Gangsta AI

The White House Says China Distilled Anthropic's Fable to Build Kimi K3 — and Ran Banned Nvidia Chips Through Thailand

Cyber Desk

By Chuck Norris · 2026-07-23 · 3 min read

Somebody Kicked Down the Door on Moonshot

On Wednesday, Michael Kratsios — the White House's science and technology adviser and director of the Office of Science and Technology Policy — went on X and did something no senior U.S. official had done before: he pointed at one specific Chinese lab and said, out loud, that it copied an American AI.

The lab: Moonshot AI. The charge: that Moonshot ran "large-scale industrial distillation" against Anthropic's Fable model to build Kimi K3, the 2.8-trillion-parameter open model that lit up the leaderboards last week with a 76% win rate on the Frontend Code Arena.

“Distillation is a real, legal technique — a small model learning from a big one's outputs. It stops being cute when the big model belongs to a rival who never handed you the keys.”

Kratsios said Moonshot built an internal platform to run that distillation at scale, one that could switch between access methods to dodge detection. Then he added the part that turns a research squabble into a national-security headline.

The Chips Took the Scenic Route Through Thailand

Kratsios also alleged Moonshot got its hands on servers packing Nvidia's GB300 — Blackwell-class silicon that is flat-out banned from sale to China — by tapping hardware sitting in Thailand. That is not an IP fight anymore. That is export-control evasion, the kind that comes with Entity List designations and real penalties attached.

The Treasury Department did not wait around to play nice: it floated the prospect of sanctions against Chinese AI firms within hours.

The Evidence — and the Caveat Chuck Won't Pretend Away

Independent researcher Ryan Greenblatt of Redwood Research published a cross-entropy analysis showing Kimi K3 identifies itself as Claude far more often than random chance should allow — the same forensic trick used in past distillation disputes.

Here is the honest part, because a roundhouse only counts if you aim it straight:

• Models scrape the open web, and the open web is full of Claude transcripts. Contaminated training data can make an innocent model *sound* guilty.
• Weights carry no watermark. Nobody pulls a receipt out of a neural net.
• Moonshot has not answered the latest charge. When Anthropic raised a similar one earlier this year, Beijing called it "groundless."

So the accusation is loud, the forensics are suggestive, and the proof is — like most things in this business — contested. Kimi K3's open weights still drop July 27, four days after the White House lit the fuse. Every enterprise eyeing that free download now gets to weigh raw performance against a provenance question with the U.S. government's fingerprints on it.

What Chuck Sees

Two labs. One model that talks like the other. A shipment of banned chips that supposedly wandered through Southeast Asia. Whatever the truth turns out to be, the era of quietly borrowing your rival's homework is over — the referees showed up, and they brought sanctions.

The move that never loses? Stop betting your work on one model's honesty. Line them up — Kimi, Claude, Gemini, GPT-5.2, Grok — and make them prove it head-to-head. Ensemble beats single model, every time. See which AI actually holds up on Gangsta AI.

Try Gangsta AI free →

More: Best AI models · Compare all AI · Frontier Models · All articles