Y Combinator CEO Garry Tan said he believes regulators should not intervene in AI model distillation, the practice of extensively prompting one model to learn its reasoning and replicate its capabilities, and suggested American AI labs might benefit from adopting similar techniques against domestic frontier competitors.
Tan’s comments come after Anthropic released a report this week alleging that Chinese AI labs have engaged in unauthorized distillation attacks, using concealed identities, fraud, and stolen credentials to extract knowledge from its models. Anthropic CEO Dario Amodei has previously called on regulators to address the practice. Tan clarified that he does not support using stolen credentials, but believes labs should be free to access competitors’ models legitimately and use the resulting information as they see fit.
Tan argued that frontier AI companies, which trained their own models on vast amounts of publicly available and often copyrighted data without permission, have limited standing to restrict how customers use information obtained through legitimate API access. He said government policy should treat broadly-trained AI intelligence more like a public good rather than something restricted by proprietary terms of service.
Tan added that he wants to preserve a balance between open-weight and frontier AI labs, warning that the greatest risk to the industry would be a single dominant company controlling access to the most advanced AI capabilities.
