SamKeeper

Then Eve, Being A Force

Laughed At Their Decision



patreon

games, comics, and books


pfp by @girlpillz


vogon
@vogon

great research with some cool methodology that confirms what everyone should've already suspected: if you ask someone to generate text on amazon mechanical turk, there's an almost 50% chance that they're just going to ask a large language model to do it

haven't really read much of the machine learning literature so I don't know how widespread this is any more, but also I suspect we're going to get a lot of follow-ups where people discover that their research using mturk as a proxy for human performance can't be replicated if you actually pay people to do the task


ireneista
@ireneista

we've been deeply concerned about the way in which generative language models destroy a lot of the data sets, such as wikipedia and web scrapes more broadly, that everyone uses as ground truth for all manner of research, which is going to make research a lot more general

this paper identifies an aspect of that problem we hadn't even thought of, namely that even studies which think they're paying people to do things are often getting ML output instead


garak
@garak

Pre-2022 text datasets are now the information equivalent of low-background steel.


SamKeeper
@SamKeeper

I'm still stuck back on the horrifying realization that apparently a lot of research relies on a digital underclass on mechanical turk, which is already some cyberpunk dystopia shit. it increasingly feels like large language/image models are essentially just covid for computers, this sudden (semi) exogenous shock that makes visible all the ways that the system was already astonishingly fucked.

though hey this is another causal factor on the "why does everything feel so technologically stagnant" pile I guess! jeeze!


You must log in to comment.

in reply to @vogon's post:

I haven't worked much with mturk but I believe they let you use a standard input for the HIT result or provide your own little embedded web app, and I suspect they did the latter for this

in reply to @garak's post:

we backed up a lot of stuff we had space for. then we decided we need like a hundred tebibytes more space to back up the stuff we actually want to, so that's been blocked on various life constraints because this stuff is rather expensive really. :/

in reply to @SamKeeper's post:

It's not just Mechanical Turk. In fact, entire industries are founded on the kind of cheap survey labor you'd find on Turk. How do advertisers, that backbone of the modern Internet, determine which demographics want which products? According to them, it's through an all-powerful algorithmic wizardry, that's just organizing the data after the fact. Dig down far enough and you'll find the data itself comes from tightly monitored tasks that consist of watching an ad and filling out how it made you feel - and considering they accept neither "I, like every other rational human being, am opposed to advertising as a matter of principle" nor faked enthusiasm as valid answers, these can be surprisingly difficult tasks to stomach long enough to get paid. All this, for $.03. Assuming you weren't rejected halfway through the survey because you didn't fit whatever demographic they were looking for.

Marketing isn't the only field that relies this much on shitty surveys; public opinion polling and certain scientific studies do as well. The former is just the one I have the most direct experience with.

Back when I was turking, academic research was one of the best paying categories of work on mturk. Content moderation was typically the lowest, AI training and "what if your computer was just a guy"-type data entry also paid miserably but was fast and high volume.

it's kind of astonishing--though I suppose it probably shouldn't be given Everything Else--that huge swaths of research are sustained by simple, what, a gentlemen's agreement that no one is going to bring up the awkward fact that none of the data is representative of much of anything and a bunch of it is just straight up fake?