“By 2024, many AI models had surpassed human-level performance on benchmarks including MMLU, Big Bench, GPQA, and MATH.”