OpenAI’s Math Proofs Fall Short of Community Standards

OpenAI recently published hundreds of claimed solutions to some of the most difficult open problems in mathematics. The lab said it consulted an advisory group of leading mathematicians to avoid the criticism that followed its previous high‑profile claim.

The Advisory Group on Mathematics and Artificial Intelligence (AGMAI), hosted by Princeton’s Institute for Advanced Study, released guidelines in late September. Their first recommendation was to stop testing advanced problems on proprietary models, yet OpenAI’s release explicitly evaluates its closed models on open research problems.

Only ten of the 719 manuscripts include the model’s chain‑of‑thought, and just 42 % of the proofs have been formalised in a language such as Lean. The advisory group stressed that human understanding must follow any automated result, a responsibility that remains unclear in OpenAI’s approach.

A new paper from Cambridge and King’s College London documents discrepancies between the natural‑language explanations and the Lean code for a Navier‑Stokes‑derived problem, illustrating the translation gap that can undermine confidence in AI‑generated proofs.

Why it matters for GPU / AI infrastructure

Training and verifying large reasoning models demands massive GPU clusters. If the output cannot be reliably checked by humans or formal tools, the compute investment yields questionable scientific value. Cloud providers like AiGpu must ensure that their hardware supports not only raw throughput but also tooling for formal verification and reproducible experimentation.