Yesterday, Anthropic alleged that it had detected what it described as “an industrial scale campaign” by DeepSeek and two other prominent Chinese AI labs, Moonshot AI and MiniMax, to distill its Claude models. Distillation is the term AI researchers use to describe a method of boosting the performance of smaller, usually weaker AI models by fine-tuning them on the outputs of a larger, stronger model. In this case, Anthropic claims the three Chinese AI companies created 24,000 fake accounts in order to generate 16 million exchanges with Claude that they then used to train their own models, in violation of Anthropic’s terms of service. (Of these exchanges, DeepSeek was only responsible for 150,000 of them, according to Anthropic, but DeepSeek-linked accounts seemed particularly interested in distilling Claude’s reasoning capabilities.)
Also yesterday, Reuters reported, citing an anonymous senior U.S. government official, that the U.S. believes DeepSeek trained V4 using Nvidia’s latest generation Blackwell AI GPUs, in likely violation of U.S. export controls that were supposed to prevent Chinese AI companies from acquiring Nvidia’s most advanced chips. The story said the U.S. believed that DeepSeek has a data center in Inner Mongolia stuffed full of Blackwells–although it said the U.S. was unsure exactly how it obtained them.
