Ashley Capoot, reporting for CNBC: U.S. Treasury Secretary Scott Bessent on Tuesday said the Trump administration will look into whether Chinese artificial intelligence models have been distilled from American models, stating that the government does not support “IP theft.”
Chinese open-weight models are gaining steam against leading offerings from American companies like OpenAI and Anthropic, sparking concerns from tech executives and government officials about the durability of the U.S. lead in the AI race. Moonshot AI, a Chinese startup, released a model called Kimi K3 earlier this month that outperforms those companies across some industry benchmarks. Open weight refers to models whose final trained parameters are publicly released for download, while the underlying code and training data remain private.
“If we see, especially that overseas models are stealing from our great companies, we have the ability to sanction them because of this theft,” Bessent told Fox Business’ “Mornings with Maria” on Tuesday.
Bessent said the technical term for this theft is called distillation, which is an AI training method where a smaller, less capable model is built using outputs from an existing, stronger model. Anthropic sent a letter to the U.S. Senate Committee on Banking, Housing, and Urban Affairs last month alleging that the Chinese tech company Alibaba had carried out the “the largest known distillation attack” against it to date.
If distillation — training on data that doesn’t belong to the Chinese labs — is a sanctionable offense, then why isn’t pre-training? How is it fully legal for Chinese companies to use training data published by American citizens on the internet? This whole argument is bogus, and it’s quite evident that Bessent knows very little, if anything, about what he’s talking about. AI models are built on the idea that training a computer on the open internet is fair use — that pre-training is transformative enough under fair use doctrine, not “stealing.” American companies can train their models on Chinese text and vice versa because pre-training is inherently transformative, an argument I’m quite sympathetic to. I think that if we built a time machine back to 2022, a permission or compensation structure should’ve been created, but alas, this is the system we’re stuck with. The world’s copyright system is not suited for AI pre-training. Back to the Bessent argument. Anthropic, and perhaps OpenAI, are perturbed that Chinese models — like Moonshot’s Kimi K3 open-weight large language model — were trained via distillation. Kimi K3 is cheaper than Claude Opus 4.8 and GPT-5.6 Sol, yet almost matches GPT-5.6 Sol in overall intelligence and beats Opus 4.8 entirely. Kimi K3 comes just shy of Claude Fable 5, a model over three times the price. While Anthropic can’t even be compelled to offer Fable 5 access to some of its paying customers, Kimi K3 offers similar performance at a fraction of the cost — but it was built by distilling Anthropic’s models. Anthropic’s view, which it has surely expressed to the government, is that this is an adversarial nation subsidizing AI development, stealing the secrets of American AI labs, and building a compelling product for much cheaper. From that perspective, Moonshot and China look guilty.
But that’s a pacifist view; Anthropic is wailing for pity from a government that has treated it abysmally. Just Tuesday, a judge approved a billion-dollar settlement between Anthropic and a litany of authors after it pirated books to train its models on — is that not “stealing?” Is Anthropic not one of the most blatant copyright infringers of the modern era? It’s quite obvious here that Anthropic is crying wolf not because China gained some unfair advantage in its training process, but because Anthropic is so upset that it wasted perhaps billions of dollars in subsidized tokens all for its users to prefer a foreign competitor’s clearly superior model. The problem here was never about morality or copyright or distillation; the AI labs are morally bankrupt, and they know that. If they didn’t, they wouldn’t be cutting billion-dollar checks to authors.
So where do we go from here? I think the solution is quite obvious: Anthropic must start distilling its own models. Train a massively expensive competitor to Fable 5, distill a new version of Opus on that model’s reasoning traces, and ship the new Opus at a much lower cost. We’re reaching a point in the AI industry where cost is a massive factor in the models enterprises and individuals use, as reasoning traces become more expensive. Yes, pre-training is currently the most expensive part of releasing a new model, but we’re already seeing the shift toward inference and post-training. New model releases nowadays are not new pre-training runs; GPT-5, the last “major” AI model from OpenAI, was a heavily post-trained GPT-4o from spring 2024. As LLMs mature and reasoning tokens become more valuable, inference will be the biggest operating cost, not pre-training. Train one model and distill three others, then offer inference at a competitive price. Distillation — from China and by the American labs themselves — is in the AI industry’s best interests. It saves on pre-training costs and leaves more room to make healthy margins on inference.