
White House Office of Science and Technology Policy director Michael Kratsios accused Moonshot AI of using Anthropic's Fable model to train its Kimi K3 model. He also claims Moonshot built an internal platform to bulk-call US models and evade detection. No public evidence was provided, and Moonshot hasn't responded.
Then Teknium, co-founder of Nous Research (maker of the Hermes Agent), copied the exact phrasing. He said Anthropic scraped his GitHub, X, Reddit, Discord, and Facebook posts to train Fable. Then added: "They're basically distilling every human on Earth."
It's obvious sarcasm. Web scraping and model distillation aren't the same thing. Anthropic has publicly acknowledged that its training data includes public internet information. Teknium's point: when US models train on everything, it's called "training"; when Chinese models learn from US models, it's called "theft."
https://twitter.com/Teknium/status/2079966815913161091