Comment by kbos87 | Hacker News Reader

Comment by kbos87 | Hacker News Reader

I was an autopilot skeptic until I wasn’t. As a user of the system it’s 100% clear that it’s my responsibility to be able to take over at any time. Autopilot regularly restricts the places it will engage and speed it will operate at based on road conditions. The safety of autopilot concerns me a lot less than the safety of other drivers on the road, the majority of who seem to be using their phones any second they think they can steal while behind the wheel.

Autopilot also gets noticeably better every passing week or two that I’ve been using it. Just as one example it now alerts you and slows down when it sees an emergency vehicle on the side of the road. (To be clear I’m not commenting on the specific accidents pointed out in the article, only my experience using it and seeing the pace of improvement.)

On the other hand, Tesla’s history of how they have marketed autopilot to buyers seems downright misleading. It’s clear now that autopilot isn’t 100% autonomous, but I’d be pretty upset if I made a buying decision a couple of years ago based on the hyperbole coming from Elon Musk and on their website about what was supposedly right around the corner.

> it’s 100% clear that it’s my responsibility to be able to take over at any time.

There's no way you can actually be ready to take over at any time though. That's just setting you up to take the blame when autopilot fails.

I mean, sure, sometimes you'll detect the error and avert it. But sometimes you won't.

The phrasing of GP's comment implies that GP is an actual user of the system. Has actual experience. You shouldn't say "there's no way" etc to people who have experience when you have none.

I'm curious what an actual user of the system thinks the benefit of Autopilot is, compared to other cars that also require the driver's attention all of the time.

> ou shouldn't say "there's no way" etc to people who have experience when you have none.

What makes you think I have no experience?

Also, it's perfectly fine to dismiss claims that seem impossible without supporting evidence. I don't need to believe every psychic until I disprove them. There's no way someone can be as good at taking over randomly from an AI as they are at performing the action in the first place.

Your comment makes it sound like “taking over from AI” is some complicated action. In reality that means turning the wheel reflexively to a road hazard or hitting the brakes in the same way that I would if autopilot wasn’t engaged. My attention is still on the road.

If you are constantly turning the wheel and applying pedals in a way such that autopilot wasn't engaged, how does autopilot do anything?

Alternatively, if you aren't constantly turning the wheel and applying pedals, how many seconds do you think it takes you to start? And before that, how much slower is your base reaction time when your "attention is on the road" but you're not driving?

You have your hands on the wheel at all times. You’ll already have an intuition of situations in which the autopilot might make a mistake, and you viscerally “feel” the car’s decisions, as well as its level of confidence. In tight situations, where you cannot afford an inch of a steering mistake, you’ll grab the wheel harder, in a way that if autopilot does try to deviate from the path you deem safe, it will disengage itself against the pressure of your hand.

I very much feel like an airplane pilot must feel using the “autopilot” in a Boeing 747: The system helps me reduce my cognitive load, and allows me to focus more on situational awareness versus micromanagement (e.g. staring at the taillights in front of me and missing the motorcycle coming up from behind). But at no point would I ever delude myself into thinking that the system absolves me of responsibility.

That makes it even more scary. In an aircraft, you usually get a few minutes to determine the problem, disengage autopilot and correct it manually. Autopilot is also disengaged during periods of uncertainty like takeoff or landing (unless it's very well proven). In case of a car, you get mere seconds to react. And that reaction should come from your intuition or mental model of the autopilot. That implies that you should have faced a similar dangerous situation before hand, so that you know what may go wrong.

Another way I read it, it's like you being a standby driver ready to take over when the inexperienced driver makes a mistake. I've been in a similar situation with a driving instructor before. While nothing bad happened, the time it took for him to react convinced me that humans are bad at this. The only reason I find autopilot justifiable is if it's quantitatively better than the human driver.

Have you ever used Autopilot yourself?

I agree that if somebody played with their phone while on Autopilot, it would be dangerous. If you have to context-switch and put your hands and feet back into position, it might be too late.

But Autopilot is designed for you to keep your hands on the wheel at all times. I use it on a daily basis. And while I have not measured my "Autopilot disengagement reaction time", I am sure it is in the range of mere milliseconds.

If you had to react to a notice on the screen, or even just to warning sounds, it all would be horribly slow. But when you use Autopilot, it all works through tactile feedback. You feel each tiny movement the car wants to make, be it steering left or right, braking or accelerating. And while a driving instructor has to reach over to take the wheel, all movements the car makes are "piped" through your own muscles.

> The only reason I find autopilot justifiable is if it's quantitatively better than the human driver.

That's the point. And to make that happen, Tesla needs this current generation of "supervised Autopilot". Because somebody's got to train and fine-tune that neural network. You need humans to make the system super-human.

> I agree that if somebody played with their phone while on Autopilot, it would be dangerous.

I know people who use auto-pilot like this. The technology enables it.

> while I have not measured my "Autopilot disengagement reaction time", I am sure it is in the range of mere milliseconds.

That seems impossible. The human reaction time to decide the appropriate response to driving stimulus is 500 ms (when an engaged driver). The human reaction time to apply a change once the decision is made is 200 ms. Add to that however many ms it takes to notice that the Tesla is not responding the way you want (say it's as immediate as any other stimulus, so another 500 ms) and you can see how it's not as safe.

> to make that happen, Tesla needs this current generation of "supervised Autopilot". Because somebody's got to train and fine-tune that neural network.

I see no reason why my commutes should be more dangerous or my driving should be made riskier to train an AI I am not being compensated for nor consenting to improve.

It's such a great auto pilot but you're on the hook if anything ever goes wrong seems contradictory

Reply on news.ycombinator.com