Transcript of the talk, from a Whisper transcription of the audio. Auto-generated: speaker names in particular are unreliable. ======================================================================== # Tracking China's Growing AI Innovative Capacity Using Publications, Patents, and People Authors: Sagar V. Baviskar, Lee G. Branstetter, Cameron Drayton, Douglas Hanley, Eduard Hovy, Lei Li, Xibao Li, Prasanna Tambe, Jiarui Wu, Liujie Wu, Huiyan Zhang Discussant: Daniel Trefler Video: https://www.youtube.com/watch?v=pwODfPHl7WY&t=13050s ## Talk (03:37:30 – 03:59:40) [03:37:30] I'll ask you to all sit down ready for our final paper presentation. [03:38:03] All ready to go with so many coffers, but please take it away. [03:38:09] All right, I guess I'm waiting for the slides to come up. But while I do, can I just ask everybody in the room to join me in thanking the organizers for this incredible. [03:38:20] Convening. [03:38:21] It's been so thought for looking. So thank you very much. [03:38:24] I also want to acknowledge the effort of my co authors, my many co authors who've done incredible work. We're still kind of at the beginning stage. [03:38:34] There we go. [03:38:37] But, you know, and we've got a long way to go. [03:38:40] I also actually want to thank the earlier panel, right, the one that immediately preceded me. I mean, what a thoughtful dissection of this notion of an AI partner race and, you know, what a great way to sort of think about whether we're actually went in one or not. [03:38:56] But my starting point would be to observe that there are lots of people less sophisticated. [03:39:03] China's winning hands down. So here the reddish orange line counts the number of AI patents granted by China's patent office by year, the overwhelming majority of which go to Chinese patentees. [03:39:22] The blueish line counts the numbers of AI patents granted granted by the US PTO by year, and the overwhelming majority of those go to Western inventors with Americans in the majority. [03:39:32] And the lines are not even close. [03:39:35] Now when we narrow our focus to what we call core AI patents, the aggregate numbers of designated AI patents drop substantially on both sides, but the pattern of a quickly emerging and eventually overwhelming Chinese advantage remains. [03:39:52] And a very similar picture has been generated by other leading scholars who have used very different definitions of AI patents. So Han Ming and a group of his co-authors have similar to ours. [03:40:12] So is the picture painted by the work of Josh Lerner and his co-authors who've also come out with a paper in recent months that tracks Chinese patents granted across a range of so-called critical technologies including AI. [03:40:24] So different research teams, sorry, I need to do this. There's Han Ming's. There's Josh's. Different research teams, substantially different approaches, strikingly similar pictures. [03:40:42] And Lerner concludes that these pictures suggest the possibility of Chinese sputnik moment, a Promethean burst of Chinese innovation dramatically outstripping anything in the West. [03:40:55] We respectfully disagree. The main thrust of our paper is that in China, AI patents are a great example of Good Heart's law, which is interventions have induced a patent surge that is at least somewhat disconnected from underlying inventive capacity. [03:41:18] And after 2017, these interventions focused more intensely on AI. [03:41:23] We know that there's impressive AI innovative capability that's grown rapidly in China, but how do we separate the wheat from the chaff and patent data? [03:41:31] We're going to rely on the fact that advances in AI technology have built closely on advances in AI science and the skills of AI scientists. [03:41:40] And we're going to use publications and conference participation to identify an AI expert network in both the US and China and measure the connection of different patenting entities to that network. [03:41:52] And we're going to find that a large fraction of Chinese AI patents are obtained by entities which have little or no connection to this network, but face strong policy induced incentives to patent in priority areas anyway. [03:42:06] And this is less true for the US. [03:42:09] So should we care about AI patents at all? And I see Abe shaking his head when I pose this question. [03:42:16] Right now, some of the audience may be thinking that, well, what's really interesting right now in the AI space are the recent changes in the performance of frontier foundation models, the most powerful general purpose LLMs. [03:42:29] That's evolving very quickly, right? More quickly than patents can really catch up. [03:42:33] And the most exciting firms in the space have either been patenting very selectively, at least until recently, or taking an open source approach. [03:42:40] Our partial defense is that we want to take a somewhat broader view of AI, including some of its enabling hardware technologies. [03:42:47] We want to look over a broader time period than that which has elapsed since the emergence of chat GPT. [03:42:54] And we would argue that given that, that longer time horizon, that more broader approach, patents might be a little more relevant than