologies are created equal. GPTs are special. Economists and economic historians deem them to be engines of growth. They often precede waves of huge productivity growth. Those are some of the key characteristics that economists highlight about GPTs. But what I want you to really take away here is that there's this regular pattern by which GPTs make their mark on the economy. And this is Stanford economist Paul David. That's this extended trajectory of gradual and protracted diffusion into widespread use. And you, and there needs to be all these complementary streams of innovation, all these different application sectors to take advantage of a GPT. So you don't see the impact of electricity on the earliest adopting country, the US, until 40, 50 years after the initial debut of the electric dynamo. So if we care about GPTs, we should care much more about that diffusion stage. Innovation is when a, when someone first debuts or commercializes a new technology. Diffusion is the process by which that technology spreads across the population of potential users. What's really important about GPTs is the population of potential users is essentially the entire economy and all these different application sectors. So, of course, being an innovation leader, a country that's on the front edge of innovation is going to help with adoption, but it's not determinative. Oftentimes, there's even latecomer advantage when it comes to adoption. And I think this is a bit counterintuitive, but when we're talking about competition among the advanced economies, all of these countries are going to have firms and universities at the technological frontier. So think about AI today. Chinese firms and Chinese universities are pretty close to the technological frontier. I would say they're pretty fast followers. So the main impact of AI will actually come through who can imitate and diffuse faster.

Second part of the argument: I've showed you different technological trajectories. They map onto different institutional adaptations. So how do you best adapt to the demands of GPT diffusion? For me, I focus on these institutions that widen the engineering skill base with new GPTs, and these solve two of the demands that GPTs pose. One is a pretty obvious: technology races ahead, you have to have the skills in human capital to catch up. I think the second is more interesting. Right? What, for diffusing electric vehicles or diffusing, um, these specific leading sectors like chemicals or steel, you don't necessarily need a lot of coordination between the electric vehicle industry and say, the construction industry or the education industry. But to diffuse AI across the entire economy, you need the, the construction industry, and then the tutoring services industry, to all, you need all these different application sectors to know what's happening, what's, what's, what are the latest advances in large language models, what are the standards that are being set, the benchmarks that are being used to judge whether this multimodal model is better than the other. And so that type of knowledge actually sometimes comes, is systematized and standardized in engineering disciplines. So it's no surprise that after a new GPT, we often get a new engineering discipline. Steam engine and mechanical engineering, electricity and electrical engineering, computer science is very much an engineering-oriented discipline, even though it's not in the name. Okay.

So quickly to sum up, this table gives you the two different mechanisms I'm going to trace through the historical case studies: leading sector product cycles versus GPT diffusion. They give opposite predictions on all these different dimensions that we've talked about, and then they also suggest different institutional adaptations. So we're going to trace to see through the Second Industrial Revolution whether it followed the leading sector story or the GPT diffusion story.

Second Industrial Revolution, 1870 to 1914. I'm going to go really relatively quick through the historical session so we get more time for what this means for US-China competition in AI, because I think that's what most people are interested in. But just to lay the groundwork, this is so many new scientific and technological advances. People call it the Second Industrial Revolution. And you have also an economic power transition, right? One great power is now sustaining economic growth at higher rates than its rivals. Germany and the US are catching up and overtaking the UK, and Britain faces its relative economic decline at this point. And I want to emphasize, a lot of these sources are not, I'm not going back to all these different industrial revolutions and collecting original sources. I'm relying on the great work of historians of technology and economic historians to see whether they match up to different theories or not. The one key point from the background setup to this case is in the time period we're talking about, 1870 to 1914, it's the US that becomes the clear economic leader, especially if you care about that productivity leadership that I started this talk with. So US is the blue you see on all these measures of economic efficiency. The US overtakes Great Britain before 1914. Same cannot be said for Germany.

So let's go through all these different technological dimensions on impact time frame. Leading sector product cycle says we should see the impact pretty immediately in terms of an economic power shift. GPT says, hey, more gradual and projected timelines. And I think that's what we do see. I'm only going to highlight the electricity evidence, uh, because there's a lot of qualitative and also quantitative evidence that supports that long decades, multiple decades diffusion story. Uh, here's some evidence from Sergio Pesta's work where he collected patents in other fields like the mechanical field and other fields that were starting to contain electricity-related vocabulary, and you see that the real takeoff doesn't take place until after 1920, where now a bunch of these other industries are starting to adopt, uh, electricity in their stream of innovation.

