Y Combinator CEO Garry Tan says US open-weight labs should distill frontier models

Garry Tan, the CEO of Y Combinator, has taken a position directly opposed to Anthropic in the escalating fight over model distillation: he wants regulators to do nothing, and he thinks American open-weight labs should be distilling from US frontier models themselves.

Asked by CNBC how regulators should respond to Chinese labs extracting knowledge from frontier models, Tan said, "I would do nothing," adding, "We could argue that there should be an American distillation regime." He later explained to TechCrunch what he means by that: smaller US open-weight labs should be free to apply the same training techniques to American frontier models, so the United States ends up with a stronger set of open-weight options that are not Chinese.

What distillation is and why it is contested

Distillation involves heavily querying a model to work out how it behaves and reasons, then using what you learn to train another model. As TechCrunch notes, it is a routine and legitimate technique inside AI labs. The dispute is about who is allowed to do it to whom.

Anthropic published its second report on the issue this week, alleging that Chinese labs are carrying out "illicit distillation attacks" — concealing their identities and, in some cases, relying on fraud and stolen credentials to query models without permission. Anthropic CEO Dario Amodei has publicly urged US regulators to step in.

Tan's disagreement matters because of where it comes from: the head of one of Silicon Valley's most prestigious and prolific startup accelerators is breaking with the frontier labs on a policy question they are actively lobbying on.

Tan's two-part argument

Tan is not endorsing stolen credentials. His position, as he framed it, is that US labs should be able to "come in the front door" — distilling openly rather than deceptively.

His case rests on two points. First, he believes it is overreach for AI labs to dictate what paying customers do with the outputs their models generate. Second, he points out that the proprietary labs built their systems by ingesting enormous amounts of human knowledge — including plenty of copyrighted material — without asking the permission of the people who created it.

"Controlling what users and customers do with API calls to closed weight models feels constraining," Tan told TechCrunch, arguing that intelligence trained on broadly accessible public data should be treated "more a form of a public good than something locked away behind restrictive terms of service."

Not a war on the frontier labs

Tan is not arguing against the frontier labs' existence or their economics. He told CNBC that those labs are "at the frontier and driving it forward" and that he wants that work to remain fundable as an ongoing business, while open-weight models exist to "give people freedom and access."

The outcome he fears is concentration of power. In his telling, the genuine doomsday case for AI is not runaway capability but a single proprietary provider — with the best access to capital and the best researchers — pulling so far ahead that nobody can catch up. One monolithic company holding the full power of frontier AI, he said, "would be bad."

Why it matters

Distillation is becoming a fault line in US AI policy. On one side, Anthropic is documenting alleged abuse and asking for a regulatory crackdown. On the other, the head of Y Combinator is arguing that model outputs derived from public data should sit closer to a public good than to a licensed asset. Where regulators land will decide whether frontier labs can use terms of service to restrict downstream training, whether American open-weight labs get a sanctioned path to competitive models, and how much of the open-weight landscape ends up Chinese by default.