otherwise. [03:43:06] And I will note that my imminent discussion, Dr. Anne Trefler has used AI patent data in his own work. [03:43:11] So presumably he agrees with me that it's not a complete waste of time to look at it. [03:43:16] But there are also important people outside this room who are taking patent counts pretty seriously as a way of measuring progress. [03:43:25] And they're viewing it through the context of international geopolitical competition. [03:43:29] Some of these are powerful people in a position to make important decisions about research funding or product regulation. [03:43:35] And if their decisions are based on a fundamental misinterpretation of what the patent data say, then there could be adverse consequences. [03:43:42] So with that as justification, let's take a look at this graph, which is a stacked area diagram that breaks down China's core AI patent explosion by applicant category. [03:43:57] Now I don't think I have a laser pointer. Do I? [03:44:01] All right, so I might have to do the game show host. Yeah, okay, here we go. [03:44:06] All right. [03:44:08] Who's next on the real fortune? All right, so this deep red slides right here are small patentees, right? [03:44:16] There are entities that have been awarded fewer than 10 AI patent grants by the end of the period. [03:44:23] Now we might think that some of these are exciting emerging startups. [03:44:28] But they're actually not. [03:44:29] I mean, to the best of our knowledge, these entities have not hired any of our expert Chinese, you know, AI superstars. [03:44:37] Most of them are located far from Chinese tech hubs in places like Beijing or Shanghai. [03:44:42] And there are thousands and thousands of such marginal applicants. [03:44:47] This orange red slice are universities, right? [03:44:51] But they're not the universities you might be thinking of. They're not Qinghua or Shanghai Jiao Tong, right? [03:44:57] These are universities that in terms of their citation adjusted patent, sorry, publication portfolio, are ranked below the level of the University of Central Florida in the United States. [03:45:06] Now, if there are any UCF alums in the room, right, I want to acknowledge that it's an important source of STEM education in the Sunshine State. [03:45:14] We don't think of it as a driving force in the world of AI, okay? [03:45:19] But these universities in China are getting a lot of patents, right? [03:45:23] They're generating 93 times as many patents as similar universities in the United States. [03:45:29] This orange slice are state-owned enterprises that through best of our knowledge have hired no member of the Chinese AI expert network. [03:45:37] The orange slice are the leading universities and their capabilities are surely impressive. [03:45:43] They're now some of the top rated computer science departments in the world, okay? [03:45:47] But they're generating nine times as many patents as their US counterparts without generating nine times as many papers or citations. [03:45:54] The green slice are larger Chinese companies that either hire a member of the network or they have a product or service portfolio that at least partly validates their patent grants. [03:46:04] I don't know if you can see it, this light blue slice are foreign firms patenting in China. [03:46:09] The darker blue slice are state-owned enterprises that have hired a member of our expert network and the gray is everything else. [03:46:17] All right, now I guess I'll just leave it here. Where's Vanna White when you need her? [03:46:27] All right, so, you know, what are we taking away from this? [03:46:31] Well, prior research has already noted the surprisingly large role played in Chinese AI patenting by universities, small firms, and SOEs. [03:46:38] What we're showing is that by 2025 a majority of AI patents granted in China are being awarded to categories of patents, [03:46:45] categories of applicants that are largely disconnected from the expert network we've been able to identify. [03:46:52] Now, when we focus on the subset of Chinese applicants with stronger ties, [03:46:56] we actually see a number of AI patent grants that's not that dissimilar to what we've seen in the US, right? [03:47:02] You know, overlaid on top of this diagram, this purple line are the corresponding AI patent grants in the US. [03:47:09] Okay, let's flip our perspective and look at AI patenting in the US, right? [03:47:16] So, these are core AI patents granted by the USPTO. [03:47:19] And for perspective, that top red line shows the total of all AI patents granted in China. [03:47:26] The bottom red line shows AI patents granted to Chinese corporations with non-trivial patent portfolios. [03:47:32] Okay, and what we can immediately see is that the vast majority of AI patents in the US are awarded to companies with non-trivial patent portfolios. [03:47:41] The pink slice shows patent awards to US companies, companies based in the US. [03:47:47] The orange slice shows patents awarded to non-Chinese, non-US companies. [03:47:52] That little red slice shows the small but growing role of Chinese companies as applicants, [03:47:59] sorry, grantees in the US patent system. [03:48:02] And non-corporate assignees account for very