For me, the key GPT in this period was incubated much earlier in the mid-1800s. These machine tools like the turret lathe and the universal milling machine that allow you to cut and shape metal and wood in more precise ways, and so you can create these interchangeable, almost standardized parts. So if a bicycle breaks down, you don't have to completely build a bicycle from scratch. If a sewing machine breaks down, you don't have to, you can just replace the part that broke down. And so the proliferation of these machine tools had started to diffuse throughout the US economy. And that's what I see as, as the, the impact time frame that that matches the GPT story.

Is it about innovation or diffusion leadership? US is not leading when it comes to leading sector innovation. The US is not really even an export-based economy at this point. When the British Institute of Electrical Engineers is trying to figure out why is what's happening in this space, we're not behind on inventive genius or innovation in electricity, we're behind on the practical application to the requirements of the nation. So here, this is the US's advantage. It outpaces Germany and the UK by more than double the rate in terms of machine intensity. It's not about exclusive access to special innovations. UK sends all these study teams to the US to figure out what's happening in this space, what is it, what's the secret sauce? It's about the adaptation of these special machine tools to all branches of industry, the eagerness of diffusion. And this is broad-based economic growth. It's not concentrated in just one or two sectors because GPTs require all these complementary innovations.

Okay, so when it comes to institutional adaptations, uh, for me, this supported this idea of GPT skill infrastructure. The US was able to effectively broaden and systematize mechanical engineering skills. Britain just simply didn't produce enough. Uh, here at the University of Oxford, I think it was very late to even establish the first faculty member in engineering. And so we have engineering density measures to support that. Germany's did produce enough mechanical engineers, but its problems were with systematizing and standardizing that mechanical engineering knowledge. So just one data point on that, there were some reports about the number, the hours devoted to certain things in the curriculum for engineering schools. Representative of the US, bottom three, and Germany, excuse me, bottom two for the US and top three for Germany. The shaded lines, the black darker shaded lines show that there's much more time in the US schools devoted to practical exercises in the lab, in the shop, whereas the German schools are prioritizing oral theoretical instruction. So there's not that diffusion of mechanical engineering knowledge.

One more data point, right? US chemicals is the most scientific industry at this time. US is far from the frontier here. German universities are where the best and the brightest US students are going to study for PhDs in chemistry. American leadership is so far from the frontier of scientific research, but its leadership is based in introducing chemical engineering, not seeing leadership in chemicals as who's producing, who's producing more innovations and synthetic dyestuffs, but seeing leadership in chemicals as who is taking these chemicalization processes like crystallization, titration, and applying to all these different industries through this process, these practices of chemical engineering.

Okay, I'm not going to go through all the other cases, but I just wanted to show you the framework for all the historical case studies. It's following the same thing, right? How tracing closely how did all these past new technologies affect economic power transitions? Would go through all those different dimensions, and the skill infrastructure in the First Industrial Revolution case, the difference in the Third Industrial Revolution case, US-Japan competition, and what was known as the Information Technology Revolution, is that, uh, there is no economic power transition that takes place. Everyone thinks Japan is going to be the number one technological power, but it doesn't happen. That challenges the leading sector mechanism because all the pieces were in place. Japan, this is my favorite quote in the entire book. Japan, these political economists and historians are saying, look what Japan is doing in consumer electronics, HDTV, semiconductor components. What Japan is doing seems really similar to what was happening with Germany and chemical leadership. I think they were wrong because they were just using the wrong template for how technologies affect power, economic power transitions. Japan was not leading the US when it came to the diffusion of general purpose technologies, specifically computerization at scale.

Okay, so we've laid a lot of the groundwork. Now, what does this mean for US-China competition over AI, which some deem as this era's most important GPT? Uh, I think the, the essentially what the last chapter of the book is doing, it's saying, okay, most of our conventional wisdom, most of the way in which we research and study this space is heavily influenced by the leading sector model. And then on all these different fronts, when it comes to impact time frame, the phase of technological change that matters the most, the breadth of growth, and the institutional adaptations on all these different dimensions, GPT diffusion theory is suggesting something quite different. And so let's go through these one by one. I like to be very clear about who I'm arguing against. This is the National Security Commission on AI's final report. I was actually a special government employee and consulted on this report, but they didn't really listen to me. Um, and then this is Graham Allison, Harvard Belfer Center Professor, Eric Schmidt, former Google CEO, who is now probably one of the most influential voices in US technology policy. Kai-Fu Lee, I think talked at Oxford on his book *AI Superpowers*. A lot of their thinking is all based on the leading sector template.

