Report · theatlantic.com
A mathematician reads the AI future
theatlantic.com
Source published October 25, 2026
Summary: 2 min read
46/100
Summary
OpenAI claimed in early September that an internal AI system using about 10,000 agents over 88 hours had solved the Navier-Stokes Millennium Prize problem. On October 6, OpenAI reported that its latest model had solved nearly 400 landmark open mathematics problems, averaging three hours of computation per problem. These claims have unsettled mathematicians, particularly researchers in pure mathematics, where understanding and developing a proof can matter as much as obtaining a correct result. Cornell mathematicians held a department-wide town hall to consider how AI could change doctoral research, thesis writing, and hiring. Numerical linear algebra researcher Alex Townsend assembled about 200 open problems in his specialty and planned to release 600 long-standing problems across applied mathematics, hoping researchers could solve them—potentially with AI assistance—before companies do. He argues that mathematicians should focus more on choosing important questions, building theories, interpreting machine-generated proofs, and finding applications. Other mathematicians fear rapid automated problem solving could weaken training paths for students and diminish the human experience of discovery. Cornell logician Justin Moore said he would wait to open a bottle of cognac reserved for the solution of a 19-year-old problem until trusted peers verify and collectively understand the AI-generated proof. He said AI does not write mathematics particularly well.
Key Takeaways
- OpenAI claimed that an internal AI system deployed roughly 10,000 agents for 88 hours to solve the Navier-Stokes Millennium Prize problem.
- OpenAI reported on October 6 that its latest model solved nearly 400 landmark open mathematics problems, with an average of three hours of computation per problem.
- Alex Townsend plans to release 600 long-standing applied-mathematics problems and believes AI can help clear existing problems so researchers can pursue new theoretical and practical work.
- Pure mathematicians may value human understanding of a proof over an AI-generated answer, even when the answer appears correct.
- Cornell’s mathematics department is considering how AI could affect Ph.D. research, thesis writing, and the evaluation of future job candidates.
What the discussion said
The thread treated AI mathematics less as a distant forecasting exercise than as an immediate collision between apparent proof breakthroughs, human understanding, and academic livelihoods. Optimists argued that today’s systems already clear a threshold the 1980s symbolic-AI boom never approached: they converse usefully, generate novel leads, and may have accelerated progress on longstanding mathematical questions. Several readers expect AI to become a scientific instrument like new measurement technology, surfacing connections a human mind would miss and pushing researchers toward problems otherwise unreachable. Skeptics did not mainly deny that machines can produce results. They challenged whether unreadable or formally encoded output counts as usable mathematical knowledge when experts cannot inspect its reasoning. Others focused on the disruption: proof production underpins junior mathematicians’ training, publication records, and tenure paths, so a flood of machine-generated results could hollow out the profession before it creates a satisfying replacement. The claim that humans uniquely choose meaningful questions drew particularly sharp pushback; readers argued that models trained on human judgments, reinforced with feedback, can learn relevance and already propose new directions. A smaller group still sees an inexhaustible frontier where human guidance remains essential. There was also frustration with media coverage framed as a duel between techno-utopians and doomers rather than informed working mathematicians.
Where opinion split
Can AI genuinely take over mathematical discovery, or is it merely a powerful solver whose output still depends on human judgment? Supporters say neural systems have no principled barrier to generating worthwhile questions or reaching insights humans cannot, while critics say mathematical value and trustworthy understanding require humans to set goals and audit reasoning.
Coverage & discussions
Hacker News · theatlantic.com
Community discussions
Community Sentiment
Positives
- AI’s recent mathematical progress looks qualitatively unlike the old Prolog-era promises: systems already produce useful dialogue, leads, and advances rather than merely marketing demos.
- Models that connect ideas across fields could become a scientific instrument as consequential as new experimental measurement, exposing structures human intuition never notices.
- Teaching models human preferences through feedback undermines the claim that they cannot distinguish meaningful research questions from arbitrary ones.
- Automating routine proof search may free researchers to attack deeper scientific and philosophical problems instead of spending careers on tractable technical steps.
Concerns
- A proof is not much of a breakthrough if the model expresses it so badly that experts cannot read, audit, or convert it into human mathematical understanding.
- Machine proof generation threatens the career ladder of mathematics: novel proofs train junior researchers and buy the credentials that sustain teaching and tenure.
- Claims of solved problems look premature when advocates point to formal Lean artifacts while human mathematicians still cannot interpret the underlying argument.
- Replacing thought rather than extending a tool creates a harsher disruption than earlier scientific instruments, potentially removing humans from the intellectual work itself.
