Garry Tan, CEO of Y Combinator, has expressed support for allowing American AI labs to use distillation methods on leading AI models, a practice currently under scrutiny due to allegations involving Chinese labs. Distillation involves extensively querying a model to understand its reasoning and replicate its capabilities, a common technique used to train new AI systems.
Tan argues that US regulators should avoid restricting this practice and instead consider establishing a framework that enables smaller, open-weight AI labs to distill models from frontier AI developers. This approach, he believes, would foster a diverse and competitive AI landscape in the US, providing alternatives to Chinese open-weight models.
This stance contrasts with concerns raised by companies like Anthropic, which recently accused Chinese labs of conducting unauthorized distillation attacks using deceptive methods. Anthropic's CEO has called for regulatory action against such practices. However, Tan emphasizes that his support is for legitimate, transparent distillation rather than illicit activities.
He points out that proprietary AI developers have historically trained their models on vast amounts of publicly available data without explicit permission from content owners. Therefore, restricting how customers interact with these models through API calls could be overly limiting. Tan suggests that access to AI intelligence derived from public data should be treated more like a public good rather than being locked behind restrictive terms.
Tan envisions a balanced AI ecosystem where frontier labs continue to innovate and receive funding, while open-weight models provide broader access and freedom to users. He warns against a scenario where a single dominant company controls the most advanced AI technology, which could stifle competition and innovation.
By advocating for open distillation, Tan aims to promote a more resilient and competitive AI environment in the US, preventing monopolization and encouraging innovation across the industry.