Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

I think the fundamental issue Musk has is that his cars can’t drive safer than even a below average driver.

Oh wow, it can stay in a lane and match speed! So can every car manufacturer. They’ve got a level 2+ system and cars won’t be safer until level 4.



>I think the fundamental issue Musk has is that his cars can’t drive safer than even a below average driver.

No one in the public has enough data to say one way or another whether that is true and by presenting it as fact you are falling into the same trap that OP is warning about.


Well it's certainly true outside the limited ODD of autopilot, that much seems undeniable. (?)


> I think the fundamental issue Musk has is that his cars can’t drive safer than even a below average driver.

Agreed.

> Oh wow, it can stay in a lane and match speed! So can every car manufacturer. They’ve got a level 2+ system and cars won’t be safer until level 4.

I don't find this to be true at all. ProPilot is terrible compared to AutoPilot and many of the other manufacturers have the same limitations (or worse).


> ProPilot is terrible compared to AutoPilot

I second this - recently rented a fully-loaded Nissan Rouge and drove for around 700 highway miles - the lane centering feature was just awful compared to my Model 3's. It would often times give up on even slight curves, or fight me when I would try to regain control after it misjudged a curve. After a while I gave up on it and just used the automatic cruise control, which worked fairly well.


> ProPilot is terrible compared to AutoPilot

This is frightening. I strictly limit where I will let AutoPilot operate because it is not at all relaxing unless you're on the open road with light traffic. If this is the state of the art, FSD is decades away.


I'm curious... which version of AutoPilot do you have?


Do you have data that supports what you think?


The OP is simply stating the base case: cars without drivers shouldn't be safer than cars with drivers. The industry claims that their cars are safer without drivers because AI. The industry must provide the evidence for its very strong claim.

The fact that a majority of the public think it's the opposite way around, that skeptical opinions about self-driving cars must provide evidence, is the industry's succes- the triumph of marketing over reason.

Of course the industry has no such data. Research has shown that this data would take many decades to obtain:

https://www.rand.org/pubs/research_reports/RR1478.html

Personally the only justification I've heard for self-driving cars being safer is "computers are faster than people". Yes- but at what?


Computers don't get board and start daydreaming (when I win the lottery where will the pool go in my mansion). Computers don't get distracted reading (ie facebook) while driving. Computers can have redundancy so the equivalent of a heart attack won't be endanger everybody (this assumes of course the redundancy works - computers have a lot of failure modes of their own)

There is plenty of opportunity for computers to be better than humans. It is a hard problem though.


OTOH, humans can tell that a merging lane is a merging lane. Hard problem indeed :).


Today that is a factor. When (if...) computers figure that out though they never forget to look to see if the lane is merging.


At the risk of showing just how ignorant I am of current state-of-the-art AI, it feels like the biggest block with self-driving cars is how short their memory is and how little high-level logic is implemented. The computer doesn't feel like it's driving the car as an activity so much as simply responding to inputs as they come into sensor range, sometimes correlated, frequently not.


> Personally the only justification I've heard for self-driving cars being safer is "computers are faster than people". Yes- but at what?

Faster at making the right decision - and the wrong decision too.


Computers can "watch" every direction at once.


They'll also happily be fooled by adversarial patches (optical illusions) in ways that wouldn't fool humans. https://www.vox.com/future-perfect/2019/4/8/18297410/ai-tesl...


And humans are happily fooled by conditions (optical illusions, rain) in ways that wouldn't fool machines.

https://xkcd.com/1958/

If you want to murder people on the roadway, there are lots of ways to do this that are easier than fancy adversarial attacks (which only fool one sensor type, at best... and are then patched).


But can they remember that they just saw something that is not visible any more because it went behind another object?


The quarterly safety numbers that Tesla releases are difficult to draw conclusions from, because they do not represent a controlled experiment over the same road conditions. However here are the numbers for millions of miles per crash (higher is better).

