Garry Tan urges U.S. path for open-weight AI distillation
Y Combinator CEO Garry Tan has argued that U.S. regulators should not prevent smaller American open-weight AI labs from using distillation techniques on American frontier models. Speaking to CNBC and TechCrunch, Tan said there could be an American distillation regime, while drawing a clear line against the use of stolen credentials or other deceptive access methods.
Distillation involves extensively prompting one model to learn from its outputs and reasoning, then using that knowledge to help train another model. It is a technique AI laboratories commonly use in model development. Tan's proposal is that U.S. open-weight developers should be able to do so through legitimate access, helping create more non-Chinese open-weight alternatives.
Dispute follows Anthropic's allegations
The comments arrived after Anthropic published its second report alleging that Chinese labs carried out illicit distillation attacks. Anthropic said the alleged actors concealed their identities and used fraud and stolen credentials to distill models without permission. Its chief executive, Dario Amodei, has publicly called for U.S. regulators to act against distillation.
Tan does not share that regulatory position. His objection is not a defence of fraud or credential theft; rather, he questions whether frontier-model providers should dictate what customers may do with information supplied through API calls to closed-weight models. He also pointed to the broad collection of human knowledge, including copyrighted material, used to train proprietary systems without permission from rights holders.
Open-weight access versus frontier control
Tan said policy should recognise access to intelligence trained on broadly accessible public data as closer to a public good than something held behind restrictive terms of service. At the same time, he said frontier labs must remain fundable and retain a viable business model as they continue advancing the technology.
The tension reflects the concerns in broad restrictions on open-weight AI about broad restrictions on open-weight AI, while the present dispute focuses on how developers obtain and reuse model behaviour. Tan's aim is a balance: frontier providers continue to push capabilities forward, and open-weight models give developers freedom and access.
What businesses should assess
For companies building with AI, the immediate issue is not a settled U.S. rule but the distinction between authorised use and alleged abuse. API terms, access controls, model provenance and the permitted handling of outputs can all shape the risk of a product strategy based on closed or open-weight models.
Tan's warning is that a market led by only one capital-rich proprietary provider would concentrate too much power. Businesses should therefore evaluate model choices against contractual access, governance requirements and the availability of credible open-weight alternatives before making long-term AI commitments.