Elon: FSD Beta tweets

[am_dmd]

Anything above 35mph tesla Vision slams on the brakes! Even chill mode drives so close to the car in front and does not have enough time to react, hence slams on brakes.
Hover your foot over the brake pedal and you will notice the tapping on brake pedal for no reason. Also there is a difference in stopping profile/brake taps when not using “Apply Brakes When Regenerative Braking is Limited”.
I have learned to ease off the go pedal and use regen to come to a full stop smoothly and fsdb should be able to do the same even more precisely unless tesla vision can’t see and interpret the scenario

 

[swedge]

Learning from individual owners would be a nightmare to try and diagnose or troubleshoot issues--- especially when NNs are involved.

I am not suggesting training the NNs for individuals. It would be more like personal map data: pothole here, stay left there, take this long way home when the freeway bogs down. The NNs already do know how to merge map and vision data. I am also not suggesting that your data be incorporated into everyones NNs. Now, if everyone intervenes with a speed correction at the same place, a global map correction might be in order.

FSD, even if not engaged, could run and compare it's decisions with yours. Keeping geotagged record of these divergences and making personal map notations where the discrepancies are consistent is not rocket science, so to speak.

Google, Facebook and Twitter all do it, using your click stream to tailor the ads.

But I do agree that at this point in FSD development, such personalization would add noise to the bug feedback. So it is not time to launch this yet. But it will make FSD much more attractive when it pays attention to how to drive our car, like anyone else we let drive it.

 

[swedge]

highly publicized crash in the tunnel

Are you speaking of the Thanksgiving day rear-ender in the SF Bay Bridge tunnel? FSD beta is not active on the Bay Bridge. It is a limited access freeway and FSD Beta switches back to standard AutoPilot as soon as you merge onto the freeway.

I drove that exact route to confirm this.

The first press story quoted the driver as saying he had FSD, and quoted the CHP as saying they did not know if FSD was in fact engaged. Almost every followup media outlet just called it an FSD failure.

Don't get me wrong, I agree that the FSD releases and V11 have attracted "imperial entanglements" as Hans Solo called it.

 

[2101Guy]

 

[qdeathstar]

Anything above 35mph tesla Vision slams on the brakes! Even chill mode drives so close to the car in front and does not have enough time to react, hence slams on brakes.
Hover your foot over the brake pedal and you will notice the tapping on brake pedal for no reason. Also there is a difference in stopping profile/brake taps when not using “Apply Brakes When Regenerative Braking is Limited”.
I have learned to ease off the go pedal and use regen to come to a full stop smoothly and fsdb should be able to do the same even more precisely unless tesla vision can’t see and interpret the scenario

I think I experience phantom breaking about once in ever 200 miles.

 

[swedge]

It would be huge if it worked! People spend a lot of time driving!

I do hope Tesla turns out to be right about Robotaxi. At every single moment, a very large fraction of the vehicles in any given town, county or state are parked, taking up space and capital. And the amount of space used for parking is huge. At any given time over half of parking spaces are empty: our space at home is empty whenever we park somewhere else. Cars cost us a fortune: purchase, insurance, interest, license fees, taxes, fuel or power, maintenance, parking space, time sitting behind the wheel, and 40,000 deaths yearly in the US.

The idea of point to point, on demand transportation is very powerful. But requiring each of us to own a car is an extravagant way to accomplish this.

What got me wondering about the RoboTaxi market model is my esperience of FSD. FSD continues to get better, but I often drive without using it and my wife refuses to even try it. As a beta tester I engaged it a number of time to find problems and reported those. But even when I know where it will fail, I seldom choose to use it even where it will work OK. AP on freeways I use a lot, but FSD on city streets, not so much. I hear similar stories here on TMC. Now this could well be because FSD is simply not good enough yet. But it could also be because it just doesn't help me very much, and I enjoy driving myself anyway. Perhaps, even if I never needed to disengage or adjust the speed, it is doing something I'd just as soon do myself.

With a true fully autonomous driving system, I can see clear use cases, such as when we're too old to drive, or need a lift to the airport. But we are a very long way from that now, on the technology but also regulation and social acceptance. Oh, and industrial resistance and infrastructure adaptation.

