I agree - I view LLMs as semantic databases (of mostly stolen training data) with inference capabilities and NLG querying.