Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
Inside China’s AI Strategy: Innovation, Diffusion, and US Relations (with Jeffrey Ding)
0 Plays1 second ago
On this episode, Jeffrey Ding joins me to discuss diffusion of AI versus AI innovation, how US-China dynamics shape AI’s global trajectory, and whether there is an AI arms race between the two powers. We explore Chinese attitudes toward AI safety, the level of concentration of AI development, and lessons from historical technology diffusion. Jeffrey also shares insights from translating Chinese AI writings and the potential of automating translations to bridge knowledge gaps.
You can learn more about Jeffrey’s work at: https://jeffreyjding.github.io
Timestamps:
00:00:00 Preview and introduction
00:01:36 A US-China AI arms race?
00:10:58 Attitudes to AI safety in China
00:17:53 Diffusion of AI
00:25:13 Innovation without diffusion
00:34:29 AI development concentration
00:41:40 Learning from the history of technology
00:47:48 Translating Chinese AI writings
00:55:36 Automating translation of AI writings
Recommended
Transcript
The Impact of Technology Diffusion
00:00:00
Speaker
The simple definition of diffusion is the spread of a technology across a population of users. Why economists and economic historians pay so much attention to these engines of growth, these general purpose technologies, is because their population of potential users is the entire economy. The usage rate and the adoption rate of information and communications technologies like cloud computing, linkages, how well do different parts of your science and technology ecosystem speak with each other and collaborate with each other and and share ideas.
00:00:34
Speaker
I found that China ranked much higher on those innovation capacity indicators than on um those diffusion capacity indicators. We always think about the AI applications that are going to cause conflict.
00:00:48
Speaker
One of the first areas where we had strong AI development was in translation and translation hopefully can improve cooperation ah and lead to peace.
Introducing the Podcast and Guest
00:00:58
Speaker
Welcome to the Future of Life Institute podcast. My name is Gus Docker, and I'm here with Jeffrey Bing.
00:01:04
Speaker
Jeffrey, welcome to the podcast. Thanks for having me, Gus. Do you want to introduce yourself to our audience? Yeah, I'm an assistant professor of political science at George Washington University, where I research U.S.-China competition and cooperation in an emerging technologies such as AI.
Understanding the AI Arms Race
00:01:23
Speaker
And I started in this field back in 2017 when I joined the Center for the Governance of AI, which at that time was housed under the Future of Humanity Institute at the University of Oxford.
00:01:37
Speaker
When we talk about China, there's a lot of talk about an arms race in AI. How accurate would you say that narrative is and and what are people missing? I think the main thing people are missing is this assumption that AI weapons will be discrete, countable, bomb-like technologies.
00:02:03
Speaker
because the the the reference of an AI arms race is a callback to Cold War races involving nuclear weapons and intercontinental ballistic missiles.
00:02:16
Speaker
And so think that's the main misconception is this idea that there are going to be these discrete AI weapons that countries will need to build up more and more of. And whoever builds more will have ah some sort of military superiority.
00:02:32
Speaker
It's not to say that there aren't useful aspects of that analogy. I think there are competitive dynamics in artificial intelligence where countries want to be farther ahead ah in terms of their AI capabilities.
00:02:46
Speaker
I just think those AI capabilities will be more diffused and embedded through all these other types of military systems or economic applications.
00:02:57
Speaker
And they're there won't be this type of, there's a missile gap between the US and the Soviet Union, and the Soviet Union has X amount more missiles than us. That type of dynamic, I think, is is where the analogy gets little bit misplaced.
U.S. and China: Strategic AI Development
00:03:11
Speaker
Yeah. One worry here is that one country say the US or China will be somewhat ahead, but because AI development is so fast paced, we might end up in a place where one country has access to a system that's much more powerful than the other country.
00:03:30
Speaker
do you think Do you think this is a realistic worry? And do you think the governments of China and the US are concerned about this? I think to some extent, those types of scenarios have found more of an audience in the US government than the Chinese government.
00:03:44
Speaker
This type of fast takeoff scenario that's really based on assumptions about seamless recursive style improvements in AI.
00:03:57
Speaker
think those make sense in Google Docs and in speculation. But when you drill down it and look at what empirical evidence do we have of this, I think it's fairly unlikely that just because one country has a small lead in AI development, that means that they will have a fast improving, almost exponential increase in their in their capabilities so much that the other country can't catch up.
00:04:25
Speaker
I think well I think In the U.S. s government, to to but to clarify more my thinking on this scenario, I think among U.S. policymaking circles of people who whose takes on tech policy are very informed, I think the scenario that they often bring up is not necessarily this recursive self-improvement, but more some sort of reaching some sort of threshold in terms of cyber capabilities, where then that that becomes the decisive strategic advantage.
00:04:57
Speaker
And the country that's able to automate all these cyber operations is able to completely disable an adversary. To me, that also betrays just a fundamental lack of understanding of how cyber operations work in terms of the most the most important and critical cyber operations require a lot of on-the-ground human intelligence.
00:05:15
Speaker
And why is that? If you look at Stuxnet, the reason why you need so much on-the-ground intelligence is most of the important cyber um systems are air-gapped.
00:05:27
Speaker
ah So they're not they're not all just connected through digital networks, ah which means there's a physical layer of separation. So you need to turn someone on the ground ah so that they plug in the USB, for example, into the physical charging port of an air gap system.
00:05:44
Speaker
If Rebecca Slayton for international security calculated the cost of Stuxnet and and the cost of the software development was very, very minimal, at least when you incorporated the on the ground intelligence and analysis costs.
00:05:59
Speaker
And they had to completely rebuild replica of the centrifuge sites. Iranian centrifuge sites. So that's a lot of physical infrastructure development costs too. So these scenarios assume away lot of what actually happens with real world cyber attacks when you when you dig deeper. And I just dug one layer deeper there.
