Post #2803619
2026-05-23 22:52 UTC
Replies (2)
-
@KoboldCoterie@pawb.social 2026-05-23 23:09
Hate to tell you…
-
@Kynsey@lemmy.ml 2026-05-23 23:58
I think the very design of LLMs makes them not very useful even when used “correctly”. They are basically machines that are very good at sounding plausible. But they have absolutely 0 way of checking if info is correct or not. It’s just whatever is reinforced the most in their model is treated as correct. I think the underlying technology likely has uses. But the way it is currently being produced into products is something that, even if you tried to use it correctly, would simply end up tricking you with some plausible bullshit. Maybe you tell it to edit a paper and it decides to “fix” one of your opinions to be the most common one. Maybe you tell it to tell you the nuance behind a historical fact and it makes up a very likely sounding story that is entirely bullshit which you then repeat to someone else without realizing. The ability it has to sound plausible is its biggest flaw. Because an LLM will VERY rarely if ever say the words, “I don’t know.” You’d basically have to have gone in and coded it to respond to that specific question to respond with “I don’t know”. Otherwise it’ll just make something up.