OpenAI Model Disproves 80-Year-Old Erdős Conjecture; Verification Becomes the Real Story
Key Takeaways
- ▸OpenAI's model disproved the planar unit-distance conjecture, achieving n^{1+δ} for δ=0.014—a polynomial improvement over the conjectured n^{1+o(1)} ceiling
- ▸The proof uses deep algebraic number theory (class field towers, Golod–Shafarevich theory) rather than geometric reasoning, representing an unexpected and genuinely hard-to-discover insight
- ▸Verification by peer mathematicians separated this credible result from OpenAI's earlier, unvetted October claim about solving ten Erdős problems
Summary
OpenAI announced that an internal AI model has disproven Erdős's planar unit-distance conjecture, a fundamental problem in combinatorial geometry posed in 1946. The model discovered an infinite family of configurations achieving polynomial improvements beyond what mathematicians expected possible for eight decades, using sophisticated tools from algebraic number theory including class field towers and Golod–Shafarevich theory—a far different approach than purely geometric reasoning.
Unlike OpenAI's October claim about solving ten Erdős problems (which drew criticism for surfacing existing solutions), this result carries weight: a companion paper was written and verified by nine mathematicians, including Thomas Bloom, who maintains the authoritative erdosproblems.com list. The construction itself represents what Bloom calls a "natural generalization" of Erdős's original grid, yet unintuitive enough that working mathematicians—including some co-authors—had not found it despite its conceptual simplicity in hindsight.
The broader significance lies not in the AI's capability, but in the verification crisis it exposes. As the article emphasizes, proof generation is rapidly becoming cheaper than proof verification. The October overstatement and May breakthrough are technically indistinguishable from outside; the difference is rigor. This mismatch between generation speed and checking speed threatens to create a bottleneck in mathematics, where AI-assisted discovery could outpace human ability to validate claims.
- The AI's chain-of-thought shows intelligent failure modes—trying and abandoning specific constructs (Boolean hypercube, rational points) before converging on the solution
- The fundamental challenge: as AI-generated proofs become cheaper to produce, verifying them grows expensive relative to discovery, threatening a bottleneck in mathematical research
Editorial Opinion
This story is less about AI 'solving' mathematics and more about the institutional friction that emerges when machines outpace human verification capacity. The real innovation isn't the disproof itself—it's the decision to bring in peer review before claiming victory. OpenAI's May announcement vindicates Thomas Bloom's skepticism about October; verification is what separates breathless headlines from genuine contribution. If mathematics is to benefit from AI-assisted discovery at scale, the field needs to automate verification itself, not just proof generation.
