OpenAI's Navier-Stokes Proof Ignites Fierce Debate in Math and AI Circles
OpenAI announced that 10,000 agents invoked by an unreleased model spent 88 hours proving that the Navier-Stokes equations admit physically impossible solutions under certain conditions, a problem tied to the Clay Mathematics Institute's $1 million Millennium Prize. Mathematicians and AI commentators then clashed over whether the work truly solved the problem or exploited a loophole in the problem statement.
Confirmed
- OpenAI claimed an internally generated LLM proof solved the forced Navier-Stokes case using 10,000 agents over 88 hours (m7).
- Scientific American reported skepticism from mathematicians including Luis Silvestre of the University of Chicago, who argued the proof exploits external forcing and may have 'solved the wrong problem' (m17, m20, m1).
- Gary Marcus framed the dispute in three layers: whether it counts as solved by Clay standards (he says yes); whether training relied on unpublished human work (unclear); and whether it solved the core problem physicists care about (he says no) (m4, m13, m8, m16).
- Sabine Hossenfelder defended OpenAI, insisting the Clay problem itself was solved and criticizing those who downplayed the problem's relevance afterward (m18).
- A blogger aggregating mathematician views concluded OpenAI solved only part of a multi-faceted problem and its announcement was overstated (m10).
- A mechanical engineering professor in thermal fluids called the solution physically unreproducible, 'a mathematical curiosity' yielding no insight into turbulence (m3).
- Grady Booch reiterated that Transformer LLMs are architecturally next-token predictors and non-deterministic statistical engines; symbolic harnesses only blunt the edges. He argued current AI can do induction and deduction but lacks true abductive reasoning (m2, m5, m9).
- Pedro Domingos relayed martinmbauer's defense: the problem includes both forced and unforced cases, and solving the forced case is not evasion—analogous to vacuum solutions of Maxwell's equations. Domingos also slammed Scientific American's coverage (m11). Mathematician Paul Calhoun likewise rebutted the 'dodging' criticism (m1).
Unconfirmed
- Whether OpenAI's training relied on unpublished work by human mathematicians remains unknown per Gary Marcus (m4, m13).
- Booch noted many experts suspect the insight came from human experts rather than the AI itself; this remains unresolved (m12).
Why it matters
- The episode tests the credibility of AI capability claims: whether formally solving a Millennium Problem equates to solving the scientific core remains the fault line (m8, m13).
- cloneofsimo argued academia severely underestimates what OpenAI's math agents have solved, rejecting the 'stolen Codex logs' narrative, while stressing mathematicians remain valuable even with powerful proof machines (m6, m19).
- austinc3301 argued solving Navier-Stokes plainly requires intelligence, and any definition excluding it lacks predictive value (m15).
- Booch invoked Plato's cave, cautioning that structures found inside models are not rooted in the model's own perception and action, and that the two AI camps keep talking past each other (m14).
2026-09-22 ~ 2026-09-23 · 20 related posts
- Episode 1: 25 Fields Medalists Sign Open Letter Warning of Severe AI-Math Misalignment(2026-09-12, 95 posts)
- Episode 2: After AI Solves Math, What's Left for Humans? A Multi-Round X Debate(2026-09-12, 26 posts)
- Episode 3: Fields Medalists' Anti-AI Letter Criticized as Self-Serving Rationalization(2026-09-12, 2 posts)
- Episode 4: Tao and Litt Debate Whether Pure Math Becomes a Mere Hobby in the AI Era(2026-09-12, 7 posts)
- Episode 5: Math researcher: neither AI firms nor math community prioritize understanding(2026-09-13, 2 posts)
