Anthropic has accused three Chinese AI companies of creating over 24,000 fake accounts to enhance their own AI models using Anthropic’s Claude AI model. DeepSeek, Moonshot AI, and MiniMax allegedly engaged in ‘distillation’ by generating millions of exchanges with Claude to target its unique capabilities such as agentic reasoning, tool use, and coding.
These allegations surface as the US government deliberates export controls on AI chips to slow down China’s AI advancements. Distillation, a common training technique, allows labs to create smaller models at lower costs but can also be misused for replicating competitors’ work. DeepSeek, for instance, has been accused of mimicking OpenAI’s products using distillation.
DeepSeek notably gained attention last year with its R1 reasoning model, offering competitive performance at reduced costs. The upcoming DeepSeek V4 model is expected to surpass Claude and ChatGPT from Anthropic and OpenAI in coding tasks.
The alleged attacks varied in focus, with DeepSeek targeting foundational logic improvement, Moonshot AI concentrating on agentic reasoning, tool use, and coding, and MiniMax engaging in exchanges related to computer vision and development. Moonshot AI recently introduced the Kimi K2.5 model and a coding agent.
Source: TechCrunch