this post was submitted on 15 Mar 2026
633 points (98.9% liked)
Greentext
7949 readers
564 users here now
This is a place to share greentexts and witness the confounding life of Anon. If you're new to the Greentext community, think of it as a sort of zoo with Anon as the main attraction.
Be warned:
- Anon is often crazy.
- Anon is often depressed.
- Anon frequently shares thoughts that are immature, offensive, or incomprehensible.
If you find yourself getting angry (or god forbid, agreeing) with something Anon has said, you might be doing it wrong.
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
I don't know if I'd go that far. It's one of the easiest ways to identify troll behavior. Sure going through and analyzing every post is insane, but being able to click on someone's history and see that oh they say stupid stuff for reactions regularly is pretty helpful. I'd also argue that it's helpful when people are making intellectually dishonest arguments when they have ulterior motives. Racist dog whistles for instance, when someone makes an argument that's a pretty obvious dog whistle and then claims they're just asking questions the Post history can sometimes help cut through that nonsense.
Honestly I think having an LLM that makes it easy to automatically label users as “tankie”, “maga”, “troll”, would make social media much nicer. I would avoid making the first response to a user who is known for their stupid arguments.
On the flip side, having a “this user cites their sources” tag would also be awesome.
Yeah but how would you trust that the LLM isn't biased, or the company that licenses and puts it in a browser extension isn't either? I don't know.
I'm asking because I like the idea, it's a good one.
Because classification tasks like this do not require frontier models, they could easily be run on a cpu locally with publicly available models.
That... doesn't answer the question. How do you assert that the base model you download and run doesn't have a bias one way or the other?
By using and testing it, obviously. It can’t magically develop a bias later on.
Everyone has a different definition of what unbiased means, so this would not be a “one size fits all” kind of thing. You would simply use a model that you personally deem good enough.
Not obvious. You're right, there is no magic in this technology, and you clearly don't understand how it works.
There is not a single LLM currently available that is able to consistently provide a correct or workable solution when faced with a semi complex word problem that's able to be contained in one paragraph. They may nail it on occasion but they cannot do it consistently. The "problem" of figuring out if someone is a bit eccentric, has poor social skills, is actively trolling, only trolls sometimes, or any combination of the above, is orders of magnitude higher than that. (edit:) To say an LLM is capable of that kind of logical determination is completely ignoring the evidence to the contrary.
I disagree, parsing through buckets of text (not one paragraph, a user’s full comment history) is literally the only thing LLMs ARE good at. This is not a logic problem, nor is it something that requires 100% accuracy.
It doesn’t matter if someone is just weird or malicious, I don’t necessarily want to engage in dialog with someone who is unlikely to respect my time or words.
Right, if we're talking about "good" in regards to speed then you're correct. But if we're talking about discerning intent, seriously? I find it hard to believe you're speaking in good faith and without bias yourself here. Disguising intent is the leading method to 'jailbreak' an LLM. Half the time at least, trolls are attempting to disguise their intent (with varying degrees of success). So that would be a solid failure at worst, or miss swaths of trolls at best.
I don't want to engage with someone like that either, but I care about not skipping over the people on the fringes of behavior, people who don't just regurgitate an echo chamber. This task might not require 100% accuracy but I personally wouldn't be satisfied with anything less than 99.9%.
I think using something like what we've been talking about is very very very far off in the future for me, if I were to ever do so at all. This conversation has made me realize that.
Fair enough! It was really a hypothetical anyway. No such system has been built to my knowledge.
That might be the worst idea I've ever heard in my life.
As with all people who are apparently experiencing AI psychosis I highly suggest you just learn to do things for yourself. You can make your own tags based on your own observations and experiences.
I thought it was a good idea
Yes but I do not have the time to read through everyone’s comment history. This would easily scale out to every comment that pops up on your feed BEFORE you waste time on it.
If there is exactly one thing AI is good at, it’s text classification. Don’t let your (perfectly reasonable) disillusionment with all the other uses of AI make you think that’s it’s a completely useless tool.
Hell, you probably don’t even need an LLM for this, there are lots of AI text classifier algorithms available.
Yes you're right that's certainly does sound like a reason why we shouldn't have any more clean water. Totally worth it.
Running LLMs locally does not require water cooling. Not sure what you’re talking about.
I think a cool solution would be to aggregate all of the tags that each user has received from other users and if there are frequent enough overlaps, a suggested tag might show up next to their name.
Of course, that would require user tags to be logged, which is not currently the case, afaik. It’s also not foolproof, because I’ve got at least one user tagged as “belligerent and stupid,” which, while probably helpful for others, is not likely to come up in other users’ tags. Most of my tags are probably pretty common though: troll, occasional troll, thoughtful, insightful, etc.
And finally, it might be susceptible to brigading or worse, if someone decides to make a bunch of accounts to tag LGBTQ users or something. Using the same federation rules as in other scenarios, where users or instances can be blocked or defederated at the instance level would help that, but I don’t know if it’s possible.