Every time you open ,
an AI picks ~50 posts
out of thousands.
Here's exactly how it chooses — in plain English. No math. Eleven interactive sections, ~5 minutes.
In one breath.
Two buckets of posts.
One tab uses both.
Switch tabs to watch which posts show up. The For You tab mixes people you follow with strangers an AI picked for you. Following uses only the first.
“Thunder”
“Phoenix Retrieval”
10,000 candidates.
50 survive.
From the moment you tap the app, the algorithm shrinks a giant pool stage by stage. Drag the slider to step through.
Posts that never even get scored.
Before any AI runs, these get dropped from the pile. If your post hits any of these for a given viewer, it's invisible — no matter how good it is.
19 ways you might react.
That's the entire AI.
For every post, the model predicts how likely you are to do each of these things. Hover any one to see what it does. Weights are illustrative— exact production values aren't in the public release.
Like
Default engagement. The cheapest positive signal — fast and frequent.
One report can outweigh many likes. The AI doesn't have a notion of "quality" — only "engagement likely". It's predicting what people like you will do, not judging your post.
Tap reactions.
Watch the score move.
The 19 predictions get added with weights into one number. Higher = more screens. Here's a live version. One report ≈ many likes lost.
ILLUSTRATIVE — EXACT WEIGHTS NOT IN THE PUBLIC RELEASE
You can't dominate.
Each post by the same author
counts less than the last.
Even if all your posts are great, only the first one or two get full credit per viewer. This is hard-coded — the AI doesn't get to override it.
By post 3, you're at ~⅓ effective score. Each one cannibalizes the last.
Video gets extra credit
If your post has a video longer than a threshold, it gets an extra scoring signal (video-quality-view). Longer ≠ trivially better — but trivial clips don't qualify.
The system has a tutorial mode
If you have few interactions (~<100), your feed is routed to a different cluster — more topic-based, less personalised. First-time users get training-wheels ranking.
From "Post" to feed.
About 7 stops, ~50 ms.
Press play and follow a single tweet through the system.
You hit Post
Your post is published. The clock starts.
Same 12 posts.
Three different feeds.
Watch the same set of posts get re-ordered as filters and ranking turn on. Tap a mode.
Will this post go viral?
Answer 8 questions.
A diagnostic, not an oracle. Each yes/no maps to a behaviour the algorithm rewards or punishes.
Your followers will mostly see it. Out-of-network reach modest.
What the open code
does NOT show.
Things people assume but the public release doesn't support. Tap to expand.