We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
OpenAI Blog·5 min read
An OpenAI model has disproved a central conjecture in discrete geometry
TL;DR
For 80 years, mathematicians conjectured that points in a plane require at least a certain density to avoid large empty regions—a foundational problem in discrete geometry (the study of point arrangements and distances). An OpenAI model discovered a counterexample disproving this conjecture by identifying a configuration of points that violates the predicted constraint.
✦ Why It Matters
AI can now tackle unsolved pure math problems, expanding the frontier of what automated reasoning can discover beyond engineering applications.
Key Takeaways
- Neural networks can explore mathematical solution spaces to find counterexamples to long-standing conjectures.
- AI-assisted mathematics is now viable for pure theory problems, not just applied domains.
- Verify AI-discovered mathematical results through traditional proof methods before publication.
- Consider machine learning for exhaustive search problems in geometry and combinatorics research.
Full Summary
Discrete geometry studies properties of finite point sets in space, including distance relationships between them. A major unsolved conjecture, standing for eight decades, proposed a lower bound on the number of distinct distances that must exist between any set of points in a plane.
OpenAI trained a neural network model to explore geometric configurations and search for counterexamples to this conjecture. The model successfully identified a point arrangement that violates the conjectured bound, disproving the 80-year-old hypothesis.
This result was verified mathematically and published, marking a significant milestone where machine learning directly contributed to resolving a fundamental open problem in pure mathematics. The achievement suggests AI systems can augment human mathematical reasoning by exploring vast solution spaces efficiently.