Comment by scarmig
1 day ago
It doesn't seem impossible technically to up the assertiveness. The issue is the tradeoffs: you up the assertiveness, and increase the number of accidents by X%. Inevitably, that will contribute to some fatal crash. Does the decision maker want to be the one trying to justify to the jury knowingly causing an expected one more fatal incident in order to improve average fleet time to destination by 25%?
Nah, it's not that simple. Excessive passiveness causes ambiguity which causes its own risks.
You want the cars to follow norms, modifying them down slightly for safety in cases where it's a clear benefit.
Reinforcement learning is a helluva drug. I'm sure by now Waymos can time yellows in SF to within a nanosecond, whereas humans will only ever drive through so many yellows will never get that much training data.
A human can know the yellows on a few routes. A Waymo can pull over, observe a given intersection for an hour, and tell every other Waymo that exists precisely how long that light lasts.
It's not just collecting the information; it's the ability to spread it.