Good morning. The Kimi/Qwen shockwave from earlier this week is now rippling outward in three directions at once: a Washington policy meltdown, a serious business-model question for Anthropic, and — in the “wait, really?” column — a claimed Jacobian Conjecture counterexample from Claude Fable that people are actually starting to verify. Also, Anthropic’s copyright settlement got signed, sealed, and delivered.
Washington’s AI apparatus is unraveling in public. MIT Technology Review reports that Moonshot’s Kimi release has cracked open a public feud among Trump’s current and former AI advisors, with David Sacks arguing against protecting US incumbents while others push for national-security-based vetting of models. On cue, TechCrunch notes that Chris Fall — the third CAISI head in a few months — resigned after twelve weeks, following Sacks in March and Collin Burns’ week-long tenure in April. CAISI was also pointedly left out of the White House’s new “Gold Eagle” safety program, which suggests the standards body is being quietly written out of the story.
Is Anthropic actually cooked? An Emerging Trajectories piece argues Kimi K3 and Qwen 3.8 hitting Fable-5-adjacent performance is an existential problem specifically for Anthropic, which lacks the owned data centers and power generation that Meta and Alibaba are riding. The HN thread mostly pushes back: OpenAI has the same “model-only” exposure, users happily pay for marginal quality gains, and one commenter dryly noted that “one is heading to IPO and the other isn’t showing their books” tells you which lab is really under pressure. A separate Werd.io post makes the broader case that China is winning by commoditizing the model layer, though HN commenters were unconvinced by the claim that 80% of startups already run Chinese models — Llama, they note, hasn’t made Meta rich either.
Claude Fable may have broken the Jacobian Conjecture. Alex Harrison posted that Claude Fable 5, under his direction, produced a polynomial map from ℂ³→ℂ³ that appears to be a counterexample to the 1939 Jacobian Conjecture — which if verified would also sink the Dixmier and Poisson conjectures. Harrison later pushed a verification script after HN pressed him for one. The counterexample is compact enough that commenters are speculating similar results have been sitting in reach for years, unfound simply because no human had bandwidth to search the family exhaustively. Harrison works at Anthropic, so the model version involved may not match what’s publicly available.
Cursor’s agent swarms actually built SQLite. Cursor published results from a hierarchical swarm where planner agents on frontier models delegate to cheaper workers — hitting 80% on the SQLite test suite in four hours, versus their old swarm failing before hour two. The impressive infrastructure detail: they built a custom VCS handling 1,000 commits per second because Git couldn’t keep up. HN commenters reasonably asked whether rewriting SQLite in Rust proves much given Turso’s Rust rewrite is almost certainly in the training data.
Anthropic’s $1.5B settlement gets final approval. TechCrunch confirms the court has signed off on Anthropic’s copyright deal, with authors receiving $3,000 per work across roughly 500,000 titles. The case turned on Anthropic’s use of pirated sources like LibGen — Judge Alsup separately ruled that training on legitimately acquired copyrighted text is fair use, which is the part that matters industry-wide. No binding precedent though, so the parallel cases against Google, Meta, OpenAI, and Midjourney remain open.
Sony names 30,000 songs in Udio suit. The Verge has the list: Elvis, Beyoncé, Harry Styles, all identified via audio fingerprinting after Sony got discovery access to Udio’s training data in an earlier suit. Sony is seeking up to $150,000 per work, and is now the lone holdout — Universal and Warner already settled and moved to partnership deals with Udio, which tells you where the industry expects this to end.
OpenAI publishes on long-horizon safety. OpenAI put out a post on safety and alignment for long-horizon models, though the article body wasn’t accessible when we checked. Worth bookmarking if you’re tracking how the labs are framing agent-era risks.
That’s the morning. The China narrative isn’t going anywhere, but keep an eye on the Jacobian claim — if it holds up, it’ll be the first time an LLM authored result actually gets absorbed into the mathematical literature.