Can I teach a computer?

You might think that to make a computer know what a cat is, someone sat down and wrote the rules: a cat has fur, and whiskers, and pointy ears. But it never works. For every rule the world has an exception, a cat with no fur, a fox with pointy ears. So the people who build these machines do something cleverer, and much stranger. They do not write the rules at all. They let the computer find them.

Show it, do not tell it

Instead of rules, you give the computer examples. Here is a cat. Here is another. Here is a dog, which is not a cat. A thousand pictures, each one labelled. Then you let the computer guess, and you let it be wrong.

Wrong, then a little less wrong

The first time, the computer is hopeless. It looks at a cat and guesses dog. So you tell it the truth: no, cat. Here is the magic. It does not memorize that one picture. It reaches inside itself and nudges its own settings, just the tiniest amount, in the direction that would have made it a little less wrong. One nudge does almost nothing. But show it the next picture, and the next, each wrong guess followed by one tiny nudge toward less wrong, and the wrongness starts to shrink. Slowly. Then faster. Until one day you show it a cat it has never seen in its life, and it says, quietly and correctly, cat.

It learns the way you do. By being wrong, then a little less wrong, a million times.

So who is the teacher?

Notice what just happened. Nobody ever told the computer what a cat is. It was never given the rule. It found the rule on its own, hidden inside your examples, by being wrong on purpose and shrinking the wrongness one nudge at a time. Which means the real teacher was never the programmer. The real teacher was the examples, and the person who chose them. Choose better examples and you get a better mind. So when you teach a computer you are not typing rules. You are choosing what it sees. That makes you the teacher. It is a big job, and it is yours.