00:06:21
Speaker
You can dig another layer deeper and look at how AI would also enable the defender to have better protective capabilities. Alan Defoe and Ben Garfinkel have this article about how The offense-defense balance, when it scales, when you put more and more resources or when you when you allow both sides to take advantage of AI, the defender is able to cover the attack surface.
00:06:46
Speaker
And actually, offense, defense scaling favors the defender in all of these future AI scenarios. So that's going another layer deeper. And I think if you just start picking away at these assumptions, this idea that AI will be this nuclear bomb-like type of capability starts to to go away.
00:07:07
Speaker
And to me, the better analogy than an arms race, the better analogy is comparisons to electricity. and other general purpose technologies that are still very, very transformative, ah but their impact doesn't come in sort of this one step decisive strategic advantage.
00:07:25
Speaker
One day, Anthropic opens a box and we have AGI capabilities that rule the world. Their impact comes more through ah gradual protracted timeline of diffusion across a wide range of military and economic application sectors.
00:07:42
Speaker
How does AI development differ ah between the U.S. and China? I think one of the main differences, probably the most obvious difference, is the U.S. is in a compute-rich environment and China is in a compute-constrained environment.
00:07:59
Speaker
So if you look at interviews from DeepSeek's CEO, ah he mentions that one of the biggest bottlenecks to their research and development is access to computing power.
00:08:12
Speaker
ah in the form of high-end and video chips. And so we've seen the U.S. develop these export controls to ah constrain and restrict China's access to ah computing power.
Export Controls and AI Development in China
00:08:23
Speaker
On a lot of these other fronts, I see a lot of parallels. In some ways, the U.S. and China are more alike than the U.S. and the European Union, or China and the European Union.
00:08:35
Speaker
ah Both countries have these tech giants. that are increasingly influential in shaping AI standards and regulation. In a lot of ways, these tech giants are similar in the sense that if you go into DeepSeq or Minimax or Zero One Point AI, all these Chinese large language model startups, the culture and the vibe is very similar to a Silicon Valley startup.
00:09:02
Speaker
And so there's a lot of similarities, but I think that the key difference ah in some ways shaped by the October 2022 export controls is access to comp compute.
00:09:13
Speaker
And it's not just shaped by by those controls. It's also ah China doesn't have these homegrown chip design companies like NVIDIA. How does China make up for being relatively compute poor?
00:09:27
Speaker
I think that's that's partly what's driving Chinese companies to go down this other track of training smaller size models in terms of in terms of the number of parameters.
00:09:42
Speaker
Using innovations like DeepSeek did with their mixture of experts approach or a company like Model Best is is also taking a different track, but but also... trying to develop more efficient models or or denser models, packing packing more and more into less and less compute and and smaller and smaller size models, ah but still squeezing out comparable levels of performance.
00:10:07
Speaker
So we've actually seen Chinese companies really embrace open source models as a corollary to this this path of smaller size development. And so even before DeepSeek's V3 and R1 models came out in the fall of 2024, even late summer of 2024, Alibaba's Tony Chan-Wen Q1 open source model was outperforming Meta's Lama model, a lot of these different benchmarks.
00:10:37
Speaker
and And again, the premise there is smaller models that not only cost less compute to train, but also less compute to use for the demand side and the application side.
00:10:52
Speaker
So that's that's one of the the main differences between China's AI ecosystem
Global AI Safety Practices
00:10:57
Speaker
and the US. s In respect to China's and the US's efforts on the safety side, how would you say, or attitudes to safety, how would you say they differ?
00:11:09
Speaker
It's a really important question. I think this is an area where a lot more research needs to be done. i think Scott Singer at the Carnegie Endowment of International Peace did this deep analysis of Chinese companies' commitments to AI safety and compared all those to the commitments that industry players have agreed to at various global AI safety summits, starting with the UK AI safety summit back in November 2023.
00:11:39
Speaker
And he found there there were a lot of parallels. And I've looked at different Chinese company, blue papers disclosing their AI safety and security practice. And lot of times they are looking at what Google, Anthropic, OpenAI are doing and trying to trying to translate that into their own context.
00:12:00
Speaker
One more point on similarities, and I'll get to differences, but I think Even in draft laws, such as the the draft AI law developed by experts at the Chinese Academy of Social Sciences, you'll you'll see planks that are that are very similar to the EU AI Safety Act.
00:12:19
Speaker
Or China's algorithm registry, by by some accounts, this registry that requires generative AI services to do a security assessment, safety and security assessment.
00:12:33
Speaker
and disclose information about their training data and other sort of transparency requirements. That algorithm registry proposal by some accounts was drawing on an initiative that the New York state government implemented.
00:12:49
Speaker
And so there's a fair degree of learning and transfer and convergence in this area. I would say terms of key differences, sometimes when i translate a Chinese term, 家治对戟,
00:13:07
Speaker
joger twitchy as value alignment that might generally have some parallels to what we mean by value alignment in the English speaking Western AI safety community, how to get super powerful AI agents to follow the designs of their human programmers.
00:13:28
Speaker
In the Chinese context, sometimes that term value alignment is used to refer to making sure large language models produce content that is not politically sensitive. in the sense of and in that sense of value alignment.
00:13:42
Speaker
that That's complicated also by the fact that the Chinese term for safety, anquan, means both safety and security. Whereas in the English language, Western context, a lot of people are trying to draw an important fine-drained distinction between safety issues, preventing AI accidents and risk scenarios like that, versus security issues, preventing unauthorized use of powerful AI systems or or misuse of these powerful AI systems.
00:14:12
Speaker
And so there's still some some points of divergence as well. To what extent do you think there's a shared culture between the talent working at the Chinese companies and the talent working at US companies?
00:14:26
Speaker
You've described or you described the some of the Chinese companies as having a Silicon Valley culture. And I'm assuming that the people who work and at the Silicon Valley companies and at the Chinese companies are and very in-demand workers. they're They're kind of internationally oriented and so on. Do you do you think...
