Musk Submits X Algorithm Homework, Releases Ad Blending Code Four Months Later

According to Dongcha Beating monitoring, after Musk promised to open-source the X new algorithm in January 2026, the X official algorithm repository saw its second major commit on May 15, 2026. Compared to the initial version on January 20, this update was significantly larger in scale, involving a total of 187 files with 18,263 insertions and 926 deletions. The core increment shifted from "Explainable Recommendation Architecture" to "Completing Executable Inference Pipeline and Ad Insertion Logic".

The most critical change is that the Phoenix recommendation model finally has an end-to-end demo. The newly added phoenix/run_pipeline.py can run recall and ranking in sequence based on the exported checkpoint, user behavior sequences, and precomputed corpora: first, it identifies candidate content based on user history, then predicts interaction probabilities such as likes, replies, forwards, and dwell time, and finally synthesizes ranking scores. This is a step closer to the actual recommendation process compared to the January version, which only provided descriptions of the retrieval and ranking modules.

This update also introduced a mini Phoenix model artifact of about 3GB for out-of-the-box demo inference. However, there is a parameter conflict in the repository documentation: the root README states a 256-dimensional embedding and 2-layer Transformer, while the Phoenix document and parameter table specify a 128-dimensional embedding and 4-layer Transformer. The specific configuration should refer to the config.json file unpacked from the artifact.

A more notable aspect is the advertisement section. In January, Musk promised to open-source the code related to natural content and ad recommendation, but the initial version lacked details on ad insertion. The May update filled in the gaps with home-mixer/ads/, revealing that ad insertion is not a fixed position insertion but is influenced by safety margins, adjacent content risks, author accounts, keywords, and brand safety rules.

In addition, X has also introduced a new grox/ content understanding pipeline, covering junk content identification, post categorization, policy security assessment, and multimodal embedding. Overall, this update truly complements the recommendation system's peripheral production pipeline: how candidates are sourced, how ads are inserted, how security is enforced, and how results are feedback. It is still not a complete production code, but it resembles more of a X For You recommendation system sample that researchers can dissect compared to the January version.