X For You Feed Algorithm

gist

X’s open repository describes how its For You feed combines recent posts from followed accounts with out-of-network retrieval from Phoenix and SimClusters. Phoenix predicts a viewer’s probability of taking several actions; weighted predictions form a ranking score, and a separate visibility pipeline decides what can appear. The capture covers the loaded README, including its documented omissions, rather than the repository’s complete code and history.

ideas

• Retrieval mixes social and learned candidates. Thunder supplies recent posts from followed accounts, while Phoenix and SimClusters find posts outside that network.
• Weights scale predictions rather than engagement counts. A report weight multiplies one viewer’s predicted probability of reporting; it does not cancel a fixed number of likes.
• Ranking and visibility are separate systems. One pipeline orders candidates, while label-driven rules allow, hide behind an interstitial, or drop each post.
• Reranking adds diversity. Repeated-author decay, an out-of-network discount, a new-author boost, and VMRanker adjust the initial weighted score.
• Open code has an explicit boundary. Some anti-gaming rules, prompts, and deployment infrastructure remain unpublished, while an account-label report exposes selected outcomes.