If we had an open source algorithm for Mastodon/Pixelfed that learned based on the words in the post and image/video we could have a Following + For You feed that showed you all the posts from people you follow and you could choose to see, say, 1 recommended For You post after every 3 posts from your Following feed. With the option to disable For You posts completely or tweak how often you see these.
Discovering new people to follow on mastodon/pixelfed isn’t great (hashtags are rarely used and make posts look ugly) so I still occasionally use twitter because I often discover new indie animators/gamedevs showing off their project making it really nice to browse the For You feed.
Stuff like this could be added to misskey, Firefish, or whatever other fediverse platforms there are. No need to implement it on Mastodon & piss off the entrenched community that wants no algorithm.
As for the algorithm, im imagining a few dedicated relay servers to sift through mastodon server data using a dedicated llm (thinking like llama or hugging face? I don’t know much about the technicals) to sift through and rate posts by engagement, word choice (evil, demonic, eugenics, crypto, etc) and rate each post topic collectively based on how virtiolic, how much engagement, how many posts, how similar to other posts (for spam), & the topic category & subcategory based off the respective language.
Users who activate the feature would have likes, saves, & etc stored on the client & a hash sent or something with the potential profile focus of the person.
The server weighing the details will prefer to return less seen & posted content very specifically tailored & occasionally more notable & engaged posts that are more general & less charged.
I think the general idea is interesting, & I would try to build it myself (for those who say stop whining & make it yourself), but im here grinding through kaggle learning & udemy so it’d be a while lol