00:14:43
Speaker
Do you think everyone knows what everyone else is doing and there's a sense that they're part of of a community? think that's very much true in the direction of Chinese AI researchers are very plugged in to what is happening in Silicon Valley and in the US AI ecosystem.
00:15:03
Speaker
The different WeChat public accounts I follow for the but China AI newsletter, these are versions of blogs or Chinese versions of MIT Tech Review, for example.
00:15:14
Speaker
On a daily basis, every new development in DeepMind, Anthropic, OpenAI, that is on the front page of those WeChats. public account portals.
00:15:25
Speaker
And so yeah think information flows in that direction is very strong. I think increasingly, and what China AI is trying to do, but I think a lot of other organizations like DigiChina or Center for Security Emerging Technologies Translation Pipeline, there are efforts to try to improve that information sharing in the other direction. I think there's a language asymmetry there that most of the top Chinese AI developers are also fluent in English.
00:15:51
Speaker
You don't have that same hold true for the US AI developers. But I think to your broader question about how much circulation of talent and ideas Is there? I think the clear answer is there's a really, it's very robust.
00:16:07
Speaker
Annalise Saxenian's work on brain circulation discussed this phenomenon of Chinese returnees going going back to Hanzhou or Zhongguancun in Beijing to start up the the the Chinese high-tech ecosystem.
00:16:24
Speaker
And so returnees are people who go to the US or the UK to study for a graduate degree and then, or or they work at a tech company and then they return to ah China to start their own ventures.
00:16:38
Speaker
And so even there was some discussion about how DeepSeek, DeepSeek's talent team was mostly homegrown in the sense that they went to Peking University, Tsinghua University, instead of MIT or Carnegie Mellon.
00:16:51
Speaker
And I... I think if you, I haven't done this in-depth research, but I think if you looked at the trajectory of deep seek researchers, I wouldn't be surprised to find that a lot of them still worked in U.S. companies that had bases in China, such as Microsoft Research Asia, which which to what is Microsoft's second most important lab based in Beijing and has been there for 20 plus years.
00:17:16
Speaker
And so I've done some work with Matt Sheehan at Macro Polo, where we track the talent flows of people who did their internships or had PhD fellowships at Microsoft Research Asia. These are people at Chinese undergraduate universities.
00:17:31
Speaker
And then many of them go to the U.S. to study or to work and to contribute to the U.S. ecosystem. But many of them also stay in China start to start their own ventures like DeepSeek.
00:17:44
Speaker
And so there I think there's there's a lot of information and talent flows between these two ecosystems. You've researched how technologies diffuse in society over time.
00:17:58
Speaker
ah You mentioned electricity as an example of this.
AI as a Transformative Technology
00:18:01
Speaker
Perhaps explain that concept of diffusion of technology, and then we can talk about your notion that China may have a a diffusion deficit when it comes to AI.
00:18:11
Speaker
I think the simple definition of diffusion is the spread of a technology across a population of users. What makes a general purpose technology like electricity or potentially AI distinctive.
00:18:25
Speaker
Why economists and economic historians pay so much attention to these engines of growth, these general purpose technologies, is because their population of potential users is the entire economy, right? So for all these important developments in quantum or electric vehicles and all these other emerging technologies, they're exciting, but oftentimes it's a single purpose.
00:18:51
Speaker
technology. There's one application sector, transportation. What's different potentially about AI is that there could be countless application sectors. And so it's not just going to transform the transportation sector. It could also transform education sector.
00:19:08
Speaker
could also transform all these different services industries, all these different manufacturing verticals. And so that's why i focus on the diffusion stage for for a general purpose technology.
00:19:20
Speaker
That stage when after GPT is pioneered, that long grinding process by which it spreads across all these different application sectors of the economy.
00:19:31
Speaker
And GPT in this context is a general purpose technology. Yes, not a generative pre-trained transformer. But it's a nice coincidence that they use the same acronym because, yeah, a generative pre-trained transformer is a GPT in that sense, but it could also be a GPT in the sense of a general purpose technology.
00:19:53
Speaker
and And so that's why I focus on the diffusion stage as opposed to the innovation stage, that that phase when ah technology is initially pioneered.
00:20:04
Speaker
Yeah. Do you think these stages can be separated cleanly? is Is the development of technology such that first you innovate and then you diffuse? Or isn't it rather that you have diffusion and innovation coinciding over time and influencing each other? Are these stages cleanly separable?
00:20:23
Speaker
No, it's a great point. Innovation, diffusion, there's overlap, there's mixing. It's hard to disentangle the two phases. So what I tried to do with that diffusion deficit paper you alluded to is take indicators that are more closely aligned to a country's innovation capacity and compare them to indicators that are more closely aligned to a country's diffusion capacity.
00:20:48
Speaker
And so it's hard to separate out. I agree. But indicators such as what are your top three companies' average R&D spend, right? So take your top three companies in terms of R&D expenses. So for China, that might be Huawei, Alibaba.
00:21:07
Speaker
And what is their average R&D budget? And compare them to another country's top three R&D spending companies in terms of their average R&D spending.
00:21:19
Speaker
right So that that might have some connection to diffusion, but that's mostly about the frontier firms and who can pioneer new to the world advances.
00:21:29
Speaker
Another innovation centered indicator that I used in that paper was your top three research universities in terms of global ranking and reputation. that So that might have some connection to a country's ability to diffuse AI at scale throughout the entire economy.
00:21:46
Speaker
But I think it's an indicator that's much more closely aligned to innovation capacity. And so I did the same thing with diffusion capacity, which was finding indicators like the usage rate and the adoption rate of information and communications technologies like cloud computing.
00:22:02
Speaker
or even computerization. And then linkages, how well do different parts of your science and technology ecosystem speak with each other and collaborate with each other and and share ideas.
00:22:14
Speaker
And so that's where this term diffusion deficit came from. I found that China ranked much higher on those innovation capacity indicators than on um those diffusion capacity indicators.
