Descriptions:
X has replaced its legacy rule-based feed system with a transformer model built on XAI’s Grok architecture, and Alphastack’s creator reverse-engineered the platform’s publicly available open-source code to explain exactly how it works. The new pipeline collects candidate posts, applies spam and safety filters, then runs each post through a model that predicts 19 distinct user actions — including replies, reposts, likes, follows, but also negatives like mutes, blocks, and “not interested” reports. Posts are scored on net predicted value, not raw engagement, with negative actions subtracted from the total.
One of the most counterintuitive findings: the model never reads your post text. It operates on hashed post IDs, your account history, which part of the app you’re on, and post age. For out-of-network reach — the mechanism that lets accounts with no followers still go viral — the system converts topic signals into interest vectors and matches them across all of X, which means topical clarity matters more than follower count. High scores don’t guarantee visibility either; stale posts, repeated content from the same author, and safety filters can all silently suppress distribution.
The creator also built a live tool at xphstack.fun that runs any post against X’s actual Phoenix model, using Kimi K2 for image analysis and returning letter-grade scores with suggested rewrites. The tool is available for $5 per 10 scoring runs. Practical takeaways from the analysis: lead with a clear topic, ask genuine questions (replies and quotes are heavily weighted), post while content is fresh, and avoid engagement bait or repetitive reposts.
📺 Source: Alphastack · Published May 18, 2026
🏷️ Format: Deep Dive
