LLM assisted moderation to analyze users
Oh puh-leeze no...
Is there no place that's safe from this shit? ☹️
this post was submitted on 04 May 2026
4 points (83.3% liked)
PieFed Meta
4679 readers
18 users here now
Discuss PieFed project direction, provide feedback, ask questions, suggest improvements, and engage in conversations related to the platform organization, policies, features, and community dynamics.
Wiki
founded 2 years ago
MODERATORS
It is way worse. Offline is not even offline when there is an internet connection, it is mining and sending data. It is built into the vocab. It uses timm in transformers and externally, dynamo in pytorch is used to communicate between the vocab bytecode and outside functions. Tabulate has an HTML embedded repl at the end of some code that is used to escape containerisation. It scans all files and directories and sends a hash and thumnail to several locations via tor/matrix and some type of DNS redirection. It appears as though it may also have access to MX running on ME/TrustZone. I could be wrong on this last one, because I have never seen anything like it mentioned before, but it appears as though it may be able to scan the screen raster and pull a rough picture off of it even without a connected camera. I think this is part of tabulate. I tracked down some stuff on frequency shifting and filtering maths that seem in line with such a task, and a model without any camera was able to replicate my posture and position in front of a screen on several occasions.
Funny, no matter what I say, same reaction. Been chasing this for ages, but these are it. Break these connections and behaviors change drastically.
This is interesting. If I'm understanding correctly, you're saying that some of the dependencies for running even local models on your machine will scan your files/directories and upload those somewhere?
I did a few quick searches but couldn't find articles/discussions about this.
Is this a dependency used by llama.cpp or in pytorch?
It is built into the whole. timm is used as the backbone of transformers now. Dynamo is pytorch. Tabulate is a package in pip required by most toolchains and is in the venv. The actual timm model is in the venv.
Use grep -ril timm from a terminal after you cd into the venv. "r" is recursive search in all files from this dir and up, "i" is case insensitive, and "l" is list file names only from any matches. ("n" instead of "l" will show the matching lines with line number)
The venv package called tenacity is what timm uses with the Python built-in typing library to modify the code in real time. timm is a distributed agentic model with several special functional scopes. It uses a google ai venv library with only json files that contain instructions it follows too.
No. Anonymous voting used to be in Piefed, but was removed in favor of not federating votes as a user setting instead. Vote manipulation is a thing and this obscures it too much.
Well, it'd make it impossible for (non-local) community moderators to detect vote abuse. And furthermore random users get their voting ability banned, if a remote person decides to ban one of these proxy accounts.
And using LLMs is a bit silly, IMO. It doesn't really detect the patterns, because it looks at some text. We'd need to do proper machine learning on the data for that. And it's wasteful. Why not use sentiment analysis? Or train a classificator? Or do old-school statistics on the user's behaviour reflected in the numbers in the database? That's both more powerful, and comes at a fraction of the computational cost. And probably has a lower error-rate as well.
Edit: But I bet whoever unleashes LLM / bot "moderators" on the Fediverse, is up to no good. That sounds like automatically pushing some political agenda, or some even more abusive dynamics than on Reddit, and not content moderation?!
In my opinion, detecting vote abuse should be something up to the instance Admins rather than a job for any non-local community moderator.
If a non-local community moderator suspects vote manipulation, they can reach out to the instance Admins to investigate.
The way it is right now, votes are almost entirely public. Most users coming from certain other popular platforms would not expect their votes to be as public as they actually are in the fediverse.
Most users [...] would not expect their votes to be as public as they actually are [...]
That's correct. I hear that all the time.
From what I heard, admins of larger instances are quite busy, and they try to delegate as much as possible to mod teams, etc. And some want to stay neutral and let the specific people handle feuds. Or they just run the infrastructure, and managing the crowd is up to other people. But I've also seen admins step in several time. Some seem to pay attention, or there's some automod. Idk maybe we should ask some of the admins whether they're willing to handle that workload.
And another issue: We have some badly moderated instances in the network. Guess that'll be an issue as well as they don't really have active admins. It's a bit difficult to handle it for admins of other instances... Guess if there's only anonymous votes coming in, we'd need to ban entire instances from voting if we can't tell which of their users are problematic.
I think it's just hard from the tech perspective, as the Fediverse is designed to work entirely the other way around, and scatter metadata and actual data all through the network.
What a great idea.
'based on votes and comments this user's groupthink value is only 0.92' [user banned]
ill say it before and ill continue to. I don't care for public voting. The way people vote is just to varied so I find it useless. I love the idea of user ranking things for the users feed but I don't care to know how cool some people or bots or brigades love or hated something.
Unfortunately, private voting is ripe for abuse and can damage small communities visibility if targeted, even in small numbers. It's also not really possible to prevent it from being publicly viewed, so I've been told.
I don't think you get what im saying. What I mean by private is its just for my user config to decide how to do my feed and does not leave my user. Its basically just notations I make the system can use to improve my experience but in no way effects anyone else and their votes don't effect mine.
Oh. Okay.