On the impact time frame dimension, this idea that China's going to overtake the US on AI leadership within, within a decade, right? Think back to leading sectors, near immediate impact of these fast-growing industries. If you take the GPT diffusion model, let's say we date the incubation of this deep learning paradigm to AlexNet's submission, 2012, to the ImageNet competition. If it's three, four decades before we see AI making its mark, there's no way we see the economic impact through productivity, economy-wide productivity payoffs until after 2030. Now, a lot of people, and maybe we'll get into this in the Q&A, a lot of people have argued this time could be different. Maybe you don't need the physical infrastructure to diffuse AI like you did for electricity. Maybe technologies are just diffusing faster today. But I think I'm almost more of a technological determinist in this sense. I think there are regular patterns by which GPTs diffuse, and they usually take time. Um, and there's all these organizational adaptations, talent updates that need to take place before we'll see the main impacts from AI. If you, if you adhere to the GPT diffusion model.

Now, what does it mean on the second dimension of phase of advantage? Is it about innovation or diffusion? When we talk about China's ability to compete with the US, almost all these assessments of China's AI capabilities are about, are they doing well in cutting-edge R&D? Who has the most valuable firms? Who has the most interesting startups? If you look at it from, can China really challenge the US in AI? If you look at it from a diffusion perspective, actually, in a lot of my other research, I show that China ranks pretty middling when it comes to adoption of other information and communications technologies. Significantly trails the US in adoption rates of other digital technologies like cloud computing. This is another paper where I sorted through all these different science and technology indicators, matched the ones that look a lot more, that adhere more to innovation capacity on the left, and those that match more with diffusion capacity on the right. So the innovation capacity stuff is like, what's your average R&D spending of your top three firms, your Alibabas, your Huaweis? What's the average reputation of your top three research universities, your Bagas, your Tsinghuas? What's, what's the, how many like PhD students in STEM are you graduating? On the right, it's more things like, what's your adoption rate of past ICTs? How strong are your industry and university linkages, right, and collaborations, because that's how ideas spread from developers and implementers and so on. On innovation capacity, China ranks as one of the top countries in the world. On diffusion capacity, it's much more middling.

Breadth of growth, do we target these specific industries? So China has done a lot to just target self-sufficiency in key industries, or do we have a more horizontal approach to stimulate economy-wide diffusion? This is actually a State Council, China's cabinet body, State Council affiliated think tank that is criticizing China's focus on the LS model. And the book concludes saying that the US is pretty well positioned to be the leader in what some are calling the Fourth Industrial Revolution. One number I'll give you is the number of universities that have at least one faculty member or one researcher affiliated that's published in an established AI conference. China has 29 of those universities. The US has 159. So once you get past that first tier of Tsinghua, Peking University, Nanjing University, the US, by comparison, just has a much broader bench. And on the hybrid publications, these are linkages, right, between different pieces of the science and technology ecosystem, at least one co-author from industry, at least one co-author from academia. US leads the world with the highest number. China, uh, it faces a lot of challenges in this space.

So I've laid out some of the more specific policy proposals on, if the US, for example, were to take the GPT diffusion approach, what it should be focused on right now. It's very much based on the leading sector approach. Let's do Fortress America and prevent technological leakage. Let's try to monopolize all leading-edge innovations in AI. A GPT diffusion-centric approach would prioritize US sort of a running faster approach, and sometimes you have to eat the cost of openness to to run faster.

Okay, I'm sure we'll open it up. I'm sure we'll get more in-depth on the AI side of things, but let me just leave the main arguments there. Um, this is available everywhere you get books. If the holidays are coming up, so if, if a friend of yours really likes the color orange, you know, please consider picking up a copy and looking forward to the discussion to follow. Thank you to, uh, Oxford Martin School and GOVI for organizing this event.

