Y Combinator CEO Garry Tan Urges US Open-Weight AI Labs to Distill
Overview
Garry Tan, CEO of Y Combinator, recommends that U.S. open-weight AI labs utilize “distillation" techniques to extract knowledge from existing frontier models. Tan believes that smaller U.S. AI labs should apply similar training techniques against U.S. frontier AI labs to strengthen the options available for U.S. open-weight models.
Announcement Details
In interviews with CNBC and TechCrunch, Tan expressed the following views:
-
Endorsement of Distillation: He supports smaller U.S. open-weight AI labs using distillation techniques to learn from U.S. frontier AI labs, stating this will secure more robust U.S. open-weight model alternatives that are not made in China.
-
Stance on Regulation: He opposes regulatory intervention against the current situation where Chinese AI labs use distillation techniques to extract knowledge from frontier models, adopting a “do-nothing" stance.
-
Criticism of Usage Restrictions: He stated that controlling what users and customers do via API calls to closed-weight models is restrictive. He also touched upon the fact that proprietary AI labs have ingested vast amounts of copyrighted material without permission during model training.
-
AI as a Public Good: He suggested that access to intelligence trained on broadly public-access data should not be locked behind restrictive terms of service, and should possess the nature of a “public good."
-
Concerns Over Monopoly: Tan described the immense power of frontier AI falling into the hands of a single powerful proprietary provider as a “nightmare scenario." He warned that the concentration of capital and researchers in a single company to form a giant monolith should be avoided.
Background
Tan’s assertions come against the backdrop of an ongoing conflict between frontier model developers and those who utilize their technology. Anthropic published its second report this week stating that Chinese AI labs are conducting “illicit distillation attacks" by hiding their identities to extract knowledge without permission. Dario Amodei, CEO of Anthropic, has long publicly called for U.S. regulators to crack down on distillation technology.
In response, Tan pointed out that proprietary AI labs have ingested vast amounts of copyrighted material without permission during model training. He believes that access to intelligence trained on broadly public-access data should not be locked behind restrictive terms of service, and should possess the nature of a “public good."
Impact on Local LLM Users
This matter contains important points regarding future development freedom for developers of open-weight models and those looking to leverage them to build their own models.
- Justification of Development Methods: Tan’s recommendation of distillation could back initiatives by smaller labs to extract knowledge from frontier models and develop more efficient, superior open-weight models.
- Regulatory Trends: Because frontier model developers like Anthropic call distillation an “attack" and seek regulation, there is a risk that terms of service regarding data acquisition via APIs and model behavior will become stricter in the future. As Tan stated that “the government should intervene to normalize the aspect of a public good," the nature of regulation could dictate the freedom of development.
- Model Diversity and Access: Tan called the monopolization of frontier AI power by specific companies a “nightmare scenario," emphasizing the importance of open-weight models continuing to provide freedom and access to people. If realized, this would maintain an environment where users can choose from a diverse range of models without depending on a specific proprietary provider.
Sources
Update History
- 2026-09-14: Verified the content against the official primary source.
- 2026-09-20: Rewrote the article from re-collected sources and restored it from draft to published.