Tao: Open math problems being non-renewably mined by AI
Powerful AI systems are starting to crack major open mathematical problems, raising fears that they will “strip mine” the most meaningful questions without producing the human insight or learning traditionally gained in the process. Commenters debate whether AI-generated proofs—often long, opaque formal verifications—undermine the incentives and career pipeline for human mathematicians, or simply redefine their role toward posing deep questions and digesting machine proofs. Underlying this is concern about concentration of power in well-funded AI labs, potential scooping of human researchers, and parallels to how AI is reshaping other knowledge professions.
Impact of AI on math and “finite” open problems
- Many see AI proof systems as a shock to the existing math ecosystem, especially when they “flatten” famous problems quickly.
- Key claim: the scarce resource is no longer solutions but good problems at the frontier, which took decades to curate via community effort.
- Some argue open problems are effectively non‑renewable at the current rate of AI “mining”; others say new hard/interesting problems can always be formulated.
Solutions vs understanding
- Repeated theme: the value of a problem is in the process and conceptual tools developed, not just the final proof.
- AI‑generated Lean proofs are described as huge, opaque, or “sloppy”, often requiring human “digestion” to extract insight.
- Critics of this worry that jumping straight to answers skips the exploratory “side paths” that historically generated new ideas.
Competition, scooping, and openness
- Strong concern that rumors of progress on a problem can trigger large, secretive AI swarms that scoop human researchers.
- This discourages sharing partial progress and may push math toward secrecy or “guild” behavior, undermining open science norms.
- Some compare this to insider trading or conflict of interest when the AI provider is also a competing researcher.
Careers, training, and motivation
- Worry that AI will destroy the training ground of “medium‑hard” problems, making it harder to develop new mathematicians.
- Several question why anyone would pursue a math PhD if frontier problems can be finished cheaply with compute.
- Others counter that humans will shift from solving to posing and interpreting problems, and must let go of sole‑author ego.
Analogies to coding, chess, and science
- Parallels drawn to:
- Chess after engines, where players still exist but roles shift.
- Software, where AI risks erasing junior work and creates unreadable “slop” that seniors must later tame.
- Experimental sciences, where replication and understanding still matter beyond raw results.
IP, training data, and ethics
- Heated debate on whether using prior chats/papers to train models is “theft” or fair use.
- Some suspect labs of using user prompts and private research direction to prioritize problems and scoop results; evidence is unclear and contested.
Unclear / open questions raised
- Can AI itself generate truly new, human‑interesting conjectures and problem landscapes?
- Will math re‑organize around valuing “digestion” and conceptual synthesis as much as first proofs?
- Long‑term societal effects if humans become subpar at essentially every cognitive domain remain unresolved.