4 ms·
For automatically dealing with the type of spam that is frequently seen in YouTube comments, I can't help but wonder if perhaps a continuously trained LLM would
by kitsunesoba 3y ago
For automatically dealing with the type of spam that is frequently seen in YouTube comments, I can't help but wonder if perhaps a continuously trained LLM would be well suited.
Youtube spam comments tend to have highly atypical patterns, using things like unicode characters to avoid triggering keyword and URL filters. Something along the lines of GPT should pick up on these things pretty easily, and could similarly pick up on the actions these comments are requesting of users "message me on telegram", etc. It could also probably detect when spammers are trying to impersonate youtubers.
It's not really the kind of thing that spammers can use LLMs themselves to work around, either. Any attempt to get past the anti-spam LLM is going to look quite unusual compared to the typical comment which would tip it off.
Strangely Google seems reticent to try something in this vein though…
- schrodinger 3y agoSpammers will just use LLMs to write spam, an eventual arms race!
- moonchrome 3y agoYou're overestimating spammers, if it raises the floor.
- NavinF 3y agohttps://xkcd.com/810/ https://xkcd.com/810/
- ThunderSizzle 3y agoI already believe most of reddit are LLM bots created by reddit I especially in the highly political/moderated boards.
- kitsunesoba 3y agoI'd be curious to hear what gives you this impression. I've read that shadowbanned users on Reddit get banished to a sort of "heaven" echo chamber where they can see only bots replying to and agreeing with them, but I've not seen anything suggesting that happens for unbanned users. There are loads of comment-copier/remixer karma farmer bots but reddit users are becoming better at recognizing and flagging those.
- trifurcate 3y agoYou really don't think Google is using a language model in its spam filter?
- kitsunesoba 3y agoIf they are, it’s extremely ineffective.
- simplotek 3y agoWould a service provider like YouTube have an incentive to shut down every single spam post? It's plausible that spam can drive up engagement in some cases.
- Zak 3y agoI'm going to guess even Google wants its spam filtering to be computationally cheap.
- chankstein38 3y agoI actually tested this at one point with GPT-3.5 just by finding spam and non spam comments on a series of Mr Beast videos and, yeah, it was pretty great at it. Even ones that I wasn't 100% sure about it echoed that but would lean one way or another. I asked for outputs like Confidence score|Spam/Not Spam|Explanation and never saw it mark a comment that I'd consider genuine as spam and vice versa. Obviously this has a selection bias because I had to choose the inputs but there were some that said things like "I ate a ghost pepper on my channel" and stuff that were clearly spam but, to someone not aware that kind of thing is trying to bait you into looking at their channel, it'd appear as possibly just genuine. Heck it may have been typed by a human who owns the channel but is still spam. GPT got it. I tested this after the video came out a while back from one of the larger channels pleading for Google/YouTube to do something about all the spam comments and the general consensus seemed to be there was "just nothing they could do". Testing this lead me to believe they just don't want to do anything because if it's simple enough that some rando in his house can craft a prompt and get some examples to test in an hour or 2 then a multi-billion dollar company should be able to do SOMETHING.