Navier-Stokes Millennium claim: Lean-checked blowup, prize not claimed
OpenAI says an internal model (post-Astra, not GPT-6 Astra itself) plus ~10,000 agents produced a Lean-formalized proof that 3D Navier-Stokes can blow up in finite time under a smooth external force. That is a serious math-and-AI milestone if the formalization holds. It is not the Clay Institute awarding a Millennium Prize, and OpenAI says it will not claim the $1M.
Quick Take
- Fact (OpenAI, Sep 8): Agents on a new internal model (training since Aug 28; company says more capable than GPT-6 Astra) reached a singularity proof ~Sep 5 (~88 hours); GPT-6 Astra spent ~17 more hours on Lean formalization. OpenAI frames statements C and D of Clay’s official formulation (smooth force allowed).
- Fact (OpenAI): The company “do[es] not intend to claim the Millennium Prize.” Concurrent credit: it cedes priority on forced Euler to Tristan Buckmaster (NYU) and Levent Alpöge (Anthropic); keeps Navier-Stokes for itself; says researchers/agents did not see that pair’s work before public release, while not ruling out de-identified product-usage data.
- Fact (coverage / Clay posture): Quanta and others note Lean checking raises confidence in internal consistency; Clay’s president called evaluation “deliberately unhurried”; as of coverage, Clay still lists Navier-Stokes unsolved. Forced vs unforced is the live mathematical fight, not a footnote.
- Judgment: Separate Lean-checked blowup (strong if verified by outsiders) from “AI solved a Millennium Prize” headlines and from the messy priority spat. Company claim ≠ prize status.
What OpenAI actually asserts
The write-up describes a vortex that spirals inward and stretches “like spaghetti”: speed unbounded in finite time, energy finite, force smooth. Lean artifacts and a paper were shared so the logic can be machine-checked. The compute story is extreme (millions of agent messages; cost estimates in the millions of dollars). That is capability theater with receipts, not a chat demo.
Company claim: this “resolves” the Navier-Stokes Millennium problem via C/D. Fact: C/D permit smooth forcing; many working mathematicians treat the unforced question as the one that matters. OpenAI’s own “won’t claim the prize” line quietly acknowledges that gap between press kit and prize process.
Credit fight, then Clay’s clock
OpenAI says the Sep 1 kickoff followed rumors later tied to Buckmaster/Alpöge. After Lean verification (~Sep 6), it offered a joint announcement; learned their public result was forced Euler, not Navier-Stokes. Quanta and follow-on coverage document mutual timelines and contested conversations (who knew what when; whether product traces helped). Buckmaster has said he has not seen OpenAI’s proof and is not accusing theft; Bubeck has disputed parts of Buckmaster’s account. Treat provenance as disputed, not settled.
Independently, Clay’s rules still demand peer-reviewed publication, years of scrutiny, and community acceptance. Lean is a powerful filter for formal gaps. It does not replace that social process, and it does not auto-label the result “the” prize solution if the community reads C/D as the wrong target.
Bandwagon Check
Fairly Hyped as a Lean-checked finite-time blowup under smooth forcing from a post-Astra internal system; Overhyped if read as “OpenAI won the Millennium Prize” or as closing the unforced problem. Fact: agents + Lean + public artifacts. Company claim: C/D resolution and AI-progress narrative. Judgment: ride the formalization; wait on Clay and independent mathematicians; keep the Buckmaster/Alpöge credit fight in the open without letting it erase or inflate the math.