Y Combinator's Garry Tan: Let US open-weight labs distill frontier models too

Y Combinator CEO Garry Tan is breaking with a growing chorus of frontier AI labs calling for a crackdown on distillation — the practice of heavily prompting one model to extract and replicate its reasoning. Rather than have regulators police the technique, Tan argues the US should embrace an “American distillation regime” in which smaller domestic open-weight labs use the same methods on US frontier models, coming in “the front door” rather than through the stolen credentials and identity-masking that Anthropic accuses Chinese labs of using.

His reasoning rests on two claims. First, he sees it as overreach for labs to dictate what customers do with information their models return via API calls. Second, he points to the frontier labs’ own origins: they trained on vast amounts of copyrighted and public material without asking permission, so locking that derived intelligence behind restrictive terms feels inconsistent to him. Tan frames access to models trained on broad public data as something closer to a public good, and suggests government has a role in normalizing that view.

The stance puts Tan directly at odds with Anthropic CEO Dario Amodei, who has urged US regulators to restrict distillation and whose company just published a second report alleging illicit distillation attacks by Chinese labs. Tan’s larger concern is market concentration: he argues the real “doomer scenario” is a single monolithic company running away with the best capital, researchers, and models — and that a healthy ecosystem of open-weight alternatives is the counterweight, even as frontier labs remain well-funded and commercially viable.