Y Combinator's Garry Tan: "Do Nothing" About China Copying American AI Models — And Let US Startups Do the Same Thing
While the NSA, FBI, and CISA were publicly naming six Chinese AI companies in an industrial-scale copying advisory, the head of Silicon Valley's most influential startup accelerator gave a one-line answer on what Washington should do about it: nothing.
Here's the full picture behind that soundbite:
→ Garry Tan made the comment to CNBC at Y Combinator's Demo Day — days after the NSA, FBI, and CISA named DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI in a joint advisory alleging distillation of American models since late 2024
→ Distillation is the practice of extensively prompting a stronger AI model to learn how it reasons, then using those outputs to train a cheaper, smaller one
→ Anthropic's own numbers make the accusation hard to dismiss: it says Alibaba-linked operators ran over 151 million Claude exchanges between May and July 2026 — the largest such campaign it has ever measured — through more than 3,500 fraudulent accounts, using a fixed prompt specifically designed to extract Claude's chain-of-thought reasoning
→ Tan's response wasn't just "ignore it" — he wants American open-weight labs to be allowed to distill US frontier models too, calling it "an American distillation regime," so smaller domestic labs can build competitive open alternatives instead of ceding that space to China
→ He was clear he's not endorsing stolen credentials or covert access — his proposal is for lawful, open use of frontier-model APIs to learn from their outputs
→ He separated this from what he called the "existential risk" conversation dominating headlines this week, saying he's more focused on near-term risks like cybersecurity, with job-loss and economic disruption from AI likely "decades away"
→ Notably, this stance hasn't slowed YC's own AI bet — 149 of the 196 startups at that Demo Day were AI/ML ventures
Here's what stands out to me: Tan is making a market-structure argument, not a security one. His bet is that the real risk isn't China learning from American models — it's the US ending up with only closed, expensive frontier models and no domestic open-weight alternative to compete with China's. Whether you find that reassuring or reckless probably depends on how much you trust "let the market sort it out" when the market in question involves state-linked actors extracting 151 million exchanges in three months.
Where do you land — is distillation a genuine national-security threat, or a normal part of how AI competition is supposed to work?
#AI #YCombinator #GarryTan #AIRegulation #TechPolicy #ArtificialIntelligence
— 𝔖𝔞𝔫𝔡𝔢𝔢𝔭 ℜ𝔞𝔦𝔷𝔞