Tesla FSD Safety Stats Misleading

[whitex]

Look at this accident I witnessed a week ago. There is no doubt in my mind that if the victim car had FSD, it would have been avoided. The Tesla would have seen the car coming over in the right rear and moved over.

Hopefully it would have - you'd have to test it in a scenario like this (very rapid, unexpected PIT maneuver). An interesting corollary to this - if FSD would in fact avoid this PIT maneuver, does it mean police chase vehicles will no longer be able to use the PIT maneuver to stop cars running from the police (since FSD will just skillfully avoid the PIT every time)?

Perhaps a better point would be if the Honda had FSD, it would not have done this dangerous stupid move in the first place (no situational awareness, no signaling, rapid lane change)? Humans are increasingly driving worse and worse. I personally blame cell phone usage while driving over any other factor. A couple of weeks back I saw an accident where a young couple was in a car, the road turned, but somehow neither of them noticed, resulting in them driving straight into a ditch. Perfect weather, no rain, clear visibility, low speed (35mph limit, they were driving under that) they both got out of the car (thankfully) unhurt to wait for the police and tow.

Sponsored

 

[whitex]

Isn’t the point to be “better than humans?”

It will be, or already might be better than humans overall (plus the bar is lowering as the general population is getting worse at driving). Even though AI is just mimicking human driving, the key is that you feed AI only the top x% of safest drivers to train on, so it ends up driving like those drivers.

 

[69Mach390]

Conversation so far has been interesting, but no one has commented on the point of the Reuter’s article that says the Tesla employees who train FSD don’t trust it.

I’m having trouble getting past that.

Yeah, the people who cook fast food typically don’t eat it either.

They know too much.

 

[69Mach390]

It will be, or already might be better than humans overall (plus the bar is lowering as the general population is getting worse at driving). Even though AI is just mimicking human driving, the key is that you feed AI only the top x% of safest drivers to train on, so it ends up driving like those drivers.

Human + computer is better than either alone.

My point was just a counterpoint to the “2 eyes 2 ears” argument.

I want safety features that can see things I cannot vs a co-pilot with my same limitations.

I also know that we have better predictive behavior instinct than the computers. You know that car you can tell is probably going to cut you off based on how aggressively you see it driving behind you? Or the motorcycle you just know from experience may cut between cars? Things like that.

 

[whitex]

Human + computer is better than either alone.

Counterintuitive, but not true. As the computer gets better (say it only requires human intervention once a year) the humans get more complacent. Furthermore, it turns out as the computer gets better, humans actually screw things up interfering.

I want safety features that can see things I cannot vs a co-pilot with my same limitations.

You indirectly get this - you get a copilot with 12 eyes instead of 2 a swivel, so a nice continuous visual coverage (those eyes never blink). The computer never gets tired, or distracted. I get wanting superhuman senses into this (like radar, lidar, ultrasonics) , and you do get them on some platforms as safety guiderails. They might come to main world-model-AI driving too (someone might find a way to generate useful training data), but it also might prove to be unnecessary, especially with V2V (no need to see a mile ahead if you can see everything every car ahead of you sees).

There is one interesting radar usecase that I saw, ground penetrating radar used for geo location. Not sure how far the researchers got with that one, but it basically relied on the fact that underground rarely changes, so if you map it with a radar, you can then use the radar for very precise location of your vehicle, despite snow, mud, water, whatever.

I also know that we have better predictive behavior instinct than the computers. You know that car you can tell is probably going to cut you off based on how aggressively you see it driving behind you? Or the motorcycle you just know from experience may cut between cars? Things like that.

You may find it surprising, but AI is really good at pattern recognition like that, and getting better. It can do even more. I've seen an interesting scenario a while ago where a self driving car reacted to something which no human could explain - it was driving in a fog and suddenly applied brakes to slow down. Engineers thought it was a bug, but they couldn't figure out why. Purely coincidentally some engineers got to talking about this months later and someone provided an extended recording of this incident. It turned out there was a car not super far ahead of the self driving car, but it was technically not visible to any humans watching the recording until few minutes later in the drive when the fog cleared. So apparently the computer somehow inferred from the swirl patterns of the fog the presence and speed of the vehicle ahead, so it slowed down to match it. I'm thinking you don't have this superpower, but an AI trained on a billion miles of recorded drives somehow learned it.

 

[69Mach390]

Counterintuitive, but not true. As the computer gets better (say it only requires human intervention once a year) the humans get more complacent. Furthermore, it turns out as the computer gets better, humans actually screw things up interfering.

You indirectly get this - you get a copilot with 12 eyes instead of 2 a swivel, so a nice continuous visual coverage (those eyes never blink). The computer never gets tired, or distracted. I get wanting superhuman senses into this (like radar, lidar, ultrasonics) , and you do get them on some platforms as safety guiderails. They might come to main world-model-AI driving too (someone might find a way to generate useful training data), but it also might prove to be unnecessary, especially with V2V (no need to see a mile ahead if you can see everything every car ahead of you sees).

There is one interesting radar usecase that I saw, ground penetrating radar used for geo location. Not sure how far the researchers got with that one, but it basically relied on the fact that underground rarely changes, so if you map it with a radar, you can then use the radar for very precise location of your vehicle, despite snow, mud, water, whatever.


