Anthropic accuses Alibaba of distilling Claude, days before its own models get curbed
Anthropic told Washington that Alibaba was siphoning Claude's capabilities. Two days after the letter went out, Commerce restricted Anthropic's own newest models.
The complaint names a specific method: training a weaker model on a stronger one's outputs, the technique Anthropic calls "distillation." The thread's twist is that almost nobody buys the framing as a copyright grievance.
A self-described insider says the labs distill via AWS Bedrock, not Anthropic's restricted platform, so the company may not even hold the evidence. A UK student lays out the real pipe: Chinese resellers hawking Opus tokens at a 93% discount and selling the reasoning traces as training data.
A letter to Washington, then a curb on Anthropic itself
According to Reuters, Anthropic has accused Alibaba of illicitly extracting Claude's capabilities through what it calls a "distillation" effort: training a less capable model on the outputs of a stronger one. The company framed it as illicit and fraudulent and pressed the case with US officials.
The timeline is the part that stings. One commenter quotes the article directly: on June 12, two days after Anthropic sent its letter, the Commerce Department imposed restrictions on Anthropic's own newest Mythos and Fable models, fearing military-intelligence use abroad. The export-control campaign Anthropic helped stoke landed on Anthropic.
The thread reads it as protectionism, not theft
Almost nobody in the thread treats this as a copyright wrong. A commenter who flags himself as partly on the inside argues the data may not even exist where Anthropic implies.
"They do distill the models, but they don't go to Anthropic, they just use platforms like aws bedrock, there are too many restrictions on Anthropic's own platform." — a self-described partial insider
Another walks through the mechanics on the ground in China, where Claude is blocked and bought through grey-market resellers.
"Chinese resellers are offering Claude tokens at 70-90% below official Anthropic API prices… reselling the model output & reasoning chains to various Chinese labs." — a UK student tracking the token-resale economy
One reply with a security background splits the technique itself, separating crude black-box copying from the targeted kind that actually moves the needle.
"The latter is basically fine-tuning the model with direction from another model. Thousands of businesses do this every day… These complaints of distillation are inflating the problem." — a commenter walking through RLAIF distillation
And the irony writes itself for the crowd that remembers how the training corpus was assembled.
"So said the guys who 'extracted' knowledge from all pirated books." — a longtime HN poster
If the insider is right that the evidence lives on Bedrock, the discovery phase of any suit cuts both ways: Anthropic would have to open its own training sources to make the theft case stick.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
Slate's $24,950 EV truck has no paint, no screen, and 205 miles of range
The base 'Blank Slate' is a two-seat pickup you wrap yourself and bolt into a five-seat SUV. Then a commenter priced a real build and hit $35k.