OpenAI Model Disproves 80-Year Erdős Conjecture
Key insights
- OpenAI's reasoning model disproved the Erdős conjecture using Golod-Shafarevich theory, achieving a polynomial improvement over square grids.
- Cambridge's Timothy Gowers verified the proof meets top-journal publication standards, the first AI result to independently advance a significant open problem.
- Princeton's Will Sawin refined the exponent to δ=0.014, confirming the core argument and amplifying the mathematical result.
Why this matters
AI can now produce publishable mathematical proofs on genuinely open problems, changing what augmentation means for research and R&D workflows at every institution running long-horizon math or science programs. A general-purpose reasoning model achieved this milestone rather than a domain-specific tool, meaning the capability is already embedded in commercially deployed systems accessible to any team today. Institutions funding fundamental research face near-term pressure to define how authorship, priority, and intellectual credit are assigned when AI independently resolves standing open problems.
Summary
OpenAI's reasoning model disproved the 80-year-old Erdős unit distance conjecture, using Golod-Shafarevich algebraic number theory to construct an infinite point family with a polynomial improvement over square grids.
Cambridge's Timothy Gowers confirmed the proof meets top-journal standards, the first AI result to independently advance a significant open problem rather than rediscover known ground.
Essentially: (OpenAI, Princeton's Will Sawin) delivered and then sharpened this result.
- Sawin refined the exponent to δ=0.014, validating the core argument.
- A general-purpose reasoning model produced this, not a specialized math system.
- The conjecture dated to roughly 1946 and is a genuine benchmark in combinatorics.
General-purpose AI can now independently advance unsolved math problems, not just assist researchers who already know the path.
Potential risks and opportunities
Risks
- If peer review identifies a flaw in the Golod-Shafarevich construction, OpenAI's credibility for AI-assisted research claims takes a significant public hit before any formal correction is possible.
- Academic institutions may over-index on this single result and deploy general-purpose reasoning models in proof-verification pipelines before reliability benchmarks or failure-mode catalogues exist.
- Priority and authorship disputes between OpenAI (model output) and Sawin (δ=0.014 refinement) could establish a contested legal precedent for IP ownership in AI-assisted mathematics.
Opportunities
- Research-focused AI math startups (Lean FRO, Harmonic, DeepMind's AlphaProof team) can point to this result as external validation when seeking institutional research funding or university partnerships.
- Universities and national labs running open-problem programs could deploy general-purpose reasoning models against standing conjectures at low marginal cost, accelerating timelines across combinatorics and number theory.
- Scientific publishing platforms and professional societies (arXiv, Annals of Mathematics, American Mathematical Society) face pressure to establish AI-authorship disclosure and verification standards, creating a near-term governance tooling opportunity.
What we don't know yet
- Whether OpenAI's model required human-directed prompting toward Golod-Shafarevich theory, or arrived at that approach autonomously, has not been disclosed.
- Full peer-review submission and acceptance timeline is unconfirmed, leaving untested whether the top-journal standard Gowers cited will be formally validated.
- Whether OpenAI has run the same reasoning model against other open conjectures in combinatorics or number theory, and with what results, is not reported.
Originally reported by wsj.com
Read the original article →Original headline: OpenAI's Reasoning Model Disproves 80-Year Erdős Unit Distance Conjecture — First AI-Generated Proof to Meet Top-Journal Mathematical Standards