FSD v12.x (end to end AI)

[AlanSubie4Life]

A fleet wide basis seems a MUCH better measure, no?

And we have that. Both AP on highways and FSDb in non-highway driving are much safer than manual driving overall.... (~2-4x safer vs manual tesla driving, ~5-8x safer than the overall US fleet manually driving

There is no useful data here, as you know. Would be good to have the comprehensive data set so a comparison can be done (would be difficult even then!). Just no way to make the comparisons when aggregated like this unfortunately.

Some day we will know. Just not today.

It may well offer a safety improvement. It’s very possible. We just unfortunately cannot tell. Selection bias etc etc etc.

 

[uscbucsfan]

FWIW I find autopark useful some of the time (when I need to park but ALSO have something else to be doing with my hands at the time- putting away my phone, sunglasses, etc...)-- if I don't then manual is going to be better.

FSD I find useful, as I suggest, the vast majority of all driving time...





A fleet wide basis seems a MUCH better measure, no?

And we have that. Both AP on highways and FSDb in non-highway driving are much safer than manual driving overall....

View attachment 1016966

FSD is not safer than a human driver. There's no data to back that up.

FSD + a human driver is safer than a human driver alone. Absolutely.

The car makes you stay ultra vigilant at all times. That's not a bad thing, but it absolutely skews the data. A system that would prevent someone from driving if they look at their phones, looked away, got tired, etc. would have a similar result.

 

[Knightshade]

There is no useful data here, as you know.

Obviously I find it useful or I wouldn't have quoted it.

It was less useful when they just listed all driving on all roads into one bucket.

Breaking it out by AP mostly highway vs FSDb mostly city vs NEITHER is [B}useful[/B] data. Not sure why you'd think otherwise.

It may well offer a safety improvement. It’s very possible. We just unfortunately cannot tell. Selection bias etc etc etc.

I thinik there's a fair argument in selection bias comparing "tesla drivers" with 'all drivers" because all drivers includes folks still driving around in a 1973 pinto with drum brakes and no safety systems at all. (though even then there's not enough of those pintos left on the road to account for ALL of the 5-8x crash rate difference)

I think you'll have a lot harder one arguing the data doesn't make it pretty clear among Tesla drivers both AP in its ODD and FSDb in its ODD when active are safer than a purely manual Tesla driver in either condition.

The car makes you stay ultra vigilant at all times. That's not a bad thing, but it absolutely skews the data. A system that would prevent someone from driving if they look at their phones, looked away, got tired, etc. would have a similar result.

Given the driver monitoring that heavily enforced the no phone/no look-away thing came AFTER that published data I don't believe your claim the results would be same from JUST that can possibly be accurate.

 

[uscbucsfan]

Given the driver monitoring that heavily enforced the no phone/no look-away thing came AFTER that published data I don't believe your claim the results would be same from JUST that can possibly be accurate.

The FSD data has always had intense driver monitoring...even AP had wheel nag in that data.

BTW, I think I was obviously talking about FSD. AP is like cruise control with lane assist.

FSD is not safer than a human today because it can't complete most trips, today.

 

[Knightshade]

The FSD data has always had intense driver monitoring...even AP had wheel nag in that data.

I think you might be the first person on earth to call the wheel torque sensor that can be fooled with a fishing weight or tennis ball "intense driver monitoring"

Not really sure what you're on about with the "can't complete most trips" bit-- completing a trip or not isn't an accident.

FSDb turned on results in a lower accident rate than FSDb not turned on for mostly non-highway driving... even just among Tesla drivers (and even moreso among all drivers)

AP turned on results in a lower accident rate than AP not turned on for mostly highway driving... even just among Tesla drivers (and even moreso among all drivers)

 

[AlanSubie4Life]

I think you'll have a lot harder one arguing the data doesn't make it pretty clear among Tesla drivers both AP in its ODD and FSDb in its ODD when active are safer than a purely manual Tesla driver in either condition.

No selection bias here? I do find it hard to know what the use scenarios for FSD Beta are relative to the entire range of scenarios encountered.

Personally, for the highest risk maneuvers that I make (for example merging right to make right onto Pacific Heights and cut into the huge unnecessary slow line), I have FSD disengaged well in advance, and carefully zipper merge (really cutting in line) at high speed into slow traffic. This is something FSD would never risk. On downhill grades I often disengage well in advance since I don’t like the behavior and it seems too risky and complex and I want margin. I can select a lane more intelligently and optimize the descent at maximum speed to improve travel time.

What does no active safety mean?

What does mostly non-highway miles mean?

