Will it be a golden age or a dark age for mathematics (and science)?
terrytao.wordpress.com
October 18, 2026
17 min read
50/100
Summary
Francesco Maggi argues that large language models could accelerate mathematical and scientific discovery but could also concentrate knowledge and weaken independent human research. He cites LLM-generated work including a finite-time blow-up solution for the forced Navier–Stokes equations and a proposed proof of the quasi-Riemann Hypothesis, while stressing that mathematicians must independently scrutinize, interpret, and assimilate such results. Lean, a proof assistant, can formally verify proofs, but important machine-generated Lean proofs may run to hundreds of thousands of lines and remain difficult to translate into understandable natural language. Maggi says theorem proving alone does not define mathematical progress: mathematicians also create new concepts and formal languages through “mathematization,” as exemplified by Newton, Riemann, Einstein, Shannon, and Grothendieck. He argues that authorship, citation, peer review, confidentiality, and the evaluation and training of researchers require new standards when LLMs contribute to research. He also warns that proprietary frontier systems could give a few companies disproportionate control over scientific priorities and capabilities. Maggi calls for coordinated, publicly governed scientific LLM infrastructure, researcher access to frontier systems, disclosure and reproducibility standards for LLM-generated scholarship, confidentiality protections, and research into effects on education and training. He argues these measures could preserve an open, intellectually independent scientific culture while using LLMs to support theory-building across mathematics and other sciences.
Key Takeaways
- LLMs can generate formally checkable mathematical proofs, but very long Lean proofs and poor natural-language translations can leave their meaning inaccessible without extensive human study.
- Francesco Maggi argues that mathematical progress includes mathematization—the invention of concepts and formal languages—not only the solution of established theorem-proving problems.
- OpenAI reported in its October 6, 2026 release that it had tried 4,000 problems and succeeded on some hundreds after aggregation and selection, according to Maggi's reading of its statement.
- Maggi warns that opaque training data practices and proprietary frontier models can complicate citation, confidentiality, peer review, and independent verification of scientific work.
- Maggi calls for public and academic institutions to build and govern competitive LLM infrastructure alongside commercial systems to preserve independent research capacity.
What the discussion said
Commenters treated the golden-age question less as a forecast of raw theorem output than as a fight over who controls AI-enabled knowledge and whether humans can still understand it. Several readers think AI can make mathematics dramatically more productive: it can handle brute-force symbolic work, cross-reference huge literatures, and give more people an on-ramp to ideas once guarded by elite institutions. Some expect the bottleneck to move rather than vanish, creating demand for people who frame problems, validate machine-generated claims, and turn results into usable understanding. The sharper anxiety is that a correct answer is not the same as mathematics. Researchers learn by following proofs, intuitions, failed routes, talks, and shared conceptual machinery; opaque model output may deliver results while starving that ecosystem. Readers also worried that frontier systems and the scientific discoveries they enable could be concentrated inside commercial labs, protected through patents, or used to lock rivals out under safety rationales. That concern extends to confidentiality: researchers may withhold early ideas if proposals, reviews, or drafts can be fed into models that train on them. A smaller group rejects nostalgia for professional gatekeeping and sees AI as a chance to make mathematics less exclusionary. The thread’s qualified consensus is that the technology can produce a scientific boom, but only if public institutions retain capable, transparent, independently governed access rather than becoming customers of a few private research engines.
Where opinion split
The central dispute is whether AI-driven mathematical progress will strengthen human mathematics or replace it with inaccessible machine output. Optimists argue that automation removes drudgery, expands access, and frees people to formulate and check deeper ideas; skeptics answer that results without teachable reasoning, open access, and human expertise reduce mathematics to proprietary answer production.
Community Sentiment
Positives
- AI could turn literature-scale cross-referencing, symbolic inference, and routine verification into cheap infrastructure, letting researchers spend more time on problem formulation and genuinely new ideas.
- Independent frontier-model capacity for universities could prevent advanced scientific assistance from becoming a privilege rented from a handful of commercial labs.
- Machine assistance may crack open a field seen as socially and linguistically gated, translating dense notation and elite conventions into explanations newcomers can actually use.
- Accelerating the mechanical parts of research could increase demand for humans who pose worthwhile questions, audit claims, and connect discoveries across disciplines.
Concerns
- A model that emits a valid theorem but cannot convey the intuitions and methods behind it risks producing answers while eroding the human learning pipeline that makes mathematics cumulative.
- Private ownership of frontier scientific models could concentrate discovery, patents, and agenda-setting in firms whose incentives favor extraction over broad scientific access.
- Researchers may increasingly hide proposals and unfinished ideas if reviewers or administrators can upload them to training systems, poisoning the trust needed for open collaboration.
- AI could create a short burst of easy gains before weakened incentives to share original ideas leave knowledge concentrated among organizations that own the strongest systems.
