Tesla Launched FSD v14

[ChrisO]

I think our brains just aren’t wired to do this sort of risk analysis competently.

Carefully supervising FSD and being ready to engage immediately especially when it has done this exact drive correctly several times before is asking way too much of the Average Joe.

The EU did the right thing here. Random schmucks should not be beta testers of a device that can end lives.

I totally agree.

Studies show that the average person gets distracted about every 5 minutes and that is when they are “actively engaged”, there is no way that they can detect a problem, and get their hands back on the wheel in time if they believe the car is handling everything for them.

Sponsored

 

[E90alex]

I have a 2.5 hour drive I make at least a dozen times/year. Over 2 hours of it are on hands-free interstate with Blue Cruise. When I don't use BC, I get to my destination feeling like I drove 2.5 hours, and find a nap is useful. When I use BC (nearly always now when available), it seems like a 30 minute drive and I am raring to go at my destination. Just resting my hands on the arm rest or lap, and feet lightly on the pedal and floor does wonders for reducing the drive effort. When the road is very "tame" with no close traffic or other concerns, I also love to take in the scenery much more, up to the limit of the BC warning to keep my eyes on the road. Spending a couple seconds looking at something away from the road is so much better than a quick glance that I might allow myself while I am in control.

I, too, love driving, but prefer twisty mountain 2 lanes where BC doesn't work anyway to have my driving fun, and find interstate driving boring for the most part, gladly letting BC help.

Are you comparing full manual driving to hands free blue cruise?

Or comparing hands free blue cruise to hands on required adaptive cruise and active lane centering?

 

[bibkisbu]

I would like to see the actual statistics on this.

Whereas I don’t trust it, but in truth it is only competing against all the idiots on the road. That isn’t a very high bar to get over.

There's a lot of misinformation in this thread. Nobody has ever died from tesla FSD. Across over 4.5 billion miles, no one has died.

In the very old versions of FSD (back when there were less miles and less people using it) 2 people did die, because old versions of FSD would actually not include highway driving and the car would switch over to autopilot which is a 2017 software.

The person who died was actually someone who set his speed to 70 in a 45 (when you could do that) and a semi truck pulled out in front of him

Autopilot has completely different perception technology (similar to bluecruise but more advanced) and does not use neural networks.

Since FSD came to highways, FSD was rolled out of beta and no one has died. I believe 12.7 people typically die per billion miles so statistically this software + heavy driver monitoring prevented some deaths.

Bluecruise is actually in a probe by the NHTSA for failing to identify stationary vehicles. Tesla (even autopilot) does not have any inherent hard limitation against detecting stationary vehicles because they use vision and not radar.

 

[MacherAWD]

There's a lot of misinformation in this thread. Nobody has ever died from tesla FSD. Across over 4.5 billion miles, no one has died.

In the very old versions of FSD (back when there were less miles and less people using it) 2 people did die, because old versions of FSD would actually not include highway driving and the car would switch over to autopilot which is a 2017 software.

The person who died was actually someone who set his speed to 70 in a 45 (when you could do that) and a semi truck pulled out in front of him

Autopilot has completely different perception technology (similar to bluecruise but more advanced) and does not use neural networks.

Since FSD came to highways, FSD was rolled out of beta and no one has died. I believe 12.7 people typically die per billion miles so statistically this software + heavy driver monitoring prevented some deaths.

Bluecruise is actually in a probe by the NHTSA for failing to identify stationary vehicles. Tesla (even autopilot) does not have any inherent hard limitation against detecting stationary vehicles because they use vision and not radar.

If an accident is about to occur, and FSD bails out at the last second, do you not count that as an FSD error?

You joined today, have 1 post. T isnt releasing full data sets, and no one is requiring them. Tesla crash more than other cars, I suppose it could just be that people who buy Tesla's just suck at driving.

 

[bibkisbu]

I find it ironic that the China is leading the industry to make this safer.

This is also very false.

There was a test in china which compared all ADAS systems compared to simply tesla autopilot (not even FSD). And tesla absolutely dominated the competition even most competitors having lidar. No one else is close.