little. [03:48:06] Okay? [03:48:08] So, when we compare Chinese corporate AI patent grants to US patents, which are mostly awarded to corporations, the numbers are about the same by 2025. [03:48:20] China's levels of AI patents are growing very quickly, to be sure, but the picture is one of rough parity by 2025 rather than a blow up that favors China. [03:48:30] So it's mostly patenting by non-corporate entities in China that's producing the divergence in the headline numbers. [03:48:37] And where does that surge come from? [03:48:40] Right? Well, it comes from policy interventions. [03:48:44] Right? And these began with China's consequential accession to WTO way back in 2001. [03:48:50] WTO accession brought the ratification of trips, which at least in principle strengthened IP protection, [03:48:57] potentially allowing foreign firms to block indigenous technological development with their growing patent portfolios. [03:49:05] To prevent this, national and local governments sought to encourage indigenous inventors in China to construct a defensive great wall of Chinese patents, [03:49:15] and local officials were held responsible for meeting patent application targets in their jurisdictions, the KPIs that were referred to earlier. [03:49:23] The Shanghai government showed how much local patenting could rise if local government covered patent filing fees, [03:49:30] and soon local governments across China were copying these generous subsidy programs, now seen as essential for local officials to meet their patent targets. [03:49:40] Another inducement where the tax breaks offered to so-called high and new technology enterprises, right, [03:49:47] so smaller enterprises could qualify for a very large cut in their corporate income tax rate by taking out just a few patents in targeted areas. [03:49:56] This program has been described by research undertaken by Daniel Ashu and others. [03:50:03] Right, and finally, okay, there we go. [03:50:09] Patent-based promotion pathways for university professors along with generous subsidies for university patenting, [03:50:17] a generated wave of university patents across the country. [03:50:21] Now, the patent policies previously pointed to were not unique to AI, [03:50:32] but an important policy directive in 2017, which was mentioned during the previous panel, right, [03:50:37] that new generation or next generation artificial intelligence development program [03:50:42] made clear to all levels of government that AI was going to be a central target of technology promotion policy. [03:50:49] That policy shift has been documented. [03:50:51] It's clear that local governments took it seriously. [03:50:54] They were now under pressure to raise AI patenting and their jurisdictions. [03:50:59] And Martin, not the Peterson Martin, the Berkeley Martin, [03:51:03] showed that government-affiliated venture capital firms dumped a bunch of money on firms working in AI after this policy directive went into effect. [03:51:15] So essentially what we're arguing is that all that money and policy emphasis produced a lot of AI patents. [03:51:26] Now, even as the Chinese government at all levels was focusing more resources on AI, [03:51:32] the emerging evidence of a low quality patent bubble in China was becoming too great for the government to ignore, [03:51:39] as was evidence that firms were manipulating their patent and R&D records to qualify for that juicy H&T tax break. [03:51:48] Right, so what happens next? [03:51:51] Well, the Chinese government sought to undertake a major reform of its patent incentives in the early 2020s. [03:51:57] Now, some of this was outlined in this document produced by the Chinese Patent Agency, [03:52:03] which called for an almost immediate ban of subsidies for patent applications [03:52:08] and an eventual ban on grants to patents successfully received. [03:52:15] There was also a wave of company audits to try and crack down on abuse of the H&T tax break [03:52:22] and the Ministry of Science and the Ministry of Education tried to move from an emphasis on university patenting quantity to university patenting quality. [03:52:34] Right, you know, in general, that's how all of the instruments of government policy in China were trying to move. [03:52:42] Right, laudable. [03:52:44] Right, but the problem is that patent quality is pretty hard to assess at early stages in the innovation process, [03:52:51] even for the innovators themselves. [03:52:53] Often, we don't know how valuable our technologies really are until they're, you know, competing with one another in the marketplace. [03:53:03] So not a great time for my younger son to call me. [03:53:10] So, you know, we argue in the paper that the incentives to overpatent are still there despite efforts at government reform. [03:53:22] They're embedded in the system of qualification for the H&T tax break still. [03:53:30] Right, they're embedded in the incentives that banks have to make loans using patents and targeted areas collateral. [03:53:37] They're embedded in the ways that government venture capital funds reward patenting and targeted technologies with funding. [03:53:44] They're embedded in the ways that patents still count towards tenure and promotion for