Garry Tan Urges US Labs to Distill Frontier AI Models

Y Combinator chief Garry Tan has publicly defended the practice of AI model distillation, arguing that U.S. open-weight AI developers should actively use the technique on domestic frontier models. Distillation involves extensively prompting a powerful model to analyze and replicate its reasoning capabilities. Tan's comments, made in interviews with CNBC and TechCrunch, position him against proprietary AI giants that are lobbying regulators to restrict the practice.

The debate intensified after Anthropic released its second report detailing what it termed "illicit distillation attacks" by Chinese AI laboratories. Anthropic chief executive Dario Amodei has urged U.S. regulators to intervene, accusing foreign entities of using fraudulent credentials to bypass security and distill their proprietary systems. While Tan does not support using stolen credentials, he believes legitimate API users should face no restrictions on how they utilize model outputs.

Tan argued that proprietary model developers did not seek permission when training their systems on vast amounts of copyrighted public data. Consequently, he believes restricting what customers do with API responses is an overreach. He suggested that government intervention should instead "normalize the fact" that access to this intelligence behaves more like a public good rather than being locked behind restrictive terms of service.

For AI practitioners and startups, Tan's stance supports a vibrant ecosystem of open-weight alternatives that can compete with heavily funded proprietary systems. He warned that the ultimate "doomer scenario" for the industry is not superintelligence, but rather a single monolithic company monopolizing capital and research. Encouraging a domestic distillation regime would allow smaller American startups to build highly capable, open-weight models, ensuring technological diversity and preventing a single provider from dominating the market.