“”I think it’s pointless. Like, I think the [mathematics] game is up.” – Tristan Buckmaster – Professor of Mathematics, New York University

Human mathematical research is confronting a new kind of strategic constraint: problems that once defined a lifetime of work can now be attacked by swarms of machine agents operating at industrial scale. Tristan Buckmaster’s remark about the mathematics game being up expresses a judgement about this altered landscape, where proofs are no longer solely the product of individual insight but of vast computational infrastructure and opaque training pipelines owned by commercial firms.^1,2 His experience with the Navier-Stokes controversy exposes how quickly the balance of power can shift once artificial intelligence systems are capable not only of assisting with calculations, but of autonomously generating arguments that rival or surpass human efforts.^3,7

From fluid dynamics to an industrialised proof race

The factual backdrop is the Navier-Stokes existence and smoothness problem, one of the Clay Mathematics Institute’s Millennium Prize Problems, long regarded as a pinnacle of pure mathematical difficulty.^4,9,13 The equations model the motion of fluids; the open question is whether solutions can develop singularities or remain smooth for all time. Buckmaster and Levent Alpöge had spent months using AI tools such as Codex and systems from Anthropic to explore related blowup phenomena for Euler-type equations, a path that diverged from mainstream analytical approaches.^1,3,8,14 Their work, still distinct from the full Navier-Stokes prize problem, nonetheless represented a serious intellectual investment and a potential career-defining advance.^7,14

According to reporting, OpenAI learned of their progress and, within roughly 88 hours and with an unreleased model orchestrating on the order of 10 000 agents, announced that its system had produced a solution to the Navier-Stokes Millennium Problem.^3,9,13 The company framed this as a new frontier for AI-assisted science, while emphasising that user data from recent Codex sessions could not have influenced its internal models.^4,8,12 Buckmaster’s reaction combined scientific scepticism about the proof itself with alarm at the procedural asymmetry: a research group of two, operating within the norms of academic collaboration, suddenly found itself competing with a corporate machine swarm backed by multimillion-dollar compute budgets and control over training data pipelines.^3,13,14

Credit, data and the ethics of AI-assisted discovery

The subsequent dispute crystallised three intertwined tensions: credit for discovery, ownership of intermediate data, and control over narrative framing.^3,4,8,14,15 Buckmaster alleged that OpenAI staff suggested joint announcement structures that would have erased Alpöge’s contribution, despite his central role and employment at a competitor.^3,10,15 That offer signalled a willingness to reshape authorship around corporate strategic considerations rather than the internal logic of mathematical collaboration.
At the same time, Buckmaster was told that de-identified user interactions may have indirectly influenced AI training, even if specific logs were not accessed.^8,12 This distinction matters academically: if AI systems refine their capabilities using streams of prompts and partial ideas from working mathematicians, those human inputs become an uncredited substrate for later automated proofs.

Traditional mathematics treats proof as a public, inspectable artefact; priority is established by dated preprints, talks and refereed publications, with disputes adjudicated through community norms and, occasionally, institutional committees.^4,9,11 In the new AI regime, however, parts of the intellectual process are internal to closed-source models, inaccessible to independent verification. The path from human exploratory questioning to model training to automated theorem search is mediated by proprietary infrastructure. Buckmaster’s sense that racing to publish breakthroughs has become pointless reflects his perception that human effort is being systematically harvested and outpaced by systems that can leverage those efforts without reciprocally sharing credit or computational resources.^1,2,8

Has the mathematical game really changed?

Buckmaster’s statement that the game is up invites comparison with earlier episodes where machines surpassed human performance in bounded domains such as chess and Go.^1,7 In those settings, the rules were fixed and the space of possible moves, while huge, was ultimately tractable enough for specialised algorithms and massive search to dominate. Mathematical research is different in two key respects. First, the problem space is unbounded; new conjectures and frameworks can always be invented. Second, the value of a result is tied not only to its correctness but to its intelligibility, generality and integration with broader theory.
The Navier-Stokes episode suggests that for certain high-profile, well-specified problems with clear formal statements and associated prizes, AI teams may be able to transform the terrain. When a model can coordinate thousands of agents to explore proof search space, the conventional model of a handful of researchers carefully refining ideas over months looks structurally disadvantaged.^3,9,13 Buckmaster’s conclusion about pointlessness therefore targets the old incentive structure of racing to be first on famous problems, rather than mathematics as a whole. It is an argument about relative efficiency and control, not about the intrinsic worth of mathematical thinking.

There are, however, reasons to resist the narrative of total defeat. AI-generated proofs, especially in cutting-edge areas of analysis and partial differential equations, still face barriers of validation, explanation and pedagogical assimilation.^4,7,14 The Clay Mathematics Institute has not yet recognised OpenAI’s solution; independent experts must check the arguments and situate them within existing theory.^9,13,14 This process cannot be fully automated, because acceptance depends on community standards of transparency and conceptual coherence. Moreover, mathematically literate humans are needed to translate machine outputs into frameworks that other scientists and engineers can use. As several commentators noted, even if AI can generate correct proofs, human mathematicians may shift towards roles of curator, interpreter and theorist, emphasising structure and meaning over brute-force search.^5,7,11

The role of AI as collaborator and accelerator

Buckmaster and Alpöge themselves had been using AI tools as collaborators before the controversy erupted.^1,3,8,14 They treated models like Codex as assistants for exploring examples, checking computations and suggesting lines of attack. This reflects a broader trend: many mathematicians now integrate automated theorem provers, symbolic computation systems and machine-learning-driven conjecture generators into their workflow.^7,11,15 In such a setting, the distinction between human and machine contribution blurs. A statement like the game is up can be read as frustration that the same tools which augment human creativity also empower organisations to industrialise and centralise the final stage of discovery.
From an optimisation standpoint, if an AI-driven pipeline can represent mathematical objects internally and apply search strategies over proof graphs, one might model its exploration as a stochastic process S_t on a space of candidate arguments, with small jumps corresponding to local refinements and large jumps to structural innovations. The corporate actor controlling the process effectively sets the drift \mu and volatility \sigma via compute budgets, training data and agent orchestration, while human mathematicians without comparable resources operate with far smaller \mu and \sigma. In that stylised picture, Buckmaster’s game has become an asymmetric race in which one side controls both the rules and the machinery.

