Garry Tan Argues for US Open-Weight AI Distillation Strategy

Garry Tan, the chief executive of Y Combinator, is urging American AI developers to embrace model distillation to foster a more competitive ecosystem. Rather than penalizing the practice, Tan suggests the United States should implement an official distillation regime allowing smaller labs to leverage knowledge from proprietary frontier models to build high-performance open-weight alternatives.

The debate over model distillation

Distillation involves training a smaller AI system by having it observe and learn from the outputs of a larger, more capable model. This technique has become a central point of contention in Silicon Valley. Anthropic, under the leadership of chief executive Dario Amodei, recently published a report accusing Chinese entities of performing illicit distillation by using fraudulent credentials to scrape data from frontier models without authorization. The Scale Report understands that these security concerns have prompted calls for stricter federal regulations on how API users interact with advanced AI systems.

Tan rejects the notion that regulators should intervene, arguing that AI labs are effectively overstepping by attempting to control how their customers utilize provided outputs. He maintains that frontier companies are built upon massive datasets often harvested from public sources without explicit consent from intellectual property holders. By framing access to intelligence as a public good, Tan suggests that locking capabilities behind restrictive terms of service creates a dangerous market imbalance.

Preventing a monolithic AI future

The core of Tan's argument lies in his fear of market consolidation. He warns that a nightmare scenario for the industry would be the emergence of a single monolithic firm that controls the best talent, capital, and technology. By promoting open-weight models, he believes developers can ensure widespread freedom and technical access, preventing any one company from monopolizing the next generation of computing.

While Tan advocates for widespread access, he clarifies that he does not support the use of stolen credentials to bypass security protocols. Instead, he envisions a transparent process where the front door to model knowledge is accessible to all domestic players. This approach, he argues, protects the business models of frontier labs while providing a necessary counterbalance to keep the broader AI field decentralized and innovative.