00:22:28
Speaker
So I think sometimes when we analyze China's AI ecosystem, we gravitate towards those innovation capacity indicators that are, to be fair, easier to collect. And we miss out on these challenges that China still faces in terms of diffusing AI advances from those frontier firms like DeepSeek and Alibaba, from those frontier cities like Hangzhou and Beijing, to the small mediums size firms and in a town in an inland province that no English language newspaper has ever covered.
00:23:01
Speaker
What do you think explains this diffusion deficit? I think on some level, it shouldn't be too surprising. Countries that are less wealthy and less developed and generally just diffuse technologies at a slower pace.
00:23:15
Speaker
So in some ways, China's diffusion successes with respect to high-speed rail and consumer-facing apps like mobile payments and food delivery are are the exception rather than what we we would expect of a ah lesser developed economy.
00:23:35
Speaker
I think planned economies and economies where the party state is trying to exercise more and more control over technology development, that is not as conducive to to these fast-acting organic diffusion processes.
00:23:51
Speaker
So in that diffusion deficit paper, I looked at the case of the Soviet Union. where a lot of Western analysts fell into that same trap, where they only looked at the Soviet Union's successes and innovation capacity indicators, and they missed out that this planned centralized system wasn't able to activate those those diffusion processes that are that are more healthy and robust when countries take a decentralized approach to science and technology development.
00:24:22
Speaker
Now, I think China is much better on this front than the Soviet Union. but But still, my opinion, there's there's too much centralized control over technology development and AI development. And and in fact, we see trends that ah China is going, is sort of trying to hang on to the vestiges of this planned economy.
00:24:40
Speaker
The other point I'll mention is... The focus of my book, Technology and the Rise of Great Powers, is looking at institutions for skill formation, which country can train a broader base of engineers, engineering talent in AI to implement AI technologies at scale.
00:25:00
Speaker
And on that count, based off of my metrics, the U.S. still has a very substantial advantage in terms of the universities that can, ah in in terms of a broad base of universities that can train AI engineering talent.
00:25:14
Speaker
What does it look like to have a country that is highly capable of innovation, but not as good as diffusion? how does How does that country develop over time?
00:25:25
Speaker
and Because I'm imagining that for sustained economic growth and improved living standards, you would need to have both. You would need to have diffusion of technology so that you have economic growth, basically, that can then be used to fund new ah research and development over time.
00:25:41
Speaker
So what yeah can Can you sketch us a ah picture of what innovation without as much diffusion would look like? Yeah, I think in some ways it's a top-heavy model where you have incredibly developed regions in the coastal provinces, in the first-tier cities, and potentially an ever-widening digital divide between those regions and the less developed regions.
00:26:11
Speaker
And your point about broad-based growth is important. China's trying to bridge that divide. They've recognized that part of the promise of AI and all these emerging technologies, the potential to improve productivity as the driver of long-term economic growth.
00:26:29
Speaker
And so I think... that's the thing to look out for in the future. to me, the most one, if you could pick one variable, if you only could pick one variable to that, that was essential to, to long-term power and a country's future trajectory, I would pick productivity.
00:26:51
Speaker
And so trying to bridge this ever widening gap, between country the the provinces and the areas that are pioneering all these air advances and the regions and the companies that are struggling to keep up ah to even implement and absorb these advances, I think that's going to be essential for China going forward.
00:27:14
Speaker
What's most important for state capacity and geopolitical power? Is it is it innovation or is it diffusion? and Building on that point about why i would i would pick productivity as one of the most important components of power, the the reason why I emphasize that is historically, ah rise and in the rise and fall of great powers, countries become the economic leader first, and productivity growth allows them to sustain economic growth differentials over the long run.
00:27:51
Speaker
And then they translate that economic leadership into geopolitical and military influence. We saw that with Britain's rise in the first industrial revolution. We saw that with the U.S.'s rise in the second industrial revolution, where the U.S. becomes the preeminent economic power before World War I. And you see the U.S. bring that power to bear as the arsenal for democracy.
00:28:13
Speaker
in both world wars. And ah so that's why I emphasize productivity growth. Michael Beckley at Tufts, he's done a quantitative exercise where he finds that if you only look at total GDP, total economic output, that leads to overestimating the power of countries that are really populous, but are not as productive and not as economically efficient.
00:28:39
Speaker
and And that economic efficiency is really important in terms of measuring the true power of nations. And so that's why i favor that diffusion stage over that over that innovation stage, because it's not enough to pioneer these new breakthroughs and have the frontier firms and universities for that to translate to productivity and innovation.
00:29:04
Speaker
improved broad-based economic efficiency, you need the diffusion stage for these technologies to to spread to these small and medium-sized businesses, to the less productive sectors of the economy.
00:29:19
Speaker
There's a vision out there of some frontier AI company in the US developing a system that's that's capable of AI research and development that then improves itself recursively until you have a very powerful system such that that system is perhaps like a country of geniuses in a data center.
00:29:41
Speaker
that can then create new technology and affect the world to a very high degree. this This seems to me to be in contrast with your the the picture you're painting of the of kind of gradual diffusion of technology over time.
00:29:57
Speaker
Maybe talk a bit about that difference and and how lack of diffusion might be, in a sense, a lack of power to affect the world. It's a great connection you're drawing.
00:30:08
Speaker
think before I go into this, I do want to acknowledge I am not a technical expert. In AI, my approach to answering this question is to say, let's look at the three past industrial revolution revolutions and see what we can learn from to apply to what will happen with AI in the future.
Economic Leadership and Technology Diffusion
00:30:30
Speaker
I call that the outside view. in in kanema in in Daniel Kahneman's terminology. It's also completely valid to take the inside view where if you are Dario Amadei or you you are Jack Clark at Anthropic, you are much more plugged in on what is happening day to day with AI development.
00:30:50
Speaker
And so you use your inside view to predict and speculate. And because this this is all speculation, you use your inside view to predict and speculate what will happen with this data center of geniuses.
