Spronktonculous

This is what I drew after this situation 👇

cause like. apart from the (potential) exception of writing code, there is not really anything useful that a Large Language Model can do.

it imitates patterns in the text that it takes in

that's fundamentally how it works

Yes, haha funny, the AI is bamboozled by spelling, but seriously: I asked Google the same question 5 different times just now, and not only did I get 5 incorrect answers, I got 5 completely different incorrect answers.

The fact that they're different every time demonstrates just how much of an LLM's output is up to random chance. This is how LLMs imitate human language: each stage of constructing the text has built-in randomness, just like how human language has variation and flexibility.

This is my problem with using these things in the workplace. Like, imagine you're using an LLM to fill out a report of an accident that happened, or to write dietary recommendations for a patient, or to give feedback on a paper a student wrote...really anything you can think of.

Even if the LLM-generated text doesn't say anything technically flat-out wrong, the words and details are still randomized instead of being chosen by you for an actual reason.

Tech Bros™ think that emails need to contain information, some of which can be implied, and enough randomness to make sure it feels like it's been written bespoke for this specific interaction. What I need in an email is five distinct keywords in the same semi-polite assembly that conveys exactly all of the information I need it to to be legally correct in as few words as possible, and it needs to be the same thing for the same request every single time with little variation.