Y Combinator's president wants American open models to copy the big ones too

Anthropic last week accused seven China-based labs of using Claude through fraudulent accounts to train their own models, and called for regulatory action. This week Y Combinator's president Garry Tan argued the opposite: no new rules, and American open-weight labs should use the same method.

The technique under discussion is distillation. You send a strong model a large number of questions, collect its answers, then train a smaller model on those question-and-answer pairs. The smaller model inherits part of the larger one's ability. The cost is close to nothing, because the expensive part, the original training, was already paid for by someone else.

Tan's answer on CNBC was short: asked what regulators should do, he said he would do nothing. In an interview with TechCrunch he set out his reasoning. Policing what a customer does with API calls to a closed-weight model strikes him as constraining. He added a second point: frontier labs did not ask permission when they trained on copyrighted material.

His real concern is American open models, not China

Tan is not arguing that what the Chinese labs did was acceptable. His point is that the United States has small labs publishing open-weight models, and if those labs can learn from American frontier models in the same way, the open models the world runs will not all come from China.

Today the weight of the open-weight world really does sit in China. Qwen, DeepSeek, GLM, Kimi and MiniCPM can all be downloaded and run on your own machine. There is no American family of comparable scale.

Both sides are internally consistent. Anthropic's objection is to terms of service being broken and fraudulent accounts being used; Tan's is that the same labs showed a similar disregard for permission when they gathered their own training data. The two arguments do not meet.

Tan made the same case to CNBC at Y Combinator's Demo Day. He summarised what he wants from regulators as an equilibrium between open-weight and frontier models, with frontier models keeping enough of a price premium to sustain their business. CNBC interview.

Source: TechCrunch.