Garry Tan wants US open-weight AI labs to 'distill' frontier models, too

The largest AI labs in the U.S. may soon be legally able to use their frontier models to train smaller, open-weight models.

Garry Tan, CEO of Y Combinator, has proposed that U.S. AI labs should be permitted to "distill" frontier models, a technique where a model is extensively prompted to learn its reasoning. Tan argues that American open-weight AI labs should employ this method on U.S. frontier AI models to foster a stronger domestic ecosystem of open-weight AI. This suggestion comes amidst Anthropic's recent reports alleging illicit distillation practices by Chinese labs, prompting calls for regulatory action. Tan, however, believes that AI labs should not dictate how customers use information derived from API calls, drawing a parallel to how proprietary labs ingested vast public data without explicit permission. He advocates for treating intelligence trained on public data as a public good rather than a proprietary asset, aiming to prevent a single entity from monopolizing AI power.

Tan's proposal challenges the current debate around AI model distillation, which has been framed by companies like Anthropic as an illicit activity when conducted by foreign entities. By advocating for U.S. labs to openly distill frontier models, Tan suggests a redefinition of the practice, positioning it as a legitimate tool for fostering a competitive and accessible AI landscape. This approach could democratize access to advanced AI capabilities, preventing a concentration of power in a few proprietary hands.

This perspective frames AI development as a public good, especially when trained on publicly available data, and argues against restrictive terms of service that limit downstream innovation. If adopted, Tan's vision could lead to increased competition among AI models and greater accessibility for researchers and developers. The proposed "American distillation regime" aims to balance the advancement of frontier models with the need for widespread access and a robust open-weight ecosystem, potentially influencing future regulatory discussions and industry norms.