📰 Key Takeaways

Y Combinator CEO Garry Tan has publicly said he doesn’t want regulators cracking down on Chinese AI labs for “distilling” frontier models, and he’s even arguing the US should build its own distillation pipeline. Distillation means learning how another model works and reasons by querying it extensively — it’s a common and legal practice used to train new models across the industry. In a CNBC interview, Tan said “I wouldn’t do anything,” and he explained further to TechCrunch that he wants smaller US open-weight AI labs to be able to use the same training techniques on America’s frontier labs, so the US ends up with richer, non-Chinese open-weight model options. The comments come as Anthropic released its second report this week accusing Chinese labs of running “illicit distillation attacks” — hiding their identities and using fraudulent or stolen credentials to distill models without authorization. Anthropic CEO Dario Amodei has previously called on US regulators to crack down on distillation as well. Tan clarified he’s not arguing that US labs should use stolen credentials — rather, they should get the information “through the front door,” fair and square. His argument has two layers: first, he thinks it’s already overreach for AI labs to dictate how customers use information obtained through their APIs; second, he points out that the labs behind those closed models trained them by scraping massive amounts of human knowledge and copyrighted content without permission in the first place. He argues governments should push the idea that “intelligence trained on public data is inherently more of a public good than something locked up by terms of service.” Tan wants to see a balance maintained between open-weight labs and frontier labs, and he’s blunt that the real AI “doomsday scenario” is all frontier AI capability eventually concentrating in a single closed, monopoly company.

💬 JudyAI Lab Commentary

We tend to assume distillation is something that only happens to other people, but Y Combinator CEO Garry Tan pushing back on that narrative brings the conversation back to a more fundamental question: who does the rule actually benefit?

Distillation is, at its core, learning how another model reasons by querying it a lot — it’s been standard practice across the industry all along. Tan’s point is that if the US uses regulation to block Chinese labs from distilling frontier models, while also not letting America’s own open-weight labs do the same thing to America’s frontier models, that’s a double standard that doesn’t hold up. He also flags a contradiction that often gets glossed over: those closed models were themselves trained on massive amounts of unauthorized public data and copyrighted content. For AI builders, this is a good reminder that any “safety” or “compliance” argument is often, at the same time, a way of protecting someone’s market position — it’s worth asking who this rule actually locks in an advantage for.

Next time you see a debate over whether model capabilities should be restricted from access, ask first: if this rule stands, who does it make harder to catch up to?

📅 Source Info

🔗 Further Reading