The issue is that safety is 99.999% a software problem and not a lidar problem. And the chinese systems prioritize driving features (which gives you the appearance of an intervention free drive), over safety. So because the car can park itself or do some maneuver it gives the appearance of a 100% intervention free drive even if 1/10 drives you have a dangerous mistake.

Tesla never prioritized driving features over safety because they were attempting to do unsupervised driving. Having drives with zero intervention is not important if the car is not fundamentally safe.

What makes FSD v14 very different is that the car is 1000x safer than it ever was on v13. v13 was smoother and honestly drove smoother than it is possible for a human to drive. But v14 is more realistic in that the car no longer has the loose behaviors it did with v13. It stays 100% laser centered in the lane lines, understands what it can't see with cameras and doesn't attempt to go in situations where the cameras are occluded.

There is also this misconception that driving without lidar/radar is not possible. At scale you don't know if it is not possible until you do it.

people think that the car gets blinded by the sun, can't see at night, when it's raining, etc. but this is simply not true. These cameras for automotive are HDR cameras which have less glare and higher dynamic range than our eyes do. They're not designed for producing good looking images

people think that the car gets confused by wil-e-coyote style painted walls but that is false as well. Monocular depth estimation gets depth from motion

people have this misconception that the car is purely guessing the depth of objects from the size of them. This is false. Cameras have fixed lenses and project the image onto the sensor. Therefore every photosite (pixel) exists along a vector from which it is projected by the lense. Take overlapping camera views, intersect the vectors and you have 100% known depth of every object. And then use motion, use other visual clues as well, and it's not the guess you think it is.

Tesla certainly has the bare minimum of cameras and computers to drive, but there is nothing inherently preventing them from doing unsupervised. The only requirements for unsupervised driving is that the car meet certain driving standards and complete the entire ODT including minimum risk maneuvers.

That means if tesla cameras are blinded (say if it drives into a puddle and the water splashes over the camera), the car needs to handle that condition. The response that tesla robotaxi does is to brake to the last known position of the car ahead (slow down) and instantly put the hazards on. After the 2 seconds the water clears, it moves at normal speed again.

 

[bibkisbu]

If an accident is about to occur, and FSD bails out at the last second, do you not count that as an FSD error?

You joined today, have 1 post. T isnt releasing full data sets, and no one is requiring them. Tesla crash more than other cars, I suppose it could just be that people who buy Tesla's just suck at driving.

Yes. All accidents within 5 seconds of ADAS enabled are counted.

Tesla reports accidents to the NHTSA. FSD is getting into accidents.

There was an accident where a guy jerked his steering wheel (disabling FSD) and steered it into a tree. His car flipped over and he went to the hospital. He genuinely doesn't remember doing it but the crash data says he did it.

Tesla reports it to the NHTSA because it's required and happened within 5 seconds of FSD being enabled

Tesla actually do not crash more than normal cars. In fact they likely crash less because the auto emergency braking is #1 in the world. The issue is that tesla reports all their accidents to the NHTSA while other brands lack the telematics data to identify an accident. So tesla tries to look as good as possible by reporting every ADAS accident to the NHTSA. And the company that stated tesla had higher than average accident rate estimated the miles of used teslas when the model Y was a very in-demand car and users were selling them for a profit or no loss. So they look at low-mileage used teslas and extrapolate that the miles of teslas must be lower than they are.

There is this misconception about ADAS levels and what tesla has with consumer FSD is level 4. The SAE J0316 makes it very clear that a system with a safety monitor (no matter how bad it drives) is level 4 if it does the entire driving task. The issue is in the same document it says that levels are self certified by the manufacturer. So tesla abuses the famous "tesla loophole" where they simply claim their system is level 2 and drive on public roads. It is possible this loophole is closed if another manufacturer tries to build a system and it is unsafe.



If you want to see what emergency braking looks like, here is the euro NCAP model 3 test. This is the highest score ever for emergency braking.

