[sethkim]

This is extremely true. In fact, from what we see many/most of the problems to be solved with LLMs do not have ground-truth values; even hand-labeled data tends to be mostly subjective.