X Open Sources 'For You' Algorithm

X (formerly Twitter) has open-sourced the core recommendation algorithm for its "For You" timeline under the Apache v2 license on GitHub. The repository quickly garnered over 27,000 stars, with a commitment to continuously integrate future updates. Elon Musk stated that the open-source move aims to actively seek criticism to improve recommendation fairness and quality, noting that any government censorship requests are now clearly visible. Developers used Grok to parse the entire 370,000 lines of code, revealing that the system ranks posts by predicting user interactions rather than assigning a direct "quality score." ⏎ Confirmed ⏎ - Open Source Content: Released core ranking systems, model configurations, feed filters, content labels, and the mutual follow graph; the code covers a mix of content from followed accounts and ML-retrieved external content, as well as safety systems like NSFW detection. ⏎ - Weight Details: According to a Grok summary shared by Musk, the algorithm calculates a weighted sum of predicted interactions like likes, replies, and reposts. xiaohu identified likes as 0.5 points and copying the link for external sharing as 20 points; zivravid noted that external link sharing is the highest weighted signal, while likes are nearly noise. ⏎ - Algorithm Mechanisms: Includes 21 positive weights, 5 negative weights, as well as diversity penalties and negative downranking signals. The "mutual follow" boost mechanism discussed in July was officially explained: it increases the recommendation weight of original posts from mutual friends by raising the predicted probability of a user replying to them. ⏎ - Transparency Measures: X launched a shadowban query page where users can check labels affecting account and post visibility; the limiting mechanism is visible to followers but invisible to non-followers. ⏎ - Updates: On August 15, the official open-source account announced updates to the repository, including the actual working principles of ranking weights and new content for the Brazil 2026 election. ⏎ Why it matters ⏎ This move is viewed as a significant step towards algorithmic transparency for mainstream social media platforms. Developers using LLMs to parse the code validated the practical value of open source. The revelation that likes are nearly noise while link sharing carries the highest weight offers the first clear look at the types of interactions the platform truly incentivizes.

2026-08-14 ~ 2026-08-15 · 16 related posts

Primary sources

3 near-duplicate retellings: xiaohu · xiaohu · IgorBrigadir