Garry Tan, the CEO of Y Combinator, has publicly split with Anthropic over whether distillation should be treated as theft. Tan told TechCrunch that smaller American open-weight labs should be free to use the same training techniques on US frontier models, giving the country a more robust set of open-weight options that are not Chinese.
Distillation means extensively prompting another model in order to learn how it works and reasons, and it is routinely and legitimately used by AI labs to train new models. Anthropic released its second threat intelligence report this week, alleging Chinese labs hide their identities, rely on fraud and stolen credentials, and carry out "illicit distillation attacks." CEO Dario Amodei has previously called on US regulators to crack down on the practice.
Tan's argument is twofold. He calls it an overreach for labs to dictate what paying customers do with the information their models return, and he notes that the proprietary labs did not ask permission when they ingested as much human knowledge and copyrighted material as they could to train their own models. "Access to intelligence that was trained on broad public access data should itself also be more a form of a public good than something locked away behind restrictive terms of service," he said. He is not advocating stolen credentials, only that American labs be free to walk in the front door.