4 min read
The White House has accused China of stealing a hundred-billion-dollar company’s crown-jewel AI model. The alleged method of the heist? Talking to it. Depending on who you ask, this is either the boldest industrial espionage of the decade or the sound of America discovering that the internet works both ways.
Here is what was actually claimed. Michael Kratsios, the White House science and technology adviser, says the Chinese lab Moonshot AI built its enormous new model, Kimi K3, by “distilling” Anthropic’s Fable. He alleges Moonshot ran a “sophisticated internal platform to conduct large-scale distillation against US models,” switching between methods of access to dodge detection. The Treasury has started rattling the sanctions sabre. For good measure, officials also accuse Moonshot of getting its hands on export-banned Nvidia chips.
What distillation actually is
Strip out the spy-thriller language and distillation is mundane. You take a strong model, feed it a mountain of prompts, harvest its answers, and use those answers to train a smaller or cheaper model to imitate it. Done openly and at small scale, it is legal, common, and taught in every machine-learning course going. The White House’s objection is to doing it covertly, at industrial scale, against a rival nation’s flagship. That is a fair thing to object to. It is also monstrously hard to prove.
The timeline the whole case hangs on
Here is the thread the accusation dangles from. Anthropic’s Fable 5 was pulled offline under US export rules, then returned to public access on 1 July. Moonshot’s Kimi K3, all 2.8 trillion parameters of it, launched on 16 July. Fifteen days. The White House says that gap is “consistent with” industrial-scale distillation. Which is a lovely phrase that also happens to be consistent with coincidence, with a model that was mostly finished already, and with a dozen other explanations.
The experts are not buying it
This is where the story wobbles. Researchers who actually build these things told TechCrunch that you do not conjure a frontier model as capable as Kimi K3 by hoovering fifteen days of outputs out of a model that only went public a fortnight earlier. Distillation can teach a model to mimic style and shortcuts. It does not hand over the years of training, data and architecture that make a flagship genuinely good. The accusation, they suggest, mistakes a smoking gun for a warm kettle.
The irony nobody in Washington wants to say out loud
There is a deep and delicious hypocrisy humming under all of this. The American labs built their models by scraping essentially the entire internet, every book, blog, forum post and photo they could reach, and when anyone objected they waved the magic words “fair use”. Now that a competitor has learned from their outputs, the same industry is aghast at the theft of intellectual property. You cannot spend three years insisting that training on other people’s work is transformative and legal, and then clutch your pearls when someone trains on yours. Pick a lane.
None of which means Moonshot is clean. Covert, large-scale distillation to dodge export controls would be a real problem, and the Nvidia-chip allegation is serious on its own. But “we think they did it because the dates line up” is an accusation, not evidence, and the people who understand the technology best are the ones least convinced. For now, file it under geopolitics first, forensics a distant second.
The AI tool stack actually worth paying for
One email a week. The tools, models and moves that matter, minus the hype and with the horseshit filter set to maximum. Free.
Did you know: in machine learning, distillation is a formal, published technique, usually called knowledge distillation, where a small “student” model is trained to copy the outputs of a big “teacher” model. It is how a lot of the fast, cheap models on your phone are made. The controversy is never the method itself, only who is doing it to whom, and whether they asked first.