“The "Humanities Last Exam" benchmark currently has a time limit of 15 or 30 minutes for models to solve problems, but at least one AI lab has requested this be extended to 24 hours.”