00:31:05
Speaker
For me though, when I hear when i see that idea, it is very much falling into this trap of this mis-evaluation or this misunderstanding of what happened in past industrial revolutions.
00:31:19
Speaker
When I went to study the industrial revolution case, the layman's view of what happened there was Britain had this monopoly on heroic inventors, heroic geniuses.
00:31:30
Speaker
They had the James Watts of the world, and that's what led to Britain's economy outpacing the Netherlands and France and all these other competitors. Same thing with the second industrial revolution case.
00:31:43
Speaker
The surface level understanding was Germany came out ahead because they dominated all the Nobel Prizes in chemistry. They had all the geniuses in the most scientifically advanced field.
00:31:55
Speaker
And that's why they outpaced Britain before World War I. in When I went back and looked at what did historians using econometric methods and sort of the most intensive analysis, what did what did they find actually happen in all these cases? If you go beyond the headlines, it was actually Britain didn't have this advantage in geniuses.
00:32:17
Speaker
um You had all these macro inventions coming out of France as well. What Britain had an advantage in was this average technical literacy and applied mechanics skills. of People taking night classes at mechanics institutes and learning this general mechanics knowledge to apply iron-based machinery across all these different sectors.
00:32:39
Speaker
And it was a very slow and gradual process. Same thing with the second industrial revolution. The U.S. was actually sending all of its geniuses to Germany to study for PhDs in chemistry.
00:32:50
Speaker
The U.S.'s advantage was building this broad-based body of mechanical engineering knowledge to apply this system of interchangeable parts manufacturing across all these different application sectors.
00:33:03
Speaker
And so I think turning the page to ai maybe one day we will have these data centers of geniuses. I think even if we do have these superpowered geniuses from this recursive automated R&D system, again, that that stops at the stage where you have these new ideas and new developments.
00:33:25
Speaker
ignores the messiness and the the difficulty of then translating that into useful applications across all these different sectors. And I think that process will be a lot more jagged, a lot more troublesome than people in these AI companies believe or or want you to believe.
00:33:46
Speaker
So I think that's that's the gap, this gap. We stop at the genius stage. My story, that's where my story starts. Okay. We've got these new ideas and it's going to take time. And so other countries are also going to have their geniuses.
00:34:01
Speaker
We have this long history of parallel innovation too. Like the electric dynamo. You had... that the the first industrial generator came about and was pioneered about the same time in all these different countries.
00:34:13
Speaker
And so that's where my story starts. Now it's how do you translate that into commercial value across all these sectors where they might not be ready to adopt all these ideas from from this data center of geniuses. So I think that's where we differ.
00:34:29
Speaker
In the world as it looks today, is it realistic that advanced AI could be developed in in one country only or say in one city only, perhaps in in in one company only?
00:34:41
Speaker
Or is your notion of kind of gradual diffusion of technology also notion of widespread technology development? I think to me, it's going to be more dispersed.
00:34:53
Speaker
And that's in part because i think I have a different vision of what AGI looks like. I think the one city, one company, one country approach is one day, Anthropic or DeepMind, open up a box and we get super intelligence.
00:35:11
Speaker
but We get a super intelligent agent. For me, actually, my view was very much is very much more informed by Eric Drexler's comprehensive AI services view. I think that might have been one of the most influential pieces on my thinking during my time at the Center for the Governance of AI at Oxford.
00:35:28
Speaker
And that piece is saying likely the pathway forward will be super intelligent AI services targeted at very narrow, discrete tasks, because that's just what companies will optimize for.
00:35:42
Speaker
Users who who will but want that over continued development of just all powerful tasks. agents that can do everything. And so I think to me that like app store model where you have all these different super intelligent services, that model then lends itself more to ah to to a more dispersed development pathway where maybe one country or one company or one city will develop innovations for this particular service, for this particular app.
00:36:11
Speaker
And another country, another city, another another company will have an advantage in developing services for this other app. Yeah, although I think what we are seeing now, and it's it's early, of course, is that you have companies releasing these these general AI, so general tools where you can you can ask a chat GPT to generate an image or to write up some notes or to plan a holiday. holiday You can ask it to do ah a variety of things.
00:36:42
Speaker
and Does that at all conflict with the notion of kind of narrow services provided by by different providers? Yeah. it's a good It's a good counterpoint. I think in some ways, that is a service of an office assistant or a chatbot with more capabilities than Siri.
00:36:58
Speaker
I think you'll still want an AI service that's targeted at improving machine quality inspection on the production line for for this particular tool. You will still want a particular AI service for autonomous driving.
00:37:18
Speaker
you probably wouldn't want to use Gemini for that. So I agree with you that that we're starting to see some agents that have capabilities that might go across all these different services. But but to me, that that App Store model still holds when you go go beyond what these chatbots are doing.
00:37:38
Speaker
Although if the chatbots evolve over time into systems that are, say, also writing ah scientific papers and perhaps planning out drug development and are becoming more and more general, that that I guess would would change the picture? Or is is Drexler's view of AI as this services view, does that still hold?
00:37:59
Speaker
Maybe. that So I think a good way to think of that when you when you look at writing papers, the best services, at least to my knowledge, are not the ones that just use the foundational model, but ones that are third party integrations and deployments of that model.
00:38:16
Speaker
ah So if I were thinking, what advanced search engine technology am I going to use? should I use Gemini? Should I use OpenAI's technology? No, I'm going to go to Perplexity's offering that integrates services from all these different companies, including DeepSeek's R1 model.
00:38:34
Speaker
So even in these stages, it's early. We you agree, it's all speculative. But to me, it's... If you look at it from the comprehensive AI services model, there's plenty of evidence to support that's that's what it's going to look like going forward.
00:38:49
Speaker
in In the way of third-party integrations, specializing in particular discrete services, like give me the best advanced search engine tool for research. And that's going to be another company using third-party integrating services from foundation models from all around the world, from all these different companies.