With AutoPilot, Without AutoPilot but With Active Safety Measures, Without AutoPilot or Active Safety Measures, and NHTSA Average Vehicle

  19Q3 - 4.34 2.70 1.82 0.50
  19Q2 - 3.27 2.19 1.41 0.50
  19Q1 - 2.87 1.76 1.26 0.44
  18Q4 - 2.91 1.58 1.25 0.44
  18Q3 - 3.34 1.92 2.02 0.49


Those numbers are useless for concluding anything about what would happen if you remove the human completely, there are videos that show how the human intervened and saved the car from certain crash and I this situation should be counted as "Deaths" for the AI only column.

Driving assistants would improve safety certainly, but removing the driver and letting the assistant drive will need real world like tests, for now maybe you could have an independent group monitoring a number of Teslas and checking each time a human intervened and give a score on how bad the AI was.


Actually I’m not convinced assistants will improve safety.

Level 4 for example could be a blood bath, where cars drive all the easy situations, and hand over emergency control to out of practice drivers barely paying attention at a moments notice.

I believe our driving skills were better when we had manual transmission, and were more involved in overall operation. Hard to determine if safer, b/c of advances in crumple zones, airbags, etc as manual declined.


The sudden control swap you describe is level 3.

Level 4 requires that the vehicle safely stop in situations it can't handle.


We do have some data. E.g., the Volvo XC90 hasn't killed any passengers, ever, thanks to collision detection and emergency braking systems: https://www.telegraph.co.uk/news/2018/04/15/britains-safest-...


It is possible , we would need data,

The extremely sad part is that we could reduce the number of deaths until true AI arrives using simple tech that could detect if the driver is drunk or tired, and controversially forcing the speed limits(the car knows that this road is max 90Km/h and that the driver has less then 1 year experience so it will not let you get over 70 (local laws apply) unless you activate emergency mode - this would trigger some bad consequences if it was no emergency)


Good to see improvement, but is this controlled for the fact that autopilot is mostly on highways? I'm not sure what the crash rate per mile is on the highway vs driving in a city.


I use autopilot even on small roads even if it's not designed for it. After V10 it also handles roads without a centre line just fine.

However I do pay attention and keep my hands on the wheel, I think autopilot + human is much safer than either one, but since the latest updates I rarely have to take control if not going in roundabouts or junctions.

I also had one instance a few weeks ago where I tried to enter another lane manually and the car took control from me and put me back in my lane as there was someone in my blind spot without their headlights on so I didn't see it.


That sounds promising. OTOH, my car tried to drive into a tractor-trailer a couple weeks ago. That was a momentarily interesting experience. It seemed to become confused by the onramp merge joining my travel lane (pretty normal for autopilot to screw this up), but then it "realized" the mistake and dove for the other lane. I haven't been as happy with v10 as I was with v9.


Yes, I know especially model 3 owners are less happy with V10 than V9 in this regards. I've had the opposite experience with the X.


I would bet a body part on it being way lower, especially if we do not limit it to fatal crashes only (I read that statistic quoted above as "all reported crashes", not just those with fatalities).

I had two crashes in my lifetime. Both occurred when driving backwards out of a parking lot. No fatalities, limited damage. Good luck finding such situations while driving on a highway...


I don't know about per-mile, but there are more crashes in the city per hour.

(that's data from the highest crash rate country where you were able see so many crashes that you develop a feel)


That's not how this works. If the cars are safer, there should be independent data that shows this. As it is, we get one-sentence "safety reports" from the manufacturer with a single data point. They're not safer by default because the manufacturer says so.


The claim was that Tesla’s system is not safer, and that only level 4 would increase safety. I asked if data supported those two independent claims.

I’m ok if we assume that things are unsafe before we have evidence they are safe, but that’s not the same as declaring they are unsafe without evidence.


The current situation is somewhere between the two: There have been some crashes (e.g. Boca Raton) that strongly suggest there is a dangerous systemic problem.

In such cases, demanding statistical confirmation that it is a significant problem (as NASA did with the O-ring erosion problem) is not the right response. It reminds me of a case in the building of the 688 class submarines: a cracked weld, that had supposedly passed multiple inspections, was found in part of the torpedo-handling equipment. At that point, all welds, including pressure-hull welds, became suspect.

https://www.goodreads.com/book/show/31730.Running_Critical


Unless you are Boeing.


