In this episode of the McKinsey Global Institute’s Forward Thinking podcast, host Michael Chui speaks with Jeffrey Ding, researcher and founder of the ChinAI Newsletter, about information asymmetry in artificial intelligence between China and the West. They cover why data may not be like oil, the Chinese industry adage on products, platforms, and standards, “unsexy AI,” and more.
Anna Bernasek, co-host: Michael, there’s a lot of talk right now about artificial intelligence, or AI, and what it means for global competition. I’m really glad we’ve got a guest today that can talk to us about what’s really going on, particularly when it comes to the US and China.
Michael Chui: Yeah. It definitely is a fascinating topic—at least, I find it personally. I’m a former AI practitioner and more recently, at the McKinsey Global Institute, have been able to study the impact of AI on business and more broadly. And one of the reasons I’m so excited about today’s conversation is because it’s with somebody you probably don’t know yet but probably should. He’s famous in certain corners of the internet but his work, it turns out, is relevant everywhere.
As you alluded to, our MGI research suggests that while there’s AI happening all around the world, there are two places where the most AI development is taking place, and it’s the US and China. And what’s interesting about that is that while a lot of the Chinese AI developers are reading and even coauthoring English-language papers, very few Western AI practitioners are able to keep up with the flow of information in the Chinese language, even when a lot of it is published openly.
Anna Bernasek: It’s almost like a one-way mirror in terms of the way information flows?
Michael Chui: There’s definitely been an asymmetry, which might seem strange in a field where a lot of the work is openly available on the internet. But our guest, Jeffrey Ding, has been helping to make sure more AI information flows back from China to the West. He’s a doctoral student at Oxford doing a fellowship at Stanford, and he publishes an influential ChinAI Newsletter.
Anna Bernasek: He helps shine a light onto what’s happening in China with regards to AI development. I’m really interested in learning more. Let’s turn to the interview now.
Michael Chui: Jeff Ding, welcome to the podcast.
Jeffrey Ding: Thanks for having me.
Michael Chui: There’s a weird corner of the internet where everybody knows you, at the intersection of China and AI. And then there are a lot of other people who probably need to know more about what you do. So why don’t we start with that? What do you do?
Jeffrey Ding: For the past three years, actually—coming up on the three-year anniversary—I’ve been translating Chinese-language writings on AI and related topics for a weekly newsletter where I will have, usually, a full translation of the article, or government white paper, or blog post from a Chinese writer on AI topics that I really enjoyed reading that week. And then I’ll translate it, digest it for an English language–speaking audience, and also share some of my own reading recommendations. It’s the ChinAI Newsletter, very cleverly named.
Michael Chui: How did this come about? How did you end up writing this newsletter and doing all this translation?
Jeffrey Ding: I was doing my master’s at Oxford in international relations, and around that time, a Centre for the Governance of AI started up at Oxford, and they were looking for interns who were interested in AI policy and governance. But in their list of qualifications, they also put “Chinese language expertise preferred.”
So I just threw my hat in the ring, and they let me join and start writing a report on China’s AI development. And I was researching that report. I came across a lot of texts in Chinese language that not a lot of Western analysts had digested or analyzed.
One of those was a 500-page book cowritten by Tencent, which is a leading Chinese tech giant, and the [China] Academy of Information [and] Communications Technology, which is a government-affiliated think tank under the Ministry of Industry and Information Technology in China.
They had essentially put out this huge book on China’s AI strategy. And as I began translating chapters and sections of that book and just sending it out in emails to colleagues and peers, I got a lot of good reaction from people. And just continued doing that in a weekly email and eventually expanded it into a newsletter.
Michael Chui: You said you have Chinese language skills. Did you grow up in China?
Jeffrey Ding: I was born in Shanghai and moved to Iowa City when I was three. My parents came to the University of Iowa for grad school, and so it’s uncertain whether I think of English as my first language or Chinese, but my parents forced me to go to Chinese school as a kid. And in this dusty apartment basement for three hours every weekend, learning diction, reading textbooks, I kept up my language skills that way. And in undergrad, at the University of Iowa, I also did a Chinese language major.
Michael Chui: All right. So you go to Iowa City, you graduate, you end up at University of Oxford because Rhodes Scholarship, right?
Jeffrey Ding: Yep.
Michael Chui: And you’re studying there but you now also are a predoctoral fellow at Stanford’s Center for International Security and [Cooperation], right? Where are you physically right now?
