“Most current AI benchmarks, such as those based on multiple-choice questions, are useless because they measure solved problems and are not predictive of future capabilities.”