Tech • AI • Robotics • Game
The Mathematics Are Mad
TL;DR
OpenAI’s claimed solution of 372 open mathematics problems has intensified a backlash led by Terence Tao and other leading mathematicians, who argue that machine-generated proofs risk undermining how mathematical knowledge is created, taught and understood.
KEY POINTS
A mass release of claimed solutions
OpenAI said it had solved 372 open math problems, including questions that in some cases had resisted progress for decades. The scale of the announcement prompted talk of a “math apocalypse” and fears that machines could outpace human researchers across major areas of the field. Some of the problems were described as prestigious enough that a human solver might have been propelled toward top honors such as the Fields Medal.
Why mathematicians object
In a public letter dated September 11, 2026, Terence Tao warned that the drive by AI companies to use hard problems as benchmarks is “detrimental” to mathematics because corporate incentives and the discipline’s goals are misaligned. The concern is not simply that problems get solved, but that they are solved in a way that bypasses the community processes that normally turn a breakthrough into shared understanding, new collaborations and future research.
Proofs without digestion
In traditional mathematics, a major proof typically triggers seminars, workshops, follow-up papers and recruitment of younger researchers into the field. Over time, the result is simplified, contextualized and incorporated into textbooks. Critics argue that dropping hundreds of solutions at once short-circuits that process, replacing a generational transfer of insight with a backlog of opaque results that few people can absorb.
The role of formal proof systems
Many of the new results were tied to Lean, a formal proof language well suited to machine verification. Large language models are especially effective at producing vast amounts of structured code, making Lean a natural vehicle for AI-driven theorem proving. Not all 372 problems were accompanied by full Lean formalizations, but roughly a third reportedly were, reinforcing the sense that automated proof pipelines are now viable at scale.
Fear of brute-force mathematics
A central anxiety is that AI may prove statements without revealing the most meaningful ideas behind them. Critics say some machine approaches look less like elegant mathematical insight and more like overwhelming computation. That raises a deeper philosophical worry: if enough important results are obtainable mainly through machine search, human mathematicians may become less central to frontier discovery.
The bin-packing flashpoint
One emblematic case involved a geometric packing problem in which unit squares are arranged into a minimal area. The machine-found configuration appeared highly irregular, with no obvious pattern or conceptual beauty. For many researchers, that kind of result suggests a future where optimal answers exist but are accessible only through raw computation, not through the elegant structures mathematicians traditionally seek.
Counterargument: ugly proofs can become beautiful later
Supporters of a more cautious view note that an apparently brute-force result does not end the search for understanding. In the packing example, later work reportedly reframed the same optimum on a torus, revealing a more orderly Fibonacci-like structure. The episode highlights a familiar pattern in mathematics: a first proof may be messy, but it can still point the way toward a cleaner conceptual explanation.
Integer multiplication showed rapid human follow-up
Another striking example involved the asymptotic complexity of integer multiplication. An AI-derived bound improved on the long-standing n log n barrier by an almost absurdly tiny margin, with an exponent correction on the order of 2^-182. Yet once that barrier was broken, human researchers quickly tightened the result several times, reportedly improving the saving parameter from 2^-182 to about 2^-17, suggesting that machine results can also catalyze fresh human progress.
A wider debate about training the next generation
The dispute extends beyond pure research. Tao and others emphasize that mathematics depends on attracting and educating young researchers through understandable breakthroughs and active intellectual communities. If AI systems routinely “eat the lunch” of early-career mathematicians by solving benchmark problems first, the field could struggle to train the people needed to interpret, refine and extend those results.
CONCLUSION
The fight over AI in mathematics is not only about whether machines can prove theorems, but whether proof without human comprehension counts as real progress for the discipline. The outcome could reshape both the future of research and the pipeline that produces the next generation of mathematicians.
Ask a question