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

U.S. open-weight AI models could gain capabilities without proprietary licensing costs.

Y Combinator CEO Garry Tan is advocating for U.S. AI labs to employ model distillation techniques, mirroring practices attributed to Chinese labs. Distillation involves extensively prompting an AI model to learn its reasoning and operational methods. Tan believes smaller, open-weight U.S. AI labs should be permitted to distill larger, frontier AI models, thereby increasing the availability of open-weight options not originating from China. He argues against regulatory intervention in distillation, suggesting a "distillation regime" for the U.S. Tan's stance contrasts with calls from figures like Anthropic CEO Dario Amodei, who have urged U.S. regulators to curb "illicit distillation attacks." Tan clarifies he does not endorse unauthorized or fraudulent methods, but rather wants legitimate access for U.S. labs. He posits that AI companies that trained on broad public data should not restrict how users interact with their models, drawing a parallel to how proprietary AI labs ingested vast amounts of public, and sometimes copyrighted, information without explicit permission.

Tan's proposal suggests a strategic shift in the competitive landscape of AI development, aiming to democratize access to advanced AI capabilities. By encouraging U.S. labs to distill frontier models, he seeks to foster a more robust ecosystem of open-weight alternatives, challenging the dominance of closed, proprietary systems and potentially mitigating concerns about a single entity controlling powerful AI technology. This approach could spur innovation by allowing a wider range of developers to build upon and refine sophisticated AI models, leading to more diverse applications and increased accessibility.

This perspective challenges the current model of AI development where frontier models are often tightly controlled by a few large companies. If adopted, it could lead to greater transparency and collaboration within the AI community, as well as stimulate competition by lowering the barrier to entry for developing advanced AI. It also raises questions about intellectual property and the definition of AI as a public good, potentially prompting a re-evaluation of how AI models are accessed, used, and further developed in the future.