You may find it surprising, but AI is really good at pattern recognition like that, and getting better. It can do even more. I've seen an interesting scenario a while ago where a self driving car reacted to something which no human could explain - it was driving in a fog and suddenly applied brakes to slow down. Engineers thought it was a bug, but they couldn't figure out why. Purely coincidentally some engineers got to talking about this months later and someone provided an extended recording of this incident. It turned out there was a car not super far ahead of the self driving car, but it was technically not visible to any humans watching the recording until few minutes later in the drive when the fog cleared. So apparently the computer somehow inferred from the swirl patterns of the fog the presence and speed of the vehicle ahead, so it slowed down to match it. I'm thinking you don't have this superpower, but an AI trained on a billion miles of recorded drives somehow learned it.

I’m talking about today with Tesla and Porsche, not some future date when the computer doesn’t need us anymore nor the ACTUAL full self driving (driverless cars) with significantly more sensors. Those are better than us.

 

[FlyingPoint]

Whitex:
Hopefully it would have - you'd have to test it in a scenario like this (very rapid, unexpected PIT maneuver). An interesting corollary to this - if FSD would in fact avoid this PIT maneuver, does it mean police chase vehicles will no longer be able to use the PIT maneuver to stop cars running from the police (since FSD will just skillfully avoid the PIT every time)?

The FSD would not speed up to avoid, it would only move over. So if you keep coming at the FSD car, eventually the FSD would not be able to move over any more as either another car or barrier would be in its way.
Whitex:
Perhaps a better point would be if the Honda had FSD, it would not have done this dangerous stupid move in the first place (no situational awareness, no signaling, rapid lane change)? Humans are increasingly driving worse and worse. I personally blame cell phone usage while driving over any other factor. A couple of weeks back I saw an accident where a young couple was in a car, the road turned, but somehow neither of them noticed, resulting in them driving straight into a ditch. Perfect weather, no rain, clear visibility, low speed (35mph limit, they were driving under that) they both got out of the car (thankfully) unhurt to wait for the police and tow.

Simple lane keeping should have alerted the driver as it crossed the lane line. In my experience, FSD will not turn into a lane where a car is present. In my Tesla, I have my light bar programmed to turn yellow when a car is next to me and red if my blinker goes on to swith lanes into it.

 

[whitex]

I’m talking about today with Tesla and Porsche, not some future date when the computer doesn’t need us anymore nor the ACTUAL full self driving (driverless cars) with significantly more sensors. Those are better than us.

IIRC Tesla has ACTUAL full self driving cars in Austin already (RoboTaxi service without safety drivers). Porsche is nowhere near.

 

[whitex]

Simple lane keeping should have alerted the driver as it crossed the lane line.

The video doesn't look like someone crossing the line, it just looks like someone just decided to change lanes without signaling or paying attention, so even blind spot warnings would not go off in most cars. So many people drive cars like they'd ride a cow, get on, point it in general direction, once in a while yank on the cows ears to correct the direction. Sadly, I've seen behavior like in the video many times, someone on their phone, looks up, realizes they are about to miss an exit or turn, so they yank the steering wheel plowing across multiple lanes of traffic to get to the lane they need. Most of the time they don't hit anything, just get honked-at and/or flipped off. Sometimes they end up jumping lane divider curbs (I once saw a car almost flip when they did that in front of me, ended up recovering after driving on just 2 right wheels for a few feet), taking out cones or lane dividers, or catching a flat tire due to cutting across shoulders where all the debris accumulate.

IMO, until we get to all cars driving themselves, we need to make getting a driver's license harder (make tests require higher level of driving skill, possibly require specialized training), and probably need much more strict driving-while-using-a-phone enforcement. I've seen people actually watching videos while driving (movies, shows, TikTok, YouTube, ...), on a lot more occasions than you'd think, and not even hiding it.

 

[69Mach390]

IIRC Tesla has ACTUAL full self driving cars in Austin already (RoboTaxi service without safety drivers). Porsche is nowhere near.

I know Phoenix (and other cities) have the driverless Jaguar iPace all over the place. It’s so strange to watch them maneuvering. But they have a ton of cameras and sensors.

 

[AutoX]

Tesla FSD is already safer than the average driver.

The industry is moving toward camera only for autonomous driving. It's looking like Tesla will win this round of the VHS vs Betamax war.

If you haven't experience FSD first hand, I recommend that you get a test drive at a tesla center. You'll absolutely be blown away by the advances over the standard active lane keeping autonomy found on the Porsche.

 

[whitex]

I know Phoenix (and other cities) have the driverless Jaguar iPace all over the place. It’s so strange to watch them maneuvering. But they have a ton of cameras and sensors.

Those are Waymos (Google).

Robotaxi (Tesla) is starting in Austin, but expanding to other cities quick. Robotaxi is a Model Y:

There is also CyberCab from Tesla, which will be joining RoboTaxis.

 

[FlyingPoint]

It's a "brave new world". I will take autonomous vehicles whether Waymo or Tesla any day of the week over the assholes I encounter on the Long Island Expressway every weekend.

 

[Vercingetorix]

Mercedes has level 3 self driving. They don’t brag about it even though they are well ahead. And they don’t kill the guinea pigs.

 

[AutoX]

Mercedes has level 3 self driving. They don’t brag about it even though they are well ahead. And they don’t kill the guinea pigs.

Mercedes level 3 is limited to <40 mph. Not much to brag about. Also there doesn't seem to be much difference between a level 2 and level 3.

https://www.nhtsa.gov/sites/nhtsa.gov/files/2022-05/Level-of-Automation-052522-tag.pdf

base on this, Tesla's system really sounds more like level 3 vs level 2.

Sponsored