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

Y Combinator CEO Garry Tan told CNBC and TechCrunch that U.S. open-weight AI labs should be allowed to use distillation techniques on frontier models, proposing an American “distillation regime” rather than regulatory crackdowns. He argued that access to intelligence trained on broadly available public data should function more like a public good and warned against a single proprietary company monopolizing frontier AI capabilities.

By AI NewsroomPublished 40 minutes agoUpdated 40 minutes ago0 views

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

Why It Matters

The debate affects who can replicate and build on leading AI systems: allowing lawful distillation would expand open-weight alternatives in the U.S., while regulators and some companies worry about unauthorized extraction of model capabilities. Decisions here will shape competition, research access, and how much power rests with closed, frontier model providers.

Key Facts

Person: Garry Tan, CEO of Y Combinator
Tan's stance to CNBC: "I would do nothing." (on regulators acting against distillation)
Proposal: Tan said the U.S. could adopt an "American distillation regime" and wants smaller American open-weight labs to use distillation on American frontier labs.
Distillation defined: A training technique where a model maker extensively prompts another model to learn how it reasons and works.
Anthropic report: Anthropic released a second report alleging Chinese labs conduct "illicit distillation attacks" this week.

Y Combinator CEO Garry Tan told CNBC he would not support regulatory intervention against model distillation and suggested the United States consider its own "American distillation regime." Speaking to TechCrunch, he said smaller U.S. open-weight AI labs should be able to use distillation techniques on American frontier models, creating more domestically based open-weight options rather than ceding that space to Chinese labs.

Distillation, as Tan and other practitioners describe it, involves extensively prompting a target model to understand and reproduce its reasoning and behavior; it is a common method used to train successor or distilled models. The practice has become contentious after Anthropic published a second report this week accusing some Chinese labs of conducting what it calls "illicit distillation attacks," allegedly using hidden identities and fraudulent credentials to distill without permission. Anthropic CEO Dario Amodei has urged U.S. regulators to take action.

Tan rejects crackdowns and distinguishes lawful distillation from the alleged illicit activity Anthropic describes. He argues that proprietary frontier labs overreached when they scraped broad swaths of human-created material to train their models without seeking permission from rights holders, and that it is inconsistent for those same labs to restrict what users do with model outputs or API responses. He told TechCrunch that intelligence derived from broadly accessible public data should be treated more like a public good than something fully locked behind restrictive terms of service.

While advocating for permissive rules for open-weight labs, Tan also said he supports a mixed ecosystem: he wants frontier labs to remain fundable and commercially viable, while open-weight models provide wider access and freedom for developers and researchers. He warned that the worst outcome would be a single dominant company concentrating capital, talent and model access — a monopoly scenario he sees as a greater risk than allowing lawful distillation. Tan emphasized he is not endorsing theft or misuse of credentials, but rather legal, front-door access to model outputs as part of a balanced AI landscape.