faculty. [03:53:50] And they're embedded in the KPIs that local officials face with respect to AI patenting in their jurisdictions, [03:53:58] or at least high value patenting, which is an aggregate that includes AI patenting. [03:54:11] And AI patents to these small patentees as energetic efforts by small firms to qualify for the H&T tax break. [03:54:20] It's easy to qualify to view this massive surge of patents being granted to lower tier universities as a policy driven university patent bubble. [03:54:31] And it's easy to see the energetic investment by state owned enterprises as being a response of entities under direct government control to government policy priorities. [03:54:45] All right. [03:54:46] All right. [03:54:51] So what else can we do? [03:54:53] Right. [03:54:54] Well, it's hard for inventors even to evaluate the quality of their invention. [03:55:01] Right. [03:55:02] But there's a long tradition in economics that long tradition of economic research that validates this proposition. [03:55:16] Right. [03:55:17] And the EPO tracks international patent families. [03:55:20] And so we bought this data and we were able to match roughly 99% of our Chinese AI patents and 99% of our US AI patents to this data on international patent families and determine which of these patents was taken out in at least one country other than the home country of the inventor. [03:55:37] And how does the picture of Chinese AI patenting change when we apply this screen? [03:55:44] It changes a lot. [03:55:46] All right. [03:55:47] Now we have to be careful here. [03:55:49] Right. [03:55:50] The later years of our data are now subject to a new truncation bias because international patent law gives applicants additional time to figure out where else in the world they're going to patent. [03:56:02] So, you know, past 2020 picture of overwhelming Chinese dominance changes utterly that purple line. [03:56:17] That's multinational patents granted by the US PTO. [03:56:21] Right. [03:56:22] This are the Chinese patents that are taken out in more than one country. [03:56:26] So there's now a huge gap between the two. [03:56:29] It's actually growing over time. [03:56:31] And the distribution of AI patenting in China that looks so different from the US before now looks much more similar. [03:56:39] Right. I mean, most of the patenting is concentrated in these big Chinese companies patenting by other entities. [03:56:45] Once the screen is applied shrinks into relative insignificance. [03:56:51] All right. [03:56:52] I have like one minute left. [03:56:54] Catherine. [03:56:55] Give or take. [03:56:56] Are there other ways that we can look at patent quality? [03:56:59] Yes, there are. [03:57:00] And as usual, Hanming has beat me to the punch. [03:57:03] Right. [03:57:04] So one way to evaluate patent grants to publicly traded companies is to look at the stock price reactions to the news of the patent grant. [03:57:13] Right. [03:57:14] These are the methods pioneered by Kelly at all. [03:57:16] Hanming actually did this. [03:57:18] Right. [03:57:19] Really interesting results. [03:57:20] But when I convert his results to a common currency and you can tell me if I miss reading them, they suggest that on average Chinese patent values are substantially lower than the US counterparts. [03:57:31] You can look at citations. [03:57:33] Hanming has done this to already in his data. [03:57:37] He comes to the conclusion that Chinese patents site US patents a lot more than the reverse. [03:57:43] But of course in China, most citations are added by the examiner rather than the applicant. [03:57:47] Right. [03:57:48] So we're not absolutely sure what they say about knowledge spillovers. [03:57:51] Finally, Kogan at all have actually used text in the patent document itself and it's similar similarity or diss similarity to patent documents that preceded it or come after it to try and judge patent quality. [03:58:06] You know, I'll just kind of skip to the point. [03:58:10] We apply these methods to our different categories of Chinese AI patent applicants. [03:58:17] The big takeaway seems to be that on average, it's the elite universities and the big companies that are producing higher quality patents. [03:58:24] But what's also interesting is that for almost all of our applicant categories as the patent explosion unfolds, measure novelty actually falls. [03:58:34] All right. [03:58:35] Well, there's a lot more to say, but I have no time. [03:58:37] Right. [03:58:38] So what do we conclude? [03:58:40] Raw Chinese AI patents are unlikely to work well. [03:58:43] And this is I think is Han Ming's phrase as a ruler, right, to measure comparative inventive progress across the two countries. [03:58:51] Policy induced patenting distortions persist despite reform efforts. [03:58:57] They're not data on academic publication and personnel movements provide an independent indicator and inventive capability and these point to impressive and growing Chinese capabilities, which should not be discounted. [03:59:08] They're very impressive, but they're not well reflected in the patent data. [03:59:13] What are we going to do going forward? [03:59:15] We're going to expand and extend our data on the Chinese expert network. [03:59:18] But