Y Combinator's Garry Tan Urges U.S. Labs to Embrace Model Distillation
AI-generated image · US National Wire
The startup accelerator CEO argues that American open-weight labs should be permitted to extract knowledge from frontier models to counter Chinese competition.
Y Combinator CEO Garry Tan is calling for an "American distillation regime," suggesting that U.S.-based open-weight AI labs should be free to use distillation techniques to learn from the reasoning and functions of frontier AI models, as TechCrunch first reported.
Distillation involves a model maker using extensive prompting of another model to extract knowledge. While Tan told CNBC he believes this approach would provide the U.S. with a more robust set of non-Chinese open-weight options, the practice is a point of contention among industry leaders. Anthropic recently released a report alleging that Chinese labs are conducting "illicit distillation attacks" through the use of stolen credentials and fraud. Anthropic CEO Dario Amodei has previously urged U.S. regulators to crack down on these activities.
Tan told TechCrunch that he does not support the use of stolen credentials, but argues that proprietary labs should not be able to dictate how customers use the information provided by their models. He further noted that frontier labs previously ingested copyrighted material without permission to train their own systems.
Tan expressed concern over a "doomer scenario" where a single monolithic company controls the AI frontier due to superior access to researchers and capital. He told CNBC that while frontier labs should remain fundable businesses, open-weight models are necessary to ensure freedom and access to intelligence.