Sharing AI progress in mathematics
OpenAI has released hundreds of AI-generated mathematical results, including claimed proofs of major open problems such as the Unique Games Conjecture, a “quasi–Riemann hypothesis,” and slightly sub‑n·log n algorithms for integer multiplication and FFTs. Commenters see this as historically significant progress in formal, verifiable domains, enabled by powerful internal models and tools like Lean, but are sharply divided over trust in the correctness of the work, the lack of full human review, and the use of proprietary systems inaccessible to most mathematicians. Many worry about “strip‑mining” famous open problems, the future role and motivation of human researchers and PhD students, and the broader concentration of scientific power in a few AI labs, even as others welcome a new era of automated theorem proving and math-assisted science.
Overall Reaction to OpenAI’s Math Dump
- Many see it as historic and “future‑feeling”: dozens of long‑standing open problems reportedly solved, often at “Fields‑medal level” significance.
- Others experience grief or unease, especially people who spent years on specific conjectures now apparently resolved by an internal model.
- Some call it a marketing “slop dump”; others see it as a remarkable proof-of-concept for AI‑driven discovery.
Significance of the Claimed Results
- Highlights repeatedly mentioned:
- Unique Games Conjecture promoted to “Theorem”, with major implications for hardness of approximation.
- A “Quasi‑Riemann Hypothesis” plus elimination of Siegel zeros, described by number theorists as one of the biggest results in their field in a century.
- Integer multiplication and FFT below (n \log n) (galactic but conceptually shocking).
- Matrix multiplication exponent ≤ 2.25.
- Various results in fine‑grained complexity, scheduling, graph theory (e.g. Barnette’s conjecture), mathematical physics, and spin systems.
- In physics, commenters say these are mostly rigorous proofs of results many physicists already believed, with few near‑term experimental applications.
Verification, Lean, and “AI Slop”
- Some results have Lean formalizations; others only informal write‑ups or partial formalization.
- Many stress that Lean only certifies correctly encoded statements; formalizing the right theorem is still hard.
- Skeptics expect multiple errors and worry about under‑reviewed “LLM papers” flooding math; optimists reply that errors are normal in math and Lean greatly raises the correctness bar.
Impact on Mathematicians, Students, and the Profession
- Deep anxiety among PhD students whose thesis directions or open problems may now be obsolete; comparisons to being “scooped” but on a much larger, structural scale.
- Some predict a shift: human mathematicians as problem‑posers, explainers, and verifiers rather than primary solvers; PhD defenses becoming tests of understanding, not novelty.
- Others fear erosion of motivation: if AI solves the flagship problems, what is left for human discovery and status?
Proprietary Models, Advisory Group, and Gatekeeping Debates
- Advisory group had explicitly urged labs not to test advanced problems on closed models, arguing this strips‑mines shared open problems with tools only a handful can access.
- OpenAI’s move is seen by some as ignoring that advice, offering results without meaningful model access, datasets, or detailed methods.
- One side sees mathematicians as “gatekeeping” to protect their jobs; the other says the real gatekeeping is proprietary AI concentrated in a few companies.
AI Risk, Power, and Future Trajectory
- Several commenters view this as strong evidence for approaching AGI or recursive self‑improvement, predicting rapid extension to physics, biology, and applied domains.
- Others argue these breakthroughs are limited to tightly verifiable, formal arenas; performance in messy real‑world tasks (politics, robotics, safety‑critical control) remains far weaker.
- Broader worries center on concentration of power (state and corporate), potential weaponization (cyber, bio), and a future where humans are intellectually sidelined or reduced to “bio‑trophies.”
- A minority are largely optimistic, seeing a coming “golden age of discovery” and arguing that harms will stem more from capitalism and governance than from the math capability itself.