Following rules vs. figuring it out

We'll get to what they noticed. First, meet one of those machines yourself.

The machine below guesses animals. It asks a question, hears your answer, and makes its guess. Inside its head are thirteen rules. Every one of them was written by a person.

If it can't guess your animal, you win — and the winner gets to add a rule.

Winning doesn't end it

Did you win? Then you added a rule. And that animal won't beat it again.

But there's another animal. Beat it with that one, add another rule, and then there's another animal after that.

There are more than a million kinds of animal in the world. Could you write a million rules?

No. And the count isn't even the real problem.

• Every time something new turns up, a person has to write it in

• The more rules there are, the more they start contradicting each other

• Anything the writer didn't think of, the machine doesn't know

This machine can never get smarter than what was written into it. That's the ceiling on a rule-written machine.

So people changed the approach entirely

Today's AI isn't handed rules.

Instead it looks at tens of thousands of photos. Dog photos, cat photos, piles of them, each with a label. And the machine works out for itself what separates a dog from a cat.

That material has a name: data.

That last row looks wrong, doesn't it. Part 3 is about that.

So what counts as AI?

Plenty of things have "AI" stamped on them. The stamp doesn't settle it.

There's one test. Did a person write the rules, or did the machine find them by looking at material?

• An automatic door — it opens when you step up. One rule. Not AI

• The brown rice button on a rice cooker — a set time and temperature. Not AI

• Your photo app grouping pictures by face — it worked that out from tens of thousands of faces. That's AI

• Translating into another language — that's AI

But how does it find them?

A machine finding rules on its own sounds like magic.

It isn't, at all. Turn one value yourself in the next chapter and you'll see.