00:39:09
Speaker
In your view of AI development, this would mean that so-called wrapper companies would be more relevant. These are companies that take a foundation model and use it or present it or advertise it to a user base in a specific way where you might have the same underlying model, but ah wrapped in a way where it performs well as a psychologist, say personal psychologist or something like that.
00:39:36
Speaker
Yeah, I think that us that is one model, one pathway that that looks a lot more like comprehensive AI services. It could be another model where it's not a wrapper, but the underlying code is just open source AI systems.
00:39:55
Speaker
And then another company specializes and builds their own technology off of a open source base. So that's that's not the wrapper model, but but it's another model that we're seeing become pretty successful and and important.
00:40:08
Speaker
Again, this is starting to get out of my wheelhouse in terms of I mostly study the past and historical analogies analogies and and try to see are there any parallels today.
00:40:19
Speaker
i think my viewpoint aligns pretty closely with this recent paper that came out by Arvind Narayanan and Sayesh Kapoor called AI as a Normal Technology.
00:40:31
Speaker
And I think they do a better job of fleshing out the details on what's happening in AI today. But maybe I'll also say i'm not I'm not saying this other pathway is not possible.
00:40:42
Speaker
i think that's where I think that's where I am skeptical of people who seem completely convinced of their scenario of how we get to AGI and what will happen in the future.
00:40:53
Speaker
That's my most, I'm not, and that that's where I get the most skeptical because no one really knows. It's all speculation. We're really bad at technological forecasting. If there's anything I'm certain about, that's, that's what I'm certain about.
00:41:06
Speaker
You'll see how I'm even hedging in terms of whether AI is a GPT. I always say potentially the next GPT. And in my book, i I clarify why I do that. So I'm open to the idea that there there is this other pathway and that I might be wrong. And so I think we should two we should do things that that that safeguard humanity, that that ensure AI safety and security that that might be true across all different pathways, or we should prepare for worst case scenarios.
00:41:36
Speaker
I just see this this comprehensive AI services pathway as the more likely way forward. and When you're studying history and trying to learn from past historical technological developments, this seems to me like such such a a difficult exercise.
00:41:51
Speaker
the The fundamental assumption there is something like the future will be like the past, but how on earth do you go about this? understanding what has changed since the last end industrialched industrial revolution and therefore what you can but what lessons you can take away from history, what analogies hold over time. that's how do How do you go about this this investigation?
00:42:15
Speaker
A lot of challenges. There's a lot of holes in this approach. What I will say is the last historical case that I study is U.S.-Japan competition over computers.
Productivity and the Computer Revolution
00:42:25
Speaker
And so if you were to take this language of recursive self-improvement, we will get an army of software programmers that will just caused this intelligence explosion.
00:42:37
Speaker
And you plop that down at the height of the computer revolution, it might not be so out of place. And so we, but we had Solos paradox, which was, you saw the computers everywhere, but not in the productivity statistics.
00:42:53
Speaker
because it took time for people to adjust and shift to this computerization's trajectory, shift their organizational practices up skill. And eventually you did see the computers in the productivity statistics, but it took time.
00:43:07
Speaker
So to me, it's not that history should define the future. But I think oftentimes we're biased in the other direction.
00:43:17
Speaker
We think that the moment that we live in is this unprecedented moment that we're living at the hinge of history, that somehow it's always within the livespans of the people writing today, that will be make or break. We wanna believe that we are living in the most significant period of history. And I think we're overly biased towards the future is new.
00:43:42
Speaker
Segways will change everything. And so I think there's there's a lot to learn from history. it's It's important to take into account that things might be different. There's all these things that might be different. There's also things that are different that might only reinforce the general purpose technology diffusion theory that I outlined.
00:44:01
Speaker
So for example, one important difference in this time period versus other time periods, this is based off Andrews et al., which I believe was OECD paper.
00:44:13
Speaker
They looked at the most developed economies, the OECD countries, and they found that one of the biggest changes over time is the initial innovation gap has shrunk, which means the time that the time between and innovation that gets pioneered in one country, and then it gets adopted by a frontier firm in the second fastest country.
00:44:37
Speaker
that that innovation That innovation gap has shrunk. And that makes sense, right? The world is more globalized, information travels more and more. But I think what they the other thing that they found that really interesting was the intensive adoption gap has actually increased over time in our current age, which means the time it takes between the frontier firm in one country pioneering and adopting this new innovation and for the entire country to adopt that innovation and the time it takes for that innovation to travel throughout the economy within within a country.
00:45:14
Speaker
And so, yes, things have changed, but but that that change actually only makes it more important to take into account the lessons of the past, say, take into account this GPT diffusion mechanism. so So it's important to to recognize that that it goes the other way too. there are There are some changes that might weaken my argument, but that there are some changes that also strengthen the argument and just being upfront about what are the factors that have changed.
00:45:37
Speaker
and tracing those through, I think that's the best way to then learn the right lessons from history. I do wonder how the fact that current AI is, in some sense, a non-physical thing, and and don't get me wrong here, it it relies on on data centers for training, and of course, you you need infrastructure to to transmit the thing and
Cultural Understanding Through AI Translation
00:46:00
Speaker
so on.
00:46:00
Speaker
But it's it's if you have a new model, a new, interesting, powerful model, you can you can quickly ah provide that model to a bunch of different consumers and companies.
00:46:12
Speaker
And in the past, say, if you had an a if you had an innovation that improved the physical thing, would need to manufacture that thing, would need to build specialized factories, and so on.
00:46:22
Speaker
Does that change anything about your ah your expectations of the pace of diffusion of AI? Yeah, I do think so. i think I think that's one of the clear changes in that electrification required all of this physical infrastructure to be built terms of transmission lines, central utilities, and generating stations.
00:46:46
Speaker
So Yeah, I think that's one difference in favor of faster diffusion for AI. ah do think AI still relies on the need for a lot of complementary innovations in the hardware space.
