But is it possible to program one to anticipate the boundless stupidity of other drivers? If that doesn't happen, I see dead people.
Humans are better at analyzing the situation and predicting problems. For example, if you see a car that is barely able to stay within the lane, and this happens not just once but for the last thirty minutes, chances are that the driver is drunk or falling asleep. A computer may not make that connection, and it will not assign a "red flag" to that car. When that car finally ends up in your way, the robot will only be able to react. Perhaps it will react fast enough - but an attentive human would not be driving near the suspicious vehicle in the first place.
I do not expect to see a complete, functioning AI as a car's driver. However it may be possible to add software modules to mathematically analyze behavior of other drivers. That weaving within the lane, or leaving it, may be measured and noted. The same would apply to poor speed control, tailgating, frequent lane changes, and so on - these are not complicated facts. Then different vehicles may be assigned different danger levels, and the algorithm will be slightly influenced to stay away from those.
Sure is. Because in the end the “boundless” stupidity of other drivers falls into a handful of predictable categories: cutting you off, lane impinging, wrong way, tailgating, a few others. And all of these categories have a correct method to avoid collision. Really it’s pretty straightforward. I see fewer dead people, because every computer controlled car is one less boundlessly stupid human on the road making hazards.