Seventy years to get here

AI looks like it suddenly got good. It's actually a very old story. Along the way people announced "we've basically done it" several times, and several times they were wrong.

The fastest way to see how wrong is to look. Every card below is a prediction someone really made. Read the front and decide whether it came true.

They missed in both directions

How many did you call right? The number to look at isn't your score — it's which way each one missed.

• Three said it would happen too soon. Twenty years. Two months over a summer. Any moment now.

• Two said it would never happen. A hundred years for Go. No chess-only machine could do it.

Clever people missed in both directions. And every time they missed, people got disappointed and the money and attention drained away. That happened twice.

So why did it suddenly get good?

Here's where most people guess wrong. They assume somebody had a brilliant new idea.

Most of the ideas are old.

• The matchbox learning machine you'll meet in Part 3 is from the 1960s

• The idea of calculating the way brain cells do turned up in the 1940s. A machine built on it came in the 1950s

• A way to teach many layers stacked together turned up in the 1980s

So what changed? Two things.

① There is vastly more material. People started putting photos and text on the internet. In the 1980s you could show a machine a few hundred cat photos. Now it's hundreds of millions.

② The arithmetic got fast. And from an unexpected direction. Graphics cards were built to draw game screens quickly — and it turned out the sums they do are almost exactly the sums AI does.

Which adds up to this.

It wasn't a new idea. It was material and power.

Then we should look at those sums

If the material is photos and the arithmetic got fast — what exactly is a photo, to a machine?

That's where Part 2 starts.