Jeffrey Ding: I’m physically back home at my parents’ place in Iowa City but remotely doing this fellowship at Stanford.
Michael Chui: You started writing this newsletter based on the translation work that you had done. And then you had a bit of a seminal paper in 2018, right? In many ways, it busts some myths that are out there about China and AI. And so maybe we talk about some of those. I think it’d be interesting. I think there’s one—and this also came up in your foreign affairs paper as well—this idea that “data’s the new oil. China has the most people in the world, therefore they have the most data.”
Jeffrey Ding: I think there’s a lot to deconstruct from that “data is the new oil” myth. The first is that data is always application-specific. Having the most mobile phone users does not translate to autonomous vehicle applications. When we talk about who has the most data, it always has to be application-specific.
And that speaks to a broader point with regards to AI as a general-purpose technology: at least for people like me who study AI and politics, we often just throw around “AI” like it’s a magic word. But the different application scenarios for AI, whether it’s smart manufacturing or transportation or natural language processing, all of those will have different data needs and demands.
And then the technical landscape is changing as well with respect to the salience of data for AI applications. So in some settings, simulated data is becoming more and more relevant. For example, I believe Waymo, in the last year, for their autonomous vehicle application, they drove more simulated miles than actual road miles.
And we’re also seeing developments where you can train on a smaller batch of data as well and still get the same level of performance and efficiency. So I think those guiding principles are important to keep in mind when we think about this “data as the new oil” meme.
Michael Chui: Just because you have a lot of data doesn’t mean you’re going to win. There are all these other factors. What about privacy?
Jeffrey Ding: From reading a lot of Chinese texts, and also from a lot of good English-language coverage that has come out, there’s been a growing recognition that there are actually very robust discussions of privacy and personal information protection in the context of privacy and the protection of data from abuse by companies rather than the government.
I do think that that distinction has to be made, that there is privacy in the sense [that] Chinese consumers definitely don’t want their personal information leaked on the internet or their bank records leaked. Every human being wouldn’t want that.
I’ve also translated reports from Nandu Personal Information Protection [Research] Center that surveyed thousands of Chinese adults, and an extreme majority of them oppose sharing of facial data and think that AI poses a significant threat to privacy.
I do think that there is growing momentum towards more privacy protections. And you see that reflected in actually how Chinese tech companies have responded. For example, federated learning is a technique to train data in a way that’s more preserving or sensitive to privacy concerns. And I’ve at least seen that Huawei, JD.com, some of the big Chinese tech giants have really made an investment in building up their technical capacities in federated learning.
Michael Chui: Let’s talk about another topic that comes up a lot: this idea of competition, particularly between the US and China, in the area of AI. People wonder about what the right metaphor for it is. Somebody asked me, “Is it a race?” And I said, “I don’t think it’s even a decathlon. It might even be an Olympics. It could be more.” Because there are different ways and areas in which people are competing, or different areas of competition.
I think I saw, when someone called it an arms race, that you had a scholarly retort in Foreign Affairs. You also said in one event, “It’s insulting to previous arms races to call it that.” What do you mean by that? Why is the idea of an arms race not the right way to think about competition in AI?
Jeffrey Ding: I think the first way to combat this arms race narrative is just to acknowledge that not all technologies are created equal. A weapons technology is different from a more general-purpose technology like AI. We wouldn’t say there was an “electricity arms race,” with electricity being the quintessential general-purpose technology.
I think the second way to think about the competition angle is just to ask a really simple question: What are we racing for? When people talk about a race between the US and China, are we talking about who can take the most advantage of AI in terms of its transformative effects on military affairs? Are we talking about who can garner the most economic growth from adopting AI at scale? I think oftentimes when people refer to this race, they have maybe versions of all these different things in mind, but it’s never fully specified.
And coming from the academic world, a lot of our job is just to ask this boring question of, “What is the actual problem space you’re talking about?” And I think the people who talk about AI races or AI arms races often won’t be able to answer that question.
Michael Chui: What’s your answer to the question?
Jeffrey Ding: I think the most important and salient aspect of AI for US-China competition is in the economic realm. We know historically that general-purpose technologies bring with them huge upsurges in productivity growth. The best example of that is with electricity and American productivity growth in the 1920s.
