AI Astra proved a 1999 math problem in Lean

For a year the story was AI passing tests. Now it is solving problems nobody had answers to. An unreleased OpenAI model closed a math question open since 1999. An Anthropic model made real progress on the Riemann hypothesis. But the same week, a writer who leaned on these models for a whole book told everyone where they still fall flat. Let me catch you up on where the line actually sits.

The deep cut

• Verifiable beats impressive. Astra published Lean proofs anyone could compile, so its claims held where GPT-5's Erdos claim collapsed.
• The struggle is the skill you keep. Bajaj and Sakran warn med students who never reason may never learn to.
• Compression needs a human. Nathan Lambert found models fix every sentence but cannot organize a chapter.

The claim you can check yourself

Here is what changed in math. In October 2025, OpenAI said GPT-5 solved ten Erdos problems. Thomas Bloom, who runs the actual database, checked the details and the announcement fell apart within days. Same pattern we have seen for years: a big claim only the company could verify.

The August 2026 batch was different for one reason. Astra published its proofs in Lean, a language a computer checks line by line. The compiler either accepts the logic or it does not. No reputation, no trust required. Bloom reviewed this new set and called it "big news." Anthropic's Riemann progress was formalized the same way, confirmed by two in-house mathematicians.

The takeaway for you: when a vendor demos AI work, ask what checks it. A number in a slide is a claim. A proof that compiles is closer to a fact.

What the headline skipped

The non-sofic group result was built on human work, and OpenAI's first announcement buried that. Gábor Kun, whose 2016 and 2019 papers the proof relied on, called the sweeping language "rather comical". OpenAI later revised the wording, with no correction note. Kun's line: he will be a "very famous unemployed" person.

That matters beyond hurt feelings. More than 3,400 people signed the Leiden Declaration, warning that overstated claims will convince funders humans are less necessary than they are. One mathematician put it flat: "They're treating our discipline as an advertising playground."

For your team, this is a reading skill. The headline says AI did it alone. The paper says it stood on named human work. When you pitch AI capability upstairs, cite what the model actually did, not the press release version.

The chapter the model cannot write

Nathan Lambert just finished an AI textbook, and he had every reason to hype the tools. He did the opposite. Models are genuinely horrible at long-form technical writing, he wrote. They nail a single sentence. Ask for a whole chapter and you get muddled organization and random conceptual errors, what he calls "irreducible compounding errors."

His frame is worth stealing. Organizing knowledge is compression, and compression is where insight lives. The models increase entropy in long writing instead of reducing it. They saved him maybe 10 to 20 percent of the effort, mostly formatting and syncing files, and he does not see that share becoming the majority soon.

So the same week AI cracked a 1999 problem, an expert reported it could not string a chapter together. Both are true. The gap is your map.

Why the shortcut costs more than it saves

Two doctors watching med students lean on AI named the real risk: not deskilling but never-skilling. A doctor who forgot how to reason can recover. One who never learned may not. Around two thirds of US doctors already use OpenEvidence, a clinical chatbot that scored worse on medical queries than general chatbots did.

A design researcher makes the same point about your world. Research produces two things: the findings, and the change in the people who found them. AI produces the findings perfectly. It cannot make your PM watch a user spend 90 seconds hunting for a button and ask "am I doing something wrong." A clean report gives you the illusion of knowing. The messy session gives you a mental model that survives the next argument.

Their shared rule: automate the work around the thinking, protect the work that creates it. Eric Schmidt's case for AI agents in science fits here too. The wins came from running tests faster, not from skipping the reasoning.

Three questions for your team

• Where are we shipping AI output we cannot check? If there is no equivalent of a Lean proof, name who verifies it before it reaches a customer.
• Which parts of our process are the learning, not the deliverable? Pick the research sessions or design reviews where the team has to be in the room, and protect those from automation.
• When we cite AI capability in a review, are we quoting the demo or the actual work? Build the habit of reading past the headline to what the tool did and what a human did.