Teach it yourself

In 1.1 you fought a machine whose rules were written by hand. Beat it, add a rule — and then there's another animal. It never ended.

This time you do the opposite. You teach it without writing a single rule.

What you did and what you didn't

Here's what you did while teaching it.

• Held up an object and pressed a button

• Typed a name

• Pressed Teach it

Here's what you did not do.

• Write down "an eraser is rectangular"

• Write down "a pencil is long and thin"

• Write down "a mug has a handle"

That last row is the whole chapter. Rules only grow as fast as a person writes them. Photos you just keep taking.

That line that came down

When you pressed Teach it, a line came down. You've seen it before.

It's how wrong it is — the same line from 1.2, where you turned the weights by hand. Turning one made the number shrink or grow.

This time the machine did the turning. Only one thing was different.

Nobody can turn three thousand weights by hand. Turn one and the others slip. So the machine turns them. That's what learning is — fixing it a little for however wrong it was, over and over and over.

Nobody knows exactly

Now back to the strange row you were left with in 1.1.

Who knows the rules — nobody knows exactly

When the machine said "that's an eraser," ask it why. The best answer anyone has is this: "3,075 numbers ended up set this way, and that's what came out."

Look through those numbers one by one and you won't find a sentence like "an eraser is rectangular." Not even the person who built it can find one.

An AI can produce answers its own maker can't predict. That checkers program in 1.2 beat its author for the same reason.

Did you fool it?

In the last step you put the object somewhere else, against a different background. How did that go?

Usually it misses. So is the machine bad at this?

No. It learned what you showed it. If you always shot on your desk, always in the same spot, the machine may have learned "the thing on the left side of a desk" rather than "eraser."

Same machine, same method — but change the photos you feed it and it learns something else entirely.

That's why shooting more photos and teaching it again made it better. You didn't add a new idea; you changed the material. That story came up in 1.3 too.

So what's going on in there?

Three thousand and seventy-five weights got set, we said. Where exactly are those weights attached?

Next chapter you take the lid off.