Great. Thanks so much, Jeff. I'd like to invite also Kayla Bloomquist up to the stage, who is a DPhil candidate at the Oxford Internet Institute and, uh, also has been an affiliate of GOVI and, uh, previously a US Foreign Service Officer, in fact. So we'll also join in asking some questions. And Kayla and I are going to ask a few questions, and then hopefully open it up to all of you who will have, uh, lots of questions. Um, maybe, maybe I'll, uh, just take the opportunity to ask, ask one and then send it to Kayla. Um, so I'm curious a little bit about how you see the ways that, uh, the things that determine when are good at diffusion and when they're not good at diffusion, because you, you turn to that sort of at the end a little bit, but I, I feel like I could understand that a lot better, and that's super interesting and seems to relate to policy recommendations. And as you were talking, I remember the story of Lord Nuffield. Some of you may know, he made a lot of money in the car industry, and then he tried to give it to engineers at Oxford, and I think this was in 1937, and he was told that's not something that can be studied at Oxford. You may not do that. And as a result, he had to give his money to social sciences, for which we are still benefiting. That's one of the largest gifts. Um, so I mean, that's just, it's very interesting to me, and I'm wondering, you know, are these sort of factors that determine what technologies can diffuse, where are they kind of static things? Are they things that no matter what the technology is, country A is going to be be very good at diffusing that technology? Or are they sort of culturally specific to certain technologies where you know you might say, you know, engineering is not what we do, or something? Or what, you know, how do you see the factors that determine it, and what does that mean in terms of policy recommendations?

So I think it's a really good starting point. The factor that the book focuses on is, does your country have, uh, a strong institutional basis for widening the base of engineering skills, uh, in that particular GPT? And so that's as far back as I go down the causal chain. I don't come up with an explanation for why one country has, uh, a better GPT skill infrastructure than others. I have some ideas for what accounts for that. Uh, I think in general, uh, decentralized systems, uh, are are better suited to, uh, having the education and training systems that can adapt to producing these engineering skills. Um, in some of the cases, for example, in US-Japan competition, Japan's approach to cultivating software engineering talent was very centralized and top-down. They channeled resources to like centers of excellence, and we're saying this is where we're going to do our software engineering training. They weren't as flexible in adapting to the curriculum demands of computer science and software engineering, uh, at least as well as the US. Uh, so I think that's one factor is, uh, the, the level of decentralization. Uh, and also other factors would be looking beyond just the skills that I, like I focus on human capital formation, but there's all these other factors that can change, aren't static, that will, that will shape which countries are going to be better at diffusion. Um, I mentioned some of them there, like the system of industrial organization. Is your, are, is the way you organize business through like, uh, sort of big business networks, or is there more of a competitive landscape? The level of intellectual property protection. Uh, so there, there's a direct contrast between if you have really stringent patent protections, that might help with the leading sector innovation, but it might hamper diffusion. And we're having similar debates in AI about open source as well, right? Uh, I would say, uh, having a really healthy open source landscape is conducive towards diffusion, but it might not be as conducive towards this sort of like leading sector innovation type model. Um, so, so a few starting points there.

Okay, I have lots of questions, but I will refrain and turn to Kayla. Please.

Well, first, I wanted to say thank you so much for helping expand the conversation on, you know, what does it mean to be leading and winning in whatever this so-called race is. Um, and I think those of us that are in and around this field should continue to strive to expand that conversation. I know it's definitely helped my work on like concepts of government legitimacy, leadership legitimacy within the US-China AI race. And to further expand that, I know it's not so much a core focus of the book, but I was, I'm curious your take based on the research and then broad understanding of where the field is now. Talks a lot about the book talks about domestic diffusion and the importance of that for economic leadership, but how do you, and how would you moving forward, factor in global diffusion of these technologies, especially in this broader US-China tech competition space? There's so much, uh, discussion about competition about digital infrastructures, telecommunications, smart cities, and especially with Chinese government efforts to, uh, promote the export of Chinese technologies. In that sense, this becoming a core part of the foundations of this race. So how would, how would you consider or, you know, expand these arguments around the domestic versus global diffusion?

