What possibilities actually exist for 4Chan to utilize Large Language Model technology for the purposes of assisting with moderating the website? My idea here isn't so much a plain autojanny doing all the work, as much as a special tool for assisting us with doing some of the most important work, which falls under GR1.Say that someone like the Babyshitter is subtly altering the image so that it doesn't get caught by the autoblock thing, like he's currently doing, what if we had an AI bot which had used the content put into the autoblock as its training data for the purposes of identifying subtly altered versions of it?Then that bot patrols a board and then scans through every currently active thread for if someone posts something like that, and if someone does, it immediately makes a global GR1 report and maybe also pings the moderation channel on Discord. A crucial question here is if the bot should be given any operational autonomy for this task, and thus be able to request/issue bans and delete posts. AI can be very fallible, so I'm inclined to say probably not.What would be the practical and logistical constraints in developing this kind of bot, and then operating it? My primary idea with this is to deal with those posting illegal content like CSAM, and possibly certain people operating their own spam bots (ergo fighting fire with fire).Finally, if this is at all possible to implement, what do we name the thing? Murphy? Fagballs?
>>10389I think intentionally training an AI with GR1 material would be a legal grey area that we wouldn't wanna wade into, even if the material passed to it was just anything that was GR1 BR'd on 4chan.Handling GR1 materials for reasons other than processing their immediate deletion and notifying the necessary authorities seems like a one way ticket to a legal assfucking.I'm sure facebook et al get away with having done so, but they're multibillion dollar corps and we're the asshole of the internet.
>>10390Yeah, that sounds like a nightmare to wade into, and even if you manage it the cost would have to be pretty high running such a model against every image posted here.
>>10391at least one of the 4chan archive sites runs an LLM analysis on most images posted to give them "descriptions", but I assume they're just using some currently free public-facing image describer instead of paying for the time on their own buck. Plus that's not instantaneous. I've not done any AI stuff myself so I don't know what the costs are like, especially at the scale of "content analysis/comparison of every image posted to 4chan"
>>10390>>10391>>10392Well, best I understand, the training data for LLMs would not directly contain the content, thus it shouldn't by itself contain anything illegal. Someone who understands this technology better than me is free to chime in, however. I also don't think that it would even need to run on every board, at most it might run kind of regularly on places like /b/ and /v/, high traffic boards where there's known troublemakers of higher consequence, and only ever get rolled out termporarily for if someone starts shitting up some other board with botted spam.Combative shitposters with a grudge are already weaponizing things like bots and altering image hashes, and I think that it would be in our own interest to try to see if there's any potential use in LLMs assisting with jannying.
no
>>10389c'mon, dude, you're going to put me out of a job
>>10390>I think intentionally training an AI with GR1 material would be a legal grey area that we wouldn't wanna wade intoDoesn't literally every other social media company do this already? This doesn't sound like something unprecedented to me at all.
>>10397It's probably one of those things where you have to be REALLY sure you have all your ducks in a row, and potentially have lawyers and stuff involved to make sure that remains the case, so I imagine it's a big hassle
