[adamc]

I've asked it for things like book recommendations and gotten:

  - completely made up books
  - real books that were only marginally related
  - real books with really bad reviews

I'd estimate that only 30-40% of the time did I find the results at all useful.

[airstrike]

stop using it like a database that you can query

[bigfudge]

Agreed this is a bad idea in the case you are replying to, but I love ChatGPT as a way to recover the name of a book or film I’ve forgotten. I recently prompted for “a book about nuclear wasteland dominated by a church” and it gave me A canticle for Leibowitz (which is great). I’m not sure how easy that would be any other way.

[airstrike]

absolutely, I've been using it for that at lot as well and it's remarkably well suited for it. there's really no better tool for the job.

it's just that every thread about LLMs in AI invariably has someone complaining about best results from a query best described as `SELECT * FROM...`

[cruffle_duffle]

I wonder how many people are promoting it correctly. You can’t just query it like you might for google or something. It works best with lots of context and back and forth. And yeah, for many things you are going to get directional answers not exact ones (esp with “rote memory” like exact quotes from a book or something.)