Scoring every video for you would take 46 days

transcript

Picking your next video the obvious way takes 46 days. You open the app. You get about 100 milliseconds.

There are more than 20 billion videos. Your best model scores one in about zero point two milliseconds. Score them all for one person opening Home? About 46 days of compute.

So split it in two. Rohan opens Home. Cheap sources fan out in parallel, about 5 milliseconds each. Together they narrow 20 billion down to about 500. Only those 500 meet the expensive model. Rules drop what he's watched and cap any one channel. Twenty videos.

Nearest neighbours of his taste, from a precomputed index. New uploads from channels he follows. That's how a brand-new video gets a chance. And videos people often watch after the one he just watched.

Now the math. 500 candidates. Times zero point two milliseconds. That's 100 milliseconds of compute. Spread across cores, it fits in about 30. From 46 days.

But here's the catch. What should that model predict? Clicks are easy. But clicks reward misleading thumbnails. You click, then leave. So rank by expected watch time. When YouTube first did, in 2012, views dropped 20 percent. Later, weighted by satisfaction surveys, likes and shares. Valued watch time.

Never run the expensive model on everything. Narrow it cheap. Rank it expensive.

Full breakdown in the lecture.