January 19, 2026
Open-source or open-ish? You decide
X algorithm has been open sourced
Fans call it ‘open‑ish’—missing pieces, Rust surprise, and meme mayhem
TLDR: X posted code for its For You feed using Grok AI to mix posts you follow with global picks. Commenters love the peek but roast the missing model “weights,” incomplete bits, and déjà vu with the old repo, sparking debate over real transparency versus show—and whether competitors can capitalize.
X cracked open the hood on its “For You” feed, bragging that a Grok-powered AI now picks your posts by predicting what you’ll like, reply to, or share. The crowd’s reaction? A chorus of “open‑ish.” As swyx snarked, the code ships “without weights,” the secret sauce that makes the model actually smart. X says it ditched hand‑tuned rules so the AI learns purely from your clicks and swipes.
Curious devs poked around like it was a mystery box. [rapsey] sighed that parts look “not meant to be built,” with missing files and omitted bits—and yes, a surprise cameo by Rust sent eyebrows up. Meanwhile [internetter] asked what everyone wondered: how is this different from the old twitter/the-algorithm drop?
Fans loved the diagrams and the spicy names—Thunder (people you follow) plus Phoenix (the global firehose)—but skeptics asked if this is transparency or theater. [moneywoes] wanted “anything interesting?” while [chistev] fretted: are they handing rivals a blueprint?
Memes flew: “Open‑source soup, hold the seasoning,” “Thunder vs Phoenix sounds like a Marvel crossover,” and “Grok reading my likes like a diary.” Bottom line: cool peek behind the curtain, but the missing puzzle pieces keep the drama alive—and the “For You” debate far from settled.
Key Points
- •X open-sourced the core “For You” feed recommendation system repository.
- •The system combines in-network posts (Thunder) and out-of-network posts (Phoenix Retrieval) into a unified ranking.
- •Ranking uses a Grok-based transformer (adapted from xAI’s Grok-1) to predict probabilities of user actions (e.g., like, reply, repost, click).
- •A Weighted Scorer aggregates predicted engagement probabilities into a final score using configured weights.
- •The design removes hand-engineered features and most heuristics, relying on user engagement history and learned modeling for relevance.