A user opens the app, watches a video for 1.4 seconds, and swipes away. That single swipe has to update what they see next — not in the next session, not in the next hour, in the next scroll. Most recommendation systems built in the 2010s couldn’t do that. They retrained overnight, served a static ranked list, and called it personalization. TikTok’s “For You” page broke that model by treating every watch, pause, replay, and skip as a live signal that reshapes the next candidate pool within seconds. That’s the part worth understanding — not the neural net architecture, which nobody outside the company has verified, but the systems problem of making feedback loops that fast without falling over.
The Core Problem: Candidate Generation Under a Latency Budget
A “For You” feed request has roughly 100–200 milliseconds to go from “user opened app” to “here are 10 ranked videos.” In that window, the system has to narrow a catalog of hundreds of millions of videos down to a few hundred candidates…
