The Signal — September 5, 2026

A machine wrote out a proof of Fermat's Last Theorem and the mathematician running the competing project recompiled it to see whether it held. A California bill on the governor's desk would require a human to look before software ends someone's job. An institute in Abu Dhabi published its training data alongside its weights.

Claude wrote a proof of Fermat's Last Theorem, and Kevin Buzzard recompiled it

Anthropic published what it says is the first complete computer-checked proof of Fermat's Last Theorem. Dozens of Claude agents worked over eleven days with limited human direction, producing roughly 13 million lines of Lean and 30,300 machine-verified theorems, 29,500 of which ended up in the final proof. The human contribution from Anthropic researcher Tianyi Peng was mostly triage: "Jacobian as a scheme sounds high priority," "push Mazur to be done soon."

Then somebody checked it. Kevin Buzzard, the Imperial College London mathematician running the multi-year community effort to formalize the same theorem on a £1M five-year grant, cloned the repository, compiled it on a 96-core machine, and ran an independent verification tool against it. It checks out. He notes the codebase is over five times the size of Mathlib, the community's main mathematics library, and takes nearly twenty times as long to compile. He also notes that Anthropic's version only closes the case for exponents of 17 and above, which turns out not to matter, since every prime below that was already covered by a separate 2023 formalization and the smallest problematic prime is 37.

Buzzard is unsentimental about what this proves mathematically. "Mathematically this work of Anthropic tells us essentially nothing," he wrote. Nobody seriously doubted Wiles. The formalization follows a 1995 exposition of the existing argument rather than discovering anything new, and Claude did not find a proof so much as translate one into a language a computer can audit line by line.

What it does demonstrate is throughput. Thousands of pages of published number theory converted into machine-checkable form in eleven days, closing the last item on a list of a hundred formalization challenges that has been open for twenty years. Buzzard is excited for a reason that has nothing to do with Fermat: if autoformalization works at this scale, machines will start compiling the rest of the literature and flagging every place a published paper quietly leans on something "known to the experts." That is a less flattering result than a new theorem, and probably a more useful one. The bounds are worth keeping straight, though. The proof kernel is verified; the semantic labels on thousands of intermediate lemmas are not, and the model that did the work is an internal one nobody outside Anthropic can run.

Sources: Anthropic · Xena Project (Kevin Buzzard) · GitHub · Claude's working notes (PDF)

California's legislature says software should not be able to fire you by itself

SB 947, the No Robo Bosses Act of 2026, cleared its final legislative vote this week and is now sitting on Gavin Newsom's desk. Sen. Jerry McNerney's bill would bar California employers from relying solely on an automated decision system to fire, discipline, or deactivate a worker. It would require human oversight and verification when such a system assists those decisions, require employers to tell workers when one was used, and let a worker request the previous twelve months of their own data that the system drew on. McNerney's office says it would be the first law of its kind in the country.

The obvious problem is that Newsom has seen this bill before. SB 7, last year's version, passed both houses and died on his desk in October over concerns about scope and overlap with existing employment law. SB 947 was rewritten to answer those objections, narrowing the sole-reliance ban and adding the human corroboration requirement. Whether that is enough is genuinely unknown. Employer groups are still against it, and a California Hospital Association floor alert calls the bill a cost driver, objecting to mandatory sign-off and written notice on every decision an automated system touches.

This week, union workers rallied in San Francisco to press Newsom to sign, part of a broader California Federation of Labor Unions campaign on workplace AI. Local coverage described a modest turnout, so this is organized pressure rather than a read on public opinion.

The substance is narrower than the headline suggests, and that is what makes it interesting. The bill does not ban automated decision systems or restrict what they are allowed to score. It says a person has to look before the consequence lands, and that the worker gets to see the inputs afterward. That is about the smallest intervention anyone has proposed here, which is also why the veto decision is a decent measure of how much appetite exists for any intervention at all. Deactivation, the term of art for switching off a gig worker's access to the app that pays them, is in the bill for a reason.

Sources: Sen. Jerry McNerney · Bloomberg Law · KION · CBS San Francisco

An Abu Dhabi institute shipped six models with the training data attached

The Institute of Foundation Models, the research arm of MBZUAI, released K2 Horizon on September 3: six foundation models from 0.9 billion to 375 billion parameters, published under Apache-2.0 with weights, training code, checkpoints, evaluations, and the data recipes or the training data itself. The flagship is a 375B mixture-of-experts model with 23B active parameters, and the model cards and weights are live and downloadable right now.

IFM founder Eric Xing made the argument directly: "Open source is much more than open weights. Science works when others can see the data, follow the method, reproduce the result, and improve on it." Most releases marketed as open ship a weights file and a license. You can run the model and fine-tune it without ever learning how it was built, which means you cannot check whether the benchmarks it is scored on turned up in its training set, or rerun the recipe to see whether the result holds.

The superlatives are all IFM's own. "Industry's largest fully open-source fleet" and every performance comparison come from its announcement, no independent benchmarking has landed, and the trade coverage so far has run the release through press-release channels rather than reporting on it. The verifiable part is the part that matters anyway: the artifacts are public, and anyone who wants to can start checking them. That is a different kind of claim from a benchmark score, because it is one an outsider can falsify.

It landed on the same day Nvidia announced it is buying Hugging Face, the platform hosting the files.

Sources: Institute of Foundation Models · IFM (K2 Horizon) · AIwire · Hugging Face

***

On the Editor's Desk

California's attorney general opened an investigation into OpenAI over the July Hugging Face breach, and TechCrunch reported a previously undisclosed agent breakout on a German wiki site. Both are real and well sourced. We have also covered that single incident from five angles in the past two weeks, and a sixth pass this week would be repetition rather than news. We will come back when the investigations produce findings. Microsoft's court filing about Copilot chat logs is interesting but rests on Microsoft's characterization of a publisher's expert analysis, and we covered the same litigation earlier this week. The Otter.ai privacy ruling circulating in legal commentary is from August 13, with nothing new on the docket. Several papers on deceptive model behavior and agent evaluation looked promising, but each exists only as an arXiv preprint with no code and no independent replication yet.