training, perception, hardware
It is not a tiny person in the dashboard. It is a huge pile of adjustable numbers that learned a mapping.
Think of a graphic equalizer with millions of sliders. Nobody sets them by hand. You play a song, measure how wrong the output is, and nudge sliders until the song sounds closer to the target. Do that across millions of songs and the board starts generalizing to songs it has not heard.
A driving network is that board. Camera pixels go in. Some mix of scene description and control requests comes out. During training, the sliders — called weights — move so the outputs better match examples. During driving, the sliders stay frozen and the same math runs many times per second on the car’s computer.
Nobody sets them by hand.
Two consequences matter for readers. First, there is no readable sentence inside the network that says “if ball then child.” The behavior is distributed. Second, the network only knows the kinds of songs it was trained on. A totally new instrument can still confuse it.
Calling this “the AI” is fine as shorthand and sloppy as engineering. The network is one component. Data selection, training code, onboard compute, and the evaluation harness around it decide whether that component is useful. The next posts separate practice from game day.