How Twitter (X) Actually Works
How does X sort through 500 million tweets every day and pick the ~30 that appear in your feed? This is the full breakdown of X's open-source recommendation algorithm — from the raw tweet firehose to the neural network scoring, invisible behavioral signals, and the 145,000 secret communities you already belong to.
Every concept is visualized from scratch using custom motion graphics.
Key takeaways:
• X processes ~6,000 tweets per second, narrows 500M candidates to ~1,500, then ranks ~30 for your feed
• A neural network with 48 million parameters scores every candidate tweet
• A reply is weighted 27x more than a like — an author reply is worth 150x
• Your scroll speed, dwell time, and mute/block signals all feed the model
• SimClusters maps every user and tweet into 145,000 invisible communities
• Cosine similarity in this high-dimensional space is how strangers' tweets reach your For You tab
• The full pipeline runs in under 1.5 seconds
📚 Sources:
• X Algorithm Source Code (open-sourced): https://github.com/twitter/the-algorithm
• "Twitter's Recommendation Algorithm" — Official Blog: https://blog.x.com/engineering/en_us/...
• Heavy Ranker Neural Network Architecture: https://github.com/twitter/the-algori...
• SimClusters Paper: "SimClusters: Community-Based Representations for Heterogeneous Recommendations at Twitter"
• Real Graph (interaction scoring): Twitter Engineering Blog
🎬 Chapters:
0:00 - Intro
0:13 - The Question
0:44 - The Firehose: 500M Tweets a Day
1:32 - The Neural Network: 48M Parameters
2:38 - The Invisible Signals
3:41 - SimClusters: 145K Secret Communities
5:03 - The Final Mix
5:34 - The Reflection
6:09 - What's Next
About Algo Thinker: We decode the algorithms that shape your digital life — recommendation systems, search engines, content moderation, and AI models. Every video is fully animated with custom motion graphics.
🔔 Subscribe for more: / @algothinker
Next episode: How Tesla FSD drives with nothing but cameras — 48 neural networks, zero LiDAR.