AI 科研／形式驗證
OpenAI Releases Navier–Stokes Singularity Proof and Lean Certificate, but Mathematical Verification Is Just Beginning
OpenAI claims that an unreleased multi-agent system more capable than GPT‑6 Astra has constructed a proof that the three-dimensional Navier–Stokes equations lose smoothness in finite time under smooth external forcing. The team also released a Lean 4 formalization, but formal verification does not mean the mathematical community has accepted it as a solution to the Millennium Prize Problem.
On September 8, OpenAI released a Navier–Stokes proof generated by an internal model, claiming to complete the counterexample routes C and D in the Clay Mathematics Institute’s problem statement. The construction begins with a smooth, finite-energy, three-dimensional incompressible flow and applies an external force that remains smooth, causing the vortex core to contract continuously inward while stretching along its axis. Although the velocity and several terms in the equations become unbounded, the convection, pressure, viscosity, and other components are finely balanced so that the forcing itself need not become singular. If the argument is correct, the result would show that at least some admissible data cannot maintain a globally smooth solution—not provide a closed-form solution for general fluid simulation.
Another major focus of the technical announcement is the form in which the work was delivered. The GitHub repository contains separate Lean proofs for the full-space and periodic-torus cases, pinned to Lean 4.34.0-rc2, Mathlib, and Lake. Third parties can download the dependency cache and run `lake build`; a Comparator challenge is also available for independent verification. This makes it easier than a natural-language paper alone to detect missing lemmas, silently introduced axioms, or type inconsistencies. However, the Lean kernel can confirm only that the code derives the theorem as formalized. It cannot independently determine whether the formal statement fully corresponds to the original Millennium Prize Problem, nor can it replace scrutiny of the mathematical construction and cited sources.
OpenAI says the system consisted of a large number of coordinated agents using cached web data and programming tools, which began a concentrated search after hearing rumors of related research. The underlying model has not been disclosed, while the training data, agent trajectories, search costs, and failed attempts are also insufficient for reproduction. For now, the work should therefore be regarded as a major candidate proof accompanied by an executable certificate, not as a confirmed resolution. Under the Clay Mathematics Institute’s rules, a result must first be published in a qualifying venue, remain published for at least two years, and gain general acceptance throughout the global mathematical community before it can receive formal consideration. The most important developments to watch next are independent reconstruction of the Lean proofs, line-by-line checks by analysts that the formalization matches the paper, and whether public review uncovers implicit assumptions excluded by the formal statement.