Garry Tan argues American open-weight labs should distill American frontier models

A team that fine-tunes a small model on the outputs of a paid frontier API is already distilling. TechCrunch describes the practice as extensively prompting another model to learn how it works and reasons, and reports that labs use it commonly and legitimately when training new models ^[4]. The argument this week is about who grants permission for it.

Garry Tan's answer is nobody. "I would do nothing," the Y Combinator CEO told CNBC. "We could argue that there should be an American distillation regime" ^[2]. He told TechCrunch that means smaller American open-weight labs running the same training techniques on American frontier labs, so the country has a set of open-weight options that are not Chinese ^[3]. On the government's role, he told TechCrunch that "access to intelligence that was trained on broad public access data should itself also be more a form of a public good than something locked away behind restrictive terms of service" ^[10].

The other side of this is a vendor report. Anthropic released its second report this week alleging that Chinese labs are running "illicit distillation attacks," concealing their identities to distill without permission and using fraud and stolen credentials ^[5]. Anthropic CEO Dario Amodei had already called publicly on U.S. regulators to crack down on distillation ^[6]. So the two loudest voices want opposite things from Washington on the same practice ^[14].

Tan's case has two parts. He wants American labs free to come in the front door ^[7]. He objects to labs dictating what customers do with what a model tells them ^[8], and he points out that the proprietary labs did not ask permission either when they ingested copyrighted material to train ^[9].

For the person shipping something on Monday, nothing in the contract changes. TechCrunch's account names no American open-weight lab that has agreed to distill a frontier model, and it contains no draft rule ^[15]. The pitch is a position in an argument. The practice is governed by whatever your provider's terms already say, and by whether that provider can tell your traffic apart from the traffic Anthropic is describing.

So sort by dependency. Two questions per model you call. First, does the roadmap require training your own weights on that model's outputs, or do you only need inference? Second, if the answer is yes, does the product still ship if that permission narrows? A team that answers yes and no is not making a philosophical bet on public goods; it has a single-supplier risk with a contract clause as the failure mode. A team that answers no to the first question is unaffected.

Tan's stated fear is concentration. "The nightmare scenario, the doomer scenario for AI is that there's just one company," he told CNBC. "It has the best access to capital. It has the best AI researchers. It runs away with it and suddenly there's one company that's monolithic. And that would be bad" ^[12]. He also told CNBC he wants the frontier labs to stay fundable: "We want that to be fundable, and be a great business model ongoing" ^[11].