Elon: FSD Beta tweets

[AlanSubie4Life]

Because the point of these nag systems is to make sure you're paying attention. Every other company offering similar L2 systems (Ford, GM, etc.) uses eye tracking _exclusively_. And it works. Nobody complains about it. It's a better, more reliable system than "just right, not too hard" tugs on the steering wheel.

Sure. The tug system sucks. I’ve said it many times.

I was just saying that eyes on the road may not be good enough. Your claim was that if eye tracking worked they would just do that…evidence being that other manufacturers do it…which is…not evidence of anything.

But it’s certainly possible Tesla could enforce hands on the wheel without the tugs! They just don’t seem to have tried. (It would be easy for a human, is the evidence.) They actually may already do it, partially - not sure…with torque as a backup when there is uncertainty. Would have to experiment - I don’t think they do, but it is possible.

The torque requirement is silly. Just check my hands are on the wheel (note that neither hands nor wheel can be seen - but of course that is not needed). No additional hardware required!

Anyway the discussion was whether the cameras can do eye tracking. You seemed to think not. There is a lot of evidence they do, with limitations. Having horrible wheel torque requirements is not related, unfortunately.

Fortunately with attention monitoring it can actually be pretty hit and miss and the system could still be decent (a lot more forgiving than wipers and driving in that regard)!

Every other company offering similar L2 systems (Ford, GM, etc.) uses eye tracking _exclusively_.

I think on the freeway eyes only might be ok most of the time. Usually there is a little time to react.

However, for the highly flawed City Streets product, hands on might be a good idea which other manufacturers haven’t had to deal with just yet. Sometimes there is very very little time to react. Hands on is good.

 

[dramsey]

I think on the freeway eyes only might be ok most of the time. Usually there is a little time to react.

However, for the highly flawed City Streets product, hands on might be a good idea which other manufacturers haven’t had to deal with just yet. Sometimes there is very very little time to react. Hands on is good.

That's a good point: that existing eye tracking systems are only used for highway auto driving.

I don't know if any existing auto driving systems use a "hands on wheel" detection, but I'd guess that would involve making the wheel rim a capacitive touch sensor. These don't work with gloves, but maybe there's another solution.

 

[dramsey]

Sure. The tug system sucks. I’ve said it many times.

I was just saying that eyes on the road may not be good enough. Your claim was that if eye tracking worked they would just do that…evidence being that other manufacturers do it…which is…not evidence of anything.

Well, what are we trying to achieve here? We're trying to make sure that the driver is paying attention, because they might need to take over at a moment's notice. Eye tracking seems to me to be much better than "hands on the wheel", since it's easy to get distracted (talking to a front seat passenger, for instance) while keeping your hands on the wheel.

The evidence that eye tracking is sufficient (at least for highway driving) is indeed that other manufacturers use it, and that nobody's brought up any fatal flaws in the system, and that it appears to work better and more consistently than Tesla's system.

I think the only reason Tesla uses and continues to use the torque-based system is that the nominal hardware cost is zero-- they just use feedback from the steering system. Many things Tesla does (removal of radar, removal of USS, remove of steering wheel stalks, going from chrome to black trim, eliminating most options except paint, interior, and wheels, etc.) are simply because they make the product cheaper. If dooming users to years of crappy, quasi-functional "rain sensing wipers" to save < $5 on a commodity rain sensor is considered acceptable, then saving more than that on an IR eye tracking camera on the steering column is a no-brainer.

 

[dramsey]

Except when Teslas version of eye tracking is crap

That's because the camera's positioned for cabin monitoring, rather than eye tracking. If they'd put a camera on the steering column like everyone else, it would work fine.

 

[EVNow]

I think on the freeway eyes only might be ok most of the time. Usually there is a little time to react.

However, for the highly flawed City Streets product, hands on might be a good idea which other manufacturers haven’t had to deal with just yet. Sometimes there is very very little time to react. Hands on is good.

Sometimes I feel - FSD is the one that needs to pay more attention. Not I !

 

[spacecoin]

That's a good point: that existing eye tracking systems are only used for highway auto driving.

I don't know if any existing auto driving systems use a "hands on wheel" detection, but I'd guess that would involve making the wheel rim a capacitive touch sensor. These don't work with gloves, but maybe there's another solution.

I don't know how you do hand-on-driving in the city using a Yoke. Hands-near perhaps... Generally, I think L2 is a stupid idea among VRU:s.

 

[uscbucsfan]

I don't know how you do hand-on-driving in the city using a Yoke. Hands-near perhaps... Generally, I think L2 is a stupid idea among VRU:s.

The same way you do with a wheel...or I do. I never completely hold the wheel when it's turning, just hover above, but I've seen some keep their hand lightly on it like others do with the wheel.

 

[Goose66]

The asymptote may make some people quite sad unfortunately!

Math humor. Fun!