Don’t get me wrong - it might improve safety (on City Streets - I don’t care about anything else for this discussion); seems entirely conceivable based on my experience. I just don’t know, and it also depends on the comparison you want to make. And we don’t have the data.

Maybe I am missing something and there is a way to understand the improvement based on these four numbers. It just seems extremely complex.

This is 2022 data too. So that was before single stack?

What I care about is whether my accident risk is higher or lower if I use FSD on City Streets. And I can’t tell from these numbers. I have no idea how I would compare to other Tesla drivers. Can’t compare FSD City Streets number to the No Active Safety number. Not the same exposure, AFAICT. Am I wrong?

I think it likely improves safety for ME because I would not do risky things if I just left it engaged as much as possible. It would just plod along. But 1) that is not clearly the case - maybe it puts me in bad situations - and 2) it is not clear that is the desired comparison since it is not the same drive (it might take 10% longer).

 

[uscbucsfan]

I think you might be the first person on earth to call the wheel torque sensor that can be fooled with a fishing weight or tennis ball "intense driver monitoring"

I didn't say that...re-read what I wrote.

 

[uscbucsfan]

Not really sure what you're on about with the "can't complete most trips" bit-- completing a trip or not isn't an accident.

FSDb turned on results in a lower accident rate than FSDb not turned on for mostly non-highway driving... even just among Tesla drivers (and even moreso among all drivers)

It's disengaged because it cannot complete the task. People disengage before it causes an accident...or like the events posted here, they allow it to cause an accident.

FSDB + an active human monitor is safer than just the human left unmonitored...There is no argument that FSD is safer than a human, because FSD cannot operate without the human monitoring.

 

[Knightshade]

I didn't say that...re-read what I wrote.

Not only did I read em, I quoted em.

Here they are again

The FSD data has always had intense driver monitoring...even AP had wheel nag in that data.

If you did not intend to suggest the wheel nag was intense driver monitoring you might want to re-write what you wrote

Because the FSD data has not always had intense driver monitoring other than the wheel nag... the camera was entirely optional originally.

Most S/X drivers don't even have an interior camera.

 

[powertoold]

V12.2 is coming out, it's the end of Waymo

 

[MP3Mike]

FSDb V12.2 going out to employees:

https://twitter.com/x/status/1756064444835369091

 

[uscbucsfan]

Not only did I read em, I quoted em.

Here they are again



If you did not intend to suggest the wheel nag was intense driver monitoring you might want to re-write what you wrote

Because the FSD data has not always had intense driver monitoring other than the wheel nag... the camera was entirely optional originally.

Most S/X drivers don't even have an interior camera.

So you took a leap to just include less than 10% of the FSD fleet (at an extreme high estimate) as a troll or what? All MY/M3 and all MS/X refresh have camera monitoring in FSD since the safety score testers were allowed. We already know that MCU1 and any car without HW3 cannot use FSD (there are a few with a limited older FSD), but irrelevant to the numbers. I even said, "even has wheel nag"...not that wheel nag was included in that intense driver monitoring...just that it's different than completely ignoring a driver.

You can do better than this.

 

[Usain]

Where did you hear "much smaller and faster"? We all know about the elimination of the 300,000 lines of code, but that doesn't mean that the neural network implementation is smaller or faster.

Elon and Ashok were talking about it during their drive. I wasn't able to find the exact quote but I'm pretty certain they were saying that V12 is smaller and faster. I could be misremembering that one though. Maybe someone here can dig up the quote.

 

[JulienW]

V12.2 is coming out, it's the end of Waymo

.....well maybe Cruse.

 

[Usain]

Have you ever experienced crowded Indian roads? I have. No way. There is a large demographic of skilled human western drivers that could never do it. No way will Tesla FSD do it without a huge strategy shift.

Auto wipers anyone? lol, this thread cracks me up. Still wonder what y'all are smoking.

No, but I've ridden with guys from India who try to drive here. It's a lot scarier than FSD!

 

[TK211X]

Very excited. Especially for those on .8 and .9 who haven’t gotten pushed to 2024 branch.

 

[uscbucsfan]

Very excited. Especially for those on .8 and .9 who haven’t gotten pushed to 2024 branch.

You're an employee or Whole Mars?

 

[JulienW]

Very excited. Especially for those on .8 and .9 who haven’t gotten pushed to 2024 branch.

"Very excited" about what?

 

[TK211X]

V12.2 is based on 44.30.15

 

[uscbucsfan]

V12.2 is based on 44.30.15

But it's not going out to customers at this point.