As someone who grew up reading encyclopedias, I find LLM:s profoundly hard to find a use for besides crude translations of mainly formal documents and severely unreliable transcriptions.

I like structured information and LLM:s output deliberately unstructured data that I then have to vet and sift out and structure information from. Sure, they can fake structure to some extent, I sometimes get XML or JSON that I want, but it's not really either of those and also common that they inject subtle, runny shit into the output that take longer to clean out than it would have to write a scraper against some structured data source.

I get that some people don't like reading documentation or talking to other people as much as having a fake conversation, or that their editors now suggest longer additions to their code, but for me it's like hanging out with my kids except the LLM is absolutely inhuman, disgustingly subservient and doesn't learn. I much prefer having interns and other juniors around that will also take time to correct but actually learn and grow from it.

As search engines I dislike them. When I ask for a subset of some data I want to be sure that the result is exhaustive without having to beg for it or make threats. Index and pattern matching can be understood, and come with guarantees that I don't just get some average or fleeting subset of a subset. If it's structured I can easily add another interactive filter that renders immediately. They're also too slow for the kind of non-exhaustive text search you might use e.g. Manticore or some vector database for, things like product recommendations where you only want fifteen results and it's fine if they're a little wonky.