Hacker News Reader

Comment by semi-extrinsic | Hacker News Reader

I agree. The other day I was driving and thinking about this, and I started noticing "how many times during this drive have I looked another human (driver or pedestrian) in the eyes in order to determine their intentions?" It was a significant number even in a 15 minute drive.

I was still thinking about it during my bike ride later in the day, and I noticed that more than a dozen times I exchanged looks and nods with drivers to ensure they had seen me and that it was safe to do something.

No self driving car is able to look a human in the eyes and determine their likely course of action. And it most certainly won't be able to signal to the human a message as complex as "I have understood your intentions".

View on HN

I got bear-ish on self-driving cars when I realized that, in order to be able to drive in Bucharest, you _have to_ break the (letter of the) law fairly often[1]. Knowing when to do so requires "common sense" which is a notoriously difficult nut to crack - computer vision may be advanced enough, but it's simply not enough. I still think it will evolve in time and reach mass-market, it's just that it will take lots of time - it's not something that will have a purely-technological solution, it will likely require legal and social changes too, and those take time.

[1] E.g. Car stopped in front; do you cross the continuous line to go to the opposite lane, or just wait a bit? Or merging from a side street into a busy main road.... some drivers will let you, but which ones? If you wait until "it's safe", well, good luck.

This is another good point. There's a hardware store near to where I live where it's literally illegal (if you observe the signs) to enter or exit the parking lot of said store. Everyone just understands that someone forgot to remove a road sign saying "buses only, cars forbidden" for the lane where you enter this parking lot. But how would a computer understand this?

I’m not sure that figuring out when to break the law in your example is any harder then figuring out when turning left is legal and safe on unprotected intersection. This is not to say that it's easy, but autonomous vehicles need to solve problems exactly like that one on a regular basis.

Yes - and I was saying that I realized that solving these problems require the elusive "common sense". If we get that figured out, even just roughly, I think we'll break the barrier of general-purpose AI. I'm not sure we're close to doing that... But, maybe I'm wrong.

This argument always seems a little strange to me. I think I can count on one hand the number of times, in 14 years, I've ever made eye contact with another driver while driving, or used some kind of hand signals. And in all of those cases, if the signaling had failed, the situation would have resolved itself just fine in less than a minute anyway. Most of the time, I can't even see the other drivers through reflections on their windshields. Sure, sometimes I'll stick an arm out the window to wave my thanks when somebody lets me merge, but that's not a critical communication.

Where do you live?

It is literally impossible to drive in Rome, Buenos Aires, Lima or Rio without looking at drivers for their intentions.

Reply on news.ycombinator.com