The open-source X algorithm

The X algorithm repo was opened up, and it showed how the For You feed really works. No secrets, just code. As creators, here's what stood out from the Phoenix model and the ranking logic.

The algo is purely machine-learning driven. Grok's Phoenix transformer predicts your actions (like, reply, repost, video view) based on past behavior, then scores every post. No manual boosts or hidden rules. It's learned from data. For us: focus on content that sparks real interactions, not tricks. High replies and reposts mean conversations, and conversations are rocket fuel.

Videos, GIFs, and images get explicit and improved scoring. The model checks the media details (duration, type) for better ranking. Lean hard into dynamic visuals. Loops and GIFs are perfect. Static posts or low-motion videos underperform. Motion keeps people longer, and that lifts the score.

Early signals matter hugely. The algo uses initial likes and replies to decide if a post escapes your followers' bubble into out-of-network feeds, via semantic similarity. Reply fast (within hours), encourage challenges, and use quotes and threads to build momentum. Your people's quick responses are what push impressions beyond your circle.

Dwell time. Threads increase dwell time (people scroll through), which the model loves. Author diversity filters prevent repetition, so vary your series. Keep a daily series if you want, but mix humor, lore, and visuals. Variety signals quality.

Bottom line: the algo rewards authentic engagement over volume. Quality replies, media, and conversations win. No need for hacks. Create what your people love and interact with. Test, iterate, and watch the analytics.