 

[Bladerskb]

Elon agrees that nags need to be removed.

https://twitter.com/x/status/1723648029096382862

Hilarious because he was so against driver monitoring before and when he was questioned he responded with you don't know what you're talking about and that people probably think neural nets are fishing nets. And this was 4 and a half years ago!

"I think it will require detecting hands on wheel for at least six months"

All responses on being asked to add Camera Driver Monitoring System

"I think this was all really going to be swept, I mean, the system is improving so much, so fast, that this is going to be a moot point very soon. No, in fact, I think it will become very, very quickly, maybe and towards the end this year, but I say, I'd be shocked if not next year... having human intervene will decrease safety. DECREASE!"

"Driver monitoring/vigilance is mute"
"mute"
"mute"
"elevators"
"mute"
"Game. Set. match."

- April 2019 Elon Musk on Lex Podcast

 

[EVNow]

The same way you do with a wheel...or I do. I never completely hold the wheel when it's turning, just hover above, but I've seen some keep their hand lightly on it like others do with the wheel.

I keep my hand on the wheel when other cars are around ! Otherwise just hover above it.

 

[diplomat33]

Elon tweet on FSD V12:

Supervised FSD is vastly safer than human driving. Unsupervised FSD is trending well. Over time, Version 12 (end-to-end neural nets) will far exceed human safety even when unsupervised.

https://twitter.com/x/status/1727484899374899687

 

[nikolajz1]

From the sounds of that tweet seems like V12 isn't as good as V11 right now but has a way higher ceiling.

 

[willow_hiller]

From the sounds of that tweet seems like V12 isn't as good as V11 right now but has a way higher ceiling.

I interpreted it a little differently. Both "supervised" and "unsupervised" FSD in the above Tweet are talking about V11.

So he's claiming that with an attentive driver, V11 is vastly safer than human alone. But without a driver, V11 is only "trending well." He doesn't actually make a judgement about the current state of V12; supervised or unsupervised.

My question about this tweet is this: does this mean that Tesla is testing level 4 autonomy with safety drivers in locations where they don't need to disclose it? How else would he know how unsupervised FSD is trending?

 

[diplomat33]

My question about this tweet is this: does this mean that Tesla is testing level 4 autonomy with safety drivers in locations where they don't need to disclose it? How else would he know how unsupervised FSD is trending?

No, I don't think so. Tesla is not testing a separate L4 system, they are testing the same FSD beta we have. My guess is Tesla simply looks at metrics to see how FSD beta would perform if there were no interventions. And Tesla can use simulations to see how FSD beta would have handled a situation if it were unsupervised.

 

[Yelobird]

No, I don't think so. Tesla is not testing a separate L4 system, they are testing the same FSD beta we have. My guess is Tesla simply looks at metrics to see how FSD beta would perform if there were no interventions. And Tesla can use simulations to see how FSD beta would have handled a situation if it were unsupervised.

Actually, he has already stated in the V12 video that they Are testing the new software (which we Don’t have) in several places and locations with their appointed teams.

 

[diplomat33]

Actually, he has already stated in the V12 video that they Are testing the new software (which we Don’t have) in several places and locations with their appointed teams.

New software yes, called V12. But V12 is just a new software version of FSD beta. It is not a separate L4 system.

 

[Yelobird]

New software yes, called V12. But V12 is just a new software version of FSD beta. It is not a separate L4 system.

Actually you said “the Same FSD Beta software we have”. Last I checked nobody here has that software.

 

[JHCCAZ]

New software yes, called V12. But V12 is just a new software version of FSD beta. It is not a separate L4 system.

My interpretation of the situation with hired driver/testers is the following:

1. For expansion of any version of FSD to many regions outside of North America, Tesla has to use hired drivers because they may not be permitted to deploy the supervised version of FSD the way they have done here.
2. For v12, it's all about training on video and telemetry of "very good" human driving (also simulation, but the real thing is required to a large degree). The business of training on, or writing corrective software after studying interventions using the latest release, is no longer so useful as the primary methodology for moving forward. Therefore, they need to collect at least a major baseline of manual (but very good) human driving, and this is accomplished by hiring drivers even in North America.

 

[AlanSubie4Life]

I just want to know more about v11 safety. It would be hugely important if it is safer as currently implemented and utilized by the public.

But it is so frustrating that Tesla has never provided anything to help show the world how much safer the roads are as a result of v11 use.

Just such a difficult analysis to do - this may be why it has not happened.

 

[spacecoin]

Elon tweet on FSD V12:

https://twitter.com/x/status/1727484899374899687

Where are Community notes when you need them?

See Safety Argument Considerations for Public Road Testing of Autonomous Vehicles:

https://users.ece.cmu.edu/~koopman/pubs/koopman19_TestingSafetyCase_SAEWCX.pdf

TLDR; Supervised autonomy isn't getting to "safer than a human" by having a human when the system gets from hardly working to unsafe..