New Delhi: On January 20, X quietly did something no major social platform has fully dared before. It released the core logic of its new recommendation algorithm as open source. The announcement was followed by a blunt comment from owner Elon Musk, who admitted the system is ‘dumb’ and still evolving but argued that public transparency itself is the point.
The move is being positioned as radical openness, but it also signals a deeper shift. This is not a cosmetic update or a partial reveal. X has rebuilt its ranking engine from scratch. The rewrite fundamentally changes how content is judged, promoted, or buried across the platform.
We have open-sourced our new 𝕏 algorithm, powered by the same transformer architecture as xAI’s Grok model.
Check it out here: https://t.co/3WKwZkdgmB https://t.co/nQ5GH1a42e
— Engineering (@XEng) January 20, 2026
From manual rules to AI-driven judgment
X previously open-sourced its algorithm in 2023. That version, known as Heavy Ranker, relied on traditional machine learning. Engineers manually defined hundreds of features post age, follower count, presence of images, links, and more. Each feature was assigned a weight. The system worked like a spreadsheet with endless tuning.
The new version is called Phoenix. It abandons feature engineering entirely. According to X’s own documentation, every hand-crafted rule has been removed.
At its core is Grok, a transformer-based model similar in architecture to systems like ChatGPT. Instead of scoring posts based on what they contain, Phoenix predicts how each individual user is likely to react.
Algorithm no longer judges posts-It judges people
Phoenix does not ask whether a post is well written or optimised. It asks a single question: What will this user do when they see it?
To answer that, the model analyses a user’s behaviour history. Likes. Replies. Time spent reading. Profiles visited. Accounts muted or blocked. From this sequence, the model predicts 15 possible reactions to any post.
These include positive actions such as liking, replying, reposting, profile visits, long dwell time, and follows. They also include negative actions like “not interested,” mute, block, and report.
Each reaction is assigned a probability. The algorithm then multiplies those probabilities by internal weights and adds them up. The result is a final ranking score.
Higher score, higher reach. Lower score, less visibility.
Why engagement now matters more than virality
The logic itself is simple. But the implications are not.
Under this system, a post that triggers strong reactions—especially replies—can outperform a polished post that people scroll past quietly. The algorithm values interaction, not approval.
Old data from X’s 2023 release shows how extreme this can be. A single meaningful reply exchange was worth more than 100 likes. One report could outweigh hundreds of positive signals. While the exact weights are no longer public, the framework remains unchanged.
This explains why outrage-driven posts sometimes spread—and why they can also permanently damage an account’s reach.
What the code tells creators to do differently
The open-source repository reveals several practical lessons for creators.
- Replying to commenters is critical. Author replies carry the highest positive signal. Even a short response boosts distribution more than dozens of likes.
- Avoid pushing users toward muting or blocking you. Negative actions carry heavy penalties and affect future posts as well, not just the one that triggered them.
- External links should be placed in comments, not the main post. The system deprioritises posts that send users off-platform.
- Posting too frequently can backfire. Phoenix includes an author diversity system that downranks consecutive posts from the same account.
- The old obsession with “best posting time” is effectively obsolete. Time-based features no longer exist in the model.
What X didn’t reveal and why it matters
Despite the openness, key elements remain hidden.
The exact weights for each action are not disclosed. The internal parameters of the Grok model are closed. The training data is unknown.
In short, X has revealed the structure of the system, not its inner values.
Even so, this is more transparency than any rival platform has offered. TikTok and Instagram reveal almost nothing at this level. X has shown how ranking decisions are made—even if it hasn’t shown every number behind them.
For creators, the message is clear. Optimisation tricks are fading. Behavioural impact is everything. The algorithm no longer asks how you post. It asks how people respond.
And now, for the first time, you can see that logic in plain sight.