Romane missed the target seventeen times before hitting it.
She was counting. She really wanted to count. After each miss, she watched where the bottle cap had landed, thought for a second, and tried differently.
On the seventeenth throw, the cap hit the middle circle. She raised both arms.
How an AI learns to stop getting it wrong
An AI doesn't learn the way you study a lesson. It learns by missing, getting corrected, and adjusting.
At the start, it knows nothing. If you ask it to recognise a cat in a photo, it answers at random. Then you tell it whether it was right or wrong. It tweaks, very slightly, the way it looks at images. You try again. It misses again. You correct it again. It adjusts again.
It's called backpropagation. The idea: every mistake says something about how to improve. Missing isn't the problem. It's the engine.
The thing that changes everything is the number of tries. A child learns to aim in a few dozen throws. An AI might need millions of examples to recognise a cat. What takes us an hour can take it weeks of computing.
The activity
Draw a target on a large sheet of paper: three circles, a dot in the centre. Lay it on the floor or hang it on the wall.
Each player gets ten identical objects to throw, one at a time, from the same distance.
The rule: after each throw, look where it landed and decide how to adjust the next one. No rapid-fire throwing. Observe, think, try again.
Note how many attempts it takes to hit the target, or measure the distance to the centre after each throw.
Halfway through, ask the question: "What do you change each time? How do you know what to change?"
Most answer without thinking: too hard, too far left, aimed too high. That's exactly what you give an AI when you correct it.
Second round: blindfolded. Throw without looking where the objects land. No feedback at all. The number of attempts before hitting the target explodes. Or you never hit it.
That's an AI no longer being told whether it's right or wrong. It keeps answering, but it stops learning.
What actually happened
Romane had a strategy. She threw softly on the first go to get a reference point, then adjusted progressively. She said it was "like learning to ride a bike but longer". I couldn't have put it better.
The blindfolded round frustrated her. She thought it was unfair. "If I can't see where I'm missing, I can't get better." Exactly.
Meryl joined in his own way. He'd gathered all his bottle caps into both hands and launched them all at once in one wide sweeping gesture. Some hit the target by chance. He was thrilled. He did it again.
I tried to explain that it only worked if one had happened to land right, and that next time he wouldn't know which one that was. He shrugged. "The red one." He was pointing at a red cap that had landed near the centre.
Maybe he had a strategy too.
At the end, Romane asked how many attempts an AI would need to learn to recognise a dog.
I said: millions.
She looked at her target. "That's a long time."