❗️The U.S

❗️The U.S. just accused Moonshot AI of secretly using Anthropic’s Fable model to build Kimi K3. Michael Kratsios, Director of Technology at the U.S. Department of Commerce and Deputy Assistant to the President for Science and Technology, says the U.S. has information indicating that Moonshot AI carried out a large-scale distillation campaign using Anthropic’s flagship model. According to the statement, Moonshot allegedly built an internal platform specifically designed to distill American AI models while constantly switching access methods to avoid detection. The company is also accused of acquiring GB300 AI servers and accessing additional GB300 compute resources in Thailand to train its models. The U.S. emphasized that AI distillation itself is not the problem. Distilling models to make them smaller, faster, and cheaper is a common and legitimate technique across the industry. The issue, they say, is industrial-scale, covert distillation aimed at copying proprietary frontier models without permission. Here’s where it gets even more interesting. Researchers have spent years developing statistical “fingerprints” that can be embedded in language model outputs. They’re invisible to users and don’t affect quality, but they can potentially reveal which model generated the text used for training. If independent testing of Kimi K3 were to uncover clear evidence that it was trained this way, it could become one of the biggest AI intellectual property cases we’ve seen so far. @aipost 🏴