for now, I'm going to be quiet and let my discuss and take the stage. [03:59:23] Thank you very much. [03:59:31] Oh, sure. And now I have to hold it. [03:59:34] I can figure out how to do that. [03:59:36] I guess I should start by checking my timers. ## Discussion (03:59:40 – 04:10:20) [03:59:44] Thank you. [03:59:48] This is about China. So I want to thank the gang of four. [03:59:52] Adjaye and the three of them over there. [03:59:57] So this is a terrifically interesting paper. [03:59:59] You saw the richness of the institutional detail and you saw the enormous amount of work that particularly went into putting together those measures that he used the puzzle. [04:00:10] I don't really have to review it. [04:00:12] You the raw data suggests that China is just patenting way more than United States in core AI patents. [04:00:21] And it's a number that suggests that China is way better than the US in this space. [04:00:28] And after his numerous adjustments, all of which are very sensible and I fully endorse all of them. [04:00:35] I received that the Chinese number of 25,000 falls to 300 patents and the US number of 6000 falls to just 2000 patents. [04:00:43] And now the picture is maybe China is really bad. [04:00:47] I'm not sure. [04:00:49] But clearly there's important adjustments to be made. [04:00:52] And then as we go further into this project, where is Lee? [04:00:57] As we go further into the project, Lee has a ton of other ideas for how to make these adjustments. [04:01:01] Looking forward to seeing them. [04:01:04] This network of experts is a really difficult thing to put together. [04:01:08] And when are you going to publish that data? [04:01:13] I've been trying to put together that network of experts. [04:01:16] It's just a ton of work. [04:01:19] All right. [04:01:21] There's other puzzles when we're dealing with measurements and puzzled in the air that I'm working on now, [04:01:25] which is the market for large language models. [04:01:29] If you take a look, I just randomly picked something in early 2025. [04:01:35] What are the most popular models as ranked by the standard most common metric, which is number of tokens process? [04:01:43] You see that chat GPT is by far the dominant player. [04:01:48] You see that with the exception of deep seek, nobody else is even close. [04:01:52] Everybody else is American. [04:01:54] I'm sorry. [04:01:55] What I meant to say is everybody else is American. [04:01:57] Let's fast forward to today. [04:02:01] And you see that the Chinese are so dominant. [04:02:05] So totally dominant by number of tokens processed. [04:02:08] Okay. [04:02:11] Then the question is, is a token a token? [04:02:13] And the answer is no. [04:02:17] Some models are way more efficient than others in the way they encode and decode these tokens. [04:02:23] And in particular, as an average Chinese models require about 10 tokens per US token. [04:02:30] So let's just divide by 10 and sometimes it gets as high as 40 tokens per US token. [04:02:35] So divide by 10 and suddenly it looks like the US is dominant again. [04:02:41] The nice thing about this exercise is that I can go to the computer science literature and pull out an engineering number for how to adjust tokens for patents. [04:02:51] It's obviously much, much more difficult. [04:02:54] So here's a proposal for that. [04:03:01] That we could consider doing that Lee has already is already going to do. [04:03:07] It's to train a neural network classifier on US patent filings just on US patent filings. [04:03:12] And the outcome is did the did the patent filing get was it granted. [04:03:18] And then you take the Chinese data, filter it through that model and predict whether the Chinese filing would have met the US criteria. [04:03:29] And now you have a predictor for what share of Chinese patents would have been granted had they been applied for in the US. [04:03:36] And that would be another way of getting at this. [04:03:39] And after talking to Lee, we decided that we're starting up a new company, which is going to be a consulting company for patent patent lawyers. [04:03:53] And I have term sheets in my in my briefcase. [04:03:57] So after this, if anybody wants to come talk to me. [04:04:01] Great. We already have some in bed. [04:04:05] All right. [04:04:07] Why does this matter is a is a big question. Thank you. [04:04:14] And I'm going to say things that I think everybody in this room has already sort of processed in the last few sessions. [04:04:25] As I'm in a trade economists, that's my background. [04:04:28] And when I think about innovation in China or anywhere on the planet, I think it contributes to a pool of innovation for the whole whole planet to benefit from. [04:04:36] And as a result, I don't care where the innovation happens. I am very happy that it's going to make all of us better off. [04:04:43] That's a little bit close to there are gains from trade. [04:04:48] So let's not forget that piece. [04:04:52] Of course, there's other people who would sort of disagree and say that this all this technology from China is a Trojan horse that's going to destroy the US economy. [04:05:01] And if there are people who will suffer, you know, we can go to our tour door and enhance as