Y Combinator’s Garry Tan wants US open-weight AI labs to ‘distill’ frontier models, too

As Anthropic alleges illicit distillation by Chinese labs and CEO Dario Amodei urges regulators to crack down, Y Combinator CEO Garry Tan is staking the opposite pole. In a CNBC interview he said he “would do nothing,” and floated an “American distillation regime.” To TechCrunch he clarified: smaller U.S. open-weight labs should be able to use the same training techniques on American frontier models—without stolen credentials—so the country has stronger open options that aren’t Chinese.

That puts Silicon Valley’s most visible accelerator chief against the closed-lab security narrative in the same news week as Anthropic’s second distillation report. The stakes are who gets to inherit frontier capability: a few proprietary providers, a wider open-weight ecosystem, or whoever is willing to cheat.

If Tan’s frame wins, terms of service that forbid wholesale capability extraction look like overreach. If Anthropic’s frame wins, API access hardens and “legitimate distillation” shrinks.

We’re sympathetic to the nightmare Tan names: one company with the best capital and researchers running away with the stack. Monopoly intelligence is a real doomer path, and healthy open-weight competition is one of the few structural hedges.

But equating customer learning with industrial distillation attacks collapses important distinctions. TechCrunch notes Tan is not calling for stolen credentials; he wants front-door freedom. Anthropic’s allegation is the opposite pattern—fraud, concealment, and illicit extraction. Policy that “does nothing” about the latter while blessing the former needs a bright line, or thieves hide inside the reform.

Tan’s reciprocity jab lands: frontier labs trained on broad public data and, famously, plenty of copyrighted material without permission. Asking those same labs to police every API customer’s downstream training use does feel constraining. His public-good argument—that intelligence trained on wide data should not live only behind restrictive terms—is the strongest part of the case.

Our read: useful open weights deserve room to grow, including legitimate distillation and synthetic data pipelines that labs consent to or that law clearly allows. Covert account farms stripping chain-of-thought traces are not the same project. Pretending they are rewards the attacker and poisons the open-source brand.

I’m watching whether Washington treats distillation as trade enforcement, IP/ToS enforcement, or a competition policy lever—and whether any U.S. open-weight lab actually announces a clean, front-door distillation partnership with a frontier provider.

TechCrunch’s Julie Bort piece (September 11, 2026) situates Tan’s comments against Anthropic’s latest report on alleged Chinese illicit distillation and Amodei’s prior call for a regulatory crackdown. Distillation, as the article defines it, is extensively prompting another model to learn how it works and reasons—and is also a common legitimate training technique.