Garry Tan wants US open-weight AI labs to distill frontier models too

Y Combinator CEO Garry Tan says he wants US regulators to leave AI model distillation alone, and thinks American open-weight AI labs should get to play the same game Chinese labs are accused of playing. Asked by CNBC this week about Chinese labs using distillation to extract knowledge from frontier US models, Tan said: "I would do nothing." He added that there could be a case for "an American distillation regime."

He told TechCrunch what he meant by that: smaller American open-weight AI labs should get to use the same training technique on American frontier models, building a more robust set of open-weight options that are not Chinese. Distillation, in this context, is when one model maker extensively prompts another company's model to learn how it works and reasons; it is a technique already used commonly and legitimately across the AI industry to train new models.

His comments land right after Anthropic released its second report this week alleging that Chinese labs are running "illicit distillation attacks": hiding their identities, distilling without permission, and relying on fraud and stolen credentials to do it. Anthropic CEO Dario Amodei has previously called publicly on US regulators to crack down on distillation. Tan, who runs one of Silicon Valley's most prominent startup accelerators, is taking the opposite position.

Tan draws a careful line: he is not asking for American labs to use stolen credentials, only to be allowed to distill through legitimate access, "the front door." His argument has two parts. First, he thinks it is overreach for AI labs to dictate what their customers do with the information a model shares with them through the API. Second, he points out that the proprietary labs never asked permission either, when they trained their own models on as much human knowledge as they could gather, copyrighted material included. As he put it to TechCrunch: "Controlling what users and customers do with API calls to closed weight models feels constraining, and there's a role government can play here to normalize the fact 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."

Tan, who has described himself as such a heavy AI user that he once said he had "cyber psychosis," wants a balance between open-weight labs and frontier labs rather than a win for either side. "They are at the frontier and driving it forward. We want that to be fundable, and be a great business model ongoing," he told CNBC, adding: "You want open weight models to give people freedom and access." To him, the real worst case for AI is not distillation but concentration: a single proprietary company cornering all the capital and the best researchers. "The nightmare scenario, the doomer scenario for AI is that there's just one company," he said. "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."

Key facts

Garry Tan told CNBC "I would do nothing" about regulating AI model distillation, and said there could be a case for "an American distillation regime."
Tan wants smaller American open-weight AI labs to distill American frontier models using the same technique Chinese labs are accused of using, to build open-weight alternatives that are not Chinese.
Anthropic released its second report this week alleging Chinese labs run "illicit distillation attacks" using fraud and stolen credentials; Anthropic CEO Dario Amodei has previously called for a US regulatory crackdown on distillation.
Tan explicitly excludes stolen-credential access from his proposal: he wants only legitimate, front door distillation through normal API use.
Tan calls a single monolithic proprietary AI company, one with the best capital and researchers, the real doomer scenario for AI, not distillation.

Why it matters

This is a direct, public split between two influential figures over how the US should treat AI model distillation, right as Anthropic pushes regulators toward a crackdown. Garry Tan runs Y Combinator, one of Silicon Valley's most prominent startup accelerators, and he argues that stopping only Chinese labs from distilling frontier models leaves American open-weight labs locked out of the same technique, which keeps whichever company already sits at the frontier further ahead. That ties his comments directly into a live policy question: whether the government regulates distillation at all, and if it does, whether the rule applies evenly rather than singling out one country's labs.

Who it affects

American open-weight AI labs would gain a case for treating legitimate distillation of frontier models as fair game rather than something to shut down. Frontier and proprietary labs, Anthropic chief among them here, see their push to restrict API based distillation directly challenged, along with the closed-weight business model Tan calls constraining. Chinese AI labs sit at the center of Anthropic's report, though the source does not name which ones or say how many. US regulators are the audience for Tan's ask, whether that means staying out entirely or writing any future rule so it applies to American labs too, not only foreign ones.

How to use it

There is no product or service here to adopt; this is an argument aimed at policy. The practical line Tan draws is between two kinds of access: distillation through stolen credentials or fraud, which he does not defend, and distillation through ordinary, legitimate API use, which he wants normalized as a public good rather than something a closed-weight lab's terms of service can simply forbid. For an American open-weight lab, that is the boundary his proposal would move: today's ordinary API access, used more aggressively to learn from a frontier model's outputs.

How solid is it

The core of the story rests on Tan's own words, given directly to CNBC and then again to TechCrunch when asked to elaborate, which is about as solid as sourcing gets for a personal policy opinion. What is thinner is the material he is responding to: the source describes Anthropic's second report only as alleging "illicit distillation attacks" by Chinese labs, without naming a single lab, giving a count, or citing the report's title or methodology. Neither the CNBC interview nor the Anthropic report carries a date more specific than "this week" or "earlier this week."

Risks and caveats

Tan's proposal is his own advocacy, not an announced policy: the source gives no timeline or mechanism for how an American distillation regime would work, and no sign that regulators or Congress are actually considering one. His position should be read narrowly: he is defending legitimate, front door distillation, not the fraud and stolen credentials Anthropic's report describes, and the specific American open-weight labs he has in mind are not named. The source also does not report any response yet from Anthropic, Dario Amodei, or the Chinese labs referenced in that report.

“Controlling what users and customers do with API calls to closed weight models feels constraining, and there's a role government can play here to normalize the fact 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.”

— Garry Tan, Y Combinator CEO, to TechCrunch