Draw a loop
Nobody tells a car how to drive. Nothing was trained beforehand. Each car has a tiny brain full of random numbers, and the only thing that ever happens is: drive, keep the best, vary them, drive again.
Seven distances to the wall, like feelers fanned out in front of it, and its own speed. Eight numbers. It is never shown the track, and it does not know where it is.
A network of 16 neurons joined by 68 numbers called weights. The eight sensed numbers go in on the left. Each neuron adds up what reaches it, every input multiplied by the weight on its line, and passes the result on. Two numbers come out: how much to steer, and how hard to press the pedal. Sixty times a second.
Fifty cars drive, each with different weights. They are ranked by how far they got. The best four go through unchanged. Every other seat is filled by picking good drivers, mixing the weights of two of them, and nudging a few at random. Most nudges make a car worse. Some make it better, and those are the ones that get kept.
When progress stalls the nudges get bigger and a few complete newcomers join. Driving backwards, spinning on the spot or crawling scores nothing. This is called neuroevolution. It runs entirely on this device: nothing is downloaded, nothing is sent, and the code is open.