And these scenarios most american/korean automakers would easily fail. It involves seeing vehicles/pedestrians not in front of you and predicting their trajectory. Most cars do not have this technology.

you typically need to go with german luxury brands to get access to this and only in higher trims. Tesla is the only one doing this with the same FSD computer as they use to drive

 

[ChrisO]

Yes. All accidents within 5 seconds of ADAS enabled are counted.

Tesla reports accidents to the NHTSA. FSD is getting into accidents.

There was an accident where a guy jerked his steering wheel (disabling FSD) and steered it into a tree. His car flipped over and he went to the hospital. He genuinely doesn't remember doing it but the crash data says he did it.

Tesla reports it to the NHTSA because it's required and happened within 5 seconds of FSD being enabled

Tesla actually do not crash more than normal cars. In fact they likely crash less because the auto emergency braking is #1 in the world. The issue is that tesla reports all their accidents to the NHTSA while other brands lack the telematics data to identify an accident. So tesla tries to look as good as possible by reporting every ADAS accident to the NHTSA. And the company that stated tesla had higher than average accident rate estimated the miles of used teslas when the model Y was a very in-demand car and users were selling them for a profit or no loss. So they look at low-mileage used teslas and extrapolate that the miles of teslas must be lower than they are.

There is this misconception about ADAS levels and what tesla has with consumer FSD is level 4. The SAE J0316 makes it very clear that a system with a safety monitor (no matter how bad it drives) is level 4 if it does the entire driving task. The issue is in the same document it says that levels are self certified by the manufacturer. So tesla abuses the famous "tesla loophole" where they simply claim their system is level 2 and drive on public roads. It is possible this loophole is closed if another manufacturer tries to build a system and it is unsafe.



If you want to see what emergency braking looks like, here is the euro NCAP model 3 test. This is the highest score ever for emergency braking.

And these scenarios most american/korean automakers would easily fail. It involves seeing vehicles/pedestrians not in front of you and predicting their trajectory. Most cars do not have this technology.

you typically need to go with german luxury brands to get access to this and only in higher trims. Tesla is the only one doing this with the same FSD computer as they use to drive

If it is so safe why does it still have a level 2 rating?

I’m not making snide remark here, I’m genuinely interested.

 

[bibkisbu]

If it is so safe why does it still have a level 2 rating?

I’m not making snide remark here, I’m genuinely interested.

Levels of autonomy (1-5) do not indicate safety or ability with the exception of level 5 since it is complete driving. And that is not even possible.

Level 4 is not better than level 3 and level 3 is not better than level 2, etc. These are manufacturer assigned levels.

Tesla doing robotaxi in austin is level 4 because they say it is. Tesla using consumer FSD is probably still level 4 but tesla uses a loophole to claim level 2. By claiming level 2, tesla can enable this tech on consumer cars.

There are about 20 states tesla can operate consumer unsupervised driving if they had a system good enough. You can literally drive from montana to key west, florida, completely unsupervised

Level 3 even though it is "unsupervised" is an inherently dumb and dangerous system that puts all of the risk on the lead car in front of you. It's like a traffic jam chauffer and it's an entirely different branch of technology not based around self driving. A level 3 system could never be improved to be a level 4 system. Tesla could build a level 3 system in less than a week if they wanted to.

And you can have a level 4 system that requires human intervention every mile and you can have a system like bluecruise that could do 60 miles on the interstate. This is no metric of safety or performance

ADAS systems found in consumer cars are extremely dangerous. They don't see stationary vehicles, they don't detect vulnerable road users, they do not have the ability to dodge accidents. They do not see and anticipate cut ins, cars pulling out in front of you, etc. They don't work in blinding sun, in heavy rain, around steep curves, etc.

These are a primitive technology and building a self driving system (even supervised) has very strong implications even if you do not intend to enable it on anything more than highway driving.

GM tried to build an ADAS (ultra cruise) which would enable city streets driving based on old school ADAS techology. It was a dumb and dangerous idea and they abandoned it.

