An OpenAI model challenges a classic Erdős geometry conjecture
Reasoning models are beginning to contribute to research by searching strange corners of problem spaces, not just by summarising papers.
The interesting bit is not only that a model produced a better unit-distance construction. It is that the result sits in the awkward middle ground between software output and mathematical work: a generated object that still needs human interpretation, verification, and cultural acceptance.
For a daily reading feed, this is the kind of AI research signal worth tracking because it shows models being useful in open-ended exploration. They are not replacing proof culture, but they are becoming tools that can propose candidates at a scale and weirdness level people may not naturally try first.
The practical question is what tooling grows around this: reproducible search traces, automatic verification, interfaces for inspecting candidate constructions, and publication norms for model-assisted discoveries.