How the World Got Wired · Two AI Internets

How Chinese AI Caught Up

There is a question you should ask, after reading the section above. If American chip export controls were supposed to prevent Chinese companies from training competitive models, how did DeepSeek manage to release something this good?

The answer is, partly, that Chinese engineers are good engineers, and that with enough cleverness you can train models more efficiently than the brute-force American approach suggested was necessary. DeepSeek’s researchers genuinely did innovate, in the actual technical sense, on how to train models with less compute. Their papers, which they published openly, have been studied closely by every serious AI lab in the world.

But the answer is also, in part, a particular technical reality that does not get discussed much in the public conversation, and that I want to explain to you carefully.

The reality is called distillation.

If you train a smaller, cheaper artificial intelligence system to imitate the outputs of a larger, more expensive one, the smaller system can become surprisingly capable. You give the larger model a question. It produces an answer. You record both. You repeat this, automatically, millions of times. Then you train the smaller model to give the same kinds of answers the larger model would give. The smaller model has, in effect, learned from the larger model — without ever needing to do the original work of becoming as smart as the larger model.

If you can access a larger model — for example, by paying for its commercial application programming interface, or API — then you can, in principle, distill from it. The American AI companies sell access to their models through APIs. Chinese researchers, like every other researcher, can pay to use those APIs. The companies’ terms of service usually prohibit using their outputs to train competing models. The terms of service are, in some cases, easier to write than to enforce.

A useful analogy for a young reader: imagine that the smartest student in your class will not let you see her textbooks. But she will, kindly, answer any question you ask her. Over the course of a year, you ask her ten thousand questions. You write down all her answers. By the end, even though you never saw her textbooks, you have, in your notebook, a very good approximation of what she knows.

This is, roughly speaking, one of the ways Chinese frontier AI has stayed close to American frontier AI, despite the chip restrictions. The American models are accessible, through APIs, to almost anyone in the world. Once they are accessible, they can be — quietly, partially, with various amounts of dispute about whether anyone has actually done it — learned from.

I tell you this not because I think there is a clear villain in the story. The American companies, after all, chose to make their models commercially available to the world. The Chinese companies have, in many cases, innovated genuinely on top of whatever they may or may not have distilled. The American government, by writing export controls only on hardware and not on access to software, made an implicit choice about which kinds of barriers were possible. There are many actors in this story, each making decisions that look reasonable from their own point of view.

But the result is that the AI race is, today, run on rules that are partly hidden. The American advantage is real but smaller and shrinking faster than most American policy discussions admit.

All 24 chapters of How the World Got Wired · Library · All courses