Garry Tan wants an American distillation regime

Garry Tan, the head of Y Combinator, wants US regulators to let American open-weight labs do the same thing Chinese labs are accused of: distill knowledge out of the top closed models. Distillation here means prompting a strong model heavily to learn how it reasons, then using its outputs to train a smaller model. Speaking to CNBC, Tan said he would "do nothing" to restrict it and that "there should be an American distillation regime." That puts him directly against Anthropic's Dario Amodei, who has asked regulators to crack down after Anthropic reported that labs including Alibaba and Moonshot AI ran large distillation campaigns against Claude, using stolen credentials and fraud to stay hidden.

Tan's case rests on a few points. He argues that telling paying customers what they can do with API calls to a closed model is overreach. He notes that the frontier labs themselves did not ask permission when they trained on huge amounts of copyrighted material without paying the people who made it. And he warns that without strong open-weight competitors, a single company could end up controlling frontier AI, which he calls the real nightmare scenario. You can read TechCrunch's writeup here.

The disagreement is less about the technique, which both sides use, than about who is allowed to point it at whom, and whether US policy should treat American copying of American models as fair game.

Why it matters

If you build on open-weight models, the legality of training on closed-model outputs is about to become a policy fight, not just a terms-of-service question. Whether Tan's "American distillation regime" gets any traction will decide whether US open labs can catch up the cheap way or have to pay for their own frontier runs.