training, eval, hardware
The car that is learning and the car that is driving are doing different jobs on different clocks.
A team can run the same play a thousand times in practice, pause the film, argue, and run it again. On Saturday there is no pause. The play happens once, in weather nobody scheduled.
Training is practice. Servers replay driving clips, compare the network’s output to a target, and move weights. That can take days and megawatts. Nothing in that building has to steer a live axle.
On Saturday there is no pause.
Driving is game day. The onboard computer must answer now, with the cameras it has, the compute it shipped with, and no teacher whispering the right label. Weights do not update while you sit at a light — not in the usual design. The model is frozen for that trip.
Mixing the two words up creates fake mysteries. “Why didn’t it learn from my street?” Because learning happened earlier, on a sampled set, in a data center. “Why did a new version feel different overnight?” Because a new trained model was loaded, the way a team installs a new playbook. Practice and game day share a sport. They do not share a clock.