4 ms·

I'm not convinced that 'autonomous' driving will ever be better than non-distracted/drunk drivers. Suppose this is a task that, short of overhauling all our roa

I'm not convinced that 'autonomous' driving will ever be better than non-distracted/drunk drivers. Suppose this is a task that, short of overhauling all our road infrastructure, ultimately requires GAI? I think we generally have way too much confidence that cars will ever drive better than people, at least on the current road system. I am not saying that they definitely won't, only that I'm not sure we should take the attitude that anything that gets in the way of autonomous driving is sacrificing human lives. We don't know that. I also worry that mixing autonomous systems with people will result in more accidents, both in the case of drivers nodding off instead of supervising the car, and in the case of other drivers being surprised by the autonomous car's behavior. On top of this, it would not be surprising if autonomous car accidents were more psychologically devastating - they would likely commit different categories of error than human drivers. Conversely, if the tech industry wanted to save lives on the road, easily the most impactful thing they could do is shut off any phone screen moving faster than five miles an hour. This would inconvenience passengers, but we did after all survive for many years without phones in the car.

ben_w 5y ago
> Suppose this is a task that, short of overhauling all our road infrastructure, ultimately requires GAI? I would be surprised if level 5 self driving requires an AI smarter than a mouse or an eagle — freeform environment, arbitrary weather, hostile others, camouflage(!), long range navigation. The rest of your points I broadly agree with though.
- davrosthedalek 5y ago
  How many birds have crashed into your windows? At my parents house, this happens a couple of times a year.
  - ben_w 5y ago
    One, over my lifetime. I don’t think this a great comparison. Firstly, sheets of glass are not a common road obstruction, and even if they were cars have the added advantage of more sensor types (doesn’t require much more brain, it’s equivalent to just another ‘colour’); Secondly, I’ve also been hit by a car driven by a human. The driver had stopped at a give-way (major-minor junction) [0], then failed to look in my direction and pulled out into me as I was turning into that road. [0] right where this unrelated white van is in this Google street view photo: https://goo.gl/maps/Xe87aruFbbUazXhp7 https://goo.gl/maps/Xe87aruFbbUazXhp7
slfnflctd 5y ago
> at least on the current road system I came to the conclusion years ago that we're going to need RFID tags on everything (or something similar) to make FSD work safely enough without AGI. We'll probably also need dedicated lanes in some places. Nothing I have read or seen yet has given me any reason to think differently. Edit: The common counter-argument I see to this is that tags could be altered maliciously. My response is that we've already been building systems which scan for and filter out malicious changes for a long time, and we keep getting better at it. It seems to me like this kind of system would be far easier to design & update than any FSD software I know of.
- javagram 5y ago
  RFID is no solution at all, how do you ensure a toddler running into the road is properly RFID chipped? If you need every possible obstacle to be tagged with a locator beacon, you’ve essentially just reinvented 1970s autonomous subway or PRT technologies that require the track to be completely fenced off from surprise animal or humans or debris ended up on the track.
  - slfnflctd 5y ago
    The idea is not that it would be RFID only, but to reduce the number of non-tagged objects & people so false positives happen less often. Obviously you still need a way to deal with random, unexpected obstacles, but it's a much smaller problem space when the surrounding environment is mostly properly tagged.

ben_w 5y ago

> Suppose this is a task that, short of overhauling all our road infrastructure, ultimately requires GAI? I would be surprised if level 5 self driving requires an AI smarter than a mouse or an eagle — freeform environment, arbitrary weather, hostile others, camouflage(!), long range navigation. The rest of your points I broadly agree with though.

davrosthedalek 5y ago
How many birds have crashed into your windows? At my parents house, this happens a couple of times a year.
- ben_w 5y ago
  One, over my lifetime. I don’t think this a great comparison. Firstly, sheets of glass are not a common road obstruction, and even if they were cars have the added advantage of more sensor types (doesn’t require much more brain, it’s equivalent to just another ‘colour’); Secondly, I’ve also been hit by a car driven by a human. The driver had stopped at a give-way (major-minor junction) [0], then failed to look in my direction and pulled out into me as I was turning into that road. [0] right where this unrelated white van is in this Google street view photo: https://goo.gl/maps/Xe87aruFbbUazXhp7 https://goo.gl/maps/Xe87aruFbbUazXhp7

slfnflctd 5y ago

> at least on the current road system I came to the conclusion years ago that we're going to need RFID tags on everything (or something similar) to make FSD work safely enough without AGI. We'll probably also need dedicated lanes in some places. Nothing I have read or seen yet has given me any reason to think differently. Edit: The common counter-argument I see to this is that tags could be altered maliciously. My response is that we've already been building systems which scan for and filter out malicious changes for a long time, and we keep getting better at it. It seems to me like this kind of system would be far easier to design & update than any FSD software I know of.

javagram 5y ago
RFID is no solution at all, how do you ensure a toddler running into the road is properly RFID chipped? If you need every possible obstacle to be tagged with a locator beacon, you’ve essentially just reinvented 1970s autonomous subway or PRT technologies that require the track to be completely fenced off from surprise animal or humans or debris ended up on the track.
- slfnflctd 5y ago
  The idea is not that it would be RFID only, but to reduce the number of non-tagged objects & people so false positives happen less often. Obviously you still need a way to deal with random, unexpected obstacles, but it's a much smaller problem space when the surrounding environment is mostly properly tagged.