Garry Tan Wants US Labs to Distill Frontier AI Models Freely

Y Combinator CEO Garry Tan told CNBC and TechCrunch this week that he wants US regulators to stay out of AI distillation disputes and instead encourage American open-weight labs to freely distill knowledge from frontier models. His position puts him directly at odds with Anthropic CEO Dario Amodei, who has called on US regulators to crack down on distillation. Anthropic published its second report this week alleging Chinese labs are conducting illicit distillation attacks using stolen credentials and fraudulent identities.

What happened

Detail	Fact
Who	Garry Tan, CEO of Y Combinator
When	Interview published September 11, 2026
Position	US open-weight labs should be free to distill frontier models
Opposing view	Anthropic CEO Dario Amodei, who wants regulators to restrict distillation
Anthropic report	Second report alleging Chinese labs use stolen credentials to distill without permission

Garry Tan made his position plain in two sentences to CNBC: “I would do nothing. We could argue that there should be an American distillation regime.” He then expanded on that to TechCrunch, saying he wants smaller, American open-weight AI labs to apply the same training techniques to US frontier models, building out a richer set of open-weight options that are not Chinese-made.

Distillation, for readers less deep in AI tooling, is the process of extensively prompting one model to learn how it reasons and responds. The goal is to use those outputs to help train a smaller or different model. It is a standard, legal practice across the AI industry, though Anthropic argues the Chinese version involves hiding identities and using stolen credentials.

To be precise about what Tan is and is not saying: he is not endorsing fraud or credential theft. He wants American labs to walk in the front door and distill openly.

Why it matters

Tan’s argument cuts in two directions. First, he questions whether AI labs have the standing to tell paying API customers what they can do with information those models return. Second, he points out that frontier AI labs themselves scraped vast amounts of copyrighted human knowledge without asking permission from the original rights holders.

“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.”

That framing matters for any business using AI APIs today. If a regulatory norm were established that labeled distillation a legitimate activity, it could reshape what developers and companies are permitted to build with model outputs, including fine-tuning pipelines, synthetic training sets, and agent orchestration systems.

For businesses following debates about AI regulation in Congress, this is a concrete policy fault line: one camp wants distillation treated like IP theft, the other wants it treated more like a public good.

Tan also laid out his long-term concern. He believes the worst outcome for the AI ecosystem is not open distillation, but a single company monopolizing frontier AI. “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.”

He wants frontier labs to stay well-funded and keep pushing capability forward, while open-weight alternatives keep power distributed. Whether that balance is achievable through distillation policy is the open question.

Our take

Tan’s point about the hypocrisy of frontier labs restricting distillation is hard to dismiss. Labs that built their products on unlicensed internet data now want regulatory protection from the same kind of appropriation applied to their own outputs. That is a notable inconsistency, and it is worth watching whether courts or Congress treat the two situations differently.

For businesses actually building with AI, this debate is not just academic. If distillation gets legally restricted, using model outputs to build fine-tuning datasets or synthetic data pipelines could become a compliance risk. If Tan’s framing wins, the ecosystem of open-weight models gets stronger, which generally benefits smaller operators who cannot afford frontier API costs at scale.

Teams thinking through their AI integration strategy should pay attention to which way this lands. A world with robust open-weight American models is a different cost and capability environment than one locked to a small number of proprietary APIs. Right now, the policy direction is genuinely unsettled.

Watch Anthropic’s next regulatory push and whether Congress picks up Amodei’s framing or Tan’s. The outcome will have direct consequences for what developers can legally build.

Source: TechCrunch · AI