Y Combinator CEO urges US open-weight AI labs to distill frontier models without restrictions

Garry Tan’s proposal reframes distillation as a legitimate practice for open-weight AI labs, rather than a regulatory or ethical violation. He positions it as a counterbalance to the dominance of proprietary frontier models, which he warns could lead to a monopolistic AI landscape. This perspective shifts the debate from unauthorized access to the broader question of who controls AI knowledge derived from public data.

The argument hinges on the idea that frontier models were trained on publicly available human knowledge, often without explicit permission from copyright holders. Tan implies that restricting distillation, while frontier labs themselves relied on unlicensed data, creates an unfair asymmetry. For engineers, this could mean fewer legal barriers to experimenting with or improving upon frontier models, but it also risks escalating tensions with proprietary labs over intellectual property.

Tan’s stance contrasts sharply with Anthropic’s recent allegations of illicit distillation by Chinese labs, which involved fraud and stolen credentials. While Tan does not endorse such tactics, his call for a regulated but permissive distillation regime in the U.S. could complicate efforts to curb unauthorized access. Engineers working with open-weight models may face uncertainty about whether their distillation practices will be deemed acceptable or face future legal challenges.

The proposal also highlights a divide in the AI industry over the role of government in regulating model access. Tan suggests that regulators should normalize distillation as a public good, while others, like Anthropic, advocate for stricter enforcement. For engineers, this debate could influence the availability of open-weight alternatives, the terms of service for frontier models, and the legal risks of building on top of proprietary systems.

If adopted, Tan’s vision could lead to a more diverse AI ecosystem, with open-weight labs offering alternatives to proprietary models. However, it also risks fragmenting the industry, as frontier labs may tighten access or impose stricter terms to protect their models. Engineers would need to weigh the benefits of open access against potential trade-offs in model performance, legal exposure, and the long-term sustainability of open-weight development.