@jcrabapple @thatandromeda Wow. I think one might extend this work and parlay it into an Ig Nobel — research that first makes you laugh, then makes you think. Why does ChatGPT fail on this question? What other similar questions does it fail on?
Top-level
@jcrabapple @thatandromeda Wow. I think one might extend this work and parlay it into an Ig Nobel — research that first makes you laugh, then makes you think. Why does ChatGPT fail on this question? What other similar questions does it fail on? 3 comments
@takloufer @jcrabapple @thatandromeda Huh, I guess I underestimated the popularity of the joke? I would have thought that actual advice would vastly outnumber iterations of this sort of tired joke in the real world, but I haven’t seen the corpus! @emjonaitis @jcrabapple @thatandromeda and that could still be the case ; all we know is that *someone* got an variant of the joke for that query. Perhaps thousands did get actual advice instead. |
@emjonaitis @jcrabapple @thatandromeda Its just regurgitating an old type of joke it could find in a number of places online, and therefore in its training set. Exactly as you'd expect a language model to behave, in other words.