Hacked Tesla FSD computers disclose alarming raw data on deadly accidents
notebookcheck.net
notebookcheck.net
This is obviously true. However, I never heard these critics say that about cars without autonomy features. Way over a million people killed each single year around the world by that technology. Where was the concern for human life?
To anyone looking at it rationally, without emotion, it's clear that humans are not good at this task and that computers will eventually make roads safer than ever. The question is not if but when. Currently that data is debatable but some argue this point has already happened and cars with computer-assisted driving are already safer than those driven entirely by humans.
It’s obvious that SOME humans are bad at this task.
It’s not at all obvious that a computer will ever be better. I’ve yet to see a single example of a computer that doesn’t trip all over itself on a snowy road, or lock up when someone intentionally marks up the roadway with graffiti that a human wouldn’t think twice about.
if the best driver in the world dies because they used FSD, then fsd is the cause so it must be as good as the best driver in the world to have not killed that human. q.e.d.
This could be due to Dunning-Kruger, age based trust of new things, socioeconomics, etc.
We are a LONG ways off from FSD being better than the average driver. Anyone saying otherwise is apparently living in the Bay Area or one of the places an influencer lives that Tesla tries to game the system, because it is trash everywhere else. It still can't handle navigating a road that has the option to continue straight or go around a curve which even a novice driver can navigate without pause.
How to I objectively test if I'm a good enough driver to be better off without FSD?
In the future, vehicles could have their FSD augmented by direct comms with each other, in a trustless way. If the entire traffic network ended up interconnected, it'd behave as a single system, approaching 100% self-awareness.
Obviously, no ordinary human drivers could get anywhere near that level of situational awareness, nor ability to coordinate in parallel. All they have is a small cone of vision/hearing, a bit of sensor augmentation, and little else.
You’re not even going to build an actual strawman before claiming to have knocked it down?
We can't get Amazon and Microsoft to agree on a single object-storage standard, and you think you're going to get every major auto manufacturer, including ones regulated by opposing nations to agree on an automobile communication standard? That who controls? And it won't be considered a national security risk to allow a foreign manufacturer to be able to potentially gridlock the entire traffic pattern(s) of a nation?
>All they have is a small cone of vision/hearing, a bit of sensor augmentation, and little else.
Which literally no amount of compute power has been able to replicate or surpass to date.
Right....
For instance, we all use the metric system… we all use Nema plugs into the wall… we all speak English… we definitely all eat beef… ok, I’m confident we all at least use the same ammunition in our weapons several hundred years later? No?
There are a few Tesla crashes on camera, and they look like something even a fairly bad driver could avoid. See for example the SF tunnel crash: https://abc7news.com/tesla-sf-bay-bridge-crash-8-car-self-dr...
Besides, other vehicles have now active safety functions that seem to work extremly well. There are car models out there with ZERO fatalities so far thanks to this technology. So FSD being an improvement compared to really bad drivers is not really an improvement.
"... Tesla Model S was in "full self-driving mode" when it suddenly deployed the brakes. A total of eight vehicles crashed, nine people injured, including a 2-year-old boy."
From the video it looks like there would be ample time to do just that, even before the first crash.
If I were given a choice between one in five chance of death by cancer, or a one in six chance of death by Russian roulette, I'm going to take the Russian roulette it doesn't matter that it's a dumb game. What matters are my better odds.
This isn't to say that we should accept people playing Russian roulette it's just that if Russian roulette is a cure for cancer then we got to accept those stats.
Now if we're going to compare active safety functions to Russian roulette with rubber bullets, then maybe the FDA should mandate Russian roulette with rubber bullets.
Most people don't drive like crazy and taking a ride in my car is statistically not as dangerous as playing Russian roulette.
True
> and that computers will eventually make roads safer than ever
Nothing suggests this is true, so far. Computers are much less able than even humans to devise the right reactions to the complex environment of the road.
This has "Tesla" in the headline.
In the laws that assign liability to the at-fault driver and in the most heinous cases charge them with vehicular manslaughter. I'll accept autonomous driving on public streets when the vehicle manufacturers accept liability for accidents caused while their self-driving functions are active. My understanding is that currently Tesla always says "actually the driver is at fault because our Terms Of Service say they should have intervened when the computer didn't know what to do." That's not autonomous driving and they don't get credit for vehicle-miles driven as long as the system is allowed to throw up its hands and say "your problem now, human."
I feel like this will be a never-ending battle with a camera-based system. There will always be some visual situation not seen before. Shouldn't the default behavior be to slow/stop or throw up its metaphorical hands and release control back to the driver?
If I remember right it was lidar overriding the camera as the set up when Tesla was using lidar...
Look at what he did at twitter. As soon as he was put in charge he removed all safety measures to save money and now twitter is full of bots and scammers. Why do people assume he runs Tesla any differently??
Tesla is selling a driver assisted system that can't reliably turn the car or park it without manual intervention a decent percent of the time.
Plus with LIDAR being an active system you're more likely to see nothing or garbage where the ground and a road is supposed to be. Good idea to stop if there's a hole in reality!