And the key challenge for China right now is how to sustain economic growth when their demographic dividend is declining and when they’re trying to climb the value chain in all these different areas, most notably in manufacturing. And so a general-purpose technology like AI provides a potential way for China to continue to sustain really high levels of economic growth, which feeds into all the other domains I’m talking about, whether it’s performance legitimacy for their style of governance, whether it’s how economic developments will also undergird military prowess. That’s really the most important part of how AI will affect the US-China power balance, at least in my opinion.
Michael Chui: Because the size of your economy is the number of people times the productivity, or the number of hours worked times your productivity. And in China, the number of workers actually is starting to decline, which we’ve catalogued at the McKinsey Global Institute, too.
And so unless China raises its productivity—but that’s also true in the West as well, right? The number of workers in Japan is also decreasing, Germany. And the US, but for immigration, would also be similarly challenged. So we all need to increase our productivity.
That said, if we look at the productivity statistics over the past decade or two, productivity growth has been stagnating. And you and I know AI’s been around since—the term was invented in the 1950s. So people have been asking, “Where the heck is this productivity going to come from? We don’t see it yet.” What’s your observation?
Jeffrey Ding: Economists and economic historians can answer this question better than me. I’m mostly drawing on their analysis, like Paul David, and Erik Brynjolfsson, who’s at Stanford right now. They’ve done a lot of great research on this. And obviously McKinsey Global Institute has also looked into this phenomenon.
I think it goes back to the question—I think Robert Solow quipped, “You can see the computer age everywhere but in the productivity statistics.” And it’s a consistent pattern with general-purpose technologies: they take decades, a prolonged period of gestation, before we get the complementary innovations, before we get adjustments in human capital to adapt to structural changes that they bring.
When electricity arrived in manufacturing settings, at first they were just using electric motors as a replacement for the steam engines that were driving this central steam engine that was powering all these shafts and belts that were then coordinating all the individual machines in the factory. At first, they just tried to substitute the electric dynamo, the motor, for that central steam engine. Then, because electric motors allowed for decentralization of energy supply, they did something called “group drive” where they had to power a group of machines.
And then eventually they realized the best way to capture these productivity gains, where you use these electric motors to power individual machines. And it required a complete change in how the factory was laid out, from this belt-and-shaft system to a system of individual, electrically driven machines.
And that process takes a long time. People are used to the patterns of how things work. People have to learn new skills. And so with computers, we did see after a while that there was an increase in productivity. That was one of the key reasons why Japan never completely overtook the US in productivity, because the US adapted computers and information communications technology across manufacturing, across services industries. And it maintained a good rate of productivity growth.
I think it’s right to ask that same question about AI. And Professor Brynjolfsson and his team have done a paper about how one of the issues contributing to why we don’t see the productivity growth that we imagine, or that we expect to see, is because we have a hard time measuring productivity that’s contained in intangible assets, like human capital upgrading that’s happening right now in the AI space. So the hope is that we’ll see those productivity increases come in the next couple decades or so.
Michael Chui: You talked about also productivity being an area for competition, if you think about the deployment of these technologies. I want to pull on that thread a little bit. You’ve also talked about “unsexy AI.” There are lots of times when people think about AI and they think about science fiction, and androids, or Westworld, or whatever. What do you mean when you talk about unsexy AI?
Jeffrey Ding: I first started talking about unsexy AI when I was doing a translation on intelligent manufacturing in China. And I translated an article about Shuzhilian—this is a Chinese company I don’t think any listeners have ever heard of. But they describe themselves as a data industry chain integrated-services company, and they’re based in Chengdu, China.
And this article from jiqizhineng, which is one of my favorite platforms to follow, talked about the unsexy details about the production line for making knives, and the manufacturing workflow for making knives, and the potential for computer vision. And what’s called machine quality inspection could improve the efficiency of these manufacturing workflows for making ordinary things like knives.
And if you’re able to do better machine visual inspection, you can significantly speed up the production process. You can also use machine learning models to identify the type, location, and size of defects so as to make the whole manufacturing process more efficient.
And the context of the piece in terms of the bigger picture (relating back to your question about productivity) is that for Chinese leading companies, the defect rate in their production lines is about 1 percent, while if you compare that with the defect rate for similar products in Germany, South Korea, Switzerland, it can be as low as 0.2 percent or 0.3 percent.
We’ve talked about moving up the value-added chain in terms of manufacturing as a way to escape what’s often called the “middle-income paradox” for China. That’s a very significant driver behind their ambitions in AI. And it’s stuff like this, the unsexy AI of making knives better. It’s never going to make the front page of The Wall Street Journal, but I think it’s just as important as the more consumer-facing, obvious AI applications like facial recognition.
