Nav Toor@heynavtoor
Elon has open-sourced the algorithm that decides what you see on X.
Not a blog post explaining it. Not a press release describing it. The actual source code. With the actual model weights. On GitHub. Every line. Every signal that determines whether your post reaches 100 people or 100 million.
This is the most powerful recommendation engine on Earth. And now you can read it.
Here is what the algorithm actually looks at:
→ Fifteen actions. The model does not score "relevance." It predicts 15 separate probabilities for every post: favorite, reply, repost, quote, click, profile click, video view, photo expand, share, dwell, follow the author, not interested, block author, mute author, report. Your final reach is the weighted sum of all 15.
→ Negative actions hurt more than you think. Block, mute, report, and not-interested are weighted negatively. Every single mute on your post subtracts from your score. A post that triggers one mute and ten likes does not net ten likes. It nets ten likes minus one heavy penalty.
→ Bookmarks are not tracked. Bookmark is not in the 15 predicted actions. Every creator guru telling you to "write bookmark-worthy posts" is wrong, at least according to this released code. Bookmarks may matter for the user privately, but they do not feed the ranker.
→ Dwell time is a silent vote. If someone stops scrolling and lingers on your post without engaging, that counts as a positive signal. The model has a P(dwell) prediction. The algorithm knows when people pause.
→ Profile clicks are tracked separately. P(profile_click) is its own prediction. People clicking from your post to your profile is a distinct positive signal. Your post sells the click, your bio closes the follow.
→ Quote is worth more than repost. The model predicts P(quote) and P(repost) as two different actions with two different weights. A quote tweet, where someone adds their own commentary, carries information a silent repost does not. The algorithm rewards conversation, not amplification.
→ Follow-after-engagement is tracked. P(follow_author) is one of the 15 predictions. A post that converts a reader into a follower is rewarded directly by the ranker. Not because of a hardcoded rule, but because the model learned the signal.
→ The Author Diversity Scorer punishes spam posting. There is a specific component in the code called Author Diversity Scorer. Its job is to attenuate repeated scores from the same author. Posting twenty times a day cannibalizes your own reach. The algorithm enforces diversity automatically.
→ Candidate isolation. Each post is scored alone against your history. Posts cannot see each other during scoring. Your post is not competing with the post above it in the batch. It is competing with the user's taste model.
→ No hand-engineered features. There is no rule in the code that says "boost verified" or "suppress external links" or "promote subscribers." The Grok transformer learns everything from raw engagement sequences. Whatever the model has learned, it learned from user behavior, not from a hardcoded boost.
→ Out-of-network discovery. The Phoenix retrieval system uses a two-tower model. One tower encodes you. One tower encodes every post on X. They meet in a vector space. The closest matches surface in your For You feed even if you do not follow the author.
→ Visibility filters run last. After ranking, a final pass strips spam, violence, gore, and deleted posts. Ranking decides reach. Visibility decides whether you see it at all.
Here is what this means for you:
If your posts are not reaching people, it is not because the algorithm is broken. It is because the algorithm is working exactly as designed. It rewards:
1. Posts that get reactions across multiple action types (a like AND a profile click AND a follow beats five likes alone)
2. Conversation depth (quote tweets are worth more than reposts in the math)
3. Dwell time (write posts people stop to read)
4. Posts that convert viewers into followers (your bio is part of your post)
5. Variety from each author (post less, post better)
And it punishes:
1. Mutes
2. Blocks
3. Reports
4. "Not interested" clicks
5. Same-author saturation
For twenty years, the algorithm was a secret.
