Choosing where to look

In 4.2 you saw why counting hit a wall. It could only see three words back.

People don't read like that.

Long ago a kind woodcutter lived in a small village deep in the mountains, and he went out every morning.

You knew who "he" was at once — even though the answer sits eleven words back.

How? You read the whole sentence and picked out one word. The rest you could ignore.

The machine draws arrows too

There they are. Thicker means it looked harder there.

Choosing where to look in a sentence like this is called attention.

Stack attention in many layers and you get a transformer. Nearly all of today's language AI is built that way.

It goes back to 3.2

It sounds strange. How can a machine "choose" where to look?

There's no rule for it. It's weights — the same weights you turned like dials in 3.2.

Which word should look at which was settled by training. Nobody wrote down "a pronoun points back at a person."

Work through all eight

How many times did you and the machine land on the same word?

About six. Not all of them.

A sentence that splits people too

Look at the last one again.

Tom lent Sam his own book, and he said thank you.

Who is "he"? Tom, or Sam?

The sentence doesn't tell you. The one saying thanks is probably the borrower, so Sam — but nothing in the words settles it.

If people split on it, no wonder the machine wavers — ours doesn't even land on Tom or Sam. That isn't the machine being broken.

But looking well isn't knowing

It can choose where to look, and it can guess the next word.

So does this AI know things?

Next chapter: an AI that is confidently, completely wrong.