Wednesday belonged to one man in Washington. Michael Kratsios, the White House's science adviser, spent it doing two things at once: accusing China of stealing American AI, and announcing more than $5 billion to push American AI deeper into science. Taken together, they read less like two stories than one doctrine.

Washington accuses Moonshot of distilling Anthropic

In a post on X, Kratsios said the United States has "information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model," describing "a sophisticated internal platform to conduct large scale distillation against U.S. models" designed to rotate between access routes "to avoid detection." He also alleged Moonshot reached banned Nvidia GB300 chips through servers in Thailand.

The accusation is pointed, but it is a conclusion, not a case: Kratsios presented no evidence that any Fable output actually ended up inside Kimi K3, and researchers question whether the timeline even allows it. Distillation is also a normal, legal technique in general — Kratsios said as much himself. China's embassy called the claim unfounded. We pulled apart what was and wasn't shown in a full piece.

The same adviser, $5 billion for AI science

Hours from the same office came the Genesis Mission: more than $5 billion in federal commitments to aim AI at scientific discovery, across more than 15 agencies. The domains are strikingly broad — drug discovery, pediatric cancer, and veterans' health alongside energy grids, materials science, weapons design, and biological threat detection. It's a bet on the layer above model-building: that the frontier is now good enough to be a scientific instrument. The standalone has the full domain list and the caveats — it's a commitment, not yet a result.

The two announcements share a through-line. America will spend heavily to wire AI into its most important research, and will treat unauthorized access to the models that make that possible as a national-security problem rather than a commercial dispute.

Anthropic goes to the regulator's sandbox

In the UK, the Financial Conduct Authority named the second cohort of its "Supercharged Sandbox," and Anthropic is providing the tooling — Claude, Claude Code, and Claude Cowork — to 21 participating organisations, among them Scottish Widows, the Money Advice Trust, and TrueLayer. The FCA said it received 199 applications this round, up 51% on the first. It's a quieter counterpoint to the distillation fight: a frontier lab embedded inside a regulator's testing environment, by invitation.

Bloomsbury's cut of the Anthropic settlement

A footnote to Monday's final approval of Anthropic's $1.5 billion copyright settlement: Bloomsbury — the Harry Potter publisher — has 14,087 titles in the class, worth roughly $3,000 each, or about $42.3 million before legal fees, to be split evenly between the publisher and its authors. It's a concrete reminder that "$1.5 billion" resolves into real cheques to real rights-holders.

Also worth a look

A project called GigaToken made the rounds for claiming "~1000x faster" tokenization. The honest version: it's about 1,000x faster than Hugging Face's slow Python tokenizers, and a more useful ~140–680x against the practical baseline, OpenAI's tiktoken — CPU-only, Rust, BPE-focused, and still pre-release. Not magic, but a real speedup on a step that quietly taxes every training and inference run.