Anthropic is accusing three Chinese companies—Deepseek, Moonshot, and Minimax—of a "coordinated campaign" to steal their AI capabilities through model distillation. 

The process of model distillation involves a student model learning the chain of thought from a teacher model, and is usually used to train a company's slower models with it's faster ones (for example, this is how Anthropic trained Sonnet using it's Opus model).

But in this case, overseas AI companies allegedly used a competitors model to train their own. 

The entire timeline with a lot more details are in my full YouTube video, linked here.

Matt Wolfe

Anthropic published a blog post accusing three Chinese AI companies — DeepSeek, Moonshot, and Minimax — of running a coordinated campaign to steal model capabilities through knowledge distillation. The technique involves querying a target model millions of times to capture both its final responses and chain-of-thought reasoning, effectively training a new 'student' model on the 'teacher' model's outputs without permission. The piece highlights the irony that Anthropic itself trains on internet data scraped without explicit consent, raising questions about the credibility of their complaint.

Chinese AI Companies Are Using This Trick To Steal Model Data