[aeschenbach]

Shouldn't the evaluation of ChatGPT as a language model focus on how human it sounds, not how good it is at information retrieval? (Not an MLE)