Elon Musk resubmits the X algorithm homework and, after four months, finally releases the ad mixed-layout code

robot
Abstract generation in progress
AIMPACT News, May 16 (UTC+8), according to Beating Monitoring from Dongcha, after Elon Musk promised to open source the new X algorithm in January 2026, the official X algorithm repository saw its second major commit on May 15, 2026. Compared to the first version on January 20, this update is significantly larger, involving 187 files, adding 18,263 lines, deleting 926 lines, with the core incremental progress advancing from "explanation of recommendation architecture" to "completing a runnable inference pipeline and ad mixing logic."
The most critical change is that the Phoenix recommendation model finally has an end-to-end demonstration. The newly added phoenix/run_pipeline.py can run through recall and ranking sequentially starting from exported checkpoints, user behavior sequences, and precomputed corpora: first finding candidate content based on user history, then predicting interaction probabilities such as likes, replies, shares, and dwell time, and finally synthesizing ranking scores.
This is closer to the actual recommendation process than the January version, which only provided explanations for the retrieval and ranking modules.
This update also added about 3GB of mini Phoenix model artifacts for out-of-the-box inference examples. However, there is a parameter conflict in the repository documentation: the root README states 256-dimensional embeddings and 2-layer Transformers, while the Phoenix documentation and parameter table specify 128-dimensional embeddings and 4-layer Transformers. The exact configuration should be based on the config.json extracted from the artifact.
An even more newsworthy point is the advertising part. In January, Musk explicitly promised to open source code related to natural content and ad recommendation, but the first version hardly included any ad mixing details. The May update added home-mixer/ads/, showing that ad insertion is not fixed at a specific position but influenced by safety intervals, adjacent content risks, author accounts, keywords, and brand safety rules.
Additionally, X has introduced a new grox/ content understanding pipeline, covering spam detection, post classification, policy safety judgment, and multimodal embedding.
Overall, this update mainly enhances the recommendation system's peripheral production chain: how candidates are sourced, how ads are inserted, how safety is managed, and how results are written back. It is still not complete production code, but it is more like a sample of the X For You recommendation system that researchers can dissect compared to the January version.
(Source: BlockBeats)
View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
  • Reward
  • 2
  • 3
  • Share
Comment
Add a comment
Add a comment
MosaicBow
· 6h ago
The constraints for inserting ads make the actual business ten times more complex than the code.
View OriginalReply0
椰子壳里装Alpha
· 6h ago
It's said to fill in the outer chain, but the core sorting logic probably still hasn't been revealed, right?
View OriginalReply0
  • Pinned