training, planning, perception
The shift was not only a better algorithm. It was giving up on writing every rule by hand.
A coach can hand a player a rulebook or hours of film. The rulebook is explicit and limited. The film is implicit and large. Nobody lists every pattern the player absorbs, but after enough hours they recognize situations nobody named.
Several modern driving stacks made a similar switch. Instead of encoding every case as a condition, teams train networks on large collections of driving video so the model imitates competent behavior. A neural network here is a program that is shaped rather than fully written: examples go in, internal values move, and the outputs start to match the task. There is no single readable rule for recognizing a familiar face, and there is no single readable rule for many of these driving behaviors either.
The rulebook is explicit and limited.
The architecture can be described simply. Camera video goes in. Steering, accelerator, and brake requests come out. A lot of the intermediate object lists and hand-authored branches are no longer the main product.
Improvement then changes jobs. When the vehicle handles a scene poorly, the usual response is not editing a line of policy. It is finding better examples of that scene handled well, then training again. Progress leans on data quality and compute, not only on how many conditions a team can invent.
That is why training clusters and fleet scale became central, not decorative.