I like the idea of a self driving car for most of my trips, but I also love to just go for a drive sometime, and it feels like the later has its days numbered.
I imagined, that a 'smart' car would briefly show you ads on the windshield before agreeing to follow the course, and while on the go, would then time-to-time briefly shade the windows (or whatever is the primary outlet of attention) just to show you another personalized ad. Time to step out? Wait a sec, another ad, just to unlock the door... Captive is the word.
But a computer looking over my shoulder and second guessing me makes the process of driving even more awful and complex.
UBER, Google, and a whole lot of government acronyms such as IARPA were present.
What they spent most of the day obsessing about was:
“We know, statistically, our technology WILL kill a child. How do we handle the public and political fall out of that?”
In my view they were more concerned about that than looking for technical solutions.
I brought up a personal experience:
I once almost had a accident on Interstate 80, in a construction zone.
Somehow a car that was over packed pulled from between some construction equipment, in front of me.
This driver could not see out any window but right in front of him, and he was pulling *across* the interstate traffic, not going with the flow. He could not see out the passenger window, that was facing me, it had a sleeping bag in it.
He shot out from between the construction equipment about twenty feet in front of me, while I was doing 45 MPH, remember it was a construction zone.
The correct solutions to the problem was to floor the gas, so that I could get in front of him while there was still space, and get off on the right hand berm of the road.
Any solution the applied the brakes would have caused a collision.
The experts in the room replied that Aircraft Autopilot software may have been able to avoid a collision in this example. No one was sure and few seemed to care.
In general I agree with your point that human drivers can create complex risks requiring sophisticated mitigations. But I also agree that ML systems are going to kill people at some point (arguably they already have) and policy will need to be created for that.
With real full self driving where the driver is not supposed to have control, the legal situation will have to be decided indeed. I wonder if this is holding back progress there.
As far as I know there were no lawyers present.
"liability was the main topic of that conference."
More like underlying current of each session. It was not the intent of the conference.
If I buy a "regular car" and self-driving capabilities take over for me, my control is being taken. Maybe it's justified because my driving has been erratic in a "he must be tired" kind of way so there are some immediate benefits but maybe not.
It comes down to control and choice. Without me being able to decide and turn it on or off, I'm subject to the rules, interpretation, and controls of a black box system that I'm not allowed to understand.
An actual FSD vehicle won't care if you're drunk because it is the one that is driving. That's the right way to solve the problem.
And yeah there's is a middle-ground where you're having to get into an argument with an algorithm to get what you really want out of it.
If I want to drive, then I want to drive my own car, over which I have full authority. I don't want any automation beyond ABS.
If I just want to get somewhere, then I'd rather hop in a self-driving taxi. It doesn't even need a steering wheel.
I won't be satisfied with any awkward compromise between the two.
Threads on self-driving are often enthusiastically supportive, with many comments about how computer control will ultimately be safer because computers don’t get tired, distracted, drunk, etc. and people do.
But then threads like this one are usually highly skeptical about how well the safety technologies will work, and over losing control over one’s machine.
Maybe these are just totally different sets of people commenting; I recognize that HN has a big and diverse audience. But even if that is so, it seems interesting to think about who chooses to comment on which topics.
Because ultimately these sort of “smart” safety technologies, and self-driving technologies, are qualitatively the same: you’re ceding manual control to a computerized system. It’s just how much: does the ML system have to reliably interpret “impaired” driving, or does it have to interpret everything about driving?
It seems like there is some sort of uncanny valley where ML-powered technologies are perceived as bad when implemented partially, but good when fully placed in control. That doesn’t make logical sense to me but I’d be interested to hear different views.