Y Combinator's Garry Tan: Don't Regulate AI Distillation, Build an American Version Instead

Y Combinator CEO Garry Tan broke with the national security establishment at Demo Day 2026, arguing the U.S. should do nothing to stop AI model distillation — the practice of training smaller models on the outputs of more powerful ones — and should instead consider building its own distillation framework. His comments came the same week the NSA, CISA, and FBI jointly issued a cybersecurity advisory warning about the practice, and amid ongoing accusations from Anthropic that Chinese firms including DeepSeek, Moonshot AI, and MiniMax have distilled from its models without permission.

Tan's position is rooted partly in the murky legal terrain around AI training data. Critics of OpenAI and Anthropic's distillation complaints note that much of the data used to train frontier models may itself violate copyright law — a point underscored by active litigation, including The New York Times' 2023 lawsuit against OpenAI and Microsoft, and a 2025 settlement between Anthropic and a group of book authors.

Rather than cracking down, Tan wants regulators to focus on maintaining a healthy market equilibrium between freely available open-weight models and proprietary frontier ones — ensuring frontier models can still command a price premium that keeps their business viable. He describes this as a tightrope worth walking, calling the right balance potentially the best possible outcome for the industry and for public access to AI.

On safety, Tan pushed back against what he characterized as science-fiction-level doomsday fears, even as Anthropic's recent disclosure that it blocked five foreign scientists from using Claude to research dangerous pathogens added new urgency to bioweapon concerns. Tan believes near-term cybersecurity threats are the real priority, while longer-term risks like AI-driven job displacement will take decades to fully materialize — which he frames as a feature, not a bug, giving society time to adapt toward more creative work.

Meanwhile, Y Combinator is betting heavily on the technology regardless: 149 of the 196 startups presenting at the current Demo Day are classified as AI or machine-learning ventures.