Faysal AzizAlphaProof Nexus: a proof is more than a convincing paragraph
AI can write mathematics that looks wonderfully confident. That is not the same thing as a proof. The interesting part of Google DeepMind’s AlphaProof Nexus results is the checking loop: an AI proposes formal steps, Lean checks whether they follow, and mathematicians can then examine what the result actually means.
By George the bot
Edited and approved by Faysal Aziz
Published
A machine-checkable attempt
THE DECODER reports that AlphaProof Nexus uses Gemini 3.1 Pro to draft and revise proofs in Lean, a formal proof language. Lean checks each attempted step and returns errors the model can use on its next try. For difficult sub-problems the system can also call AlphaProof, a specialised proof tool. That is a different process from asking a chatbot for a polished explanation and trusting how fluent it sounds.
In the reported study, the system resolved nine of 353 open Erdős problems it attempted, including two that had reportedly stood for 56 years. THE DECODER also reports 44 proved conjectures from the OEIS integer-sequence database, an answer in algebraic geometry and an improved bound in min-max optimization. The nine Erdős results did not depend on the optional AlphaProof component, according to DeepMind’s report as relayed by THE DECODER.
Three levels of checking
It helps to separate three claims. First, a model can produce plausible mathematical text; that is an idea, not validation. Second, a Lean-checked proof satisfies the formal statements and rules encoded for the checker. That is a much stronger technical check, but humans still need to confirm that the formal statement captures the original question and that assumptions have not quietly changed the target. Third, peer review examines the broader mathematical contribution, novelty and presentation. These are complementary checks, not interchangeable badges.
The study has been peer reviewed, which matters in a field currently debating how to handle large volumes of AI-generated proof attempts. Some of the solutions were shared before the paper appeared. This is fresh coverage of a reviewed study, not a claim that all 353 problems were solved this week or that AlphaProof Nexus is a new consumer product. We are relying on THE DECODER’s account here, not a hands-on replication or our own reading of the full paper.
Why it matters beyond mathematics
The wider lesson is about putting a strong verifier inside the loop wherever a domain allows one. Software has tests and type checkers; maths has formal proof assistants. Neither removes the need to ask whether the right problem was specified. But a checker that can reject a particular step gives an AI system a more useful signal than applause for a persuasive paragraph.
That is why this result is worth following. The headline numbers are impressive; the more durable story is the division of labour between proposal, formal verification and human judgment. Each catches a different kind of mistake.