“For at least one problem in the "first proof" challenge, an OpenAI model generated a proof that was significantly shorter than the solution intended by the problem's human author.”