Hacker News Reader

Comment by maypeacepreva1l | Hacker News Reader

Tesla’s software is dangerous. I own one and I have to be a lot more alert than when I am actually driving. Just yesterday when I was driving rural town, it was stopping every time it sees a truck in the incoming lane. I suspect if something like that happened in this case as well.

View on HN

I can't imagine having a Tesla with the drive-assist software in its current state. It's like having a mentally ill spouse at home, you always wonder what irrational thing is going to come next, you have no peace any second of the day. Now try that during something safety-critical, like driving. The constant hypervigilance is exhausting, and when you aren't looking, the Tesla drives itself into a traffic bollard. Why is such crap road-legal?

Because our current form of government can't keep pace with the rate of new things.

Maybe the lawmakers, are asleep at the wheel?

and paying $10k for it, too. I imagine a number of customers paid for it, so they want to use it, no matter the state it is in.

that is the only way it will ever get better.

I have a Model Y, and I don't feel comfortable with autopilot or lane assist. Both of them do "dumb stuff" regularly. The day I got the car, lane assist attempted to steer me into the center median of a highway when the car was going around 50 miles per hour. After that, I realized I'm not interested in "beta" features that put my life in danger. Autopilot drives like a paranoid grandma. It brakes constantly when it shouldn't. All that said, the car is a lot of fun, and I really enjoy it, but I choose to maintain full control at all times.

> The constant hypervigilance is exhausting

This is exactly it, IMO.

Whether you like AP or not largely comes down to how your driving style compares to the computer's. If you are a defensive driver, you won't really care for the kinds of obviously dumb situations AP will happily drive you into. Eventually it reacts, but it takes a lot longer for the sensors to notice what a human brain can easily predict is about to occur.

I found AP to be an interesting toy, but it never made driving more relaxing for me, because I had to be more aware, not less, of everything around me.

I agree. Plus it is a lot more stressful to imagine that the car could stop randomly rather that hitting an object in front of you (which is already quite bad). The former is literally mentally exhausting.

Autopilot is not as bad as you think it is, for one it does not stop abruptly. Thousands of drivers are using it regularly. Drivers report less fatigue and paying better attention to things that matter rather than keeping the car in line. Of course, there are some drivers that are not comfortable with it yet.

EU seem to have more stringent regulations - for example, to ensure that AI maneuvers will be less unexpected EU regulations put a hard rate limit on steering angle. Trouble with regulating at such level is that in some cases (curved roads, hazard avoidance) it makes the car less safe.

Overall, Level 2 systems, like Autopilot, are not autonomous and their performance depends in large part on drivers judgement. I think regulations need to focus on human - AI interface requirements a bit more.

As a non-owner it seems kinda scary with the regular updates that the behavior would change. One day it stops for trucks in small towns, the next day it doesnt.

Does Tesla force OTA updates, or can the owner turn they off and accept them manually?

There are a couple controls. You can opt-in to get releases earlier, or wait until they're a bit more fully baked. And you can refuse the update when it pops up. I don't recall it ever being forced, but I could be mistaken.

Your experience is not universal. In my case the combination of me+software is certainly safer than me alone, and also less taxing for me.

Can you explain how you come to the conclusion that it is safer?

There's been a number of occasions where I would have been in an accident had it not been for the software saving me. Obviously it is possible that one day the software will result in an accident that would not have happened had I been driving alone, but thus far that has not happened, and so, thus far, I can conclude that the combination is better than me alone.

It’s fairly rare for a driver to be involved in “a number of accidents” in just a few years. Most drivers go decades with zero collisions, so if you’re experiencing multiple per year, saved only by Tesla software, you might try to see if there are other avenues that you could explore to reduce your risk to be more like the population average.

I would say close calls are fairly common, especially in urban areas with lots of traffic. It’s not just your driving but the people around you. It takes two to get into an accident.

Right. I suspect that the net effect is that the Tesla software transforms these what would have been close calls into…still close calls where the Tesla software gets credit for a “save”.

If a Tesla “saved” a driver 5 times in 20K miles, my first question is always going to be “how many collisions did they have in the prior 20K miles in their other car?”

This is a good point, and I agree. If there was a close call you can't actually say for sure if it would have been a collision without the software.

This effect has a large influence on the overall A/B analysis. If pre/post test analysis suggests that Tesla saved you from only 0 to 1x collisions vs a previously estimated 4x-5x collisions, your tolerance for newly introduced collisions would naturally be much lower.

I get what you're trying to say, but if anything, you're just making the case for Tesla software.

Consider that there is a range of drivers, from "good" to "bad". If most drivers go decades with zero collisions, that's great, and it probably puts me closer to the "bad" end of the spectrum. By your own admission drivers like me should explore avenues to reduce our risk. Why is Tesla software not a valid avenue?

If you're a "good" driver, you don't need it, and that's fine, you can get some other car or drive with autopilot off or whatever. But for us "bad" drivers, the software makes us safer (both personal risk and to others on the road) so why not use it? What other avenues would you suggest exploring?