many papers on the subject and find that, you know, young white males are being hammered by China. [04:05:12] And other sectors of the economy are benefiting from China. [04:05:16] So we need to think as as we've seen over and over again today that there are nuances and we need to think in more detail more granularly than just make large grandiose statements about what needs to be done. [04:05:31] There is Janice's many, many decades of important research on this topic and it has been looming large for a very long time. [04:05:41] So obviously you don't want me talking about that. You want Janice talking about that. [04:05:46] And then another thing that we might care about because we are economists, most of us is the stuff that we normally do. [04:05:53] Why do we have to have new questions? Let's stick with old questions. [04:05:58] Let's ask the basics that we would for any innovation question. Is there too much or too little innovation? [04:06:04] Should we be encouraging more innovation in the spacers? [04:06:06] There are already enough innovation in this space. [04:06:08] It's a basic question. And you know, there's always the tradeoff between increased consumer surplus versus business stealing. [04:06:15] And so finding out where we are on that would be very, very helpful. [04:06:21] The second thing is if you read through at least paper, you will see the absolutely horrifying second best policies of the Chinese government. [04:06:29] They're so distortionary as to make patenting a fool's errand. [04:06:35] So we want to think about what are the best of the second best policies. [04:06:43] Is it the US model that should provide, should guide us? [04:06:47] I don't think so personally. I think I look at the American model and think it's got its own horrible distortions. [04:06:54] So I'm not sure what the benchmark should be for if I was advising the Chinese government. [04:07:00] I could certainly tell some things to the US government. [04:07:04] But we need to think about what are the second best implementable policies in this space. [04:07:10] So these are just all I'm doing is thinking, if I'm just working in the space of innovation, what are the standard questions? [04:07:17] Thank you. [04:07:19] And then this is what I thought was going to be my most important contribution. [04:07:25] But now I think we're all talking about the same point, which is what is it that matters. [04:07:30] Lee's work is unabashedly about frontier models. He's very, very, very clear about this. [04:07:37] But there's other things that we care about. We care about adoption rates as we've been discussing today. [04:07:43] China is targeting 70% adoption rates. We have had a long enriched discussion about what exactly is the adoption rate and you know, in which industry in what context, etc. [04:07:53] We need if there is a race or two races, we need to know about the adoption race. What is going on in the adoption race? [04:08:02] So we're getting some stuff here today, which is great. And we're going to need more stuff. [04:08:10] Just one example, because I spend I spend about three weeks in China every year. The gym example always just struck me. [04:08:17] I've met a colleague who said, yeah, we use AI in the gym. And I'm like, you use AI in the gym? [04:08:24] For what? Well, when you're lifting weights, your form matters, you can hire in Toronto for 100 bucks an hour, 70 to watch your form, or you just get an AI with the camera to watch your form and suggest corrections. [04:08:41] So, you know, when I'm in China, I see adoption everywhere, right? Not in Canada. [04:08:49] All of this comes back to not all of this, but this emphasis on expert networks, which is very important for these thinking goes has a lengthy tradition. [04:08:59] It goes all the way back to Joe Mochir's work. And if you haven't read the new book, they emphasize not only useful knowledge that we're creating knowledge that can be used in applications. [04:09:17] But there have to be people on the ground to know how to implement that. In the industrial revolution, you needed, you know, people who, you know, Smith's who could work with steel that was of the right tensile strength to work a, you know, a steam engine as an example. [04:09:36] So, my sense, and this is my long term prediction that China with its huge manufacturing base, and with some of its other initiatives that Suresh was talking about in health. [04:09:48] I feel like China in the adoption race, it's not, it hasn't won. But, you know, in 10, 20 years from now, you know, if I was a betting man, I would bet it on China. [04:09:59] So, let me conclude with that. That China is going to win the adoption rate. ## Q&A (04:10:20 – 04:23:00) [04:10:24] Okay. I have a couple of questions. One is patenting abroad is very expensive. And so there's also a natural bias, you know, toward patenting domestically. [04:10:38] So I wonder whether there's a way to correct for that. And second is, you know, you read all those news articles and I know Australian National Institute published a report looking at publications, not patents. [04:10:53] And, you know, in 2020, 2015, US was dominant, you know, 57 out of the 60 plus critical technologies in 2025. I am trying to