On January 20, 2026, Elon Musk followed through on a long-standing pledge by open-sourcing the core recommendation algorithm behind the đ (formerly Twitter) âFor Youâ feed. The release confirms a major architectural shift: the platformâs ranking logic is now dominated by Transformer-based neural networks, closely aligned with xAIâs Grok model, rather than hand-tuned heuristic rules.
The full codebase is publicly available on GitHub, offering an unprecedented look into how one of the worldâs largest social platforms curates attention in real time.
𧊠System Architecture: Thunder and Phoenix #
The open-source repository reveals a four-stage ranking pipeline capable of evaluating thousands of posts in under 200 ms.
Thunder: In-Network Content #
- Purpose: Surfaces posts from accounts you already follow.
- Design: A high-speed, in-memory system optimized for sub-millisecond lookups.
- Goal: Guarantee immediate visibility of followed accounts without expensive database queries.
Phoenix: Out-of-Network Discovery #
- Purpose: Finds relevant posts from the global content pool.
- Two-Tower Retrieval Model: Encodes user interests and candidate posts into vectors, selecting matches via dot-product similarity.
- Grok-Based Ranking Transformer: Processes historical interaction sequencesâlikes, replies, sharesâto predict the probability of future engagement.
Together, Thunder and Phoenix merge social graph awareness with large-scale neural ranking.
đ Engagement Signals That Shape Your Feed #
Each post receives a Final Score, computed as the probability of an action multiplied by a weighted importance factor.
| Positive Signals | Negative Signals |
|---|---|
| Replies (â75Ă a like) | âNot Interestedâ feedback |
| Profile clicks & new follows | Muting or blocking the author |
| Video watch time & image expands | Reporting content |
| Dwell time (reading without scrolling) | Low reputation score (â128 to +100) |
| Direct-message shares | Topic fatigue from repetition |
This weighting system strongly favors conversation depth over passive consumption.
đ§ Algorithmic Insights from the Source Code #
Several newly documented rules clarify how visibility is earnedâor lostâon đ:
- Conversation Dominance: Reply chains, especially when authors respond, are the most powerful engagement signal.
- External Link Suppression: Posts with outbound URLs are systematically downranked to reduce off-platform traffic leakage.
- Hidden Reputation Score: Every account carries an internal quality score; interacting with low-quality accounts can reduce reach.
- Anti-Spam Decay: Posting too frequently triggers diminishing returns, encouraging fewer but higher-quality posts.
These mechanisms collectively push the platform toward fewer viral bursts and more sustained discussions.
âď¸ Why Open Source Now? #
The timing is not accidental. In December 2025, the EU fined đ âŹ120 million under the Digital Services Act for insufficient algorithmic transparency. By publishing the code, Musk aims to demonstrate political neutrality while inviting external scrutiny and improvement.
đ has committed to updating the repository every four weeks, allowing regulators, researchers, and developers to observe how the platformâs recommendation logic evolves in near real time.
In effect, this marks the first time a major global social network has exposed the mechanics of attention at this levelâturning its recommendation engine from a black box into a living, inspectable system.