Hacker News new | ask | show | jobs
by cwalv 1246 days ago
> Driving is not a closed domain problem. There will always be fat tail circumstances where statistical methods will break down.

It doesn’t have to solve all cases to be immensely useful. All it needs is to be able to recognize when it’s unable to safely have control with enough advance notice to delegate to a human. Remotely piloted aircraft have been a thing for a long time; seems like we could do the same with trucks, so that humans only need to do the driving a small % of the time. I also suspect that current AI will likely respond to emergency situations better than humans do on average, because they don’t get tired and have superior reaction time.

1 comments

You can't just dump control to the person sleeping or playing marvel snap in the driver seat as your out. With aircraft there are attentive people on standby and more than one. With remote aircraft people are way less likely to die if something goes wrong. You also start really lowering the training level of humans.

I'll alter one of my favorite jokes to explain.

I want to go out like my Tesla ai did, in a bit of confusion and reanimated from rom, not screaming no no no no like the people in the car.

Tho truly with a lot of these the people will go out in their sleep.

Importantly with aircraft the time you have to correct an error that requires human pilot intervention before you crash is far greater than in a car. This is basic critical thinking... But so many people say we have autopilot why not driverless cars? Well obviously driverless cars are orders of magnitude more sensitive to error conditions on open roads than a plan flying in empty space.
I wasn't comparing autopilot in an aircraft to autopilot in a car. I was comparing remote piloting of an aircraft to remote operation of a vehicle.
the paren'ts exact same point holds, how long to swap an attentive person in. Have you done support chat? it will immediately be one person to 55 cars, and add in the latency, they're not stopping that tesla move left and hit the breaks issue let alone a 30 car pileup.
I said "All it needs is to be able to recognize when it’s unable to safely have control *with enough advance notice to delegate to a human*"

IMO, there are road conditions where this is possible. For black swan cases like random wind gusts, tire blowouts, boulders falling onto the highway etc. I think AI would be better on average *currently* (again, just clarifying what I already said).

I'm not saying that you attempt FSD with the copout that a driver needs to stay attentive and be ready to intervene. I'm saying that there is a high proportion of truck driving that's completely mindless, and it's possible to have a system that recognizes the difference.