Yeah, great question. Let me answer that in two ways. I think the first way is there could be a perspective that global diffusion of technologies is just happening faster. There's so much globalization, so maybe some of the insights from these past cases don't translate to this US-China competition AI case. I, I think one of the most interesting things I found was there was this OECD report, Andrew Edel, 2015, I think, and they found that there is sort of an initial adoption gap that is shrinking. So the time from which maybe some leading-edge firm in one country develops an innovation and then the time it takes for a frontier firm in another country to adopt that innovation, that time is shrinking. We have these, as you were saying, we have these global networks of diffusion. Multinational firms have R&D labs in other countries. One of the centerpieces of China's AI ecosystem, for example, was Microsoft Research Asia, based in Beijing. So that gap is shrinking. But the really interesting thing that the report found is the gap by which that technology is spreading from the frontier firm across a broad range of the economy, that gap has actually, has actually increased over time. Uh, so I think that suggests actually this focus on, can you actually, can you, can you adopt these GPTs at scale throughout your own country becomes more, has become more important.

Uh, for the second dimension of this question, is when it comes to China, what advantages are they getting from exporting to other markets? So maybe that gives like some soft power advantages, right? Kayla, your research has looked at China's discourse power and soft power. But when it comes to like, when it comes to productivity growth, uh, I, I don't really see exports as being that necessary. Maybe like just having more powerful firms and bigger firms means they reinvest into the productive capacity of China at some point. So I think there are some mechanisms, but I'm much more interested in the process of, okay, Alibaba came up with all these really cool open-source models that now rank as highly, or if not higher, than Llama's open-source models. How does that small and medium-sized firm in China, in an inland province, actually apply that to improve China's productivity growth? That, that's the diffusion story I'm, I'm more interested in.

Thanks, Jeff. Um, well, I think maybe Kayla and I might ask another question or two, but I want to also just open it up and see who wants to ask things and chat. And why don't we go back to Sumaya and then up to Marcus after that.

Hi everyone. My name is, um, Sumaya, and thank you so much for your presentation. So I have a question around third world countries. And it's very interesting that you've actually brought out how this diffusion actually changes economic dynamics. And now that China is in the picture and we have the Belt and Road Initiatives, and we also have the Digital Silk Road, are there any like causal effects that you've seen that has any form of impact for third world countries and developing specifically, and how they can leverage this? And then the second question that I have is, you mentioned that for GPT diffusions, there are two things that are important, and one thing that a lot of developing countries at the moment actually have is capacity and talent, which is something that you mentioned. Is there ways in which that from your research and from the analysis that you've done, developing countries can be able to leverage these conditional factors that we already have in place, especially like talent and growth, to be able to leverage diffusion of technology?

Thank you. Yeah, it's, uh, it's a really important question. I think the scope of the book is focused on great powers, but I think some of the arguments and findings can extend to, I would say, middle-income countries and sort of newly industrializing countries. I'm not as sure if some of the policy prescriptions I'm giving can be as applied to, uh, sort of lesser developed countries and still developing countries, in part because I do think you need a baseline level of absorptive capacity to do the international diffusion stage. Uh, so do you have something like an Alibaba that can be a fast follower, or a Microsoft Research Asia that can, that can train up a local talent base? And so the question would be, for some of these lesser developed countries, they might not have that absorptive capacity to tap into some of the advances from the technological frontier to then diffuse throughout the country. Uh, but I think once, once you do reach that level of absorptive capacity, uh, then I think, uh, a lot of, a lot of the findings that I'm presenting here will be applicable. So there's a world in which, like, the leader in AI, technological leadership in AI, is not even in China or the US. I'm really open to the idea that Europe could be the leader in diffusion and adoption, uh, or, uh, or India, or or other, uh, newly industrializing or or middle-income countries that that have that level of absorptive capacity. Uh, when it comes to what can developing countries do in terms of leveraging their advantages, uh, I, I think the, so I think in line with that first question, I would be trying to figure out what can I do to get to that level of absorptive capacity first. I wouldn't necessarily be reading this book and trying to figure out, uh, necessarily how, how do I take advantage of GPT diffusion? I would make sure, can I get the nodes? Can I get the nodes in the system that are connected to the global innovation network? Can I train? I do, in this case, I do need a really strong like anchor tenant university or anchor tenant firm that's going to that's going to create this local cluster development, like Zhejiang University in Zhejiang and Alibaba's headquarters are in, are, uh, uh, Dang University in Hangzhou, and Alibaba's headquarters are in Hangzhou. So then Hangzhou has that absorptive capacity, and then you can diffuse ideas from Hangzhou. So if I were sort of in a lesser developed country, I'd figure out how can I get those anchor tenant institutions to tap into the global frontier in a GPT, and then we can start the GPT diffusion story. Sorry, it's a little, it's kind of, uh, abstract, and it's almost like it's weird to treat this as like a chessboard type exercise, but that's how I would think about it.

