[dreamcompiler]

Two things can be true at the same time:

1. Human-driven cars are dangerous.

2. Autonomous driving (currently) is dangerous.

AD solves some human driving problems (like inattention) but introduces new ones (like driving under trucks).

[wutbrodo]

Obviously so, and this doesn't contradict anything I said. But the types of people who work on AVs are quantitative enough to understand that these magnitudes can differ, and are able to process that difference instead of waving them away as "both dangerous" and sticking with the status quo out of blind habit.

Almost literally every decision trades off one danger for another: to pick an adjacent example, airbags introduce dangers that are different from the ones they solve. And yet, "airbags have dangers! Take them out of your car" is a woefully incomplete understanding of the risk landscape, driven more by our Neanderthal brain's superstition than our Homo Sapiens reasoning about risk.

That sort of magical thinking is a hurdle that every shift in the risk landscape needs to overcome ("what if vaccines have microchips in them??").

This is, of course, dependent on actual comparative safety stats, but I've yet to hear a safety argument against taking a Waymo ride over an Uber in Chandler[1]. If the same level of transparency and rigor is used in more complex environments, I'd take a cruise/waymo car in sf too (once they launch).

[1] https://www.forbes.com/sites/bradtempleton/2020/10/30/waymo-...

[dreamcompiler]

I don't find your analogy to airbags persuasive. The vector space of "weird situations where an airbag can kill a person who would not have been killed otherwise" is vastly smaller in dimension and size than the comparative space where autonomous driving can kill someone. That in and of itself does not make AD a bad idea, but it does make it a much harder engineering problem.

As engineers we can pretty much completely characterize the airbag space and make predictions about where airbags are dangerous, and then mitigate those. What makes AD a bad idea is that such characterization is impossible with the AD space because AD is currently done almost exclusively by feeding massive amounts of data into a neural net pattern matcher. In other words we have no idea why AD works, and thus we have no idea when AD won't work until somebody dies. This is no way to do engineering, especially when lives are at stake.