Hacker News new | ask | show | jobs
by kypro 497 days ago
I hope this isn't too off-topic, but I'm always amazed at how many otherwise smart people hold the naive belief that FSD is remotely close because 99.9% of the time it works fine.

Self-driving in my opinion will require an AI that is, if not very close to, an AI capable of general intelligence.

Why?

Because in the real world to be able to drive a car as well as a human across all of the edge cases a human can you probably need something approaching general intelligence.

Humans understand that a person isn't just something with 4 limbs, but also can be that thing that looks like a white sheet with eyes by the side of the road on Oct 31st. And its these types of weird edge cases that humans instinctively understand because they have a deep world model to reason about which cannot be reasoned about by the narrow FSD AI systems we currently have.

When you think about what humans need to do when driving it's so much beyond just watching the road and turning a wheel that it seems almost absurd to imagine our current AI is anywhere near capable of handling all of the edge cases humans currently are.

And I also don't buy this argument that the goal should be to simply to reduce the total number of accidents per mile... I'd grant that it's very possible that FSD could reduce the total number of accidents per mile driven because most miles are driven in the much more narrow environment of highway driving. And here AI probably could do better job than a human on average when you factor into the equation human tiredness and distractibility. But no one is going to be comfortable with FSD occasionally plowing into a group of kids outside a school because statistically the total number of people who die in road traffic accidents is reduced on a per mile basis.

I'd be interested if anyone strongly disagrees.

5 comments

I think you can implement self driving without general AI, but it has to be really defensive, err on the cautious side, and that means it can't travel faster that 10 to 25 kph, like a bicycle basically. That car will have a "safe zone" around it, monitored with radar and/or lidar, and if anything enters just outside that "safe zone", the car stops before hitting anything.

The market for such cars would be very limited IMO.

Totally agreed with the fist paragraph but after that you're ignoring the existence of Waymo today, which people generally feel comfortable in and around where they are. Elons marketing is so strong even the doubters feel like Tesla is the leader of the pack.
> But no one is going to be comfortable with FSD occasionally plowing into a group of kids outside a school because statistically the total number of people who die in road traffic accidents is reduced on a per mile basis.

I really don't see why not. Since those deaths must be counted too, if it still is safer with than in mind then it can't be something that happens even rarely.

But when are these edge cases, are they like the edge cases when I need 4 wheel drive? Basically never and if I did need it I just wouldn't go there or would rent a car for the day or something.
Most safety problems, including the child in a Halloween costume, can be solved by the advanced AI technique called "Don't hit anything. If it looks like you're about to hit something, slow down or stop."

Trouble is, when the company deliberately ties one hand behind its back by insisting on camera-only vision, it is never going to be perfect at not hitting stuff. Either multispectral imaging, radar, or lidar would help avoid edge cases like the Halloween costume. The camera might not even realize there's a three-dimensional object in front of it if there's snow on the ground. Stupid, stupid, stupid.

That "don't hit anything under any circumstances" is incompatible with USA-typical "stand our ground" and "right of way" philosophy of doing things.
> Trouble is, when the company deliberately ties one hand behind its back by insisting on camera-only vision, it is never going to be perfect at not hitting stuff.

I have a Y with FSD. I think that underplays it a little. Yes, they don't have LIDAR, but LIDAR is very expensive and fragile. I can understand that.

But they also removed the windscreen rain sensors in favour of using the cameras. Consequently you can be driving down the highway on a clear day, and the windscreen wipers will start. They also don't have ultrasonic distance sensors. Consequently the car won't warn you about short bollards at the corners of the car. It seems to be completely unaware of them.

Unlike LIDAR these sensors cost almost nothing, and these are high end cars. Keeping them until they have the camera version reliable would seem prudent. In fact they could have used them to improve the camera version, by say comparing what the rain sensor said to the camera output. It's seems like they've allowed an AI religion to cloud their engineering decisions.