if you watch college students use LLMs, one of the more interesting things you’ll notice is that for every student who submits a wholly LLM-written paper, there are 3 or 4 who basically write their own paper, but off the instruction of an LLM-ideated outline - ThreadSky

cozyunoist.bsky.social • 28 days ago

if you watch college students use LLMs, one of the more interesting things you’ll notice is that for every student who submits a wholly LLM-written paper, there are 3 or 4 who basically write their own paper, but off the instruction of an LLM-ideated outline

Comments

cozyunoist.bsky.social•28 days ago

this is interesting because it’s exactly the opposite of the prediction both boosters & critics would make! the rote task (line-editing, punch-ups) is the one you’re doing, the exciting one (blank page ideating) is the one you’ve outsourced. but i think this actually generalises!

cozyunoist.bsky.social•28 days ago

and, conversely, the rote task is the one you have lots of experience with. you’ve line-edited many papers before, but never thought through one on this specific topic. so it’s very high cognitive load to sit down and puzzle out de novo. but the LLM has similar puzzles in corpus

cozyunoist.bsky.social•28 days ago

i think the reason why our intuitive guesses on this are so wrong is that we think of computers as rule following & creativity as free. weird, right? good artists make works of art that are relatively internally homogeneous. they define their own problem & constraints precisely

cozyunoist.bsky.social•28 days ago

craftsmen don’t get to produce a internally homogeneous corpus because there are a lot of subtle constraints besides genre + the ones you define for yourself: materials, tools, cost constraint, client vibe. your past works are much predictive of future works than an artist’s

cozyunoist.bsky.social•28 days ago

anyway, i don’t have any big point here besides that global expressivist meta-semantics (& upstream, scepticism) make the right prediction here, for both humans & LLMs. we get good at saying the most likely next phrase before we get good at pretending to know how reference works

cozyunoist.bsky.social•28 days ago

(this being a basically separate point i made in a paper a little while ago). just thought i would post this bc i was mulling about it on the train & didn’t think i had anything larger to say on the topic!

damashummus.bsky.social•28 days ago

I guess. Wouldn’t a good complexity theorist be able to predict this just as well as a skeptic? I can say that languages are generatively entrenched phenomena. LLMs makes it easier for the users to pick a up set of language clusters as enacted practices.

nidarnukes.bsky.social•28 days ago

I always found that once I start writing the paper basically writes itself but getting started writing was always the hardest part. LLMs can be helpful for organizing and creating ideas as though it were another person to bounce ideas off of.

Comments

Posting Rules

Reply