How X Predicts Likes, Replies, Reposts & Clicks
When X decides whether a post should be recommended, it doesn't simply look at how many likes or reposts the post has already received. The current open-source X recommendation system uses Phoenix, a Grok-based transformer model, to predict the probability of different actions that a particular user might take on a candidate post. Those predictions are then used by the ranking system to compare candidates. (GitHub)
In simple terms: X tries to predict what you are likely to do with a post—not just what other people have already done with it.
The Prediction Pipeline
The current Phoenix implementation exposes separate prediction outputs for favorite/like, reply, repost, photo expansion, click, profile click, video-quality view, share, dwell, quote, follow and several negative-feedback actions. (GitHub)
1 to 7. Multi-Action Prediction and Specific Outputs
1. X Predicts Actions, Not Just Engagement: Phoenix produces separate probabilities for many possible actions instead of one generic relevance value.
2. How X Knows What You Might Do: Phoenix uses the user's engagement history and context when evaluating candidates, incorporating embedding representations for users, historical posts/authors, and candidates.
3 to 6. Predicting Likes, Replies, Reposts, and Clicks: Separate model outputs (`favorite_score`, `reply_score`, `repost_score`, `click_score`) allow the system to evaluate distinct user intentions independently.
7. Full Prediction Table: The Phoenix runner exposes scores for photo expansion, profile clicks, video quality views, direct message shares, copy-link shares, continuous dwell time, quotes, and author follows.
8 to 22. Negative Feedback, Personalization, Isolation, and Creator Takeaways
8. Positive and Negative Predictions: Negative-feedback actions like "Not Interested", "Block", "Mute", and "Report" are predicted alongside positive interactions.
9 & 10. User History & Context: Engagement sequences allow the model to tailor predictions to each viewer's specific background and interests.
11. Candidate Isolation: Candidate posts cannot attend to each other during transformer inference, making candidate scores consistent and cacheable. (GitHub)
12 to 22. Practical Takeaways for Creators: Visible engagement counts do not equal recommendation scores. The model estimates future viewer behavior, making content relevance for the right audience far more important than superficial metrics.