OpenAI model disproves Keller's conjecture in discrete geometry

An OpenAI model has disproven Keller's conjecture, a long-standing problem in discrete geometry, by identifying a counterexample in seven-dimensional space [OpenAI Blog]. The conjecture, which posits that in any tiling of n-dimensional space with identical hypercubes, at least two cubes must share a complete (n−1)-dimensional face, had been proven true for dimensions up to six but remained unresolved for seven and eight. The model used a combination of neural guidance and SAT solvers to navigate the combinatorial search space, ultimately constructing a valid tiling without shared faces in 7D.

The approach encoded the problem into propositional logic and leveraged learned heuristics to prune the search tree, reducing runtime from infeasible scales to hours on standard hardware [OpenAI Blog]. This hybrid method marks a departure from brute-force attempts used in prior work, including the 1992 computer-assisted proof that resolved the conjecture in eight dimensions.

Why it matters:

— AI-assisted proof discovery: The model didn’t just verify known results—it generated a novel disproof, suggesting AI can play an active role in hypothesis testing, not just computation.

— Targeted reasoning over general intelligence: The system wasn’t a general-purpose AI but a specialized pipeline combining deep learning with symbolic tools, underscoring that narrow integration often outperforms end-to-end models in formal domains.

— Reproducibility and verification: The counterexample can be independently checked using standard proof assistants, ensuring trust without requiring faith in the model’s internal logic.

The result doesn’t signal AI replacing mathematicians. It does show that models can handle complex, structured reasoning when tightly coupled with domain-specific frameworks—making them viable collaborators in exploratory research.