Anthropic Says Chinese AI Companies Improved Models By ‘Illicitly’ Copying Its Capabilities

Did you know that there is a way to use the output from LLM that may not involve any hacking – essentially just taking large amounts of text and reusing it as training data – which is what AI companies are so nervous about?

In a blog post on Monday, Anthropic said China-based AI companies DeepSeek, Moonshot and MiniMax broke Anthropic’s rules to “illegally extract” the capabilities of its signature AI models, the Cloud.

Distillation is a common practice used by AI companies in which a “teacher” model is stimulated with specially crafted inputs, and the answers provided allow the “student” model to rapidly improve. For example, Anthropic writes, “Frontier AI laboratories routinely distill their own models to create smaller, cheaper versions for their clients.” So to distinguish the actions Anthropic is complaining about from the use of distillation that is considered legitimate, these actions are known as “distillation attacks”.

Are distillation attacks a criminal offense in the eyes of Anthropic? Nothing of the sort is alleged here, but Anthropic says these actions were taken “in violation of our terms of service and regional access restrictions.”

Anthropic, which itself is dealing with the threat of being labeled a “supply chain risk” by the Pentagon, made patriotic comments in the post. Circumventing regional use restrictions and circumventing regulations “allows foreign laboratories, including those under the control of the Chinese Communist Party, to foreclose the competitive advantage that export controls are designed to protect through other means,” it claims.

Of the three China-based companies mentioned, Shanghai-based Minimax, maker of the viral character chat app Talky, most angered Anthropic by the scale of its distillation effort: more than 13 million reported exchanges. This compares to Moonshot, with more than 3.4 million, and the most well-known company named in the post, DeepSeek, which has an estimated number of only 150,000.

OpenAI, Anthropic’s main competitor, is also paranoid about the distillation of at least one Chinese AI company, sending a memo to the House of Representatives earlier this month accusing DeepSeek of “ongoing efforts to free ride on capabilities developed by OpenAI and other US frontier labs.”

DeepSeek is expected to release its latest flagship model, DeepSeek V4, any day now, and CNBC has warned that this release could cause chaos on Wall Street, at a time when there is already enough AI-related chaos on Wall Street.