training, eval
In a shared brain, teaching one lesson can quietly unteach another.
Tighten one string on a guitar and the others feel it. A large network is more tangled than a guitar, but the metaphor holds. When you add clips of “do not brake for that overhead sign” and train again, weights that also served “do brake for a stopped truck” can move.
That coupling is the practical tax of end-to-end learning. Fixes are statistical, not surgical. Labs therefore rerun whole suites after a training run: old cities, old weather, old parking lots. A win on the new bug with a loss on last quarter’s bug is not a ship.
A large network is more tangled than a guitar, but the metaphor holds.
This is why version notes sometimes sound oddly broad. The team did not only patch shadows. They accepted a new global compromise. Riders feel that as a personality shift — more assertive left turns, then a week later more caution in construction — even when nobody advertised a personality project.
The mature habit is to treat every retrain as a new student who studied an updated library, not as the old student with one sticky note added. Evaluation exists because that student can forget.