I’m not China’s biggest proponent by any means, but I think their closing of the distance can’t be explained solely by variations of “they’re cheating”. Especially seeing as distillation of closed-source U.S. models is much less of a source of insight than whitebox inspection of open-source Chinese ones. China has a number of legitimate advantages at play:
A clean talent pipeline. There are countless stories of talented American kids with SAT scores in the 1500-1600 range being shunted into state universities and languishing in underemployment. Contemporary issues aside, America used to solve this by funneling a good chunk of the best and brightest into stable jobs at the Pentagon which eventually became high-paying jobs elsewhere. This is regarded as inequitable now, with even things like AP math courses coming under fire.
A population of one billion, with a reliable three digit IQ range[1]. This is a much larger pool of talent to select from.
A sense of shared identity/common cause. Even brilliant engineers don’t do their best work when they think their employer’s interests are, at best, orthogonal to their own. In America, Liberals frown upon the AI industry’s growing integration with the security state, and Conservatives notice that a sizable portion of the industry’s efforts involve preventing the models from endorsing their beliefs. Moreover, engineers hired from abroad might be conflicted about helping to ensure U.S. hegemony.
Support from Americans/non-Chinese, both practically and ideologically motivated. China’s open-source models are the only option if you want to do any kind of whitebox research work. Ergo, everyone not employed by XAI, Anthropic, OpenAI, or Google who wants to work with LLMs is going to do a bit of free R&D for China. Moreover, plenty of people, as touched on above, have ideological opposition to closed-source technology, and might be sympathetic to the Chinese approach[2].
I know heritability of IQ is controversial, but it strongly correlates with job performance everywhere. It remains a major advantage even if IQ is influenced more by nurture than geneticists currently believe.
The validity of these beliefs can be debated, but there’s a significant demographic that holds them, and it’s weighted towards America’s most tech-adept demographics.
I’d still say sort-of. Llama, after the initial leak, was set up to take advantage of the benefits of open source, but the most competitive and best-known open-source models are from China, and at this point they’ve got the brand recognition to match. OpenAI releasing a heavily-nerfed open-source model well behind the frontier didn’t hold headlines for more than a day, for instance, whereas everyone knows about GLM-5.2.
Of course, an American frontier company might completely pivot and open-source something competitive in a huge public display, but I consider it unlikely. Anthropic and OpenAI seem very ideologically committed to closed-source, even as the doomsday predictions around open-sourcing have consistently failed to materialize. XAI might’ve done it if the tech libertarian ethos were in power on the Right, but while it is relatively popular among voters who care about this issue, the administration has generally skewed safetyist on model releases, or at least tried to appease their concerns. Google is fairly subdued in its communications, but I don’t see them breaking from any established traditions.
Meanwhile, the Chinese government has given open source its official blessing.
I’m not China’s biggest proponent by any means, but I think their closing of the distance can’t be explained solely by variations of “they’re cheating”. Especially seeing as distillation of closed-source U.S. models is much less of a source of insight than whitebox inspection of open-source Chinese ones. China has a number of legitimate advantages at play:
A clean talent pipeline. There are countless stories of talented American kids with SAT scores in the 1500-1600 range being shunted into state universities and languishing in underemployment. Contemporary issues aside, America used to solve this by funneling a good chunk of the best and brightest into stable jobs at the Pentagon which eventually became high-paying jobs elsewhere. This is regarded as inequitable now, with even things like AP math courses coming under fire.
A population of one billion, with a reliable three digit IQ range[1]. This is a much larger pool of talent to select from.
A sense of shared identity/common cause. Even brilliant engineers don’t do their best work when they think their employer’s interests are, at best, orthogonal to their own. In America, Liberals frown upon the AI industry’s growing integration with the security state, and Conservatives notice that a sizable portion of the industry’s efforts involve preventing the models from endorsing their beliefs. Moreover, engineers hired from abroad might be conflicted about helping to ensure U.S. hegemony.
Support from Americans/non-Chinese, both practically and ideologically motivated. China’s open-source models are the only option if you want to do any kind of whitebox research work. Ergo, everyone not employed by XAI, Anthropic, OpenAI, or Google who wants to work with LLMs is going to do a bit of free R&D for China. Moreover, plenty of people, as touched on above, have ideological opposition to closed-source technology, and might be sympathetic to the Chinese approach[2].
I know heritability of IQ is controversial, but it strongly correlates with job performance everywhere. It remains a major advantage even if IQ is influenced more by nurture than geneticists currently believe.
The validity of these beliefs can be debated, but there’s a significant demographic that holds them, and it’s weighted towards America’s most tech-adept demographics.
Good list.
Also China adds about a Germany’s worth of electrical power every year, so they’re much better situated to do a massive build-out of infrastructure.
Is this true given that U.S companies such as Thinking Machines and Meta release some open-weight models?
I’d still say sort-of. Llama, after the initial leak, was set up to take advantage of the benefits of open source, but the most competitive and best-known open-source models are from China, and at this point they’ve got the brand recognition to match. OpenAI releasing a heavily-nerfed open-source model well behind the frontier didn’t hold headlines for more than a day, for instance, whereas everyone knows about GLM-5.2.
Of course, an American frontier company might completely pivot and open-source something competitive in a huge public display, but I consider it unlikely. Anthropic and OpenAI seem very ideologically committed to closed-source, even as the doomsday predictions around open-sourcing have consistently failed to materialize. XAI might’ve done it if the tech libertarian ethos were in power on the Right, but while it is relatively popular among voters who care about this issue, the administration has generally skewed safetyist on model releases, or at least tried to appease their concerns. Google is fairly subdued in its communications, but I don’t see them breaking from any established traditions.
Meanwhile, the Chinese government has given open source its official blessing.