“The current paradigm for AI benchmarks, exemplified by Humanity's Last Exam, limits the scope of model evaluation by focusing on difficult problems that are easily gradable.”