Conversation

if the homework you assign students is to build a chair and then they have to bring the chair into class and you sit on it, that is *not fake*. if a machine solved this problem for them it would *actually produce real chairs people could actually use*

238

this essay question garbage is fundamentally fake. every student knows the goal is to write some bullshit that sounds good enough that you're willing to assign it a good grade. you are not actually training the students to perform any kind of useful work and you never were

345

here's the opening of an essay i wrote as a sophomore in college. i don't remember writing this. i don't know a single thing about fMRI or the patriot act. i bullshitted this whole thing. this is what happens when you force students to write about shit they don't care about

205

this whole thing might as well have been generated by GPT-3 for all anyone could tell by reading it; it contains none of my actual writing voice, which i would not get back for another 8 years. there's none of my personality or life experience or soul in here

180

i want to clarify here that i'm using "bullshit" in harry frankfurt's technical sense, and specifically as distinct from lying: when you lie you try to hide the truth, when you bullshit you don't care either way what the truth is www2.csudh.edu/ccauthen/576f1

200

GPT-3 is literally a bullshit engine. it does not have a concept of words as referring to things; it plays games with words *only*, pure syntax, no semantics. literally the thing it is optimizing for when it produces text can be condensed to "put words here that sound good"

307

Sarah Constantin

@s_r_constantin

how could you tell if an AI was not a bullshit engine? (not sarcasm, this is something I'm genuinely uncertain about and matters a LOT for timelines)

i agree that this is a good and interesting question and i don't think i have much to say about it off the top of my head. one potential test is how long of a correct mathematical proof it can generate? at some point i'd have to believe semantics were happening *somewhere*

Replying to

6:53 PM · Apr 16, 2022Twitter Web App

and

@s_r_constantin

i mean there are trivial long proofs you can generate by mechanically applying a bunch of syntactic transformations like commutativity or associativity or w/e. i mean the kind of high-level arguments you find in mathematical papers

Likes

Replying to

and

like e.g. if GPT-5 could generate a correct proof of the riemann hypothesis then i'd be *really* scared. i don't *think* current architectures are remotely capable of doing this kind of thing but what do i know?

Joe Cappadona

@joecappadona