US AI czar accused a Chinese lab of distilling a model that isn't a month old yet

A three-week-old model, accused of raising a copy in three weeks

Michael Kratsios, the US government's Chief Technology Officer, posted that his office has "information that Moonshot distilled Fable for the development of K3" — naming Anthropic's Fable model as the source and Moonshot's newly released Kimi K3 as the copy. No evidence was attached to the post; it read as an accusation delivered from an official government account, aimed at a specific company, with a specific technical claim.

The thread on Hacker News immediately did what the tweet didn't: checked the calendar. Access restrictions on Fable were lifted around July 1; Kimi K3 shipped July 16. That leaves roughly two weeks for Moonshot to allegedly extract enough signal from Fable to bootstrap a competing frontier model, run the standard reinforcement-learning passes, and still ship benchmarks on schedule — timing multiple commenters called implausible for genuine distillation from scratch.

The moderator stepped in before the facts did

The first reply followed a familiar script — "rules for thee but not for me," pointing at the AI labs' own well-documented practice of training on scraped web and book data without permission. Hacker News's own moderator cut in on the thread itself, not to referee the accusation but to flag the reply as exactly the kind of reflexive one-liner the site's guidelines discourage: "it has also become an entirely cliché response which pops up like mushrooms anytime 'distillation' appears... what we really want are reflective responses." It's a telling moment — the moderation happened on the meta-level of the conversation, while the actual technical question (did this happen, and does it matter) went mostly unanswered by anyone with direct knowledge.

That technical question is where the thread did its real work. An engineer with a security background laid out what distillation actually costs: "Distillation is a superficial step and doesn't need a lot of data, it's not 'stealing the model' like they want everyone to believe. 99% of work is already done by that point." Their read: K3 likely does carry Claude-family fingerprints — it reportedly repeats Anthropic's own prompt-injection test strings — but that's evidence of using the model as a late-stage teacher, not of wholesale theft.

A skeptical voice went further, calling the accusation political theater: "This is BS to pressure politicians," they wrote, citing an OpenAI staffer who'd separately dismissed the idea that a model like K3 could be trained from distillation alone, and adding that researchers they know describe distillation as "either useless or merely useful in post training to show 'correct' behavior" — not a shortcut to a competing frontier model.

Copyright law doesn't have a clean answer either

A second front opened over whether "distillation" is even theft in any legal sense. If Moonshot paid for API access to Fable and used the outputs — as one commenter noted, "assuming they did then they surely paid for them, which makes it 'not stealing'" — the accusation collapses into a licensing dispute dressed up as an act of national interest. Another pointed out that Anthropic itself has no shortage of unauthorized-training controversies of its own, undercutting the moral high ground Kratsios' post implied.

What's missing from the entire thread, notably, is anyone from Moonshot, Anthropic, or the CTO's office showing up with actual evidence. The accusation stands exactly as posted: an official claim, a three-week compression that strains credulity on its face, and a Hacker News crowd more willing to interrogate the timeline than either side offered to.