Gary Tan: American labs should distill closed models, not wait for regulators

Y Combinator CEO Garry Tan isn't losing sleep over Anthropic's warnings about Chinese distillation of frontier models. In a CNBC interview this week he said he'd "do nothing" — and instead floated an American distillation regime: smaller U.S. open-weight labs would apply the same training techniques to American frontier models, producing open alternatives that don't come from China.

Distillation, for the uninitiated, is the process where one model intensively queries another to learn its reasoning patterns. It's a standard, legitimate industry practice. Anthropic, for its part, published a second report this week accusing Chinese labs of "illegal distillation attacks," identity concealment, fraud, and credential theft. CEO Dario Amodei has already publicly called on U.S. regulators to step in.

Tan makes clear he's not defending credential theft; he wants "front-door access." His argument has two prongs. First, it's unreasonable to let closed labs dictate what customers do with the information their models return. Second, those same proprietary labs never asked permission when they vacuumed up human knowledge at massive scale — including copyrighted material — to train their models.

"Controlling what users do with API calls to closed models feels restrictive, and there's a role for government in normalizing the fact that access to intelligence trained on broad public data should itself be more of a public good than something locked behind restrictive terms of service," he told TechCrunch.

Tan, who describes himself as a heavy AI user to the point of "cyber psychosis," wants to see an equilibrium. "They're at the frontier and pushing it forward. We want that to be fundable, and a great business model over the long term," he told CNBC. "You want open-weight models that give people freedom and access." In his view the balance is zero-sum: frontier labs fund the scientific breakthroughs; open weights distribute the power.

The real doomsday scenario, he says, isn't Chinese distillation — it's concentrating all frontier power in a single proprietary provider. "The nightmare scenario, the AI doomsday scenario, is that there's only one company. It has the best access to capital. It has the best AI researchers. It runs away with it and suddenly you have one company that's monolithic. And that would be bad."

In other words: Tan prefers wild competition, mutual distillation included, over a monopoly that locks the market behind paywalls and terms of service.