X’s May 2026 feed stack uses Grok-1 and Grox for ranking
A GitHub repository details a technical breakdown of the May 2026 X "For You" recommendation algorithm, which utilizes the Grok-1 transformer for engagement prediction and a distributed async Python daemon (Grox) for real-time LLM-powered content moderation and multimodal embedding extraction. The documentation covers the system's architecture, including the Phoenix ML Engine, Rust Candidate Pipeline, Home Mixer, and Grox AI Daemon, and provides specific algorithmic playbooks.
Key Takeaways
- The repository covers 207 source files, including 139 Rust files and 68 Python files.
- X’s feed stack is described as four subsystems: Phoenix ML Engine, Rust Candidate Pipeline, Home Mixer, and Grox AI Daemon.
- Phoenix ML Engine uses a two-tower retrieval system and transformer ranking with a “candidate isolation” attention mask.
- Grox AI Daemon is described as a Kafka-driven, multi-process Python engine using VLMs for policy checks, quality scores, and text/video embeddings.
- The cheat sheets include playbooks for video reach, safety filtering, reply ranking, network reach, and banger quality scoring.
Why It Matters
This is a concrete map of how X’s recommendation feed is assembled: ranking, hydration, ad blending, storage, and moderation are split across named services rather than a single black box. For streaming and video teams, the relevant detail is that the document ties multimodal embeddings and real-time moderation directly into the feed path, not as a separate after-the-fact layer. It also shows how much of the system is expressed in Rust and Python across 207 source files. What to watch next: whether the repo’s later chapters, especially the Grox and Phoenix inference sections, add more detail on the 19 engagement weights and the 18 post-retrieval filters.
Read full article at github.com