How TikTok's Algorithm Uses On-Device AI to Personalize Your Feed

I reverse engineered TikTok to figure out exactly what makes their algorithm so addicting, and I discovered that they have over 25 AI models running on your phone to figure out exactly what'll keep you scrolling.

Before a video ever plays, TikTok servers have already labeled it with something called recommend scores. This is their calculation for how likely you are to skip a video, finish it, like it, open the comments, or follow the creator.

But to get those right consistently, they have to understand you in real time, and that's where these on-device models come in.

Some of these models work out what state you're in.

For example, one reads your gyroscope and motion sensors to tell if you're walking, lying down, or in a car, and this goes out with every request for a new batch of videos, along with your battery percentage, whether you're charging your phone,

your screen brightness, whether you're on mute, and over 50 other variables.

So every new batch of videos you get is picked for whatever state you're in at that time. In fact, even a message you type but change your mind on sending is still collected by TikTok. So the feed you get scrolling your lunch break is different from the one you get lying in bed after almost sending a risky text.

And the rest of the on-device models activate when they think they're losing There's a model constantly running that predicts if you'll skip the rest of your loaded videos, and if it does, your phone requests an emergency batch of videos to get your attention back.

And if you keep skipping anyway, TikTok marks you as having lost focus and, as a last resort, push a live video into your feed.

One model's job is to even tell if you've dozed off based on how long you haven't touched the screen and whether you're lying down.

There's even one that can predict when you'll open the app next so that TikTok can start in the background and be ready for you.

Follow me to see more breakdowns like this, and let me know in the comments what app you'd like to see me do next.