Garry Tan's answer to Chinese distillation: let US open-weight labs do it too

Distillation is when one model maker prompts another company's model at scale and trains on the answers. TechCrunch cites Anthropic's September threat intelligence report for the claim that Chinese labs have been doing it to US frontier models, and Anthropic's CEO asked governments to crack down on it in an essay published the following day, September 12. Garry Tan, president and CEO of Y Combinator, told CNBC at YC's Demo Day on September 11 what he would do about it: "I would do nothing." Then he went further. "We could argue that there should be an American distillation regime."

The front door

Tan's proposal, as he laid it out to CNBC and TechCrunch, is that US open-weight labs should be allowed to distill American frontier models the legitimate way: as paying customers, through the API, "in the front door," rather than with stolen credentials, which is how he characterizes the Chinese campaigns. Distillation is already a standard training technique, TechCrunch notes, used legally by labs on their own models. What Tan wants to change is the terms of service that forbid doing it to a competitor's.

His argument for that is about who owns the output. "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," he told TechCrunch. "Controlling what users and customers do with API calls to closed weight models feels constraining."

The balance he says regulators should aim for

Tan is not arguing that frontier labs should lose their business. Slashdot's summary of the two interviews has him describing the target as a tightrope: open-weight models stay viable, frontier models keep a price premium that makes their training runs pay, and regulators worry about that equilibrium rather than about distillation itself. The alternative he warns against is a single entity, or a pair of them, holding frontier capability alone.

That is the same week Dario Amodei asked for the opposite: enforcement against distillation as part of pacing the frontier, covered here. Anthropic's threat report is the evidence base both men are arguing from. Tan takes it as read and says the response should be openness, not enforcement. Neither interview names a specific rule Tan would write, and no US open-weight lab has said publicly that it would take him up on it.

Why a build studio cares

Every build here starts with a model choice, and the choice is usually between a frontier API at frontier prices and an open-weight model we can run where the client's data already lives. That choice depends on the open-weight models staying close enough to the frontier to be worth it. Tan's regime, if anyone adopted it, would keep them close on purpose. Amodei's crackdown, if it landed, would widen the gap by law. We do not get a vote, but we do get the bill, and the bill goes one way or the other depending on which of these two wins.

The part of Tan's argument we would push back on is the word "nothing." The distillation Anthropic documented ran, by Tan's own description, on stolen credentials, which is a security incident whatever the policy view; we covered an earlier case of distilled OpenAI and Anthropic models turning up in Chinese military use. Front-door distillation and credential theft are different problems, and a policy that treats them as one will get one of them wrong.

Next step: read CNBC's interview and TechCrunch's follow-up, then our earlier item on the startups' letter against an open-weight ban for the other side of the same argument. If you are choosing between a frontier API and an open-weight model for a client build, write to us at hello@gattyworks.com.