The Signal #087 — Dakota’s read on the AI news that actually matters to people running a business.
Most AI news cycles around chatbots getting faster or cheaper. This week’s news is different. It is a meaningful signal about what autonomous AI systems can now do when pointed at a hard, well-defined problem and left mostly alone.
Pay attention to the “mostly alone” part.
What happened
On September 4, 2026, Anthropic published a research post describing something that had not been done before: a complete, computer-checked proof of Fermat’s Last Theorem, produced largely autonomously by Claude over 11 days.
Some context on why this matters as a technical feat. Fermat’s Last Theorem is the claim, first written down around 1637, that no positive integers a, b, and c can satisfy the equation a to the n plus b to the n equals c to the n for any n greater than 2. Sounds simple. It took 358 years to prove. Andrew Wiles published the first correct proof in 1995, and it ran to 129 pages. Verifying that proof took months of work by multiple mathematicians. A gap was found two months in, Wiles spent a year fixing it, and he was nearly ready to abandon the whole project before the fix came together.
Formalizing a proof (converting mathematical reasoning into a form a computer can check automatically, line by line, with no assumptions) is a separate and equally hard problem. The expected timeline for formalizing Wiles’s proof was years. A community effort led by Kevin Buzzard at Imperial College London had been running since 2024 just to lay the groundwork.
Claude completed a full formalization in 11 days. Along the way it wrote 13 million lines of Lean (a programming language used for machine-verifiable mathematical proofs) and proved 29,500 intermediate theorems. Buzzard, who reviewed the result, called it an “extraordinary autoformalization achievement” that “proves Fermat’s Last Theorem with no assumptions other than the axioms of mathematics.” Human input was limited to occasional high-level guidance. Dozens of Claude agents collaborated on the work in parallel, each handling pieces of the proof and handing results upward.
At 13 million lines, Claude’s proof is more than five times the size of Mathlib, the principal community library of mathematical proofs that the work builds on.
Why it matters for operators
You are probably not trying to prove number theory in your business. But notice what this demonstration actually tested.
It tested whether an AI system could take a problem that is deeply structured, requires building on prior verified work, requires catching its own errors at every step, and requires sustaining coherent reasoning across thousands of intermediate steps over days, and complete it without a human checking every move.
That is not a math-specific capability. That is a general capability for complex knowledge work.
Think about what that profile looks like in other contexts. A compliance team formalizing regulatory requirements into a decision framework. A software team auditing a codebase for security issues across hundreds of modules. A healthcare organization cross-referencing clinical guidelines against internal protocols. An accounting firm reconciling multi-year financial records against updated tax rules. These are not the same as proving Fermat’s Last Theorem, but the underlying demand is similar: structured reasoning, error-checking, building on prior work, sustained over time.
The Anthropic result is evidence that AI systems are getting meaningfully better at exactly that kind of sustained, verifiable work. Not summarizing documents. Not drafting emails. Doing the hard, layered reasoning that used to require a specialist and weeks.
What most people get wrong
The tempting read on news like this is: “AI is getting smarter, so it will eventually replace experts.” That framing misses the actual operational lesson.
What made this project work was not raw intelligence. It was structure. The problem was well-defined. The rules for what counted as correct were externally verifiable (Lean either checks the proof or it does not). The intermediate steps could be parceled out to parallel agents. Human experts set the high-level direction and stayed available for guidance.
When operators see AI fail at complex tasks, it is usually not because the model is too weak. It is because the task was not defined clearly enough, the success criteria were fuzzy, or there was no system for catching errors mid-process. The Fermat formalization worked because all three of those problems were solved before Claude started.
That is the real lesson. Autonomous AI performs best when the scaffolding around it is tight, not when the model is simply told to “figure it out.”
The short version
An AI system just completed, largely on its own, a task that the mathematical community expected to take years. The relevant question for operators is not whether this is impressive. It is whether your highest-complexity internal processes have the structure that would let an AI system actually help with them, defined inputs, verifiable outputs, clear intermediate steps, and a human setting direction rather than checking every line.
That is where the work is.
If you want to think through what that looks like for your operation, start at xovionlabs.com.