Y Combinator's Garry Tan: Do Nothing on AI Distillation, Focus on Open-Weight Balance

Y Combinator CEO Garry Tan is pushing back against the growing chorus of voices demanding regulatory action on AI model distillation, saying flatly he would 'do nothing' about the practice — and even suggesting the U.S. should consider building its own distillation framework. His comments came at Y Combinator's annual Demo Day, where 149 of the 196 presenting startups were AI or machine-learning ventures, signaling just how all-in the accelerator is on the technology.

Distillation — the process of training a smaller, less capable model on the outputs of a more powerful one — has become a flashpoint in the AI industry. Anthropic has accused Chinese firms Moonshot AI, DeepSeek, and MiniMax of the practice, while OpenAI believes DeepSeek's V3 and R1 models were derived from its own GPT-4 and GPT-4o architectures. The concern is serious enough that the NSA, CISA, and FBI issued a joint cybersecurity advisory on the topic. Tan's contrarian stance: the complaints ring hollow given that frontier labs themselves trained on copyrighted data without clear authorization — a point underscored by The New York Times' 2023 lawsuit against OpenAI and Microsoft, and a 2025 settlement between Anthropic and a group of book authors.

Rather than restricting distillation, Tan argues regulators should focus on maintaining a healthy equilibrium between open-weight models and proprietary frontier models — specifically ensuring frontier models can command a price premium sufficient to sustain their business models. He calls it a 'tightrope' but says it represents the best achievable outcome: open models providing broad access and freedom, while cutting-edge proprietary models retain enough commercial viability to keep advancing.

On AI safety, Tan advocates a grounded, near-term focus over existential panic. He singled out cybersecurity as the most pressing imminent risk — particularly the prospect of AI agents being weaponized to attack critical infrastructure. His pragmatism is tested, however, by Anthropic's disclosure that it blocked Claude access for five researchers in unspecified foreign countries who were apparently using the model to study dangerous pathogens, with Anthropic suspecting bioweapon research. Tan acknowledged bioweapons as a legitimate near-term existential concern, even as he downplayed fears around AI-driven mass unemployment, predicting the economic transformation will 'take decades to percolate into society.'