Greg Slepak

It's true that algorithmic feeds will work differently on the #Fediverse than they do on either #BlueSky or #X, but that doesn't mean they have to suck.

They can in fact be better.

Here's an example of how it can work (if done properly):

A common algorithm is to show popular posts you might have missed that are "near you".

Developers can explore 3 different concepts of "near you":

  • "Near You v1": Popular posts from people you follow directly. 0-steps removed.
  • "Near You v2": Popular posts from the set of people you follow plus some sample of the people followed by some of the people you follow (perhaps not all of them, that might be too much data to process for your server).
  • "Near You v3": Popular posts from the set of people you follow, plus some sample of posts from the relay(s) you follow.

Now the hard part on the #Fediverse is figuring out which posts are popular. After all, the fact that a post has been RT'd or Like'd is not broadcast to all servers. My understanding is that there are ideas for protocol adjustments for how to make "just enough" of this information available to servers without clogging the pipes or overloading servers.

In fact, all 3 "Near You"s can be algorithmic feeds that servers like #Pleroma, #Misskey, #Mitra, #Mastodon, and others offer to their users.

crib.social

Mar 26, 2025 at 5:59pmWeb
Show threadGreg SlepakMar 26, 2025

Other algorithmic feeds can be generated by the user themselves using a small self-hosted LLM. Users will be able to generate these feeds by simply speaking them into existence:

The UI will show an input field, and the user will type what sorts of posts they want to see. E.g.: "Show me posts related to information security"

Then, the server will pass every single post that goes through the server's "global timeline" through this LLM-based filter. The ones the filter approves get sent to this "algorithmic feed" (really just a list).

This technique can make the concept of hashtags (like the ones below) obsolete.

#Fediverse #algorithmicfeeds #algorithmicfeed #ActivityPub

crib.social

Show threadArjen P. de Vries Timmers 🕊️Mar 29, 2025

@taoeffect that sounds more like a personal search engine running over toots than a local LLM?

The PSE can then also integrate email, or web pages visited, etc Not easily replaced by LLMs because you'd need to further train it. One could envision a RAG setup, perhaps with advantages over classic search, but I suspect the costs are higher than the gains.

Given that users don't "even" run personal search engines, I think it will be a while before they may (/should) consider personal LLMs.