The problem is the blame game. Human behind a wheel hits a pedestrian, if they're found at fault then that "horrible inattentive/drunk/whatever driver" goes to jail for the death of another human being. It never makes anything "right", and I won't even start on how a jail sentence regardless of length can ruin your life in the US, but the public as a whole gets a feel good that "justice has been served".

How do we handle this for autonomous vehicles? Do we just fine/sue the company that made the vehicle/developed the software? Do we send imperfect human developers to jail because they made a mistake, even if in the grand scheme of things they have saved lives compared to humans being behind every action made by a vehicle?

A big part of the public image for autonomous cars is increased safety, any deaths at their hands starts raising where and how to place the blame - a subject I think very few are prepared for right now, which is likely part of why Tesla explicitly states autopilot needs a human driver present right now, and why Google has been extremely cautious with operator-supervised tests up until recently.

People get found not at fault for hitting pedestrians all the time.

...where breaking eggs means killing people?

There's no way for autonomous driving to become a reality without people dying?

I think we could achieve autonomous driving without needless deaths along the way. It might take longer, but I'd say it's worth it to, I can't believe I have to make this argument, avoid killing people.

In some sense, yes.

But is it any surprise that Uber had the first self driving car to kill a human being? You can cook in an orderly, careful way, or you can turn the kitchen into a disaster zone and expect other people to clean up for you.

I know which I prefer when it comes to human lives.

While I agree that the AV safety record is better than the human one, how is "you gotta kill random bystanders to make an omelette" okay just for this one industry? (As opposed to e.g. medical or military testing)

Military testing is typically done by putting a few pieces of tested equipment in the field. Russia is doing exactly that in Syria with their new Su-57. Pretty sure it's where a few other of the planes got a first taste of live combat, that's part of testing. I'm sure it'll either cause some unintended fatalities.

And, medicine takes tens of thousands of lives unintentionally, if I recall correctly.

So, by this logic, AZ is now a war zone?

Medicinal testing kills tens of thousands?

Not at all. You had mentioned military testing, which is safe in early stages, and I addressed that first: the testing process does follow through into live environments. Where things can go wrong.

But no, medical testing doesn't kill tens of thousands, practiced medicine does. I may have read your post wrong an figured you meant medicine in practice separately from military testing. There's research to suggest mistakes are the third leading cause of death in the US [1]. My point wasn't so much that killing random bystanders is okay, but that there's a level of unintended death in both, and not just in testing. Society decides what's acceptable, and if for miles driven the number of deaths that self driving vehicles cause is lower than the number of deaths caused by human drivers per miles driven, well... I'd say that's a fair way to look at it.

1. http://www.bmj.com/content/353/bmj.i2139

Okay, that does sound far more reasonable than comparing dead people to a part of a recipe.

Omelet metaphor aside,that doesn't really seem fair. We're talking about accepting more than 0 deaths caused by AI because the death rate should still be much lower than with humans driving. IOW, the perfect is the enemy if the good.