With camera+AI it's the opposite. For too many things AI won't really know what they are - it's mainly looking for things that it does recognize. With unclassified pixels everywhere, of course the default is to ignore what it doesn't understand, otherwise the system would never work. You can see this with the Tesla happily driving through some unrecognized pixels.
LIDAR-based systems can operate without having to unsafely ignore things that they can't account for or that don't make sense.
Merging a camera with a camera, and a camera with a completely different kind of a sensor cannot be "the same problem". It's absurd.
People wouldn't buy a car that slows down occasionally because they'd accept it would happen to they and they'd find it annoying. They'd happily buy a car that randomly kills people because they'd tell themselves it's only happen to other people.
At some point new buyers attitude will change from "this will never happen to _me_" to "shit, this happens all the time".
Also Robotaxi customers might complain more - or sue - compared to Tesla owners.
I don't think he conflates those things in his own mind, only in his rhetoric. In other words, I think that Musk lies when it suits his purposes.
It’s a charisma that convinces other people of what’s true. Elon has that, that’s why said hard on exists. He often turns out to be right.
If you think it isn't lies, then you accept he's not very intelligent or at least illogical.
The simplest answer is he's voicing his ultimate vision for a product without basis in reality like the solar powered roadster, robotaxis, etx.
It's marketing and the result is the same whether he lies or not, only distinction is if he's mentally handicapped or dishonest.
I do not think he’s always truthful by any means. But a failed prediction is not a lie. as the great George Costanza once said, it’s not a lie if you believe it.
It's marketing. You're making things very complicated and distorting reality.
Not all that many.
He’s always overly-optimistic about cost and timeline and he’s got a whole host of issues due to his mental illness but the dude did do several extremely impressive things that nobody thought would happen.
Compared to what he has accomplished you can understand how “a car can drive itself without LIDAR” seems easily believable. It may be wrong, but it’s not as hard to imagine as Starlink was ten years ago.
Or delusional.
I think he's a smart guy. I don't think he's remotely as smart as he thinks he is, but smart nonetheless. So calling him a willful liar is, in my view, giving him the benefit of the doubt because the other explanations are even worse.
This is going to bite him in the ass w.r.t. Starship, too: Rapid reusability, return to launch site, and in orbit refueling are just the major aspects of a dependency hairball in Starship. Either it all works, or none of it works well enough to actually be more efficient than a non-reusable second stage.
It's not technically impossible to solve. Humans drive with vision, audio, and balance inputs, so in extremis if you want a self-driving car you can hire a chauffeur.
However, no matter how easy or difficult it is to drive with vision alone, more data is better: driving with vision plus lidar will be easier.
Tesla's mistake seems to be relentless optimism. If self driving were simple, then it'd be here today and the market would be competing on how to add it to cars with minimal additional cost. There, the data-light method would have an advantage. Yet self-driving is not in fact here, and for equal capital investments a company like Waymo using sensor fusion will probably get there before Tesla.
(Other advantages, such as "never gets drunk, tired, or distracted" would still apply, but are not necessarily sufficient).
And they also make mistakes. Human driving is "impossible to solve" too. And it's not even about experience and individual abilities. Visual illusions can happen to anyone if they're unlucky - one night I had to convince myself to drive "into the wall", because my brain decided the hole under a bridge was a wall and the bridge was empty space due to colours being just right. You can't avoid that whether the driving system is software or wetware.
a self driving (or even self-stopping/accident avoidant) car does not remember it didn't see any signs saying road closure, it can only go off what is in front of it.
whenever i watch the videos of the FSD screen in traffic, you see things popping in and out of existence. This, to me, is why "FSD" cannot be better than a human, even with advanced sensors. Maybe if all cars were networked, but guess what, i will never own one, i'll buy old cars until they outlaw it, and then i'll probably sit in my house and wait to die. But until "non-network equipped cars are banned nationwide for any purpose" what do you propose?
Also this was on the front page on my phone's app, but it wasn't on the first 5 pages on the website - is this whitewashing by someone at HN or what?
There is Gaussian splatting now. That seems fast enough for real-time, and would provide all that you need for this. If enough Gaussians are about to hit the windshield, slow the f* down.
* Fixed a bug where telemetry wasn't wiped after a crash
* Various fixes and improvements
If it's not a joke it seems like something that should be prevented by regulators.
It would be no different than airplanes wiping their flight data recorders after a crash.
If so I agree that that is a fallacy, they definitely are not as good as eyes, and I’d think anyone who drove a Tesla would agree.
“causes a fallacy most people don’t even realize - eyes are way better than current cameras.”
This implies “eyes are better than cameras” is a fallacy. But the rest of the comment makes me think you’re right.
It’s not hard to imagine a car having better visibility than a human with cameras. But I don’t think they do yet.
Strictly speaking, this is not at all true. Eyes are pretty bad at this, much worse than cameras have been for a long while.
Vision, however, is much better. This is because of all of the postprocessing our brains do with the raw information our eyes put out before we actually "see" what our eyes are pointing at. But even that is full of errors (think optical illusions).
You don't see that because your brain does a ton of fixing up and filling in the blanks before you perceive things. Much of what you think you're seeing is actually your brain reconstructing things.