Garry Tan Urges U.S. Open-Weight Labs to Distill Frontier Models

Model distillation trains a smaller “student” system on outputs from a more capable model, a routine technique that can cut development costs and speed deployment. The practice becomes contentious when rivals evade access controls or terms of service to extract proprietary capabilities at scale. The dispute has become a strategic question for Washington: whether tighter enforcement protects U.S. intellectual property and AI safety, or entrenches closed-model leaders while China builds a stronger open-weight ecosystem.

Y Combinator CEO Garry Tan said in an interview published Sept. 11, 2026, that regulators should “do nothing” and consider an “American distillation regime” allowing smaller U.S. open-weight labs to learn from frontier models. His stance contrasts with Anthropic’s Sept. 10 warning about five alleged China-linked campaigns totaling nearly 200 million exchanges. Anthropic said an Alibaba-linked effort generated 151 million exchanges from May through July, peaked at almost 3 million a day and used about 3,500 accounts, while other campaigns were tied to DeepSeek and Moonshot AI.