Read the claim at its actual scale
On September 8, OpenAI reported an analytical proof and Lean formalization concerning finite-time singularity in Navier–Stokes. Its account describes an internal model and coordinated agents, with a result reached on September 5 and verification work completed afterward. We have read the announcement, not independently verified the proof or established the mathematical community's acceptance. That is an important boundary for a striking claim. It also leaves plenty to discuss without declaring that every scientific problem, or superintelligence itself, has now been settled.
Source: On the Navier–Stokes Millennium Prize Problem · 2026-09-08
There are several kinds of checking
A formal proof artifact, a readable explanation and an independent expert assessment serve different purposes. A machine-checkable argument can make the checking route more explicit, while readers still need to understand what was formalized and which assumptions define the result. We would ask for the precise statement and the environment needed to reproduce its verification. The announcement's existence is news. Successful reproduction and broader acceptance require their own evidence, dates and responsible readers.
This matters outside mathematics too. Imagine a team using AI to produce a calculation for equipment sizing. The useful question is not whether the reasoning looks sophisticated. It is whether the inputs, units, assumptions and required conclusion can be examined by someone qualified to approve the design. This is a hypothetical analogy, not a claim that a mathematical proof and an engineering acceptance process are interchangeable.
Preserve the failed approaches as well
Our interpretation is that ambitious agent research also needs provenance. Retain the version of the problem, the tools available, the artifacts produced and the checks applied. If people explore several approaches, show which one supports the final claim instead of presenting the entire exploration as one effortless answer. That gives an independent reviewer a route into the work. It also reduces the temptation to treat the fluency of a final summary as a substitute for the object being reviewed.
Let capability raise the standard
The more consequential the claim, the more useful it is to distinguish announcement, reproduced result and accepted conclusion. OpenAI's page includes a September 10 update about concurrent work; that update falls inside this issue's window and is a reminder to keep the record dated. We are not using later product launches to explain what readers knew that week. Our takeaway is hopeful and demanding: impressive AI work deserves an inspectable artifact and an honest account of its limits. Curiosity can move quickly. Acceptance should travel with the evidence.