00:47:01
Speaker
Some of if you the the Chinese language sources I follow, the the most there's there's a lot of attention on not what's happening in the chips to replace NVIDIA's A100s,
00:47:17
Speaker
but in the chips that can power tablets or phones ah to run a 32 billion parameter model. And so I think we shouldn't discount the hardware needs of AI diffusion, especially the pathway that the AI companies envision.
00:47:37
Speaker
It will be very, very dependent on hardware. But no, I take the point. in that the speed at which code can be shared is a huge difference in this and this current age.
00:47:48
Speaker
Let's actually talk about your translation work because you you run an excellent newsletter translating Chinese sources into English. What are some interesting things you've learned by doing that?
00:48:01
Speaker
i i I think probably... For my readers, the most interesting stuff is the benchmarks, Chinese language benchmarks on the top AI models.
00:48:12
Speaker
The most interesting stuff is like how are Chinese policymakers thinking about these issues at the level of grand strategy? For me, the most interesting stuff is actually the human interest stories that I translate.
00:48:25
Speaker
I think oftentimes when you're caught in this vortex of great power competition, and that's largely, yeah it's an interesting it's a interesting thing to talk about. it's a consequential thing to talk about. We've talked about a lot on this pod. I think when you're caught in that vortex, you miss out on the fact, just the fact that and Chinese people are people too.
00:48:46
Speaker
Yes, they live in a political regime that's very different than us. It's not one I would want to live under. But they're people too. They have shared concerns about AI development, shared worries about AI development. They make memes about AI.
00:49:01
Speaker
I did a translation about this term, artificial challenged intelligence, where Chinese people were... ah You've probably heard about... the big jaywalking billboards in China where they use facial recognition and they put up your face on a big billboard, plaster it all over to deter jaywalking.
00:49:20
Speaker
And one of my favorite translations is ah we looked at all these examples of artificial challenge intelligence and people were making fun of one of these jaywalking billboards that had plastered a face of someone from a bus ad on onto it s jay walker as to shame this jaywalker.
00:49:40
Speaker
just an advertisement person. And so ah I can't spell out in exact detail why these stories I think are valuable to understand the heart and the core of what's happening in China's AI ecosystem.
00:49:55
Speaker
but But I think it's just the thing that that I've gained the most from is just revisiting the central premise that Chinese people engage with AI as humans.
00:50:06
Speaker
And so some other translations in that vein that have been the most surprising to to me, at least before I went into this field, was there's a lot of concerns about privacy and personal information protection.
00:50:18
Speaker
There's a lot of robust discussions about how do I delete my facial data? And you even have um surveys on that question from think tanks associated with the government ah entities and in party entities.
00:50:32
Speaker
And so those are those are the types of translations and findings that that I've learned the most from. Have translating these texts ah changed your view on which research directions are most interesting?
00:50:47
Speaker
Yeah, I think if I were to trace back this whole idea of general purpose technology diffusion, it it comes from all these different things that I'm reading, all these other people who have done work in this space.
00:51:00
Speaker
But one source was the China AI translations, where there is a lot of emphasis on how AI is actually put into practice.
00:51:13
Speaker
how one of the translations sticks out to me was about this Chinese company called Shuzhi Lian that was trying to adopt computer vision technologies on the production line.
00:51:26
Speaker
So instead of people, hundreds of people looking at screens to look at defects, you use a computer vision algorithm to detect the defect early on. And so this idea of diffusion, the diffusion deficit, a lot of it came from specific translations like that one. And then also translations from Alibaba Research Institute, where they looked at overall digitization rates in China.
00:51:52
Speaker
And oftentimes the Chinese language perspective is, ah China still has a long way to go to catch up, right? On machine quality inspections, like our defect rate is 1%. We need to get to the 0.2, 0.3% that South Korea, Switzerland, Germany have gotten to.
00:52:09
Speaker
ah Sorry, excuse me, in Japan instead of South Korea in that list. And the reports on low digitization rates, it's it's Chinese sources are often the ones saying, hey, real talk, we are pretty far behind here. And so I think that influenced a lot of the thinking for the the book, the articles on China's diffusion deficit.
00:52:30
Speaker
Do you see any basis for confusion when you're reading Chinese sources? Do you see anything that might hint at either people in the West being confused about people in China or people in China being confused about people in the West, specifically ah as as this as we're talking about now?
00:52:48
Speaker
Yeah, I think one of the the big misperceptions, and I wrote about this recently, i think I called it, it wasn't a very catch catchy name, but might have been AI policy misperception spiral.
00:53:03
Speaker
And it's this idea that Chinese policymakers think the US's approach to AI development is much more coordinated and systematic. than it is in practice.
00:53:14
Speaker
And by the same token, US policymakers and analysts perceive China's approach to AI development as much more cohesive, integrated whole of society than it is in practice. And so part of this is just the competitive dynamic. Both sides think the other is going faster when it comes to maybe like Manhattan style projects or governments coordinating everything from the top down in AI development. Both sides think the other country is more aggressive on this front than they actually are. ah So in some ways that might speak to fears, anxieties, but but that's that's one misperception I see.
00:53:58
Speaker
how How is that apparent in the original sources that that there's the perception of the US raising a head in ah in a Manhattan style project to AGI?
00:54:09
Speaker
I wouldn't. I think Manhattan style is probably the extreme of that. In in the original in the sources, I think this is a piece that translated by two China Academy of Information Communications Technology researchers And they say the U.S. s has this whole of government approach to emerging technologies. And they cite the Biden administration's list of 20 or so critical emerging technologies.
00:54:33
Speaker
And then they go through all these offices that also use a similar list of critical and emerging technologies. And for them, they They see that as, wow, the U.S. has this really cohesive, integrated approach to emerging technologies.
00:54:50
Speaker
But if you know the U.S. side of the story well, you would know, oh, actually, these lists are pretty haphazard. There's no systematic way of coming up with these technologies. as They're at very different levels of abstraction.