In the mean time, I am ready for a new release to test, have been for months.

 

[AlanSubie4Life]

Now this could well be because FSD is simply not good enough yet. But it could also be because it just doesn't help me very much, and I enjoy driving myself anyway. Perhaps, even if I never needed to disengage or adjust the speed, it is doing something I'd just as soon do myself.

Only useful if you can do something else with no obligation to monitor. Otherwise you may as well drive yourself, with background safety features (ones that are more aggressive than current ones - you could imagine a more aggressive lane assist and a background cruise control that more aggressively overrides your inputs to prevent collisions rather than just beeping and using AEB).

Fortunately we are beta testing the Robotaxi software so we are just so close.

 

[Knightshade]

Fortunately we are beta testing the Robotaxi software so we are just so close.

We aren't, of course, testing RT software-- but I admire the tenacity with which you and danielinsd troll me with the claim

 

[2101Guy]

Are you speaking of the Thanksgiving day rear-ender in the SF Bay Bridge tunnel? FSD beta is not active on the Bay Bridge. It is a limited access freeway and FSD Beta switches back to standard AutoPilot as soon as you merge onto the freeway.

I drove that exact route to confirm this.

The first press story quoted the driver as saying he had FSD, and quoted the CHP as saying they did not know if FSD was in fact engaged. Almost every followup media outlet just called it an FSD failure.

Don't get me wrong, I agree that the FSD releases and V11 have attracted "imperial entanglements" as Hans Solo called it.

The bolded part. That's the point. The perception was/is that the crash is due to Tesla's FSD feature. The general public has no clue about AP vs FSD vs FSDb vs EP vs ...... It's just "FSD".

So when you combine that with all of the other recent issues plaguing the FSD progress...made sense for Elon to spend 99% of the investor day content on everything EXCEPT FSD.

 

[ewoodrick]

Anything above 35mph tesla Vision slams on the brakes! Even chill mode drives so close to the car in front and does not have enough time to react, hence slams on brakes.
Hover your foot over the brake pedal and you will notice the tapping on brake pedal for no reason. Also there is a difference in stopping profile/brake taps when not using “Apply Brakes When Regenerative Braking is Limited”.
I have learned to ease off the go pedal and use regen to come to a full stop smoothly and fsdb should be able to do the same even more precisely unless tesla vision can’t see and interpret the scenario

I don't have phantom braking that often.
Phantom braking does not exist for "no reason" it's because the car is indeed detecting and reacting to an obstacle. It's trying to be better safe than sorry.

And from my experience it doesn't "slams on the brakes!" It applies light to moderate braking which because it is unexpected surprises and feel like heavy braking. Instead of overriding it, on roads with fewer cars, let it do it's thing. It doesn't come to stop in the middle of the road (like slamming on the brakes does) It feels a lot more like my foot slipping off the accelerator.

 

[ewoodrick]

I am not suggesting training the NNs for individuals. It would be more like : pothole here,

For the next 10 years after the pothole is repaired, the car swerves to miss it.

 

[ewoodrick]

I do hope Tesla turns out to be right about Robotaxi. At every single moment, a very large fraction of the vehicles in any given town, county or state are parked, taking up space and capital. And the amount of space used for parking is huge. At any given time over half of parking spaces are empty: our space at home is empty whenever we park somewhere else. Cars cost us a fortune: purchase, insurance, interest, license fees, taxes, fuel or power, maintenance, parking space, time sitting behind the wheel, and 40,000 deaths yearly in the US.

There are indeed areas, like NYC, where driving yourself doesn't really make sense. Everything is close by and there's mass transit pretty close.

But I suspect that the areas that meet that criteria are significantly less than the small amount of land that Tesla suggest solar would need to be deployed.

My car is my car. I keep my things in it. It's always available when I need it, NOW, not 10 minutes from now for a car to arrive. I can't walk to lunch, unless I go out and steal one of my neighbors' chickens.

