X updates algorithm with Phoenix model

X’s public recommendation code changed again on May 15, and the new version gives a clearer look at how the platform says it now builds the “For You” feed. GitHub shows fresh commits to xAI’s `x-algorithm` repository, including updates in the `phoenix` directory, model files and runnable pipeline code. The repository says X combines posts from followed accounts with posts found outside a user’s network, then ranks them together with Phoenix, which it describes as a Grok-based transformer model. ### What exactly did X publish this time? GitHub’s repository page says the code is the “core recommendation system powering the ‘For You’ feed on X,” and the latest visible commit landed hours before review. The updated tree includes `candidate-pipeline`, `home-mixer`, `phoenix` and `thunder`, alongside README files that describe how the system retrieves, ranks and filters content. The main README says the feed pulls from two sources: “In-Network (Thunder),” which covers posts from followed accounts, and “Out-of-Network (Phoenix Retrieval),” which surfaces posts from a broader corpus. (github.com) The same document says those sources are merged and ranked together by Phoenix. ### Where does Phoenix fit in the feed? The `phoenix` README says Phoenix handles both retrieval and ranking. It describes a two-stage system: first narrowing millions of candidates to hundreds through approximate nearest neighbor search, then ordering those candidates with a transformer model that predicts engagement. (github.com) The repository also says the model is “representative of the model used internally” except for scaling optimizations. (github.com) In the open-source release, xAI says it published a smaller version — a 128-dimension, four-layer transformer — trained on the same real-time engagement data as the production system, while production uses a larger model. ### Did X really replace older hand-built ranking features? (github.com) The top-level README says the final score is a weighted combination of predicted engagement actions, and adds that X has “eliminated every single hand-engineered feature and most heuristics from the system.” The document says the transformer uses engagement history such as likes, replies and shares to judge relevance. That language marks a change from the older public `twitter/the-algorithm` repository, which documented a broader collection of ranking inputs and support systems including user reputation scoring through Tweepcred, graph features, trust-and-safety models and representation scoring across users and posts. (github.com) The older repository remains public and still describes the For You timeline as part of a larger shared recommendation stack. (github.com) ### Why are creators saying follower counts and “bangers” matter more? The code visible in the public repository supports part of that claim, but not all of the creator shorthand. X’s README says ranking is driven by predicted engagement probabilities and weighted combinations of those predictions, which is consistent with creators describing the system as favoring posts likely to generate strong interaction. The repository pages reviewed do not, by themselves, provide a simple public line that says follower count is now directly prioritized above all else. (github.com) Creators and independent readers may be inferring that from supporting features, retrieval behavior or code changes beyond the README summaries, but that interpretation would require line-by-line code review or direct comment from X. ### Is Phoenix actually based on Grok? (github.com) X’s own repository says yes, in narrow technical terms. The main README says the feed is ranked with a “Grok-based transformer model,” and the Phoenix README says the sample transformer implementation is ported from xAI’s Grok-1 open-source release and adapted for recommendation tasks with custom embeddings and candidate-isolation masking. That does not mean Grok is answering users inside the feed ranking system in the same way as a chatbot. (github.com) The documentation describes an adapted transformer architecture used to predict engagement and order content, not a conversational assistant embedded in the timeline. ### What can readers check next? The `xai-org/x-algorithm` repository remains public on GitHub as of May 16, with the updated `phoenix` folder, artifact files and README documentation still accessible. (github.com) Any further changes should appear in the repository’s commit history, where GitHub lists the latest update time and file-level revisions.