Great post - I do think the race framing is reductive in so many ways, and is in some sense far more charitable to the US than the reality.
The one aspect of the thesis I don't quite buy is the claim that China has an advantage in diffusion. It seems to me that a purely capitalist economy is ideal for diffusion - finding ways to squeeze the maximum efficiency out of new technologies, without regard for possible negative externalities.
China's AI+ plan mandates 90% "AI agent adoption," but mandating something from the top is different than actually driving real change. I'm skeptical that a centrally planned approach can be more efficient than the American approach, which is effectively a full court press from VCs to stuff AI anywhere and everywhere, and see what sticks. From a skim of the Jeffrey Ding article that you cite (https://jeffreyjding.github.io/documents/Diffusion%20Deficit%20working%20paper%20August%202022.pdf), he seems to agree.
Points like the imbalance in planned nuclear projects seem to stem less from a unique diffusion special sauce, and more from the simple fact that China has far more manufacturing capability than the US. Given the dominance of American software across the globe, my prior is that the US actually has a significant edge here.
In the context of AI, 'diffusion' perhaps lumps together different things that should be teased apart:
- AI chatbot users
- AI integration into existing firms/workflows
- Labor automation
- AI-driven innovation
- Manufacture of AI-driven innovations
It seems to me that AI's transformative effects (& resulting national gains) will largely come from the last 3 things, but the current U.S. software advantage relates more to the first 2.
On labor automation - America's capitalist economy is constrained by its democratic political system, and I argue this constraint seems worse than the limitations of China's central planning (which is fairly decentralized anyways -- the plan basically creates a broad incentive for provincial & municipal officials to devise & implement relevant subplans, often including cooperation with local private firms & academia). On AI-driven innovation, a capitalist economy disregards positive as well as negative externalities, which often leads to underinvestment in innovation (as reflected by the R&D stats).
But these two points map on to the sections on political response & scientific infrastructure, so maybe it doesn't make sense to implicitly include them in the diffusion section.
On the manufacture of AI-driven innovations, I agree that China doesn't have a diffusion special sauce so much as greater manufacturing capacity (so maybe talking about 'diffusion' at all is misleading here). I actually wanted to write more about the importance of simple quantity (larger population, more resources, etc.) for national power, but I found it hard to fit into this post. But basically, if major gains come from physical deployment, then the capacity to make & install lots of stuff (even just by virtue of scale) is how you actually go from advanced models to economic growth.
On the Jeffrey Ding article - I should've put this point in the footnotes, but I thought his method of evaluating diffusion capacity (looking at Global Innovation Index indicators) seemed fairly flawed. In particular, the index seems to put too much weight on irrelevant factors -- e.g. "National feature films/mn pop." and "Graduates in science and engineering, %" are weighted equally -- but I agree I should look more into this & formalize this disagreement.
Great post - I do think the race framing is reductive in so many ways, and is in some sense far more charitable to the US than the reality.
The one aspect of the thesis I don't quite buy is the claim that China has an advantage in diffusion. It seems to me that a purely capitalist economy is ideal for diffusion - finding ways to squeeze the maximum efficiency out of new technologies, without regard for possible negative externalities.
China's AI+ plan mandates 90% "AI agent adoption," but mandating something from the top is different than actually driving real change. I'm skeptical that a centrally planned approach can be more efficient than the American approach, which is effectively a full court press from VCs to stuff AI anywhere and everywhere, and see what sticks. From a skim of the Jeffrey Ding article that you cite (https://jeffreyjding.github.io/documents/Diffusion%20Deficit%20working%20paper%20August%202022.pdf), he seems to agree.
Points like the imbalance in planned nuclear projects seem to stem less from a unique diffusion special sauce, and more from the simple fact that China has far more manufacturing capability than the US. Given the dominance of American software across the globe, my prior is that the US actually has a significant edge here.
Thanks! These are good points; some comments --
In the context of AI, 'diffusion' perhaps lumps together different things that should be teased apart:
- AI chatbot users
- AI integration into existing firms/workflows
- Labor automation
- AI-driven innovation
- Manufacture of AI-driven innovations
It seems to me that AI's transformative effects (& resulting national gains) will largely come from the last 3 things, but the current U.S. software advantage relates more to the first 2.
On labor automation - America's capitalist economy is constrained by its democratic political system, and I argue this constraint seems worse than the limitations of China's central planning (which is fairly decentralized anyways -- the plan basically creates a broad incentive for provincial & municipal officials to devise & implement relevant subplans, often including cooperation with local private firms & academia). On AI-driven innovation, a capitalist economy disregards positive as well as negative externalities, which often leads to underinvestment in innovation (as reflected by the R&D stats).
But these two points map on to the sections on political response & scientific infrastructure, so maybe it doesn't make sense to implicitly include them in the diffusion section.
On the manufacture of AI-driven innovations, I agree that China doesn't have a diffusion special sauce so much as greater manufacturing capacity (so maybe talking about 'diffusion' at all is misleading here). I actually wanted to write more about the importance of simple quantity (larger population, more resources, etc.) for national power, but I found it hard to fit into this post. But basically, if major gains come from physical deployment, then the capacity to make & install lots of stuff (even just by virtue of scale) is how you actually go from advanced models to economic growth.
On the Jeffrey Ding article - I should've put this point in the footnotes, but I thought his method of evaluating diffusion capacity (looking at Global Innovation Index indicators) seemed fairly flawed. In particular, the index seems to put too much weight on irrelevant factors -- e.g. "National feature films/mn pop." and "Graduates in science and engineering, %" are weighted equally -- but I agree I should look more into this & formalize this disagreement.