Michael Chui: That’s something that we’ve also observed in our research. There is this cutting edge of developing the technology and doing the R&D, but where you actually get value in the economy is in the deployment and adoption of these technologies, which—as you said—takes a long time, as it turns out, when you actually have to go ahead and do it.
If you don’t mind, why don’t you take me back in history a little bit from the Chinese standpoint? My understanding is the “AlphaGo moment” was a big deal. First of all, can you explain what the AlphaGo moment is and then what its impact was on China and why?
Jeffrey Ding: The AlphaGo moment, or some people call it “China’s Sputnik moment” in AI—basically a huge wake-up call for how important this technology is and how much the field had advanced—was when DeepMind’s AlphaGo, which was their Go-playing machine, their Go-playing AI, beat Lee Sedol, who was the number-one player in Go at the time, I believe March 2016.
Michael Chui: And Go is what?
Jeffrey Ding: Go is a strategy game similar to chess but much more complex in terms of the move combinations. AI had solved chess already, but Go was seen as a much more monumental challenge. The funny thing, or maybe actually the unlucky thing for people who are concerned about strategic competition in AI, is that Go has particular significance in Chinese culture as a strategy game that generals would often play.
Michael Chui: This would be like if the US Joint Chiefs all had a chess club and then Big Blue beats somebody in chess and they’re like, “Uh-oh. We have got to do something about this.” Is that totally unfair?
Jeffrey Ding: I don’t think it’s unfair at all. I think it’s one substream of Chinese reactions to AlphaGo. I think there was probably another big stream of people who had followed developments in AI and were already making investments in this space.
I think people forget that two of China’s top facial recognition companies, Megvii and YITU, were founded in 2011 and 2012, respectively. That’s four to five years before the China’s Sputnik moment. So it’s not like no one was thinking about AI. But AlphaGo definitely raised the public consciousness and raised the profile of AI.
Michael Chui: One of the myths that you have attempted to inform people about is whether or not it’s all the Central Government which is causing all of these things to happen. What have you observed?
Jeffrey Ding: One thing I would emphasize is the important role of local and provincial governments, where I think research has shown that they spend more than 50 percent of the science and technology spending that happens from the public sector in China.
We hear all about these big Central Government funds like the big semiconductor funds. But a lot of the real work of industrial policy of development planning is happening at the local government level. Two examples I come back to often, and I’ve written about them for a Nesta collection, are Hefei and Hangzhou. Not any of the first-tier cities that you’d be most familiar with, like Beijing, Shanghai, Shenzhen, Guangzhou, but these are two cities in local governments that have adapted their policy to optimize the strengths in their particular area.
For Hefei, it’s not as attractive of a location as Hangzhou, which is located on the coast, but Hefei is more inland. But they have specialized in speech recognition, so they’re known as China’s Speech Valley. And with the help of two anchor tenants—iFLYTEK, which is a natural language processing tech giant in China, and USTC, University of Science and Technology of China, which is based in Hefei—with those two as sort of the partner, they’ve developed a cluster of companies focused on intelligent speech and natural language processing.
And Hangzhou has done a similar thing, but they’ve been able to set up a more comprehensive AI ecosystem, where they’ve set up an AI town, and they’ve partnered with Alibaba, which is headquartered in Hangzhou. And they’re also benefiting from Zhejiang University, also another elite university in China, being based in Hangzhou. And using those two as the two pillars, they’ve set up stuff like cloud subsidies, office tax credits for startups, and AI companies to build and develop in Hangzhou.
Michael Chui: There’s a saying that you mentioned about products, platforms, and standards. Can you tell us what that saying is and what it means?
Jeffrey Ding: It’s a well-known saying in Chinese industry circles: “Third-tier companies make products. Second-tier companies make platforms. First-tier companies make standards.” So this idea that it’s really valuable to make something like a really cool word processing software, and it would be really great if that processing software became this platform on which other people could build stuff, or that you could continuously update, and a lot of people could join and use the platform.
But what’s really, really essential is if that platform becomes the standard for—like Microsoft and their Word formatting standard. That comes with really big stakes in the sense that if it’s adopted internationally, then when a government is making decisions about what type of software to procure for big purchases, they might have to follow that standard that’s been set because it’s been recognized as the most efficient, the most secure. So that is the context behind that saying.
