Garry Tan Pushes Open AI Distillation for US Models

Y Combinator CEO Garry Tan says US open-weight AI labs should be free to distill frontier models, arguing that broader access to model intelligence is better than a single proprietary gatekeeper. His position challenges calls for regulators to restrict the practice and puts distillation at the center of the next open-versus-closed AI fight.

What happened

In a TechCrunch report, Tan said he would do nothing to stop Chinese AI labs from using distillation techniques, and suggested American open-weight developers should be allowed to use the same approach on US frontier models. He explicitly distinguished that proposal from using stolen credentials or hiding an identity.

Distillation means repeatedly prompting a capable model and using its responses to train a smaller or more accessible model. AI companies already use the technique in legitimate research and product development.

Anthropic has called for stronger action against what it describes as illicit distillation campaigns by Chinese labs. Tan takes the opposite view: if frontier companies trained on broad public knowledge, he argues, access to the resulting intelligence should not become permanently locked behind one provider's terms of service.

Why builders should care

For AI builders, the debate is practical. Distilled and open-weight models can reduce vendor lock-in, support private deployments, and make specialized workflows cheaper to run. They also create new questions around provenance, evaluation, licensing, and whether benchmark gains reflect genuine capability or imitation of a closed model's behavior.

Teams choosing an open model should test it against their own tasks, not rely on headline benchmarks. Record which data and teacher-model outputs shaped the system, keep licensing evidence, and verify that the model can run within the security and latency limits of the deployment environment. Our earlier analysis of the open-weight AI policy fight covers the broader governance stakes.

The real fault line

Tan's argument is less about one training trick than about control of the AI ecosystem. A world with several capable open-weight options gives builders leverage over pricing, uptime, data residency, and deployment architecture. A world dominated by one frontier provider concentrates those decisions in a single commercial gate.

The counterargument is that unrestricted distillation could weaken incentives to fund frontier research or make it easier to reproduce capabilities without paying for the underlying work. That tension will not be settled by a slogan. It will be settled through clearer access rules, technical provenance, and evidence that open models can deliver reliable value without quietly inheriting a closed model's risks.

Builder takeaway: treat distillation as both an opportunity and a supply-chain dependency. Demand transparent model provenance, evaluate behavior under real workflows, and keep a fallback route that does not depend on one proprietary provider.