And software propagate learned knowledge perpetually, whereas each human has to learn it over and over again.

On the other hand, software can make mistakes at scale whereas each human makes mistakes differently.

It’s not so cut and dry.

While potentially true, it's a trade off it seems proponents are willing to take.

The same kind of trade off happens with other modes of transport with very rigid constraints on movement, things like trains. A train can keep going, stop, or maybe with enough time reverse if there's tons of time and a strong need.

We don't stand around worrying saying "BUT WHAT IF SOMETHING'S ON THE TRACKS?!?", because the answer is "well then, the train will hit it. That's the tradeoff for being able to move freight so efficiently."

Sometimes they try to drive through floods. They also seem to randomly run red lights. Also, what's with the way they keep veering out of their lanes?

I think what this boils down to is that the volume of vehicle-related injury/death will decrease, and the few remaining cases will be really strange and unlikely to have occurred with human drivers.

Some human drivers can adapt to a new unseen before situation. Many can't in my experience.