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

Y Combinator CEO Garry Tan is pushing back against calls to regulate AI model distillation, arguing that U.S. open-weight labs should be permitted to use the technique on American frontier models. This position directly contrasts with recent warnings from companies like Anthropic, which has accused Chinese labs of conducting “illicit distillation attacks” and has publicly called on U.S. regulators to intervene. Tan's stance injects a significant pro-competition viewpoint into a debate increasingly focused on security and intellectual property.

Distillation is a training method where one model learns by extensively prompting another to understand its reasoning processes. While Anthropic alleges that some labs use fraud and stolen credentials for this purpose, Tan advocates for an “American distillation regime” where smaller U.S. entities could do so openly. He argues that since frontier models were trained by ingesting vast amounts of public and copyrighted data without permission, the intelligence derived from them should be more accessible, rather than locked behind restrictive terms of service that dictate what customers can do with API outputs.

Tan's Stance on AI Competition

• Preventing Monopoly: Tan's primary concern is avoiding a “doomer scenario” where a single, monolithic company controls the most powerful AI.
• Promoting Open-Weights: He believes allowing distillation will create a more robust set of American open-weight models to counterbalance proprietary systems.
• Fair Use Analogy: He draws a parallel between distillation and the initial data scraping performed by frontier labs, suggesting it's hypocritical to restrict one while benefiting from the other.
• Public Good: Tan frames access to AI intelligence, trained on public data, as a form of public good that shouldn't be overly constrained by corporate policy.

The impact of this perspective could shift the regulatory conversation from solely focusing on preventing misuse to actively fostering a competitive domestic AI ecosystem. By framing distillation as a necessary check on the power of frontier labs, Tan challenges the narrative that such techniques are inherently harmful. His argument suggests that the greater risk is not knowledge transfer between models, but the concentration of AI power in the hands of a few proprietary companies, which could stifle innovation and access for the broader market.

Garry Tan is reframing model distillation not as an illicit attack but as a pro-competitive tool essential for preventing a monopolistic AI future, effectively pitting the principles of open innovation against the proprietary terms of service of leading AI labs.