I will organize the information into "What is written in the code," "What is only in the README," and "What is written nowhere."
Why I wrote this article
On May 15th, Elon Musk announced that the X algorithm was released on GitHub. Immediately after, a large number of posts about "how to beat the new X algorithm" began to circulate. What bothered me while reading them was that some posts asserted specific original rules not found in the code, while claiming to have "read the code." I decided to check for myself and had (an AI) read the GitHub code. In this article, I will organize what I could and could not confirm.
Structure of this article
I will write the information in three layers:
- Things directly confirmed in the code
- Things written in the README but not confirmed in the code (at the "according to official documentation" level)
- Things spreading on X that are written in neither the code nor the README
Things confirmed in the code
This update was a massive one, with over 187 files and 18,000 lines added. In the four months since the initial release in January, the content has changed significantly.
Grox now "sees" posts to judge quality
The biggest change this time is the complete establishment of a content review service called Grox. It did not exist in the code released in January (though I cannot say for sure that a similar function didn't exist in the live X environment). At the core of Grox is a VLM (Vision Language Model). A VLM is an AI that can understand not just text, but also images and video frames together. It doesn't just "read text"; it can "look at images and make judgments." Grox mainly does three things:
Post Quality Scoring (Banger Screen)
A "Banger Score" is a score from 0 to 1 where the AI rates whether a post is "valuable content likely to go viral." The code contains a threshold of "score >= 0.4," where 0.4 or higher is judged as positive.
Spam Detection
The VLM determines whether a post is spam, focusing on posts from accounts with low follower counts.
Safety and Policy Violation Detection
Grox judges the following seven categories:
- Violent media
- Adult content
- Spam
- Illegal/regulated activities
- Hate/harassment
- Violent speech
- Suicide/self-harm
In the code, there is also a mechanism to perform additional checks on violent media and adult content using higher-precision inference models. However, which posts are subject to this cannot be confirmed from the public code alone.
Mechanisms that existed since the January release
From here on, these are not "new features" but mechanisms that have existed since the initial release in January. Many "latest algorithm explanations" on X mix the January and May code without distinction. If you just have an AI summarize GitHub, features from four months ago will be introduced as "changes this time."
Blocks, mutes, and reports lower the score (Since January)
From the structure of the code, it was confirmed that the following actions are incorporated into the calculation formula as negative signals:
- Not interested
- Block
- Mute
- Report
Also, the ranking scorer (ranking_scorer.rs) in the Home Mixer includes a process to suppress bias so that posts from the same account do not appear too often in "For You." This is also a mechanism that has existed since January; when multiple posts from the same account appear in one person's feed, the score for the second and subsequent posts drops exponentially.
This is to ensure diversity—"not filling one person's timeline with your posts"—and does not limit the total number of posts per day.
Things written in the README but not confirmed in the code
"Almost all hand-crafted rules abolished" is half true
The official README states a design principle: "We have almost entirely abolished hand-crafted features and heuristics." This is not a description added in May; it is a claim that has existed in the README since the initial release in January.
However, looking at the code, the filtering layers (excluding old posts, excluding muted keywords, etc.) remain rule-based.
What was "abolished" were the criteria for ranking and scoring, not filtering. And this structure has not changed in this update.
Changes officially recognized in the May README
A dedicated section titled "Updates — May 15th, 2026" was newly established in the May 15th README, and the changes this time were clearly stated in seven items. The additions of Grox, Ad integration, Query hydrators, Candidate hydrators, and Candidate sources are all listed here.
In other words, the README answers "what are the new features this time." Mechanisms that existed since January do not appear in this section.
"How" the Banger Score affects ranking is not public
"Positive judgment at 0.4 or higher" was confirmed in the code. However, the code showing how that result affects ranking was not included in the repository released this time. We cannot assert at this point that "below 0.4 will not be shown" or "0.4 or higher will be boosted."
Engagement weight values may not be production values
The following values are written in the inference pipeline for demo purposes:
- Like: 1.0
- Reply: 0.5
- Repost: 0.3
- Dwell time: 0.2
However, the production-side code is managed with dynamic parameters (supporting A/B testing), and the values are not hardcoded. We cannot confirm if these values are the actual values in production.
Things written in neither the code nor the README
The following information, which is spreading on X as "how to beat the new algorithm," was not written anywhere in the code:
- Reach decreases if you post more than 4 times a day
- Replying within 30 minutes of posting causes a conversation chain explosion
- 3-5 images are best
- Videos of 30-90 seconds are particularly strong
The reason this information spreads is that people mix the fact that "the algorithm changed on May 15th" with their "empirical rules from their own operations" in the same post. Because readers receive it under the authority of "I found out by reading the code," personal opinions look like algorithm specifications.
Summary
If I were to summarize the May 15th update in one sentence, it would be this:
"The ranking mechanism" has hardly changed. "The mechanism for reading post content" was added in its entirety.
The core of whose posts to show in what order—engagement prediction, negative signals, and ensuring account diversity—has not changed since January.
What was added this time is the "eye" part that judges what the post is saying and whether it has value.
Therefore, the answer to questions like "how many times a day should I post" or "within how many minutes should I reply" is not in this code.
There is only the fact that "AI looks at the post and judges whether it is worth reading." Rather than being misled by algorithm rumors and chasing numbers, I think asking "is this a post that leaves something for the reader?" is more suited to the current algorithm.
How much do you believe in "algorithm growth hacks"? What did you think when you actually saw the code?
I started X (formerly Twitter) in 2017 and am now in my 10th year, and I love interacting on X. It frustrates me to no end when people are deceived by someone saying something nonsensical and decide to "limit posts to 4 times a day," which hinders interaction.
Numbers are important, but don't you think interaction is also important? No matter how the algorithm changes, I want to continue working while enjoying interaction on X. I would be happy if you continue to enjoy interacting with me on X.
