Garry Tan Urges U.S. Open-Weight Labs to Distill Frontier Models

What changed

Y Combinator CEO Garry Tan said U.S. open-weight AI labs should be allowed to “distill” American frontier models through ordinary API access, rather than face new regulatory limits, according to TechCrunch.

Distillation involves extensively prompting one model so another can learn how it responds and reasons. Tan supports legitimate, front-door access: not the fraud, hidden identities or stolen credentials Anthropic says Chinese labs have used in alleged “illicit distillation attacks.” Anthropic CEO Dario Amodei has called for U.S. regulators to crack down on that conduct.

Tan argues that AI companies should not control everything customers do with information returned through API calls. He compares that concern with frontier labs’ use of vast amounts of publicly available, including copyrighted, human-created material to train their own systems.

Why it matters

Tan’s proposal would give smaller U.S. labs a path to build open-weight models from the capabilities of closed systems without first negotiating permission. That could expand models available for local deployment, customization and inspection, while increasing competitive pressure on frontier labs.

His position depends on balancing two goals: preserving a fundable business model for companies developing frontier systems while preventing advanced AI capabilities from remaining concentrated behind a few closed providers. If regulators protect legitimate distillation, open-weight labs could gain a clearer route to catch up. If frontier companies retain broad contractual or legal control over API-derived training, open models may remain dependent on permission from the systems they challenge.

What to watch next

The key questions are whether U.S. regulators distinguish authorized API-based distillation from deceptive or credential-based extraction, and whether frontier labs seek stronger contractual limits on how customers use model outputs. The outcome will help define whether access to closed models is treated primarily as a commercial service or as a research gateway to more open systems.