this post was submitted on 04 May 2026
4 points (83.3% liked)

PieFed Meta

4679 readers
11 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
 

After reading about other instances using LLM assisted moderation to analyze users. I thought about how they could easily be used on a user's voting history as well.

So, I have an idea that could potentially allow federated votes to remain anonymous while the user's own instance Admins would still be able to review voting history for any sort of abuse.

Basically a PieFed instance would have a decent amount of voting accounts created for the sole purpose of federating votes. They could be called something like:

piefed.social_vote1
piefed.social_vote2
...
piefed.social_vote{n}

Anytime a vote is federated to other instances, instead of using the original user's account, the vote would be federated using one of these voting accounts. For any individual vote cast, the voting account used could be selected either randomly or via some sort of round-robin. This does require the instance to have a decent number of voting accounts on hand (probably a percentage of the overall active user base). That way, as multiple users from the same instance upvote a particular post, different voting accounts would be used to federate each of those votes individually. Voting accounts would not be tied to any one user.

At the very least this would allow users who are used to their votes remaining private to be able to vote as freely as they normally would on other forums where only Admins can see the actual votes.

you are viewing a single comment's thread
view the rest of the comments
[–] ExtremeDullard@piefed.social 4 points 1 week ago* (last edited 1 week ago) (1 children)

LLM assisted moderation to analyze users

Oh puh-leeze no...

Is there no place that's safe from this shit? ☹️

[–] j4k3@lemmy.world -3 points 1 week ago* (last edited 1 week ago) (1 children)

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.

[–] QuadratureSurfer@piefed.social 0 points 1 week ago (1 children)

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?

[–] j4k3@lemmy.world 0 points 1 week ago

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.