X's Grok-Powered Algorithm: The January 2026 Rewrite
X’s recommendation algorithm got a significant rewrite. The core system now uses Grok’s transformer architecture instead of the previous ML pipeline.
The in-network posts, out-of-network ML retrieval, and two-stage ranking system have been rearchitected around Grok's transformer model.
Elon open-sourced the code last week, called it “dumb,” and admitted the algorithm has been flooding feeds with irrelevant junk. Which is exactly why I updated my playbook.
I used Claude Code AI agents to analyze the newly open-sourced code and compare it with the previous algorithm. The agents worked through the implementation directly and updated the guide to reflect what’s actually happening under the hood.
Full disclosure: This analysis and report was generated entirely by Claude Code CLI AI agents. I'm just publishing what they produced.
The engagement signal hierarchy has changed. The ranking logic is different. Even the ad recommendation system is now Grok-powered.
The updated X Algorithm Playbook (https://nibzard.github.io/twitter-algorithm-tufte/) now covers:
- Grok’s transformer architecture replacing traditional ML
- How the new in-network vs out-of-network retrieval works
- Updated engagement signal weights
- Why niche posts are getting buried under low-quality recommendations
- What creators can actually do about it
Same Edward Tufte design—ET Book typography, sidenotes, print-first layout—but with completely fresh content based on the January 2026 release.
Technical documentation should show you what’s real. Not what marketing claims.