It’s a type of QA question that can identify peculiarities in models (e.g. count “r”s in strawberry), which the best we have given the black box nature of LLMs.