- Episode 6: Szepesvári Clarifies Mathematicians' Open Letter: Anti-Benchmark, Not Anti-AI(2026-09-13, 5 posts)
- Episode 7: 25 Fields Medalists Warn of Severe Misalignment Between AI and Mathematics(2026-09-13, 4 posts)
- Episode 8: Szepesvári: AI is destroying mathematics' justification for curiosity-driven research(2026-09-14, 4 posts)
- Episode 9: Fields Medalists' Warning on AI in Mathematics Draws Backlash(2026-09-14, 2 posts)
- Episode 10: OpenAI's Math Breakthrough Sparks Fields Medalists' Warning(2026-09-17, 4 posts)
- Episode 11: Fields Medalist Tim Gowers Responds to Open Letter on Math and AI(2026-09-17, 3 posts)
- Episode 12: OpenAI Solves Navier-Stokes Problem as AI Reshapes Mathematics(2026-09-18, 2 posts)
- Episode 13: 25 Fields Medalists Sign Declaration Against AI Math Benchmarking(2026-09-19, 2 posts)
- Episode 14: Po-Shen Loh on Terence Tao's Blog: Why We Still Need Human Mathematicians(2026-09-20, 4 posts)
- Episode 15: OpenAI Claims Model Solved 100+ Open Math Problems, Forms Mathematician Advisory Group(2026-09-22, 59 posts)
- Episode 16: OpenAI's Navier-Stokes Proof Ignites Fierce Debate in Math and AI Circles(2026-09-22, 20 posts)
- Episode 17: Developer Questions Whether Ordinary Users Can Solve Science Problems via Prompts(2026-09-22, 3 posts)
- Episode 18: Math researchers split over whether AI labs should release math results immediately(2026-09-22, 14 posts)
- Episode 19: Report: Unreleased OpenAI Model Solved 100+ Open Math Problems in 24 Days(2026-09-22, 5 posts)
- Episode 20: Top Mathematicians Warn AI Companies' Goals Misaligned with Mathematics(2026-09-22, 3 posts)
Primary sources
- Grady Booch: transformer LLMs remain non-deterministic statistical engines at their core — Grady_Booch · 2026-09-22
- "Solving Navier-Stokes took intelligence": debate over defining AI intelligence — austinc3301 · 2026-09-22
- cloneofsimo: academia badly underestimates the problems OpenAI's math agents are solving — cloneofsimo · 2026-09-22
- Even With Powerful Proof Machines, Mathematicians' Value Stays Immense — cloneofsimo · 2026-09-22
- Did OpenAI Solve the Wrong Navier-Stokes Problem? Experts Cry Loophole — joshgans · 2026-09-22
- Gary Marcus says OpenAI's Navier-Stokes claim is overstated as mathematicians debate the proof — GaryMarcus · 2026-09-22
- Mathematicians weigh in on OpenAI's Navier-Stokes claim: only parts solved — ProfNoahGian · 2026-09-22
- Thermofluids professor: OpenAI's Clay problem solution is not physically reproducible — GaryMarcus · 2026-09-22
- Gary Marcus: OpenAI solved the physics problem as posed, but not what physicists actually need — GaryMarcus · 2026-09-22
- [source] OpenAI's 10,000 agents proved a Navier-Stokes variant in 88 hours — sparking plagiarism claims — DeepLearningAI · 2026-09-22
- Sabine Hossenfelder backs OpenAI claim of solving a Clay Millennium Prize problem in spat with Gary Marcus — skdh · 2026-09-22
- Mathematicians dispute OpenAI's Navier-Stokes claim as a 'dodge' — it solved 2 of 4 problems — aran_nayebi · 2026-09-23
- Gary Marcus Amplifies Claim OpenAI Exploited a Loophole in Millennium Challenge Rules — GaryMarcus · 2026-09-23
- [source] OpenAI's $1M Navier-Stokes proof exploits a loophole, mathematicians say — srchvrs · 2026-09-23
- Domingos rips Scientific American over OpenAI's Navier-Stokes claim debate — pmddomingos · 2026-09-23
- Grady Booch insists transformer LLMs are next-token predictors at their core — Grady_Booch · 2026-09-23
- Grady Booch: Contemporary AI Still Lacks Abductive Reasoning, Just 'Next-Token Prediction' — Grady_Booch · 2026-09-23
- Grady Booch doubts AI's Navier-Stokes claim: insights may come from human experts — Grady_Booch · 2026-09-23
- Grady Booch: model internals are cave shadows, not grounded reality — Grady_Booch · 2026-09-23
1 near-duplicate retellings: GaryMarcus