reconcile the fact, and if you, I actually I did this purposefully, I was so struck by the numbers that opened that nature and science and just hand counted how many others have that, you know, Chinese name. [04:11:23] It's just stundingly high. And so I'm trying to reconcile some anecdotal evidence we're seeing versus the results you report. You know, which picture is more accurate. [04:11:42] Leah, you know, I am. Thank you for a great presentation. I, in terms of the market valuation for publicly listed companies, right when a firm is granted an AI patent, you can look at a stock market reaction to the patent granting news to kind of value the to get a market value of the of the patent. [04:12:04] It is true that, you know, that actually the numbers like the numbers for the stock market reaction for for Chinese company, probably listed company getting a patent granted is the same as the US except the currency is different. [04:12:19] So it's the one is in US dollar. The other one is in Chinese yen, but numbers actually very similar. Okay, so, but that does, you know, it's consistent with what you said, you know, the market values a Chinese company, probably companies granting a patent at the basic one sixth of one sixth, one seventh of the US market. [04:12:39] The other thing that's interesting is that for the non listed companies, right, including the universities and and non listed companies, their patents, we can you cannot use the market, market response. [04:12:51] So we basically look at whether the non public companies patents are decided less often than the patents are granted to public listed companies. We found that that's actually that equally likely to be cited suggesting that, you know, there's no, at least from the citation statistic, there's no clear evidence that the patents granted to non public listed companies in China are [04:13:21] of lower quality. Okay, so this is just a supplementary to what you what you mentioned. But you know, I think we also reached the same conclusion that in terms of cross border citations, Chinese patterns definitely side more the US patents, even though there's some emerging [04:13:39] subfields within AI that learning is coming in that there's a reverse direction but overall Chinese patterns definitely side more the US frontier knowledge more. Okay. Yeah, so. [04:13:54] Yeah, I wonder if I feel like what you showed us in the aggregate was that there's a lot of noise in the patent data. And then you gave us reasons for the noise which were a variety of incentives. [04:14:23] And I'm trying I was trying to think what the counterfactual is. And so in other words you can imagine a country that did not have all those incentives to patent. And you know the thing I think we've always known about the patent data is most patents are junk. [04:14:46] And and most papers are. And so in other words, there's just a very long tail. And most papers have almost no citations most patents end up, you know, not conferring rights that the owners those patents are able to extract any rents from. [04:15:03] So there's only a few things that actually matter. And, you know, in your other paper that you presented at the NBR at Summer Institute, you have in that paper, a page just got like a dozen innovations that that were the major milestone innovations in the development of AI. [04:15:27] And so if all if these Chinese policies. [04:15:31] That created a fair amount of. [04:15:34] You know, potentially bogus patents, if they if they drove even a small number of first order innovations, then it's not obvious that on balance it was a bad idea. [04:15:53] And furthermore. [04:15:55] The metapointing of the question is, is there still signal in in what in what we know in the 300 and and and you presented it as a comparison to us and I thought is that the right counterfactual. [04:16:12] You know, there's right kind of factual could be all the other countries that we're having almost no innovation in AI. [04:16:21] In other words, almost everyone else but the US. [04:16:24] And in terms of commercial AI Canada was a leader in in academic research but very little commercial. [04:16:32] And then, you know, and comparing the innovation of of those countries. [04:16:39] So, I'm not sure in other words, should we take away from your paper that we can't. [04:16:45] The patent data is junk. Should we take away that there's that. [04:16:50] We need to know that there's signal in this data, but we have to treat the data in a. [04:16:56] You know, in some kind of abnormal way for those of us who've done a lot of empirical work with patent data. [04:17:02] Like how what's the take home message from this with regards to signal in the patent data. [04:17:12] Actually, I want to know at least answer that question, but I guess before we give him that opportunity, I want to land on where a little bit where Dan landed, which is so if there's two different races. [04:17:29] And the US is running a race to the frontier and China is running a race to the diffusion. [04:17:37] Does that mean everybody wins or everybody loses, I guess, depending on how you think about it, or is maybe this is too much of an extrapolation for what you showed us in your in your paper, but you