"The AI Pulse — 2026-10-28: The Enormous Theorem, Machine-Checked in 24 Days"

Somewhere in the tens of thousands of pages that make up the Classification of Finite Simple Groups, there was a flaw nobody had caught — a gap in a 1983 proof by Jonathan Walter that had been riding along in the "largest proof in mathematics" for forty years. It took a machine-checked re-derivation of the whole edifice to find it. That's the quiet headline of a day on which the systems doing mathematics and the systems verifying it finished collapsing into the same pipeline.

The Enormous Theorem, compressed into a month

Axiom Math has released a complete Lean formalization of the Classification of Finite Simple Groups: 43.1 million lines, 2.05 million declarations, built from 715 source papers and books, with every logical step kernel-checked. The original took hundreds of mathematicians decades and was long considered too vast to ever fully formalize; Axiom's human-directed agent swarms did it in 24 days — and along the way surfaced that counterexample to Walter's published proof route. The obvious read is "AI does math faster"; the more interesting one is structural. When verification is free, "too big to check" stops being a permanent property of a theorem and becomes a temporary one, and the residual value shifts to whoever directs the agents at the right questions. Independent verification of the artifact is still pending, which is itself telling: we now live in a world where a 43-million-line machine-checked proof needs the same "trust but verify" caveat as a press release. cfsg.axiommath.ai

Your "1M-token context" was a pool of 681K — and it leaked

Cloudflare disclosed that GLM-5.3 on Workers AI was advertised with a million-token context window while the underlying SGLang KV pool held only 681,280 entries; under load, cache keys evicted and got reused, and 155 responses between August 27 and September 9 served one customer's prompt data to another. Every disaggregated SGLang deployment is affected, not just Cloudflare's. Strip the particulars and you get a fleet-wide pattern the industry hasn't priced in: long-context serving stacks silently convert resource pressure into cross-tenant data exposure, and the failure mode looks like a slightly-wrong answer rather than an alarm. As context windows keep growing faster than memory budgets, "your context window" is increasingly a marketing number sitting on top of a shared pool — a lease, not a guarantee. Cloudflare's disclosure

NVIDIA open-sources the data engine, and robot-learning economics invert

NVIDIA's Flora — the GR00T-2 robotics data engine, now open-source — auto-captions, segments, and reward-labels robot video at 57.1% annotation quality for roughly $0.72 per video-hour, against GPT-6 Astra's 53.7% at $88.49. That's a ~125× cost collapse on the single least glamorous bottleneck in robotics, battle-tested at 200K hours per day, and it's free. The second-order effect cuts both ways: the constraint on robot foundation models has quietly moved from "can we get data" to "can we fund compute and get robots into real environments," which favors exactly the well-funded players who published the tool. Generosity that commoditizes your complement is an old playbook; this is the robotics edition. research.nvidia.com/labs/gear/flora

The leaderboard operator concedes its own game

Artificial Analysis launched Intelligence Index v5, its biggest overhaul since v4, and the headline change is that a majority of the composite is now built on private datasets, explicitly to counter benchmark contamination. When the industry's most-cited independent leaderboard publicly gives up on public benchmarks, that's not a product update — it's the evaluation layer admitting the arms race is over and the public numbers lost. It lands one day after HKUST showed a mere JSON field reordering crashes a trained agent's success from 85% to 19%, and it completes a week in which harness sensitivity, standard-harness reruns, and buyer indifference all pointed the same way. Private evals fix the gaming problem by recreating the trust problem: the numbers get honest by getting unfalsifiable. artificialanalysis.ai

From "AI helps science" to molecules inside pharma pipelines

Anthropic posted early drug-development results across six biopharma partners — Generate's Claude-designed CAR T component came back 70% shorter than the standard construct with 10× receptor production in actual lab tests, and Anthropic is running its own muscular-dystrophy discovery programs in parallel. The press-release era of "AI accelerates drug discovery" has produced a graveyard of unreplicated claims, so the bar here matters: these are wet-lab-verified results sitting inside real pharmaceutical pipelines, not benchmark proxies. The under-noticed fact is who's doing the verifying — frontier labs are becoming their own downstream customers, which concentrates both the capability and the audit in the same hands. anthropic.com/research

For the archive: this is the day the largest proof in mathematics stopped being "the one nobody could fully check" — the watch-item for November is whether independent Lean verification of Axiom's 43.1M lines lands before the next Enormous-scale formalization is announced.