Y Combinator CEO Garry Tan wants regulators to stay out of the picture when Chinese AI labs use distillation to extract knowledge from frontier model makers. He suggested earlier this week that U.S. AI labs might want to play the exact same game.
Tan told CNBC in an interview that he would do nothing to stop the practice. He argued that the United States could establish its own American distillation regime.
Tan explained to TechCrunch that this approach means smaller, American open-weight AI labs should apply similar training methods to American frontier AI labs. This strategy would provide the United States with a stronger lineup of open-weight alternatives that are not Chinese.
Distillation occurs when a model maker extensively prompts another model to learn its functioning and reasoning processes. AI labs commonly and legitimately use this method to help train new models.
Anthropic released its second report this week claiming that Chinese labs participate in illicit distillation attacks. The report states these labs hide their identities, distill without permission, and depend on fraud and stolen credentials.
Anthropic CEO Dario Amodei previously asked U.S. regulators to crack down on distillation. Silicon Valley startup accelerator leader Garry Tan does not agree with that stance.
Tan is not encouraging American AI labs to rely on stolen credentials for distillation. He wants these labs to gain access openly through the front door.
Tan presented a twofold argument against restrictions. He believes AI labs overreach when they try to dictate customer actions regarding information shared by models.
Tan also pointed out that proprietary AI labs did not ask for permission when harvesting vast amounts of human knowledge for training. Those labs ingested considerable copyrighted material without obtaining consent from intellectual property holders.
Tan wants a healthy balance between open-weight AI labs and frontier labs. He noted that frontier labs drive progress forward and need a sustainable business model, while open-weight models deliver freedom and access to people.
Tan warned against a future where immense frontier AI power concentrates entirely within one proprietary provider. He described the ultimate AI doomer scenario as a single monolithic company running away with the market due to superior capital and researcher access.
