I like to think of LLMs as a compressed knowledge base that you explore, navigate via prompts. Unfortunately the heavy post training from ai labs obfuscate that way too much IMHO