00:55:04
Speaker
The one that the Biden administration came up with is very different from the one that the Department of Energy actually uses, is very, very different from the one that the DOD uses. There's not a cohesive interagency process for coordinating coordinating actions on emerging technologies and AI.
00:55:22
Speaker
And so I just found that funny that that you you you still see this information gap. as you go across the ocean and this cross-border misperception of what the other side is doing on AI and emerging technology policy.
00:55:36
Speaker
Translating from ang English to Chinese and from Chinese to to English, this seems like something we should do much more of just because it would be probably a benefit from the world if both sides had a better on understanding of what the other side is up to.
00:55:51
Speaker
Do you think we can do this with AI? Do you think Do you think we can we can do 10x or a hundred x or 1000x, what you're doing, but but powered by AI?
00:56:03
Speaker
I hope so. I think I've benefited personally in my translation work from neural machine translation. That's... been one of the biggest upgrades for all the translations I do for China AI. I plop it into Google Translate first.
00:56:18
Speaker
So that goes to think, we always think about the AI applications that are going to cause conflict or their implications on conflict. Probably one of the most consequential, one of the first areas where we had strong AI development was in translation and translation hopefully can improve cooperation and lead to peace.
00:56:40
Speaker
And so I do have a lot of hope for that. I think Center for Security Emerging Technology's scout tool is an early prototype of that effort where they do automate some of that translation pipeline.
00:56:54
Speaker
and scoping pipeline. To me, I would say for the newsletter, maybe half of the work is the process of the actual translation and analysis and and taking notes on the translation.
00:57:07
Speaker
But I would say half of the work is finding an interesting translation and and that scoping process, going through different WeChat accounts, looking at what friends are posting what they find are interesting. Other colleagues in China, what are they debating and discussing in various group chats?
00:57:28
Speaker
Getting a sense of which sources now are reliable or produce the best long-form content. That's really the hardest part. and And I found that with... I really welcome a lot of contributors to help out.
00:57:40
Speaker
But the bottleneck for contributors is also finding that right article, finding the interesting article, finding the interesting white paper. Yeah. and And that kind of intuition or tastes seems like perhaps AI could at some point help with that reading through a bunch of sources. But for now, this doesn't seem like what AI is is is is best at.
00:58:02
Speaker
And I'm i'm Assuming here, um I think that that over time you've developed this and intuition or or taste where where you can pick out what is interesting but it's interesting too to Chinese AI researchers, for example.
00:58:15
Speaker
Is there any way you can try to communicate that to me and and to our listeners? What what can what could we look ah look for if we we're interested in and trying to ah develop the intuition you have?
00:58:27
Speaker
Yeah, I love the way you put that. Yeah, in some ways you're developing taste. Right. And that is a question for AI going forward. To what extent can AI develop a useful taste of things?
00:58:40
Speaker
Yeah. In terms of my taste, I like really technical white papers on the status of AI development. And so government-affiliated think tanks will oftentimes put out these more and more so a source that I find really interesting is Chinese consulting firms.
00:58:55
Speaker
For example, like McKinsey and Goldman in English language, they put out some pretty interesting reports. And now we're starting to see Chinese language versions of those. The human interest stories I mentioned, there's different magazines that are focused on that, different WeChat public accounts that are focused on that. One of those I follow closely is this magazine called Renwu, which translates as people.
00:59:19
Speaker
And they do these profiles of individuals and they also do like investigative reports on the challenges delivery drivers face, being trapped in the algorithm and having to meet certain times and go against traffic, otherwise they get point penalties. And then I do like just what the people want in some sense, the benchmarks, how Chinese companies are performing, who are the main players, how many on the diffusion side, how many companies are actually using 100 billion tokens a day in terms of computing tokens.
00:59:51
Speaker
That type of content I'm i'm also pretty interested in. so it's... I would say it's a wide ranging taste. Otherwise, it's not the fun. It's not the most fun just to do the same thing every week. Hopefully, it's like a restaurant menu that that changes every month or so or every season or so.
01:00:07
Speaker
ah so so So that's a sense of of of the types of translations I gravitate towards. When you look at your research, specifically on the diffusion of technologies, what are some open questions that you think should ah get more attention?
01:00:21
Speaker
I think for me, part of the book and part of this research agenda was just establishing when it comes to technological competition among great powers, diffusion is at the heart of the story, whether it's for economic competition or even military competition.
01:00:37
Speaker
I've done some other work with Alan Defoe where we looked at how this general purpose technology diffusion model also applies to military power. ah Going back to that very first conversation that we had, where it's not about this decisive tip of the spear type weapon, yeah super powerful AI weapon, and it's more about AI capabilities being diffused across all these different military applications.
01:01:00
Speaker
I think that's what the research agenda is trying to open up. and In a lot of these projects, I focused on certain types of institutions that that enable diffusion, mostly skill formation institutions, education and training systems.
01:01:17
Speaker
I think there's all these other factors that could affect diffusion. We've talked about some of them, right? Open source, the the speed the the channels that spread software.
01:01:30
Speaker
For me, other cool projects that if I had the time or the patience to continue doing this same type of thing would be to to look back at those historical cases and look at the role of regulation, especially safety regulations, because You had steamboat explosions in the first industrial revolution, a lot of safety issues there.
01:01:52
Speaker
People were really afraid of electricity to the point that President Teddy Roosevelt had to install electricity in the White House as a signal that this is safe to use, that this is a safe technology. And so think there's a lot to be mined on that front.
01:02:07
Speaker
what What is the role of safety regulation and maybe prudent regulation in enabling diffusion? If you take a longer timeline, prudent regulation doesn't necessarily have to be mutually exclusive with advantage in GPT diffusion.
01:02:21
Speaker
So I think there's a lot to learn from history um on that track as well. Jeffrey, it's been a pleasure chatting with you. Thanks for coming on the podcast. Thanks, Gus. Really great to chat with you.