I think they mean "categorize" instead of "classify". That is, they are looking for some form of unsupervised clustering. In that case, I agree that LLMs are n

I think they mean "categorize" instead of "classify". That is, they are looking for some form of unsupervised clustering. In that case, I agree that LLMs are not the solution in and of themselves, but the word/sentence/documents embeddings obtained from the inner layers thereof might actually be useful. Back when BERT came out, the team I was on at the time had pretty good results using BERT-derived sentence vectors for a machine learning project.