Thanks. Um, maybe we can go to Marcus here.

Hello. Really good to see you, Jeff. Um, um, yeah, so I guess overall, I think your thesis is very plausible. If you look at the history of of GPTs, it, it sounds to me that, uh, your story is is better than the sort of leading sector theory. And then I'll, I'll try to push back and give my best case for sort of leading sector theory as, as a theory in general, and maybe in particular, like it is maybe more plausible in the AI case than in terms of other GPTs. So I guess firstly, there's, there's kind of the obvious point that probably a decent amount of why people are excited about this leading sector theory is it just has this nice correlation. You're like, okay, well, whoever developed this technology first, they're probably likely to have all these other kinds of nice things that you'll need to take advantage of it. So probably that, that's a part of the thing that's going on here. Maybe more interesting is, you talk about, or some of these quotes that you pointed out, they talked about sort of the benefits of monopoly profits. That maybe isn't that convincing because that will probably be a pretty small part of, of sort of the economy. But maybe monopoly sort of advantage is more interesting as a concept. And so in the, in the recent years in international relations scholarship, people have gotten excited about this sort of idea of weaponized interdependence. For certain technologies, you will be able to leverage them for power. For certain, for some technologies, it's that's very difficult because the innovation just spreads. So like the Transformers architecture doesn't really give you that kind of advantage, but the chip supply chain and ASML and TSMC does provide sort of geopolitical, uh, sort of strength and advantage to, uh, to their home countries. And so for certain kinds of technologies where that, that kind of diffusion of this sort of base node is hard, it seems quite plausible that you might have some sort of leading sector type theories might have more power. And then maybe lastly, and more speculatively, in AI, it might just be that there's less need for human capital to sort out diffusion. And so I guess one thing is just AI systems will be able to do that work for you. You'll be able to sort of replace some human labor with AI labor. And then also, maybe compared to other GPTs, there's this sense that you can sort of plug into existing infrastructure more than than you can in other cases. You don't need to rearrange your factory because you can just plug the AI system into the role that the human was already doing. And so maybe maybe that suggests that there's this sort of other concept that maybe you should explore, which is you talked about sort of absorptive capacity, and maybe there's something like non-talent absorptive capacity to explore. So like, if my country bans the use of AI in a certain sector, then I won't be able to do the diffusion there. And maybe those kinds of factors will play a lot bigger role in the AI case if if human capital matters less.

Yeah, these are great. So let me start with, uh, the defense of leading sector theory. Yeah, I do think there is some correlation. What's interesting is when historians have gone back in all these different cases, we think that Britain is the home to all these heroic inventors, and they are for a lot of these technologies, but like Joel Mokyr at Northwestern University, this really, uh, prominent economic historian, he says that Britain does not have this comparative advantage in macro innovations. They have this advantage in sort of tinkering, implementing, and diffusing these innovations. Actually, France had what was the, the center for a lot of these new leading sectors at that time. And from the Second Industrial Revolution case, we don't see that correlation, right? The US is not the center of leading sector innovation.

I do think I take your point on there is a monopolistic advantage when it comes to power in international relations. The weaponized interdependence text is a good example of that, just being able to control, like, if if you're able to control key nodes of a supply chain, as is the case with semiconductors, you can exercise power in different ways. Power is this really malleable concept. There's so many different forms. I'm most concerned with, uh, I think if I had to bet on one aspect of power in a long-term competition among a great power and a rising power, total factor productivity is not a bad place to start. Um, and having these key nodes, maybe you can control them for two or three years, you can have some exercise of power. Uh, I, and, uh, you can use that for bargaining in some way. I don't know if the US has really gotten China to do something that it wouldn't want to do using that. Maybe it's slowed down the ability of China to train really large powerful models by a little bit. We already see a lot of the loopholes around that. The Soviet Union had a lot of, had big businesses, had some of the monopolistic advantages, controlled some of the key nodes. Why did the Soviet Union eventually falter as a great power? Economic productivity. So there's different dimensions of power. Uh, I'm not saying that, uh, controlling key nodes with monopolistic advantage isn't important for power, but if I had to pick one, uh, I would go with long-term economic growth.

