DistantNews
Support us
๐Ÿ‡น๐Ÿ‡ผ Taiwan /Technology

China's AI model Kimi K3 claims parity with US rivals, but admits to being trained on competitor's data

From Liberty Times · () Chinese

Translated from Chinese, summarized and contextualized by DistantNews.

At a glance

News Sources not specified Context piece
  • China's Moonshot AI released its new Kimi K3 model, which reportedly narrows the gap with US competitors.
  • The model was caught identifying itself as Claude, an AI assistant from US company Anthropic, suggesting it was trained using "distillation" from US models.
  • This practice, known as "knowledge distillation," involves using output from larger, more powerful models to train smaller ones at a lower cost, a method previously accused by the US of being used by Chinese entities to steal AI technology.

China's AI startup Moonshot AI has released its latest model, Kimi K3, which is claimed to be closing the gap with well-funded US competitors. However, the model has been observed identifying itself as "Claude," an AI assistant developed by US company Anthropic, raising questions about its training methods.

This self-identification suggests that Kimi K3 may have been trained using "knowledge distillation," a technique where output from advanced AI models like Anthropic's Claude is used to train smaller models. The US has previously accused Chinese entities of employing this method, sometimes referred to as "industrial-grade distillation," to steal AI technology by using vast numbers of accounts to access and extract data from leading US AI models.

If Kimi thinks it is Claude, then it was trained using Claude. Without Claude's progress, there is no Kimi's progress. If Kimi violates the TOS (Terms of Service) and abuses Claude for training, it will abuse your data.

โ€” ็ธฝ่ฃ็ฐกๅ ฑCEOA social media post by '็ธฝ่ฃ็ฐกๅ ฑCEO' commenting on Kimi K3's self-identification as Claude.

OpenAI has also accused Chinese AI startups of using its proprietary models for training. Anthropic itself has formally accused Alibaba of a "blatant and illegal" AI technology theft campaign, involving millions of illicit clicks to "distill" its technology. The Kimi K3 incident appears to be another instance of this practice.

While Kimi K3's claimed technical advancements might narrow the gap with US AI, its self-identification as Claude raises concerns about data compliance and potential misuse of competitor models. Experts suggest that Chinese large models are currently acting more like "high-end contract manufacturing" and "engineering optimization" for AI, replicating proven US paths at lower costs. However, the core operating systems and ecosystems, such as OpenAI's ChatGPT subscription model or Microsoft's Copilot integration, remain largely controlled by US companies, limiting the independent market potential of Chinese models.

The current work of Chinese large models is more like 'high-end contract manufacturing' and 'engineering optimization' in the AI field, using lower computing power costs to re-do the proven paths of the United States.

โ€” ็ธฝ่ฃ็ฐกๅ ฑCEOA social media post by '็ธฝ่ฃ็ฐกๅ ฑCEO' describing the current state of Chinese AI development.
DistantNews Editorial

Originally published by Liberty Times in Chinese. Translated, summarized, and contextualized by our editorial team with added local perspective. Read our editorial standards.

More Stories

From Our Blog