Autopilot isn't even designed to be used without a human driver. Which confirms that even Tesla doesn't think its better than a human.

It can't even handle intersections.



That article describes a failure of Tesla’s system that had terrible consequences. It gives no data comparing with human performance.


That article describes a systematic design defect of Tesla’s system. (It’s a defect that all camera based systems have. They are designed to ignore stationary objects because they can’t adequately distinguish background objects from obstacles.) Car and Driver tested this last year, and the Tesla ran into the stationary dummy nearly every time. This is something humans have almost no problem dealing with.


It has been a problem, but it is definitely not an intrinsic defect of camera based solutions. The stopped truck example has been examined in depth with rooted Tesla's. The Tesla actually detects the truck, but fails to fully recognize the scenario and attempts to drive under it! See also, greentheonly's work on Twitter on researching this.

Ignoring some stopped objects is only a temporary limitation.


I said it was a systematic defect camera based systems share, not that it was intrinsic to camera based systems. Camera based systems have much less information about object positioning than LIDAR based systems. They filter out non-moving objects because the vision processing algorithms are not sophisticated enough to distinguish between the background and obstacles. If they stopped for stationary objects in the road, they’d routinely stop because they mistook the background for a stationary object.

That may or may not be a temporary limitation. It depends on whether vision processing becomes reliable enough where you can almost always distinguish an object in the road from an object in the background.


> I said it was a systematic defect camera based systems share, not that it was intrinsic to camera based systems.

And I'm saying that even that is incorrect. Here is a Tesla failing to stop: https://twitter.com/greentheonly/status/1134987596508143617 - The driver clearly noted that aiming for a less clear spot (dangling wires) triggered a stop.

Here is another example, in which greentheonly has used root to overlay what the car sees: https://twitter.com/greentheonly/status/1135011184439169024 - It sees the truck but is treating some of the space under it as drivable space.

Another view, with more details on what the car considers drivable vs non-drivable: https://twitter.com/greentheonly/status/1134490014321131521

Also, a clear case where it did stop: https://twitter.com/greentheonly/status/1134489977801334784

The Tesla system has limitations around object identification, but absolutely does NOT just filter out non-moving objects.


Assuming what you say is true, is the Tesla system worse than a human driver overall in its operating domain?

Humans have weaknesses too. What do the data say on their relative performance? I’m not taking a position here, just asking for data.


Look, "autopilot" works on highways, where there is a median. Everything goes the same direction, and there is a minimal amount of clutter.

Its about as easy as driving gets. This is why a level 2 "autopilot" can parade around the place as something that is both innovate and safe.

None of the data compares what happens at junctions, which is where the majority of crashes happen.

It cannot work in suburban sprawl, It can't cope with poor weather, it can only drive on highways.


Not sure why you state this. I've used autopilot in a Model 3 on city streets and suburbs, it works pretty great.


Similarly, I have used autopilot in rain so bad that I was scared to drive. It performed perfectly and was much less stressful for me.

I was fully engaged and ready to take over, but never needed to.


I'll echo that. Heavy rain, I could barely see the lines, but the car seemed to find them just fine and was driving straight and smooth.

Until traffic on the other side of the median barrier hit a big deep puddle, absolutely covering my car with a thick sheet of water. Autopilot immediately started screaming at me to take control. I already had my hands on the wheel, but it still scared me because I couldn't see either! After the water washed a way in a couple seconds, I was able to re-enable AP and keep going.



That link does not provide data either way. In fact it explicitly declares its inability to do so:

“Cars driven under traditional human control are currently involved in approximately 1.18 fatalities for every 100,000,000 mi (160,000,000 km) driven.[3] According to many automotive safety experts, much more data is yet required before any such clear and demonstrably higher levels of safety can be convincingly provided”


Drivers immediately seem to offload their responsibility into this capability.

Anecdotes:

"I might be too drunk to drive, but I'm sober enough for my Tesla to drive."

And every Tesla driver I see engaged in two-handed cell phone operation with no hands on the wheel.




Consider applying for YC's Fall 2026 batch! Applications are open till July 27.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: