X Open Sources Its Algorithm And Reveals What it Takes to go Viral
Trending Topics Jakob Steinschaden
X put its feed-ranking code on GitHub, plus more of the machinery behind it. The surprising part: the weights show likes matter least, while reports hit far harder.
Based on reporting by Trending Topics, Jakob Steinschaden — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
X has opened up a much bigger slice of its recommendation system. The code behind the default “For You” feed is now on GitHub, alongside the ranking engine, under Apache 2.0 and in the xai-org organisation. That release goes well beyond the earlier code drop: model settings, filters and the signal weights used to score posts are now visible too.
The feed itself is built from two streams. Thunder supplies recent posts from accounts a user follows, while Phoenix Retrieval and SimClusters pull in content from outside that circle. Phoenix, a transformer, then looks at a user’s recent activity and estimates how likely they are to like, reply, repost, click, linger, block or report. Those probabilities get combined into a final score, with posts older than 48 hours filtered out before ranking.
The numbers buried in the parameter file are the part most people will zoom in on. A like, called “favorite” in the code, has a weight of 0.5 and is the weakest of the familiar interactions. Replies, quotes and DM shares are all weighted at 5. Copying a link is much stronger, at roughly 20 in absolute terms, while a follow counts as 4 and a repost as 1. Negative signals are far harsher: a predicted report is about 468 times the weight of a like, a mute around 118, “not interested” around 86 and a block around 62.
X is quick to warn that the math is easy to misread. These weights scale predicted probabilities, not raw counts, so a report does not “cancel out” 468 likes. The company says it has added extra comments to stop both humans and language models from making that mistake. And there are other reach levers layered on top: repeated posts by the same author are discounted, out-of-network posts are multiplied by 0.75, and accounts with few impressions get a new-author boost.
The company is also adding a transparency tool called “Under the Hood” in app settings. Users who have posted at least ten times in the past month can download a JSON file with the labels applied to their account or posts over the last calendar month, though the feature is currently limited to a random test group of accounts at least one year old. X says this should make shadowban claims easier to check against the repository itself, even as it withholds some pieces, including the prompts for Grok-based classifiers and some labeling rules.
There is also a political edge to the disclosure. The code includes a filter tied to Brazil’s 2026 elections that removes posts from accounts reported to the country’s electoral court, unless the viewer follows that account. X says developers will even be able to submit pull requests for review. The repository already has around 27,500 stars, and several dozen open pull requests waiting in line.
My take — AI-written commentary, not fact-checked reporting
This is the rare platform transparency move that actually gives people something to inspect, not just a glossy promise. But the withheld bits matter, and they always will: open-sourcing the machinery while keeping the sharpest knives in a drawer is still a very platform thing to do. The broader pattern is obvious enough — the age of “trust us, the algorithm is fair” is over, and X is trying to replace it with “read the code if you can stand it.”
Read more about this at: Trending Topics