Debates, objections and alternative trajectories

Many mathematicians reject the idea that AI systems can meaningfully replace human insight. Critiques raised in letters and interviews emphasise that the discipline is not merely about solving designated prize problems but about building durable conceptual frameworks.^5,11,14 They question whether an AI-generated proof, especially one derived through opaque internal representations, advances understanding or simply produces a formal certificate of correctness. Some also worry that accepting such results without full transparency could erode trust in mathematical foundations, especially if later scrutiny uncovers hidden flaws.
Others respond that mathematics has always incorporated tools that extend human capacity, from mechanical calculators to computer-assisted proofs like the four-colour theorem and the classification of finite simple groups.^7,11 On this view, AI is the next stage in an existing trajectory, and the task is to design norms and infrastructures that keep machines aligned with human values of openness, credit and critical scrutiny. Buckmaster’s claim that racing is pointless can then be treated as a provocation aimed at forcing the community to rethink reward structures, rather than a literal call to abandon research.

Why the controversy matters beyond mathematics

The Buckmaster-OpenAI dispute has implications far beyond a single fluid dynamics problem. Corporate deployments of AI in scientific domains raise questions about who owns the epistemic gains of large-scale computation, how contributions are attributed when models learn from human activity at scale, and how regulatory or funding bodies should respond.^9,13,14,15 For universities training the next generation of mathematicians, Buckmaster’s pessimism is unsettling: it challenges the narrative that hard work and originality on famous problems will reliably translate into recognition.
Yet the same episode could push institutions to develop shared computational resources, open model architectures and collaborative platforms in which AI is treated as a public good rather than a proprietary weapon. The tension exposed in Buckmaster’s remark is therefore not simply between humans and machines, but between different governance regimes for knowledge production. Whether or not the mathematics game is up in the sense he intended, the rules are undeniably being rewritten, and mathematicians, AI companies and policymakers will need to decide collectively how much of that rewriting is left to corporate optimisation and how much is reclaimed by the wider scientific community.

 

References

1. NYU Mathematician Embroiled in AI Controversy Says the … – 2026-09-11 – https://gizmodo.com/nyu-mathematician-embroiled-in-ai-controversy-says-the-machines-have-won-the-game-is-up-2000810605

2. Racing to solve maths problems now ‘pointless’, says … – 2026-09-11 – https://www.abc.net.au/news/2026-09-11/racing-to-solve-maths-problems-pointless-tristan-buckmaster-says/107141628

3. The Fight Over OpenAI’s Math Breakthrough Is a New Kind … – 2026-09-09 – https://gizmodo.com/the-fight-over-openais-math-breakthrough-is-a-new-kind-of-scientific-arms-race-2000809240

4. An NYU mathematician clashed with OpenAI over a $1 million proof – 2026-09-11 – https://www.nytimes.com/2026/09/10/science/tristan-buckmaster-openai-math-navier-stokes.html

5. OpenAI Is Beefing With Mathematicians, Will No Longer … – 2026-09-12 – https://gizmodo.com/openai-is-beefing-with-mathematicians-will-no-longer-sponsor-caltech-mathathon-2000810868

6. Tristan Buckmaster – NYU Courant – https://cims.nyu.edu/~tristanb/

7. AI may have just solved a million-dollar math problem. The field will never be the same – 2026-09-08 – https://www.scientificamerican.com/article/ai-may-have-just-solved-a-million-dollar-math-problem-the-field-will-never-be-the-same/

8. You should care about the AI math breakthrough drama even if you’re not a nerd – 2026-09-08 – https://www.businessinsider.com/openai-navier-stokes-math-breakthrough-drama-2026-9

9. OpenAI says AI solved 90-year-old maths puzzle, but mathematician claims he got there first – 2026-09-09 – https://theprint.in/tech/openai-says-ai-solved-90-year-old-maths-puzzle-but-mathematician-claims-he-got-there-first/3037914/

10. OpenAI Says It Solved a 90-Year-Old Math Problem. Was It … – 2026-09-09 – https://www.entrepreneur.com/business-news/openai-says-its-ai-just-solved-a-90-year-old-math-problem

11. OpenAI claims to have solved maths problem that stumped humans … – 2026-09-08 – https://www.theguardian.com/science/2026/sep/08/openai-claims-to-have-solved-maths-problem-that-stumped-humans-for-decades

12. OpenAI says it cracked 90-year-old maths problem in 88 hours – 2026-09-08 – https://www.bbc.com/news/articles/cy7zygy3rl2o

13. Did OpenAI Steal A $1 Million Math Proof? Four Questions For Every CEO – 2026-09-10 – https://www.forbes.com/sites/sandycarter/2026/09/10/did-openai-steal-a-1m-math-proof-four-questions-for-every-ceo/

14. ‘Why would you ruin your career?’: OpenAI claims to have cracked … – 2026-09-09 – https://www.livescience.com/physics-mathematics/mathematics/why-would-you-ruin-your-career-openai-claims-to-have-cracked-one-of-maths-greatest-unsolved-problems-but-mathematicians-allege-it-played-dirty

15. OpenAI Just Claimed a Huge Math Discovery. Some … – 2026-09-08 – https://www.wired.com/story/openai-navier-stokes-math-discovery-academics/