In benchmarks of the METR AI research institute, automated judges rated some proposed tests of AI capabilities as more useful ...