One person cannot drive enough miles in their lifetime to allow making a determination that one system is definitely safer than the other. The reliability that we require from an autonomous system means you might never experience a safety failure (again, even if you drive every minute of the rest of your life), but the system is still less safe.

Exactly. I mean how many accident do you expect to participate in your lifetime? Like probably 2-3 (scratched bumber) and 0 severe ones.

True, but manufacturers can look at aggregate miles traveled and come to some conclusions about the safety of vehicles without any safety systems, with active safety systems like automatic emergency braking, and in Tesla’s case, Autopilot. They publish the statistics regularly and crashes are far less common per million miles driven when autopilot is engaged.

I think these stats aren’t very useful because 1) Tesla drivers are different than other drives as they are rich. Comparing miles driven by rich people to all miles driven isn’t useful for knowing if Teslas are safer; 2) Tesla has really low numbers so it’s hard to compare a small sample to a huge amount. It would be like comparing walking accidents by 7th Day Adventists. They may walk a million miles among the whole population but that’s nothing compared to the trillion miles of the entire population.

It's my understanding there is no independent peer reviewed scientific study showing Tesla cars are safer.

> crashes are far less common per million miles driven when autopilot is engaged.

Does this count crashes that occur after autopilot disengages?

>They publish the statistics regularly

The data is not transparent, they publish what they want. How many times drivers intervened? Does Tesla or the fans consider those as a +1 for human and a -1 for the AI in the stats? Nope.

Tesla's claims of self-driving being more safe was recently debunked:

https://mobile.twitter.com/Tweetermeyer/status/1488673180403...

There have been difficulties in independently verifying at least some of these claims. I do not know whether they have been cleared up.

http://www.safetyresearch.net/blog/articles/quality-control-...

https://jalopnik.com/feds-cant-say-why-they-claim-teslas-aut...

It comes down to your driving style. Defensive drivers will not like how AP drives, and it makes driving much more tiring, not less. If you are comfortable with the way AP drives, though, I could see it being relaxing. Hell, some people sleep when AP is cruising down the highway, so clearly there is a spectrum.

AP somehow doesn't brake soon enough for my comfort when approaching slower cars and it doesn't accelerate quick enough when traffic speeds back up.

I agree, and I think most Tesla owners feel this way, given that the AP outrage seems to manifest itself exclusively on hacker news.

My wife forbade me from using AP [with her in the car] after a few phantom braking occurrences. I got sorta used to it and could jab the go pedal pretty fast, but it scared the daylights out of her. Can't really blame her. I get nervous enough as a passenger when there's a human driving, much less when a computer is driving that mistakes shadows for obstacles.

Exactly correcting software errors is a whole lot more effort than just sticking to your own way of driving.

Tesla/Elon Musk have this weird irrational obsession with using only cameras for their autopilot. Yes, humans can drive using only two eyes (and also actually ears, touch feedback and acceleration sensors), but they also have a lot of accidents. It's not obvious that you can do better with software only, and in any case you're crippling yourself by not using additional inputs for no good reason.

> irrational obsession with using only cameras for their autopilot

it's so they can upsell FSD with the highest margin using the cheapest hardware (they're not even good cameras, 1280×960 resolution because they want to push pixels directly into their neural net, and dynamic range is poor, I will never understand how they didn't spring for infrared night vision, cadillacs had it 20 years ago)

I think you could also say that people seem to have an irrational obsession with Tesla’s decision to only use cameras. Camera only systems aren’t uncommon among other automakers (Subaru Eyesight, for example.)

Your comparison does not apply. I do not believe Subaru is selling Eyesight as a hands-off driving system. AFAIK its purely marketed as a safety system for emergency braking, lane keep assist and the like. On Tesla's side they market it as a the holy grail "autopilot".

And it’s lane assist isn’t “keep the car centered”… it’s the steering equivalent to emergency braking, a sudden jerk away from the line it thinks you’re about to cross.

The newer versions have lane centering which isn’t quite equivalent to Autosteer in a Tesla, but it’s much closer to it than the older versions that only have lane keep assist, which only nudges you away from a lane line you are about to cross.

The reason I’m making the comparison is that there are reports of phantom braking when Teslas aren’t in Autopilot.

The same vision-based system is used for AEB in Teslas. Subarus use a vision-based system for AEB but don’t seem to experience phantom braking.

I think you could reasonably conclude that autopilot aside, a vehicle can safely use a vision-based system for AEB without phantom braking being a huge issue.

There is nothing irrational about being concerned by the fact that complaints of phantom braking incidents increased significantly after Tesla dropped radar from its data sources.

https://arstechnica.com/cars/2022/02/teslas-radar-less-cars-...

Our human "cameras" also scan constantly, have crazy dynamic range, employ dynamic shades (hands) and move around several feet, inside a protective weather proof shield, as well employ "sensor fusion" with "microphones" and "vibration detectors".

Reply on news.ycombinator.com