Loïc Cabannes comments on Which AI Safety Benchmark Do We Need Most in 2025?