Reverse-engineered TikTok: 25+ on-device AI models and 50+ real-time state signals
dotey · x · 2026-11-06
A reverse-engineering breakdown (shared by vista8) reveals TikTok's extreme recommendation design: 25+ on-device AI models, pre-computed "recommendation scores" for every video before playback, 50+ real-time user state variables (gyroscope/motion sensors detecting walking/lying/in-car, battery, brightness, even typed-but-unsent messages), churn interventions like emergency video batches and inserting livestreams when users keep skipping, drowsiness detection, and predicting when you'll next open the app to pre-cache videos. Claims are from the original author, not yet independently verified.
More from Companies & People
- Kai-Fu Lee: AI is too smart for CEOs to waste on customer service — kaifulee · 2026-11-06
- After 43-1 data center moratorium wins, OpenAI launches community charm offensive — TinfoilTricorn · 2026-11-06
- 7-person team Krea ranks #3 in Multi-Image Edit Arena, CEO discusses training and open-source plans — arena · 2026-11-06
- CFO: We Have 20 AI Use Cases We Can't Execute — Four Questions to Cut the List — alex_verem · 2026-11-06
- SF Systems Reading Group (SFSRG) named after 15 sessions at Greylock — _shreya_s · 2026-11-06
- Stealth startup's Cortex scores 73.2% on τ³-bench banking, beating frontier LLMs by 18 points — multiply_matrix · 2026-11-06