X open-sources For You feed recommendation algorithm
X published the source code for its For You feed ranking system in the xai-org/x-algorithm repository on GitHub. The release makes publicly available the mechanism that selects and orders posts from followed accounts and external sources. Creator accounts on the platform described the move as the first time a major social network has disclosed its core recommendation algorithm.
X has released the latest "For You" feed algorithm to GitHub.
Moving the feed closer to a Grok-style AI ranking system than a rule-heavy social media sorter.
The system first gathers posts from accounts you follow through Thunder, then adds outside posts through Phoenix, which searches for content that looks relevant even if you do not follow the author.
Ranking now behaves less like a checklist of likes, reposts, and keywords, and more like a transformer model that reads the user, the post, and past behavior as one prediction problem.
Instead of asking “is this post good,” the model predicts multiple possible actions, such as whether you might reply, like, repost, watch, or skip.
Thunder is the fast “people you follow” system.
It keeps a live memory-store of recent posts, replies, reposts, and videos from users, then quickly pulls posts from accounts you follow when your For You feed is built. It is basically the in-network candidate source.
Phoenix is the AI recommendation system.
It has 2 jobs: first, Phoenix Retrieval finds outside posts from the global X corpus using similarity search, and then Phoenix Ranking scores both followed-account posts and outside posts using a Grok-based transformer model.

The latest 𝕏 algorithm has been published to GitHub https://github.com/xai-org/x-algorithm
X has released the latest "For You" feed algorithm to GitHub. Moving the feed closer to a Grok-style AI ranking system than a rule-heavy social media sorter. The system first gathers posts from accounts you follow through Thunder, then adds outside posts through Phoenix, which searches for content that looks relevant even if you do not follow the author. Ranking now behaves less like a checklist of likes, reposts, and keywords, and more like a transformer model that reads the user, the post, and past behavior as one prediction problem. Instead of asking “is this post good,” the model predicts multiple possible actions, such as whether you might reply, like, repost, watch, or skip. Thunder is the fast “people you follow” system. It keeps a live memory-store of recent posts, replies, reposts, and videos from users, then quickly pulls posts from accounts you follow when your For You feed is built. It is basically the in-network candidate source. Phoenix is the AI recommendation system. It has 2 jobs: first, Phoenix Retrieval finds outside posts from the global X corpus using similarity search, and then Phoenix Ranking scores both followed-account posts and outside posts using a Grok-based transformer model.
reiterating: anyone claiming specifics about the algo like "4x weighting on replies" is *literally just lying*, none of the information is in the repo at all