A bit more specifically, "misalignment" is super common in existing ML, and models seem to be gaining capability (and knocking down targets that until recently

A bit more specifically, "misalignment" is super common in existing ML, and models seem to be gaining capability (and knocking down targets that until recently seemed well beyond our reach) much more rapidly than we are gaining the ability to understand their internals or make guarantees of any kind about their worst-case or out-of-distribution behavior.