X Open-Sources Its Algorithm: Copying Links Matters 40x More Than Likes
The mysterious math behind the "For You" feed is finally out in the open. X dumped its actual production code on GitHub, exposing exact ranking multipliers and confirmation of stealth shadowbans.
The production recommendation codebase for the "For You" timeline has landed on GitHub under the Apache License 2.0, exposing the inner workings of the Phoenix engine, its Rust backend, and the exact scoring formulas used to dictate feed visibility.
When a user opens their feed, neural transformers predict the statistical likelihood of specific actions. In the published scoring hierarchy, copying a link to share it externally carries a massive +20 multiplier, making a copied link forty times more impactful than a basic like, which sits at a humble +0.5. Reposting awards +1, replies and quote posts grant +5, while clicking an outbound link gets throttled down to a modest +0.2.
Negative reactions inflict immediate mathematical damage. A report slashes the score with a -234 penalty, muting an author subtracts -58.8, and marking a post as uninteresting deducts -43.2. Rather than tallying raw complaints, the algorithm multiplies these figures by the predicted probability of an action, preventing organized brigading through direct chat links from sabotaging organic reach.
To prevent feeds from becoming stale monoliths, consecutive posts from the same author face an aggressive decay factor down to 0.25. Meanwhile, a dedicated New-Author Boost automatically elevates creators with fewer than 1,000 followers and under 1,000 impressions straight toward the 15th and 16th feed slots during their first 24 hours.
The code also details the mechanics of audience throttling through its Visibility Filtering layer. System tags like Spam High Recall, Civic Integrity warnings, and emergency suppression flags quietly cut posts out of non-follower feeds while keeping them visible on the author profile. Beyond engagement metrics, the scoring pipeline prepares data points including geographic coordinates, inferred user gender, and lists of apps installed on the device.
Feeding personal device inventories and inferred gender markers into a predictive pipeline just to rank provocative posts shows that modern digital discourse is entirely governed by cold probabilistic formulas.
Source: TechCrunch
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.