OpenAI model disproves Erdős’ planar unit-distance conjecture
Key Points
- AI-generated proof disproves Erdős' unit-distance conjecture
- Construction yields n^{1+δ} unit distances (δ≈0.014 refined)
- Uses algebraic number theory (class field towers, Golod–Shafarevich)
Summary
On 2026-05-20, an internal OpenAI general-purpose reasoning model produced a computer-checked proof that disproves the long-standing Erdős conjecture about the planar unit distance problem. Historically the best constructions achieved about n^{1+o(1)} unit-distance pairs; the new result gives an infinite family of n-point configurations with at least n^{1+δ} unit-distance pairs for some fixed δ>0. A subsequent refinement by Will Sawin exhibits δ = 0.014. The proof was independently checked by external mathematicians and accompanied by a companion paper and an abridged chain-of-thought.
Key Points
- Problem: maximize unit-distance pairs among n points in the plane (the planar unit distance problem of Erdős, 1946). Previous best explicit constructions grew like n^{1 + C / log log n}; the widely believed upper bound was n^{1+o(1)}.
- New result: a model-constructed infinite family achieves at least n^{1+δ} unit-distance pairs (δ>0); refinement yields δ = 0.014. This disproves the conjectured n^{1+o(1)} optimality of grid-like constructions.
- Method: an autonomous, general-purpose reasoning model produced the proof. The construction uses advanced algebraic number theory (number fields, infinite class field towers, Golod–Shafarevich theory) to produce many unit-length differences in the Euclidean plane.
- Verification and artifacts: the formal proof was checked by external mathematicians; there is a companion paper with context and a published abridged chain-of-thought. The work includes experimental evaluation of model performance on a suite of Erdős problems and notes sensitivity to test-time compute.
- Significance for engineers: demonstrates that general-purpose AI can generate novel, verifiable research-level proofs; suggests AI systems can bridge distant domains (algebraic number theory → discrete geometry) and surface non-obvious constructions. It underscores the importance of reproducibility, external verification, and tooling for checking long formal arguments.
Practical notes
- Reproducibility: proof, companion paper, and abridged chain-of-thought are publicly linked by the authors; use external verification and independent mathematical review to validate large automated proofs.
- Implications for ML workflows: when using models for research, retain provenance (model version, prompts, compute), run independent checks, and expect value from models that are not domain-specialized but have strong reasoning abilities.
Key references
- Original proof (model-generated) and companion paper (external mathematicians) are available from the publication. The refinement giving δ = 0.014 is due to Will Sawin.