On your second point, I think this is really important, and I'm out of my depth here in terms of, yeah, if if AI has less need for human capital, why do we care about GPT skill infrastructure? And I think the thing I'll say here, because I know there's people in the room here who are much more plugged into what is happening in AI and what is happening at the frontier firms, but I think what I'll say here is, um, I imagine similar arguments were being made with the steam engine, and I imagine similar arguments were being made with machine tools.

Tools and I, and I have read text saying, "Wait, what's going to happen with the human worker?" for when electricity is going to come. And I imagine very, very similar arguments were made with computers. Um, so my perspective is, I'm taking an outside view on how to approach this problem by looking at what AI is like. This approach, which is saying that, and we're all speculating here, but this approach was just saying, "Hey, human capital might matter that much," is taking the inside view of like, "Okay, AI is going to be different, different." Um, I'm not saying my view is the only one that should be in the room, um, but we should have both of these perspectives in there. And and there's a marketplace of ideas, and we can figure out which one is eventually right. Um, so, so that would be, that would be myo, that would be my answer to that question. You know, uh, we'll take some more questions.

And I just want to ask a quick follow-up, which is just in relation to that. I wonder if you think the diffusion dynamics apply more to narrow versus general purpose AI? Because I can imagine, given the things that you're saying with narrow AI, you might say, "Okay, you know, we really need the human capital in a particular sector." You know, you, you had the example, I think, of in the bio space to figure out how to adopt AI. But in other sectors, if it's general purpose, maybe you could make the argument that it's just replacing human capital or will in a different way from the industrial revolutions of the past. So I'm curious if you make that distinction or you think it's similar across those areas.

I think it's similar. Like, my view of transformative AI is actually developed here at Gov AI. Um, I got interested in AI policy after reading *Superintelligence* because, um, Obama recommended it on his reading list, and I like follow his Spotify recommendations too. So, um, that's how I got interested in it. But my view of transformative AI was really shaped by Eric Drexler's work, um, uh, of this view of comprehensive AI services, where it's almost more, it's not like one day DeepMind opens a box and we get AGI, but it's more like, um, we get this app store of narrow but still super intelligent targeted services for things that humans want. Um, so, uh, for me, yeah, maybe I just don't have the imagination to to think about something that that wouldn't require a lot of human input. I like, for me, sometimes it just helps to look at how is a random sector going to adopt AI. So, um, I do these weekly translations of Chinese writings on AI-related topics. One week I devoted to bird watching, um, and how Chinese nature reserves are using computer vision to count the number of, uh, woodpeckers or Bell-piped herons or some, I'm, I'm butchering all these birds, I don't know birds, but, um, but it was just like, it's this really slope. You have to take eight months of collecting data to differentiate this one one beak from the other. You're in an area where there's not that much connection. So, some in some places there's not good infrastructure. Um, you have to have cameras that can weather all this different, that can adapt to all this different weather con. So all these complimentary things that have to be done. There's a lot of human expertise that still has to be involved. Um, so at least what's happening now, I, I, I'm not as imaginative, I think, and I don't think in the way that a lot of people think in terms of those who, who see this really fast acceleration and takeoff. Uh, but, but that's my view.

