Garry Tan Urges US Open-Weight AI Labs to Distill From American Frontier Models

Y Combinator president Garry Tan is pushing for American open-weight AI labs to adopt distillation techniques that train smaller models on the outputs of larger domestic frontier systems, arguing the approach would give the United States a broader set of competitive open models that aren't traced back to Chinese developers.

Tan's argument lands as open-weight releases from Chinese labs, including DeepSeek and Alibaba's Qwen family, have accumulated large followings among developers worldwide. Those models are often cheaper to run and freely downloadable, which has made them a default starting point for startups that can't afford to train their own systems from scratch. Tan's concern is that this dependence hands Beijing a form of soft influence over the global developer ecosystem, even if the underlying weights carry no strings attached.

Distillation is not a new technique. It works by training a smaller model to imitate the outputs of a larger, more capable one, and it has become a standard shortcut for teams without the compute budgets of a frontier lab. What Tan is proposing is essentially a national-strategy framing for a practice that has mostly been discussed as an efficiency trick: use American frontier models as the teacher, and American open-weight labs as the student.

That framing runs straight into unsettled legal and commercial territory. Several U.S. frontier labs restrict using their outputs to train competing models in their terms of service, and the question of whether distillation from a closed API violates those terms has not been clearly tested. If Tan's idea gains traction, it could force a public reckoning over what open-weight development is permitted to look like inside the United States.

The pitch also raises a practical asymmetry. Chinese open-weight labs have released models with permissive licenses and detailed technical reports, making them easy to build on. American frontier labs have largely kept weights closed, so any distillation-based strategy would depend on their willingness to tolerate or actively support downstream training. That cooperation is not guaranteed, and without it Tan's proposal amounts to asking smaller labs to work from APIs whose terms may prohibit exactly this use.

For now, Tan's comments are a signal about where parts of the U.S. startup world think the open-weight fight is heading. The next thing to watch is whether any American frontier lab publicly endorses distillation by domestic open-weight teams, or whether the terms of service stay as they are. If they do, the gap Tan is warning about may widen regardless of how many people agree with him.