X Open-Sourced Its Algorithm Again — Here's What Actually Changed

On May 15, 2026, X pushed the largest update yet to its open-source recommendation algorithm — over 18,000 lines of new code across 187 files. If you write on X and haven't looked at any of it, you're not behind; almost nobody has. But a few things in this release are worth knowing, because they explain a lot about why some posts take off and most quietly don't.

What's actually in the repo

The system is called Phoenix, and it's built on the same transformer architecture as xAI's Grok models. Instead of the old method of hand-coded rules deciding what counts as a "good" post, the new system reads every post and watches every video, then predicts how likely you are to engage with it — across roughly 15 different types of engagement, not just likes.

This latest release added a runnable end-to-end pipeline, meaning anyone with a terminal can now actually run a small version of X's real ranking model locally, not just read about it. It also shipped a new content-understanding service that screens for spam, safety, and what the code reportedly refers to as "banger" quality — yes, that's apparently a real internal label.

One honest caveat: the architecture is public, but the exact production weights — the precise numbers that decide how much each signal matters — are not. What you're seeing is the structure of the engine, not the dial settings X is currently running. Treat any specific multiplier you read online as a well-informed estimate, not a confirmed production constant.

The part that matters most for writers: replies

If there's one consistent theme across everyone who's dug through this code, it's that replies are worth dramatically more than likes. Multiple independent breakdowns of the scoring formula put replies at over 13 times the weight of a single like, with retweets weighted around 20 times. The single highest-value action, according to several analyses, is a two-way reply chain — someone replies to your post, and you reply back to them.

This reframes what "good content" even means on X. A post that gets quietly liked by 500 people is, in the algorithm's eyes, worth less than a post that starts an actual back-and-forth with 20 people. If you've been optimizing for likes, you've been optimizing for the wrong number.

What changed besides the ranking model

A few other things in this release are worth a mention: an integrated ad-blending system was added (how promoted content gets woven into the same ranking pass as organic posts), and a batch of new "hydrators" — components that enrich a post with extra context like author history and media details before it gets scored.

xAI has also kept a public promise to update this repo roughly every four weeks, and the cadence has actually held since January. If you want to keep half an eye on how the rules are shifting, that repo is genuinely the most direct source available — more reliable than secondhand summaries, including this one.

So what do you actually do with this

Nothing here demands a complete rewrite of how you post. But it's a useful nudge to stop treating likes as the goal. If a post isn't generating any replies, it's probably not getting much algorithmic credit even if it's racking up likes — and the fix isn't "post more," it's "post something someone feels compelled to respond to."

---

*If you're trying to write posts built around starting a conversation rather than just sounding good, TweetGem has a reply-bait style mode built specifically for that.*