What people don't realize is most consumer cars have steering column control modules (how you send steering commands to the car) that would not allow you to turn the steering wheel quickly to dodge an accident or even do something as turn the wheel more than 90 degrees.

 

[ChrisO]

Yes that makes sense.

Well one thing I definitely agree on is any system that relies on the to be both a non-active occupant and to be alert to take over at a moments notice is unsafe. That just isn’t realistic.

I would even venture to say a large part of the population is are more dangerous than Tesla’s FSD.

My gripes are definitely more clouded by the person running Tesla than anything else.

If they had better PR I think people would be a whole lot less negative on the company and technology.

 

[bibkisbu]

Well one thing I definitely agree on is any system that relies on the to be both a non-active occupant and to be alert to take over at a moments notice is unsafe. That just isn’t realistic.

here's FSD v14 in torrential rain. In same video you also have sun glare. Car gives a red hands warning 3x but it never disengages.

For "unsupervised" it would just not display the warning but still do the minimum risk maneuvers it is doing.

This is superhuman and it seems to see better than a human does. But at no point does it require you to take over at a moments notice because it isn't doing dumb stuff. There are zero adas systems that would perform in half this amount of rain

 

[Jerrytball]

Yes that makes sense.


I would even venture to say a large part of the population is are more dangerous than Tesla’s FSD.

You mean those people that just have the need to pass you in a school zone double yellow line and you’re already doing 15 miles over the school zone speed limit which seems to be the norm for anybody around where I live.

 

[Kamuelaflyer]

Tesla just launched a major update to FSD, v14. It now has the ability to avoid road debris, handle unique edge cases, and park in garages, parking lots, etc.

Future revisions of the Mach E should maybe consider getting a license to FSD from Tesla.

 

[ChrisO]

here's FSD v14 in torrential rain. In same video you also have sun glare. Car gives a red hands warning 3x but it never disengages.

For "unsupervised" it would just not display the warning but still do the minimum risk maneuvers it is doing.

This is superhuman and it seems to see better than a human does. But at no point does it require you to take over at a moments notice because it isn't doing dumb stuff. There are zero adas systems that would perform in half this amount of rain

I don’t know all the details about FSD, my comment was about “level 2 driving” in general where there is the unrealistic assumption that the driver can take over in time while doing the normal human response to a situation where they are actively participating (Not paying attention)

In reference to what FSD does from your clip, I think that the real telling point would be if Tesla put their pocket book where their mouth is. As long as a person has to take responsibility for what it does I don’t think you are going to convince people it is better than them being in control even if they are in fact a worse driver.

 

[ChrisO]

You mean those people that just have the need to pass you in a school zone double yellow line and you’re already doing 15 miles over the school zone speed limit which seems to be the norm for anybody around where I live.

Running red lights at busy times a day is so common around here it is now “normal”. And drivers determine their speed by the bumper in front of them on the highway whether that be 1 mph in crammed traffic or 100 mph.

 

[bibkisbu]

In reference to what FSD does from your clip, I think that the real telling point would be if Tesla put their pocket book where their mouth is. As long as a person has to take responsibility for what it does I don’t think you are going to convince people it is better than them being in control even if they are in fact a worse driver.

That's what unsupervised driving is for. But everyone has doubts whether it will come or whether it is real.

I bet some car brands are literally worried about impending self driving.

you know brands are scared when RJ Scaringe of Rivian is telling pretty blatant lies about autonomy.

he knows people wont buy R2 if there is a brand that has self driving and you can road trip overnight while sleeping.

People look at isolated incidents of FSD messing up and blame it on using cameras only. Tesla is compute limited and debris on roads is detected using image recognition. Lidar doesn't change that

it doesn't help also that most tesla owners are on HW3 so have not actually seen what FSD can do

And consumer FSD is easily 1.5 years behind tesla's advancements.

New AI5 computer has 8x memory and 8x compute. There will be a giant boost to FSD just from a hardware upgrade

If you look past the lies that elon has told, FSD progress has been very good and about what you would expect.

Sponsored