New secret math benchmark stumps AI models and PhDs alike arstechnica.com 4 points by amichail 18 hours ago