Garry Tan's Answer to Chinese Distillation: Do It Here Too
Asked at Y Combinator's Demo Day what regulators should do about Chinese labs distilling American frontier models, Garry Tan gave a four-word answer. I would do nothing. Then he went further: we could argue that there should be an American distillation regime.
That is not a neutral position and Tan knows it. It lands days after Anthropic's threat intelligence report accused DeepSeek, Zhipu and Xiaomi of industrial-scale distillation by relay, with a 12.1 million exchange figure attached to DeepSeek alone, and after Dario Amodei called for a regulatory crackdown. Tan's counter is that frontier models were trained on publicly available human knowledge in the first place, so access to intelligence built on that base is more a form of public good than something locked away behind restrictive terms of service. He also notes the labs did not ask permission when they ingested copyrighted material, which is the argument every defendant in every AI copyright case has been making, now repeated by the person who funds a meaningful slice of the industry.
The practical version of his proposal is interesting and mostly unremarked. Small American open-weight labs should distill American frontier models, on purpose, so the US ends up with a robust set of open-weight options that are not Chinese. Right now the serious open-weight bases people actually build on are Kimi K3, Qwen, GLM and DeepSeek. Cognition shipping SWE-2 on a Chinese base this week made that concrete in a way no policy paper could. Tan's fix is not export controls, it is domestic supply.
He draws one boundary: frontier models need to keep a price premium so the frontier business model stays feasible. The equilibrium he wants is open weights trailing closed frontier by enough margin that both survive. Whether that equilibrium exists is an empirical question nobody has answered, and it is the same question hanging over every open-weight release this year.
Read TechCrunch's writeup at https://techcrunch.com/2026/09/11/y-combinators-garry-tan-wants-u-s-open-weight-ai-labs-to-distill-frontier-models-too/ . The line that will get quoted back at him: Tan says the real doomer scenario is monopolistic concentration of AI, not runaway capability. Coming from YC, in the same week a researcher quit Anthropic over the other doomer scenario, that is a deliberate choice of enemy.
Related reading: https://clauday.com/article/50ea9194-e05b-41cf-9b28-d1ac3157f0d0 and https://clauday.com/article/b4b237cc-378d-41e1-9c1d-102a2c04e91e
← Back to all articles
That is not a neutral position and Tan knows it. It lands days after Anthropic's threat intelligence report accused DeepSeek, Zhipu and Xiaomi of industrial-scale distillation by relay, with a 12.1 million exchange figure attached to DeepSeek alone, and after Dario Amodei called for a regulatory crackdown. Tan's counter is that frontier models were trained on publicly available human knowledge in the first place, so access to intelligence built on that base is more a form of public good than something locked away behind restrictive terms of service. He also notes the labs did not ask permission when they ingested copyrighted material, which is the argument every defendant in every AI copyright case has been making, now repeated by the person who funds a meaningful slice of the industry.
The practical version of his proposal is interesting and mostly unremarked. Small American open-weight labs should distill American frontier models, on purpose, so the US ends up with a robust set of open-weight options that are not Chinese. Right now the serious open-weight bases people actually build on are Kimi K3, Qwen, GLM and DeepSeek. Cognition shipping SWE-2 on a Chinese base this week made that concrete in a way no policy paper could. Tan's fix is not export controls, it is domestic supply.
He draws one boundary: frontier models need to keep a price premium so the frontier business model stays feasible. The equilibrium he wants is open weights trailing closed frontier by enough margin that both survive. Whether that equilibrium exists is an empirical question nobody has answered, and it is the same question hanging over every open-weight release this year.
Read TechCrunch's writeup at https://techcrunch.com/2026/09/11/y-combinators-garry-tan-wants-u-s-open-weight-ai-labs-to-distill-frontier-models-too/ . The line that will get quoted back at him: Tan says the real doomer scenario is monopolistic concentration of AI, not runaway capability. Coming from YC, in the same week a researcher quit Anthropic over the other doomer scenario, that is a deliberate choice of enemy.
Related reading: https://clauday.com/article/50ea9194-e05b-41cf-9b28-d1ac3157f0d0 and https://clauday.com/article/b4b237cc-378d-41e1-9c1d-102a2c04e91e
Comments