[joelwilliamson]

The author didn't justify that statement at all. If we have gone from an error once per minute (cruise control) to once per 30 minutes, and think that level of improvement can be repeated twice more, we will be at one error every 450 hours. A third time will put us at one error every 13500 hours.

Is continuous, gradual improvement the best way to fully autonomous cars? I don't know. But the author's argument is simply that we aren't there yet.

[makomk]

The trouble is that if they're not continually involved in the driving process, drivers can't actually concentrate well enough to be able to step in quickly when something happens that the automated systems can't handle. If I recall correctly, there have been studies on this and it takes tens of seconds for drivers to be able to respond to an unexpected situation sensibly if they haven't been actively driving. Mostly-automated cars that rely on drivers to step in when something goes wrong are probably not an option.

[ghaff]

This interaction of computer and human decision-making is actually the subject of a fair bit of active research. e.g. at Duke http://hal.pratt.duke.edu/

To your basic point--yep, at some point you need to stop automating unless you're willing to hand over control entirely.