know, [04:17:51] or is there a different story here that one of these races is the right one for whatever particular objectives people have and that's sort of a question about to leave it's also maybe more question to the foot economy panel we had earlier today and for the whole race. [04:18:06] So, I think you've got some big questions to ask in a amount of sit in about two minutes. [04:18:20] Oh, first of all, Dan, thanks. And yeah, anytime you want to meet to get that consulting firm started, I'm already go. [04:18:28] Thanks. Thanks for these great questions. [04:18:31] You know, yes, on publications, right, we see strong and really incontrovertible evidence of rising capability. [04:18:39] You know, we actually had a slide that I didn't have time to present, which looks at elite AI conferences and counts the fraction of papers presented at these conferences that are authored by people with a Chinese institutional affiliation, not a Chinese name, but a Chinese institutional affiliation. [04:18:58] And that number has gone from almost zero to about 50%. [04:19:03] Right. In just a few years. [04:19:04] So, yes, there are incredible capabilities in China. [04:19:08] But where are the residents? Right. I mean, the resident and elite universities, right, the resident and leading companies. [04:19:16] And, you know, those capabilities absolutely have to be respected, right. And to the extent that some of them emerge from, you know, Chinese government policy, we can't say that it's all bad. [04:19:27] Right. But if we take the Chinese policy metrics seriously. [04:19:32] Right. Then I think and in appropriate rulers being applied within China itself, real resources are being dispersed in an inefficient way. [04:19:44] And there are social costs. Okay. But don't get me wrong. Right. I mean, the bottom line of the. [04:19:50] There's no AI innovative capability in China that's totally wrong. Right. They're incredible capabilities. They're growing quite quickly. [04:19:57] You just can't make much inference about them based on the patent data. [04:20:02] So, I mean, I think we're basically, we don't, we don't disagree. Right. You know, the stock market, you know, has kind of a discount factor of about one seventh. [04:20:12] You're suggesting that this would also apply to, you know, the non-corporate patenting. Right. But again, you know, it's hard to make inference about knowledge spillovers of the vast majority of citations are coming from examiners rather than applicants. [04:20:24] Right. I mean, maybe the examiners can see it, but maybe they can't. [04:20:28] Um, Avi, you know, who wins and what's the right race? Wow. I definitely don't have time to address that. But, but here's the thing. Right. [04:20:37] Um, if the diffusion race is won in China, then what would be kind of the acid test of that? [04:20:45] What we would expect to see is that the rate at which China is Chinese macro productivity is converging to US levels would accelerate. [04:20:54] It really accelerates sharply. And the heartbreaking fact of the Chinese economy over the last 15 years is that that rate of productivity convergent seems to have decelerated rather than accelerated. [04:21:07] Now, maybe it's too early. Right. Maybe later this year. Right. There will be a sharp acceleration that's identifiable in the data. [04:21:15] But if it doesn't emerge, right, then it would be hard to conclude that whatever is going on in terms of, you know, AI adoption is really generating the ultimate economic fruit we would be looking for. [04:21:28] Um, I think I might be out of time to address your question. Yeah, more or less. How yeah, well, you get in seconds. [04:21:35] Yeah. I mean, what's the counterfactual look the thing that is awe inspiring about China over the last 40 years or 50 years has been its ability to mobilize the human capital of its people. [04:21:48] Right. And, you know, ultimately this could be China's great gift to all of humanity. Right. I mean, what fraction of the humans capable of inventing at the technological frontier in any field are currently resonant in China. [04:22:01] It's a huge fraction. Right. And that's an amazing accomplishment. Right. One for which China deserves full respect. But how are the efforts of those people being allocated. Right. How effectively are they being mobilized in the cause of invention for, you know, either China or the rest of the world. [04:22:19] I do have concerns that the efficiency of that mobilization is being compromised. And that's a problem for the entire world, not just for China, but for the entire world. Right. I mean, we're kind of counting on those people to come up with a cure for Alzheimer's right and help us build data centers in space and all that kind of good stuff. [04:22:40] Right. If what's motivating them and guiding them are, you know, distorted of policies to some extent. That's a human problem. And I think on that somewhat somber note, we have to conclude right, Catherine. [04:22:53] We do. What a lovely place to end. It's a human problem. And I think that is going to be definitely our kickoff for our reception and dinner. So the way this is good. So first of all, let's have a round of applause just for everything that's happened.