Technology, AI & Cyber
White House official accuses Moonshot AI of copying Anthropic model
A top White House official accused Chinese AI startup Moonshot AI of covertly copying Anthropic's most advanced model to build its Kimi K3 system.
Michael Kratsios, director of the White House Office of Science and Technology Policy, said in a post on X that the United States had information that Moonshot AI used distillation of Anthropic's Fable model to develop Kimi K3.
Distillation is a technique in which a smaller AI model is trained to mimic the outputs of a larger and more capable one. The method is widely used in the AI industry, but US officials say Chinese companies have used it at scale to illicitly copy proprietary American systems.
Kratsios said Moonshot built a sophisticated internal platform to carry out large-scale distillation against US models. He said the system allowed the company to switch quickly between several access methods to avoid detection.
He also said Moonshot had acquired servers equipped with Nvidia's advanced GB300 chips and had accessed GB300-equipped servers in Thailand, likely to train its AI models. The source report said this implied a possible workaround of US export controls that restrict Chinese access to advanced American semiconductors.
Kratsios did not say what consequences Moonshot may face.
On Tuesday, Treasury Secretary Scott Bessent said foreign companies could be sanctioned for stealing US technology. He said US authorities had the ability to sanction overseas models if they were stealing from US companies.
Bessent referred specifically to distillation and said US authorities were finding watermarks of US large language models on many Chinese models. He said authorities would be looking at the issue in the coming days or weeks.
Uncertainty notes
The accusations against Moonshot AI are presented as statements by US officials; the source does not include a response from Moonshot AI.
Kratsios did not indicate what consequences Moonshot AI may face.
Kratsios described the alleged use of GB300-equipped servers as likely for training AI models.
Source
AFP news report published on .