Michael Chui: And how are Chinese companies going about trying to become the standard setters?
Jeffrey Ding: I’ve written about this in the context of China’s attempt to have more discourse power, or more of a right to speak in international standard-setting forums, especially with regards to strategic technologies like AI. And I think there’s an underlying motivation behind this in the sense that some Chinese policy makers think that China was excluded from setting any of the rules for the internet, and they don’t want that to happen with AI technology.
China has a different approach to standard setting than the US where in the US, it’s much more industry-led from industry-level alliances, with some level of support from [the] Department of Commerce’s NIST [National Institute of Standards and Technology].
But in China, it’s much more government-led and top-down-driven, where you have the [Standardization Administration of China] that coordinates a lot of the standard setting. And so, for example, I’ve translated white papers on AI standardization where there’s a Chinese government body that is convening this, and they’re bringing in a group of university and company stakeholders to write out their plan to increase influence in international standard setting.
Michael Chui: You’ve been at this awhile, but things have been changing. You already made an observation about something when you wrote Deciphering China’s AI dream, that your position has changed. Any other thoughts? What have you learned in the time that you’ve been—how many subscribers do you have to your newsletter now?
Jeffrey Ding: A lot has changed. We started from just an email sent to ten or so friends, and now we’re up to about 8,500 readers. And then I introduced last year an option for people to pay in, and you wouldn’t get any exclusive content, but it’s like a donation or a tip, like you would make to The Guardian or Wikipedia, just to keep the content going. And that’s helped me. At least probably one of the biggest changes with the newsletter is trying to let more people have ownership over it, so letting other people contribute translations, contribute their own analysis. So hopefully it’s a more sustainable model now.
Michael Chui: That’s what’s changed about the newsletter. What about the world of AI in China, and the US, and the rest of the world? How do you see things developing now?
Jeffrey Ding: The coverage of China’s AI development has gotten a lot better in recent years. So, for example, Protocol. The launch of Protocol, other news sources like Quartz and Reuters. You’ve seen a lot more people with Chinese-language skills mining and reading the best Chinese-language coverage. I think it’s gotten much more nuanced.
I think we also have just more diversity of stories about what’s happening in China’s AI development. The Center for Security and Emerging Technology (CSET) did this analysis of different rhetorical frames in news coverage about AI competition, and they found that the AI competition narrative has actually decreased in terms of the proportion of all the articles that it shows up in.
I do think we’re getting more nuanced, more comprehensive coverage of what’s happening in China’s AI development. I think for me what’s changed is also just the type of things I’m interested in. So I started out covering a lot about great power competition. And obviously that will continue to be a huge theme for US-China developments in AI.
But now the things that really interest me—one of the favorite translations that I’ve done recently is about delivery drivers in China and how they’re reacting to the pressures put on them by the algorithms of these big companies like Meituan and Ele.me.
Just trying to figure out, rather than thinking about AI as this thing in a box, this thing in a vacuum, trying to think about these human-machine interaction systems that involve AI. And I think that type of analysis and that type of thinking will only mature in the future.
Michael Chui: Which is interesting because these questions about ethics, and purpose, and the use of these technologies, those are things which are true outside of China, inside of China. They are, in some ways, human concerns as opposed to necessarily only national concerns.
I also think it’s interesting, this observation that the China coverage is getting better. I still think there’s a massive asymmetry. I think Chinese researchers read a lot more of the Western stuff, particularly in English, than vice versa.
We at MGI have even noted that. But when we write something in English, before we are able to do an official translation in Chinese, it shows up on Weibo (the “Chinese Twitter,” in quotes) way before we’re able to even do the official translation. And so it does feel like there’s an asymmetry. Does that feel true to you?
Jeffrey Ding: That’s the whole bet behind the newsletter. Once that asymmetry is gone, there’s no use for me. I’m very cognizant that there is that asymmetry. I do think the gap is closing a little bit in some of the places that I mentioned, like Protocol, and CSET is investing in a lot of translation work and doing a lot of translation work.
If you look at any of their reports on China’s AI development, you’ll see 30-page appendices of translated excerpts from Chinese-language sources. Places like New America DigiChina are doing this work, but obviously not at the scale of the pipeline the other way in terms of English-to-Chinese translation.
Michael Chui: Wow. All right. I want to respect your time. I know you have a dissertation to write. Jeffrey Ding, thank you for joining us on this podcast.
Jeffrey Ding: Thanks, Michael. Thanks for having me.