The posts that hold your attention the longest are about to get a bigger push on X. On Tuesday, October 20, Elon Musk announced two updates to the platform's timeline recommendation system: a new AI ranking model built to predict more accurately how long users will spend on a post, and a new retrieval model trained to surface posts that people read all the way through.
The two changes work at different stages of the feed. Retrieval is the step that gathers candidate posts for a timeline, and the updated model favors posts with high completion — the ones readers tend to finish rather than abandon halfway. Ranking then decides the order, and its new model is built around dwell time, the industry term for how long someone lingers on a piece of content. In plain terms, a post that keeps people reading now has two new ways to earn its place, beyond the familiar counts of likes, reposts and replies.
According to Crypto Briefing, the updates sit inside X's broader ranking framework, known as Phoenix, which runs on a transformer architecture adapted from Grok, the AI model tied to Musk's xAI venture. Phoenix does not give a post a single verdict; it calculates probabilities for a range of possible user actions and folds them into one overall engagement score that determines where the post lands. X shipped major updates to the model in January, May and August of this year, and on August 13 the company open-sourced its ranking weights, filters and model-training code under the Apache 2.0 license.
The direction fits a strategy X has described since 2023: optimizing for what the company calls unregretted user seconds — time people spend on the platform that they do not later wish they had spent elsewhere. Predicting dwell time and rewarding completion are both attempts to measure that idea directly, rather than treating a quick like or an angry reply as proof that a post was worth showing.
For creators, the practical takeaway is straightforward, even if the details are not. Posts that are built to be finished — a clear thread that pays off, a breakdown that answers its own opening question, a story told in full instead of teased — are the kind the new retrieval model is trained to find. Whether dwell time is a good stand-in for quality is a separate question: a post can hold attention because it is useful, and it can also hold attention because it is baffling or provocative. The models, as announced, measure the lingering, not the reason for it.
Musk paired the announcement with a striking claim about the platform itself, saying that active user seconds on X are growing exponentially and hitting new all-time highs nearly every day. That claim comes from Musk himself, and, as Crypto Briefing noted, no independent figures accompanied the announcement to verify the scale of the growth. What can be verified is the change in the machinery: from now on, finishing the post is part of how the post gets found.