“The GPQA benchmark, a set of PhD-level science questions, is expected to be solved by AI models in the next one or two generations.”