Garry Tan wants an American AI distillation regime

A day after Anthropic accused Chinese labs of harvesting Claude, Y Combinator CEO Garry Tan told regulators to stay out of it — and then went further. He wants American open-weight labs doing the same thing to American frontier models. AI model distillation just went from a security complaint to a policy proposal, and the outcome sets the floor price on every cheap token you buy.

What actually happened

Asked by CNBC what should be done about distillation, Tan's answer was "I would do nothing." Speaking to TechCrunch the same week, he sketched the alternative: "We could argue that there should be an American distillation regime."

The reasoning is competitive, not charitable. Tan argues that access to intelligence trained on broad public data "should itself also be more a form of a public good than something locked away behind restrictive terms of service," and that the real risk is concentration — "the nightmare scenario, the doomer scenario for AI is that there's just one company."

This lands directly against Anthropic, whose September 10 threat report described China-based labs running distillation campaigns with fraudulent accounts and stolen credentials, and whose CEO has called for a regulatory crackdown. Tan is not disputing the conduct. He is disputing the remedy: if distillation is how a competitive open-weight tier gets built, ban it and you hand the frontier labs a permanent moat.

Nobody has legislated anything. This is two positions in public, which is exactly when it is cheap to plan around both.

Why your model layer matters for your business

Your cheap tier is not a product decision anybody at your vendor made for you. It exists because distillation produced open-weight models good enough to price against the frontier. That is the whole mechanism. Fourteen-cent agentic coding tokens are downstream of a legal gray zone currently being argued in interviews.

Two futures. In one, a US distillation regime arrives and you get American open-weight models that are cheap, inspectable, and self-hostable. In the other, export-style controls land on model outputs, the open-weight tier stalls, and your realistic options narrow to frontier pricing or a Chinese open-weight model you may not be allowed to run in a regulated deal.

You do not get to pick. You do get to make the switch cost near zero. Put every model call behind one internal interface — your own function signature, not a vendor SDK spread across forty files. Keep two providers wired and tested, one frontier and one open-weight, and run a scheduled eval against both so you know the quality gap in your workload rather than on someone's benchmark. Write down what a migration actually costs today. If that number is a quarter of engineering time, you do not have a vendor, you have a dependency.

Sources: CNBC, TechCrunch.