FSD v12.x (end to end AI)

[aronth5]

It would have to have vehicle-to-vehicle communication.

Even the best drivers don’t know when the truck will move. So nothing in the training data, even if it contained all of historical human behavior, would help here.

I don’t think telepathy is an emergent behavior.

And how is this different than any vehicle that potentially can move into your cars space from any direction? As a driver you make decisions like this all the time. FSD could have behaved differently based purely on vision.

 

[Captkerosene]

I believe V12 is the most promising approach for generalized fsd, and it's already performing very well. You will see what I mean when you use V12.

But as we know from prior versions, any small issue can become a fatal flaw over time

As for police directing traffic and whatnot, I think it's a tractable problem with V12. I think lane selection in heavy / obstructing traffic is a harder problem than following hand-arm signals, but I could be wrong.

At the end of the day, the real answer is something like: yes, I think V12 will eventually be the approach that leads to 3-10x human safety performance (along with all the smoothness and naturalness that V12 brings) -- no, it won't be perfect and understand every little thing, but it will be safer than humans by far, and the real question will then be: will society accept a system that can fail but is 3-10x safer than humans?

As for the timeline, it's unclear, but again, if you use 12.2.1, it seems very close (2-3 years for this generalized robotaxi, perhaps the first HW5 cars?)

Seems to me FSD fits the Power Laws model pretty well. Log chart with time on the X axis and miles between interventions on the Y. Linear, S-curve and exponential graphs don't look right to me.

 

[zoomer0056]

3-10x safer than humans

That's like a garment sale saying "up to 90% off".

 

[AlanSubie4Life]

And how is this different than any vehicle that potentially can move into your cars space from any direction?

It’s not. I’m not following.

FSD could have behaved differently based purely on vision.

I don’t understand.

All I was saying is that it was not possible to predict movement of the truck, so it was inappropriate to leave no buffer zone (leaving no time to react). The car did not even drift to the left of the lane!

This is why I am curious about lane positioning and buffer zone maintenance in v12. If drivers are not routinely seeing it (it has not been addressed one way or the other what the current status is), then we have to wonder why. I don’t see how it would not show up.

 

[SidetrackedSue]

I look forward to the Rebellionaire posting the dashcam video of the near miss with the truck. The side camera view will give a better indication of how close it was and why the car refused to move over, even in the lane, for the incursion. The right hand side passengers disagreed with the driver as to how close it was. The camera footage would clear that up.

 

[JulienW]

Not a big deal but there is now *no chance of DirtyTesla testing v12 since he picked up a pickup (Cybertruck) yesterday. I guess Elon can unblock Chris now since he won't be bothering him about getting v12.

*Actually his wife Stephanie has a Y with FSD so a chance she could get it.

 

[powertoold]

12.2.1 average auto-max is driving so well, blows my mind

It's difficult to put into words the subtleties of behavior, like nudging over for cars encroaching into my lane, dealing with questionable bikers and when to change lanes vs overtake cars potentially turning into a driveway, etc.

There's so much subtlety in 12.2.1 that it makes me a believer

 

[Acisplat]

12.2.1 average auto-max is driving so well, blows my mind

It's difficult to put into words the subtleties of behavior, like nudging over for cars encroaching into my lane, dealing with questionable bikers and when to change lanes vs overtake cars potentially turning into a driveway, etc.

There's so much subtlety in 12.2.1 that it makes me a believer

Wish the auto speed limit would work on freeways

 

[FSDtester#1]

Are you referring to when you get the updates? I was thinking back to when the releases first drop.

Yes, me specifically. I think others do as well, but not sure.
I was in the 100% safety score group that got in with that score

 

[sleepydoc]

Wish the auto speed limit would work on freeways

From what I’ve read Tesla split the freeway stack on V12 so freeways are V11 and city streets are V12. I would expect them to merge the two again and with that I’d also expect the V12 auto speed limit features to follow.

 

[powertoold]

V11 feels like ancient technology vs V12, it's insane

 

[FSDtester#1]

From what I’ve read Tesla split the freeway stack on V12 so freeways are V11 and city streets are V12. I would expect them to merge the two again and with that I’d also expect the V12 auto speed limit features to follow.

I expected that Tesla would have sent v12 to more people by now.
Not just Californians and Doctors!

 

[Pdubs]

I expected that Tesla would have sent v12 to more people by now.
Not just Californians and Doctors!

Sleepy hasn’t gotten it yet. I haven’t gotten it yet. @Ramphex hasn't gotten it yet. I’m a swing Dr. @Ramphex is the Dr. of Love. @sleepydoc …not sure about him but I bet he’s learnt

 

[mongo]

Not a big deal but there is now *no chance of DirtyTesla testing v12 since he picked up a pickup (Cybertruck) yesterday. I guess Elon can unblock Chris now since he won't be bothering him about getting v12.

*Actually his wife Stephanie has a Y with FSD so a chance she could get it.

You assume he would be using one of their cars...

 

[sleepydoc]

I expected that Tesla would have sent v12 to more people by now.
Not just Californians and Doctors!

It’s not even Doctors ‘cause I haven’t gotten it yet! Stupid Elon always did like those California hippies better!

 

[bhakan]

It’s not even Doctors ‘cause I haven’t gotten it yet! Stupid Elon always did like those California hippies better!

I thought he hated CA and liked the Texas cowboys

 

[Mardak]

Wish the auto speed limit would work on freeways

Curious, what are you expecting it to do differently from 11.x? From what I can tell, it seems like AUTO basically allows end-to-end to go as fast or slow as it believes is appropriate whereas a set speed number puts a limit on how fast 12.x is allowed to go (so it can still go slower). Whereas 11.x's behavior is that the set speed is the target speed to get to, so traditional control code tries pretty hard to get to and maintain that speed.

 

[Mardak]

I've seen many cases of FSD just missing hitting something by inches

Were these many cases with 11.x or 12.x? At this phase of end-to-end, I would expect the older code to be much more consistent in precisely avoiding things by inches. The upcoming 12.3 might be significantly improved if Tesla did need to do some focused training on not hitting other vehicles, but given that you have 12.2.1, I would be hesitant to trust it to avoid hitting things.

 

[Mardak]

This sort of thing happens all the time, and people sometimes actually back into the road, but in most cases there is no accident, because humans are involved.

There's likely humans involved on both sides of biasing away from the reversing vehicle or the other driver realizing the mistake to stop reversing. Given that it seems like 12.x didn't understand the potential for the other vehicle to reverse more quickly into the road, does that suggest end-to-end is perhaps overfit on forward-facing vehicle motion and/or 11.x has been correctly giving extra space thus not triggering disengagement nor training data video to improve 12.x behavior?

Perhaps another set of issues that could be addressed with training 12.x on parking lot examples where reversing vehicles are much more common?

 

[Mardak]

If this is how one thinks one develops safety critical systems, think again

Are you suggesting any system that involves reinforcement learning cannot be done in a safe way? Or are you saying Tesla is developing 12.x in a different way?