An OpenAI model has disproved a central conjecture in discrete geometry

The critical distinction here is between a model generating a plausible conjecture and a model producing a formally verifiable disproof. A disproof requires constructing a concrete counterexample that holds under rigorous mathematical scrutiny, not pattern-matching toward a likely answer, which is a much harder bar to clear and the one that matters for the research community's trust.

This story is largely disconnected from recent activity covered on Modelwire. The closest adjacent thread is the competitive coding-agent race flagged in the Deepseek Code piece from May 20, where labs are racing to automate structured reasoning over formal systems. But that work targets software engineering workflows, not open mathematical research. The geometry result belongs to a quieter, slower-moving conversation about whether AI can contribute to fields where the output cannot be evaluated by execution or test suite, but only by peer proof-checking.

Watch whether an independent mathematician or a formal verification system like Lean confirms the counterexample within the next 60 days. Peer-verified reproduction is the only signal that separates a genuine result from a model that found a gap in the problem's informal framing.