X (Twitter) Open-Sources Its Recommendation Algorithm

X/Twitter open-sourced its For You recommendation algorithm with actual action weights. - Release includes real weight values, filtering/labeling system, and training code. - A like is weighted 0.5 points; copying a link to share is weighted 20 points; the configuration lists 21 positive and 5 negative weights. - The earlier open-source release showed scoring structure but not exact per-action values. - The model predicts actions such as like, reply, copy-link share, and report, then combines probabilities into one score. - An analysis of the released algorithm advises creators to make original repost-worthy content; avoid hook-first threads with key material in later posts; prioritize reads, replies, quotes, and follows over likes; space out posts; stay in one niche; and avoid engagement bait and negative signals like 'not interested' or reports.
Key insights
Actual Action Weights and Training Code Opened
X/Twitter's For You timeline source code now includes real action weight values, a filtering and labeling system, and training code. The configuration contains 21 positive weights and 5 negative weights. A like is assigned 0.5 points, while copying a link to share is assigned 20 points. The earlier open-source release exposed the scoring structure but did not reveal exact per-action values.
Model Predicts Specific User Actions
The recommendation approach does not produce a single relevance score. It predicts whether a user will take concrete actions on a post, such as liking, replying, copying a link to share, or reporting, and then a separate step combines those probabilities into one ranking number.
Creator Guidance Prioritizes Original and Shareable Content
A Codex-based analysis of the released algorithm says creators should make original content worth reposting. It advises against putting a hook in the first post and key material in later posts, a practice described as contrary to common understanding. The analysis also states likes do not carry high weight, and creators should prioritize reads, replies, quotes, and follows.
Posting Frequency, Niches, and Bait Behavior Are Constrained
The analysis advises creators to space out posts, avoid frequent posting, stay within a single vertical niche, and avoid changing account types frequently. It also warns against engagement bait such as mutual follows or reply requests, because actions like 'not interested' or reports can reduce distribution.