What can be checked is a few classes of problems
The Decoder reported on October 12 that a study in Science describes Google DeepMind’s AlphaProof Nexus. It solved 9 of 353 open Erdős problems, two of them unsolved for 56 years. The team posted those solutions online before the paper.
Beyond that, it proved 44 conjectures from the OEIS integer-sequence database, solved an open question in algebraic geometry, and improved a known bound in min-max optimization, a technique used to train some models. The report’s figures are these counts, not a claim about every open problem.
[1]
Lean checks the draft. The older tool can stay unused
The system is built around Gemini 3.1 Pro. The model drafts and revises proofs in the formal language Lean. Lean checks each draft after every edit, and the error messages guide the next attempt. For harder sub-problems it can call AlphaProof, a specialized proof tool.
DeepMind reports that even without AlphaProof, it solved the nine Erdős problems. DeepMind put the work through peer review.
[1]Peer review is the contrast this report marks
The same piece notes that OpenAI recently put a large batch of AI proofs on GitHub instead of submitting them to journals, and that an association led by Terence Tao called for a boycott. That is a separate episode. What this report lets a reader check is the counts above, the Lean checks, and that this work went through peer review.
[1]