A fascinating things about GothCon gaming convention this year was watching people type rules queries for boardgames they'd never played before into ChatGPT.
Their expectation was that it would be helpful. The result always sounded good, but was invariably wrong when cross-checked with the manual.
Ultimately it became an exercise in "Do you want a wrong answer, fast, or the right answer slow?"
Which felt like a pretty accurate summary of where LLMs as a Q&A service are right now
It is imperative, always, to remember that "AI" doesn't "look at" libraries of stored data and distill an answer.
It *generates* a *probabilistic* sentence in response to a prompt. Even if the factual answer is right there. 1/?
I knew probability and statistics was a confusing area of study for a lot of people, but I never imagined how relevant that confusion would become for things like searching, and in what weird ways.
