you post something. before most people ever see it, a model has already read it and written a file on it ๐
here's what the open-source portion of ๐'s algo says about that, in simple words.
[you don't have to read the repo yourself. have your AI read it and explain it back to you]
1. IT'S CALLED THE "BANGER SCREEN"
once your post picks up a little early traction, code sends it to AI for a "banger initial screen".
the model is Gemma 4, not Grok ๐ (the config names it gemma4-26b, a 26-billion-parameter open model from Google).
what gets screened:
- original posts that got some early traction
- posts pulled back in by a priority stream or a recovery stream
what skips it:
- replies
- posts from private accounts
the code doesn't say how much traction counts as "a little". that number isn't in the repo.
2. IT READS BEYOND JUST THE POST
the model gets two things side by side: you, and the post.
reads:
- the text
- media - images and/or video (they're loaded before the model runs)
- your name, @, bio and location
what it doesn't get on this screen: your follower count, your account age or your subscription. those fields exist, and they're switched off here.
so your bio sits next to every post it reads. if your bio says what you post about, the model has that context every single time.
3. IT WRITES A FILE ON YOUR POST
here's everything on that file:
- a short description of it
- tags
- topics it belongs to, each with a score, picked from ๐'s own topic list
- a "high quality?" yes or no
- flags for spam, nsfw, gore and violence
- a slop score ๐ซ
- whether Grok can edit your image
- a score for whether a minor appears in it
๐'s topic list has 32 parent topics:
sports, anime, celebrity, music, news, business & finance, crypto, technology, science, gaming, movies & TV, travel, food, health & fitness, memes, art, fashion, religion, shopping, cars, aviation, motorcycles, beauty, nature & outdoors, pets, relationships, home & garden, career, dance, education, podcasts, streaming
some of them split into subtopics (sports has the most, then music, technology, news, business & finance, food, health & fitness). the subtopic names are fetched live, so they aren't in the repo.
4. WHERE THAT FILE GOES
the flags go to a labeling system.
the rest gets saved to your post's record.
and the topics don't just sit there. the feed side of the code reads them:
- your viewers' followed topics get loaded into their feed request
- the topic labels on posts in someone's feed get counted, and the 2 most common get picked for topic feedback. that one ships switched off by default
- there's a "popular posts by topic" feed built too (top 5 per topic). it's switched off right now
what the code doesn't show:
how it decides "banger" vs. "slop".
the prompt isn't in the repo, and nothing in the feed code reads the slop score. so there's no proof yet that it changes where you rank in For You.
what's your niche btw?
so.... your niche might not even be what you say or think it is ๐
it's the topics Gemma files your posts under.
make it easy to file you:
- pick 1 or 2 of the 32 topics you actually want to be known for
- say the topic in plain words in the post itself, not only in the image
- let your images show the same thing your words say
- put it in your bio, since the bio gets read next to every post
and skip chasing the slop score. nobody outside ๐ knows what raises or lowers it yet.
this part of the code is open source. have your llm/agent read it, learn it, understand it then guide you through it: github.com/xai-org/x-algorithm
hope this helps โก
1:1 call ยท matches this post
Content & Branding Strategy
want 1-on-1 help getting clear on your niche, sharper positioning and a bio that says it in one line? That's the Content & Branding Strategy Call.
