Recent developments in frontier AI have begun producing mathematical results at speeds and scale that reshape what counts as a research problem or a career-making achievement. The Clay Mathematics Institute’s seven Millennium Prize problems—posed in 2000 as challenges to occupy mathematics for the coming centuries—have already seen early resolution: Grigori Perelman solved the Poincaré conjecture in 2002. The article reports that more recently an unnamed OpenAI model solved a forced variant of the Navier–Stokes problem, and OpenAI told the New York Times it had made “substantial progress on another Millennium Prize problem.” Rumors also suggest Anthropic is targeting additional problems. The author frames these events as compressing a millennium of mathematical time into days.
Terence Tao’s perspective: what is lost when struggle disappears
Terence Tao — described as perhaps the best mathematician alive, an informed user of frontier AI, and balanced in his assessment — emphasizes that struggling with a hard problem changes the mathematician. The struggle yields insight, methods, intuition and personal development that are part of the value of a proof. If AI rapidly delivers solutions, much of that formative value may vanish.
The piece argues that while it is useful not to reinvent the wheel (textbook access to Newton’s law, for instance), there is a real trade-off for professional mathematicians: years or decades of grinding on a problem produce deep understanding and often a career-defining result. When AI can solve such problems in days, something beautiful is gained — speed and volume of discovery — and something else beautiful is lost: the growth that comes from difficulty.
Proof abundance: too many theorems, too little judgement
The author coins the term “proof abundance.” Imagine a machine producing 100,000 novel theorems per day: truth becomes a commodity, but truth without interpretation is nearly useless. Which results open new directions, which close old ones, which are transformative and which are merely curiosities? Machines will make many claims; humans still need to judge significance, correctness, and relevance. If all answers are available, we risk a situation closer to Borges’s Library of Babel than to a coherent cathedral of knowledge: an enormous corpus that leaves us ignorant about where the important needles lie in the haystack.
A standard rejoinder is that AI will digest, interpret, rank and connect proofs. The author accepts this hope but points out a limit: unless humans are completely removed from the loop, the chain of interpretation must terminate in people who decide what to teach, fund, build, or believe. That bridge between machine outputs and human action cannot vanish.
Careers, motivation, and the ‘fallow’ analogy
If the locations of mathematical ‘needles’ are known and routinely harvested by AI, prospective mathematicians may be deterred from entering the field. The article draws a historical parallel to William Thurston, who dominated foliations to the point that students were advised against entering that subfield — Thurston’s success inadvertently emptied the pipeline. Unlike Thurston, however, AI does not retire or leave.
The author uses an agricultural analogy to illustrate a different risk: continuous harvesting depletes soil. Fallow land — leaving fields unplanted to restore fertility — is a deliberate and ancient practice. If machine fleets relentlessly harvest number theory, topology, algebraic geometry and combinatorics, they may prevent the recovery and regeneration that allow deep human creativity to emerge.
Institutional change: means of intellectual production
Another consequence is institutional. Historically, high-level mathematics required little more than pencil, paper and an exceptional mind. If frontier competition requires access to enormous compute, massive pretraining data and agent swarms, then departments may be forced to transform into quantitative shops to remain competitive. The author suggests that Perelman — the first human to solve a Millennium problem — may also end up being the last, if AI continues to dominate frontier discovery.
Possible responses: separation or new categories
Other domains have responded by separating human competition from machine competition: chess introduced human-only play and tool-assisted categories; speedrunning distinguishes human runs from TAS (tool-assisted speedruns). The article asks whether mathematics could create a similar boundary: split into “mathematics-as-research” dominated by machines and “mathematics-as-human-practice” that preserves human-driven problem solving. "No-AI" mathematics could be meaningful if the process of learning and discovery itself is valued.
Warnings from the community
The piece references a declaration signed by Twenty-five Fields medallists, including Terence Tao, titled “A Severe Misalignment of AI in Mathematics,” which warns that mass production of true/false statements at accelerating pace could destroy fertile ground instead of fostering new ideas. That statement encapsulates the central risk: AI’s success could, paradoxically, threaten the future viability of mathematics as a human practice.
Conclusion
The article neither rejects AI nor blindly celebrates it. It recognizes the technical achievements while highlighting cultural, institutional, and epistemic risks: loss of formative struggle, proof glut without interpretation, shrinking career-sized problems, and the need to preserve spaces for human-led creativity. One pragmatic approach is deliberate policy and cultural choices that maintain fallow areas for human mathematicians or create distinct categories in which human practice remains central, even as machines explore and exhaust vast mathematical territories.