Putting on my philosophical / ecological hat. many of the bigger problems that we have ARE the large cities. Literally too much crap in one place. Spread the people over a larger area and nature can start to recover the crap. Give mother nature a chance and she'll destroy anything. But put 12 million people in 12 sqmi and she doesn't have a chance. Yes, a million cars in 12 sqmi is too much.

 

[Knightshade]

There are indeed areas, like NYC, where driving yourself doesn't really make sense. Everything is close by and there's mass transit pretty close.

What's interesting there, and I've used NYC as the example of the ideal "it makes no sense to own a personal vehicle here" location.... DESPITE that, nearly half of NYC households own at least 1 car anyway... (it was ~45% that owned cars in 2019, last year I can find the data...but I did find that registrations surged by large amounts during covid so it might be higher now)

People REALLY like owning their own vehicle, even in places doing so makes little sense.... (in the US anyway-- ownership overseas varies quite a bit)

 

[OxBrew]

People keep saying FSDb has 90 million miles. The only actual Tesla sources I've ever found for the 90M number actually say "AP" has 90M miles.

Just a minor point, but it matters.

 

[Dewg]

People keep saying FSDb has 90 million miles. The only actual Tesla sources I've ever found for the 90M number actually say "AP" has 90M miles.

Just a minor point, but it matters.

Let's run some numbers. We know from the recall there are at least 350K users (in the US) on FSD Beta. If every user drives approximately 2 miles a week on FSD Beta (pretty conservative), that's over 36M miles a year. I can see how 90M is plausible.

 

[AlanSubie4Life]

Let's run some numbers. We know from the recall there are at least 350K users (in the US) on FSD Beta. If every user drives approximately 2 miles a week on FSD Beta (pretty conservative), that's over 36M miles a year. I can see how 90M is plausible.

Looks like about 12 million miles a month. Hard to say exactly.

Say it was 200k users for that slope since it did not include everyone. So 60 miles per month per user.

Works out to be 2 miles a day on FSD Beta.

Seems unlikely that that included freeway AP miles!

And looks like very low percentage of surface street miles are on FSD Beta (seems very likely to be less than 20%), as one would expect.

So it is not surprising that the safety numbers are quite good, since there are plenty of optimal scenarios to choose from.

 

[DanCar]

... So it is not surprising that the safety numbers are quite good, since there are plenty of optimal scenarios to choose from.

Agree. I only use FSD where I know it is good at it, example: stopped at a traffic light, going straight. And don't use it where I know it is bad at it, for example congestion and trying to make a maneuver.

 

[swedge]

For the next 10 years after the pothole is repaired, the car swerves to miss it.

Until your correct it. Unlike the mistakes that Tesla's current purely pertained NN + global map system makes.

 

[Supcom]

Until your correct it. Unlike the mistakes that Tesla's current purely pertained NN + global map system makes.

If the car can see the pothole (or any obstruction) the first time, then there is not need for it to mark the location on a map. It should see the same pothole the next time and every time thereafter until the pothole is fixed.

The only reason to mark something on a map is if the car could not see the obstruction. But then you have the complication of how to implement a reliable user interface to add, edit and delete the item from the map while driving the car. Since that is not practical, that leaves us to having to build some backend infrastructure to take full video from the car, find the potholes and send their locations out to the fleet. This is also impractical due the the huge amount of video data.

So, we are back to having the car identify the potholes so that only relevant video is sent to the mapping servers. But, if the car can identify the potholes, then it doesn't need them on the map. And we're back to where we started.

 

[kabin]

I don't have phantom braking that often.
Phantom braking does not exist for "no reason" it's because the car is indeed detecting and reacting to an obstacle. It's trying to be better safe than sorry.

And from my experience it doesn't "slams on the brakes!" It applies light to moderate braking which because it is unexpected surprises and feel like heavy braking. Instead of overriding it, on roads with fewer cars, let it do it's thing. It doesn't come to stop in the middle of the road (like slamming on the brakes does) It feels a lot more like my foot slipping off the accelerator.

One little but important detail. FSD is responding to a false target and is therefore being a hazard. And it happens all too often in my experience. My drive to drop the kids off to school I know two places where it happens 100% of the time and there's absolutely nothing there each and every time. That's how the condition of being useful (utility) drops off.