Fair enough. Okay, let's go to Liz. Uh, hi Jeff. Um, thanks so much for this talk. It was really interesting. I also really enjoyed the book, and, um, I echo Kayla in terms of, you know, what Kayla said about, um, thinking about the competition dynamics in different ways being really important. Um, I'm a PhD candidate in IR here, so I'm kind of approaching this question from an IR angle, focusing kind of on the implications you outlined toward the end for, um, the realm of AI and in that kind of the relationship, which is currently having a very competitive tone, I guess, to say the least. Um, I was kind of curious about, um, your experience also in practice, um, also because I know you are DC-based, um, given the fact that, um, you know, you've, you've probably seen firsthand the overwhelming impact that it seems like the NSE, like final report has had within the policy community there. And a lot of the, um, and from an IR perspective, I'm also looking at the way that, um, almost like perceptions, uh, of states, of abilities of states, um, sort of matter in discourse and in terms of informing, um, policy decision-making. Like, you know, so I'm wondering to what extent you think, uh, to what extent you think the trajectory of this is a bit path dependent now, where, um, you have enough people in DC driving decision-making that believe in leading sector theory and are driving, you know, and are pushing for that, um, and which in fact, I mean, in some ways there are, there are metrics there that, um, correspond or correlate with, uh, diffusion theory. So there are still some metrics that will get equally emphasized, but then, you know, there's also some divergence which you outlined in this presentation. Um, in the same way also from an IR perspective, that policym has tended to be, especially in this realm, a little bit reactive in the sense that you get a lot of people talking about, um, the, you know, China's new generation AI development plan, for example, after it's been issued, and then that's kind of this like Sputnik moment that kicks off, you know, responding actions or or empathies in on the US side, where, when you have this like, sort of reactive framework for like taking in messages, you know, across, you know, over in within sort of the Chinese leadership where they're saying a lot of aspirational things about wanting to drive like leader leading sectors in China, and then the US, you know, correspond by saying a bunch of aspirational things, you know, so it's almost a little one step removed from actual capabilities or, you know, along those metrics. And so I just, how do you address the, I guess, like the gap between that and like how these powerful memes are often what drive decision-making in these areas as opposed to, um, more of a almost like direct connection with these relevant metrics?

Yeah, yeah. Uh, I think on some levels, I've grappled with a lot of these points that you're making that, uh, it's hard to not do reactive policym. Um, some of these are already so baked in that it's hard to make change. So there are some days that it's like, uh, what are we doing here? Um, but, you know, for me, I think, you know, your, your question is almost more of like a broader question of how do, how does research make an impact? And I think for me, sometimes, uh, I'm not in the room. Uh, I'm not going to be the person in the room. I'm not even going to be the person that's influencing the paper in the room, but hopefully this is influencing the air and the ideas that are floating around in the air in the room, and that's sort of how you kind of convince yourself that the work that you do is meaningful on some level. Um, but no, it's, I think, I think, um, and I think for, for me, it's a unique place where I'm in DC where, like, I have briefed this book to the Secretary of Commerce. I'm going to brief this book to the State Department's policy planning committee, to all this. So, I, I think there is like, I'm not saying I'm sometimes in the rooms, I will say that too. But, um, I, I would say like, um, the, the other thing I would say is I think the tone in DC is shifting where the perception that China is this 8-foot-tall giant that's going to dominate the US and AI, I think that's starting to shift. And I think you're starting to see a lot of people realize that that's not necessarily the case. And especially in AI, Kayla, you probably have takes on this as well, not to put you on the spot, but if you have anything to say, feel free to jump in.

Yeah, my question is actually following on from that, and it is, can I just say very quickly that we have only a few minutes left, so it has to be a relatively short question and a relatively short answer? Yes, but over to you. So, um, with that take that you, you just had, and in the book, you, you know, through the theory, it predicts that China will be definitely not on the leading edge after taking into account these diffusion, uh, hesitations. What factors would make you change that take, um, most quickly, and what, what time horizons should we be looking at for those?

Yeah, I think, um, I think the trend is changing, right? China is, but I do, China is investing a lot more in sort of broad-based technical talent, um, in AI. I don't know what the time horizon is. I think some of these investments take a long time to pay off. Um, so, uh, of the, one of the most important industrial policies the US did in the Second Industrial Revolution was the Morrill Act, to sort of expand all these land grant institutions to training the mechanical and the agricultural arts, but that was by accident, it wasn't like a targeted, uh, diffusion policy. So, so I think I, I would look at, I would be focused on the GDP skill infrastructure indicators for China. I would be looking at, um, how strong are the communication linkages between different parts, parts of the science and technology ecosystem. Um, so, so those are some of the, some of the indicators I'll be looking at for the future.

I think we have to end it there, unfortunately. Um, thanks so much, Jeff, for work that, that really brings, uh, technology more into the conversation than, believe it or not, it is, even though I think all international relations scholars recognize its importance, but still treated as a kind of unknown, and it's never kind of really brought into the theory. So this, uh, this is really starting to do that, which is, which is very much, uh, welcome and overdue. But, uh, let's thank Jeff. Thanks so much for coming back.