“Existing popular benchmarks, such as MMLU, are no longer difficult enough to effectively measure the capabilities of state-of-the-art large language models.”