AI that recognises people

When you turned, the yellow box went first. The blue dots hung on much longer.

The 2001 method only ever learned what a front-on face looks like: the eye region is darker than just below it, the bridge of the nose is brighter than the sides. Turn your head and those relationships break.

Both of them sweep

But there's something more important to notice here. Switch on Watch the old way sweep.

A little window travels across the whole picture. Exactly what you saw in 2.2 — a small window stepping across a photo one square at a time.

The bright places passed several rounds. And almost everywhere on screen fails the very first round. That's why the 2001 method was fast — the moment something looks wrong, it quits.

Today's way sweeps the picture too. It's looking for something else. Both sweep — the old one hunts for a face shape, the new one hunts for where points go.

Guessing where the points go

Turn the video off and only the dots remain. These dots are all the AI works with.

One face gets 478 of them, packed along the eyes, mouth and jawline. They're called keypoints.

Switch to Hands and Body. Same job every time — guess where the points go.

The AI doesn't know what a face is. It knows where to put the dots. That's what recognition actually is.

There is no 100%

Switch to Hands and look at How sure it is which hand. Is it 100?

No. Recognition is always a probability. It means "an 87% chance this is the right hand", not "this is the right hand".

Try to lose today's way too. Cover half your face, turn out the light, turn right around. Eventually it misses.

Today's way isn't perfect — it's just better.

Sound is next

So far this has all been about seeing. But AI handles sound too.

Sound isn't a picture, so how? The next chapter has the answer, and it's a little unexpected.