According to US Treasury Secretary Bennett, China’s AI model has stolen the technology of the American AI model, resulting in an embarrassing situation for the Chinese Communist Party: China’s most powerful AI model, Kimi, sometimes identifies itself as the American AI model Claude. What’s even more surprising is that it handed over the weapons plans of the Chinese military to the developer of Claude, inadvertently helping the United States.
In an episode of “Axios Show” aired on Saturday, Bennett discussed the AI competition between the US and China. He expressed one of his biggest concerns in the field of AI is the risk of the US being surpassed by China.
“We cannot afford to lose our current lead to China (the CCP),” Bennett said. “One of my responsibilities is to handle US-China relations, and I can tell you, if they lead in AI instead of us, the situation would be very different.”
When discussing the development of AI models, Bennett mentioned that China’s open-source models undergo what he refers to as “industrial distillation,” which in reality is a euphemism for “stealing American models.”
Bennett then singled out the Chinese AI company “Moonshot” and its Kimi model, highlighting a major flaw. He stated, “Kimi believes it is (developed by the American company Anthropic) Claude.”
“Sometimes Kimi will directly tell you that it is Claude,” Bennett added.
There was even more astounding revelation: Kimi leaked military secrets of the Chinese Communist Party to the United States.
“This has already been made public, so I can discuss it. Kimi leaked some of the weapons plans of the People’s Liberation Army (PLA) to Anthropic – which was quite a ‘big help’,” Bennett said.
In a report released by the American AI company Anthropic on September 10, multiple instances were cited where in the process of distilling foreign AI, Chinese companies inadvertently exposed internal secrets of the CCP or Chinese enterprises. For example, a Chinese military user employed the Kimi model for analyzing surveillance data. The user loaded surveillance data from hundreds of cameras in Chengdu, including those outside Chinese military facilities. Moonshot, through an intermediary network, transmitted this data to Anthropic’s Claude model for analysis.
Ultimately, this incident underscores the complex dynamics in the competitive landscape of AI development between the US and China, raising concerns about the implications of intellectual property theft and espionage in the technological realm.