FSD v12.x (end to end AI)

[Mardak]

The livestream from MissJilianne yesterday was mostly going straight in LA traffic but there was an interesting segment at 17:27 where 12.2.1 quickly followed a lead van switching lanes revealing stopped/slow traffic ahead and happened again 17:41 back to the original lane:

Has 11.x been able to do these types of maneuvers this quickly? How much does the lead vehicle behavior affect end-to-end in matching that behavior whether it's these lane changes or yielding at stops, etc?

 

[SpaceNuts]

Speaking as someone who has received two tickets for doing just that over the years, may I disagree?

I saw thousands of drivers do that everyday in downtown Houston. Never saw anyone get a ticket. If you want FSD to be useful in this part of town, this is definitely needed.

 

[kabin]

Tesla, please, please don't regard it as bad behavior. In my part of town, if you don't cross the double yellow line early, you'll never get the chance to turn left.

Yep, we all weigh the risks and do it from time to time. Fortunately we can always disable FSD.

 

[RonAz]

Speaking as someone who has received two tickets for doing just that over the years, may I disagree?

As a former truck driver, this was standard procedure for me for decades. Most left turn lanes are too short for 75’ trucks and have become too short for the volume of traffic today.

 

[JHCCAZ]

As a former truck driver, this was standard procedure for me for decades. Most left turn lanes are too short for 75’ trucks and have become too short for the volume of traffic today.

One thing we don't see in the clip is the left repeater feed. I know that if I'm trying to obey the lane markings ahead of that left turn bay, but I see traffic behind me already moving over way early (which happens a lot), then I also have to move over too early, otherwise I'll get cut off from the turn lane.

I don't know if this is what was happening,l; there's no way we can see it from the forward dashcam feed.

With enough training of these situations, I think the AI would learn the same thing - attentive drivers have to claim their space in certain cases like this.

 

[Dewg]

As a former truck driver, this was standard procedure for me for decades. Most left turn lanes are too short for 75’ trucks and have become too short for the volume of traffic today.

Okay. Just remember this is Fight Club. First rule is we never talk about it. So I never want to hear anyone say to a cop when pulled over "but sir, the car was driving, it's not my fault"

 

[OxBrew]

Yeah like "if isCalifornia : region = [1 0 0 0 0 0 0 0 ... 0]" and input region to the neural network.

How many tokens will you need? And how long is each one? A weight and bias for every node in the NN? How many nodes?

I have no idea about any of that, but my hunch is a LOT of tokens (might need several per town, laws are inconsistent and variable within towns) and each one maybe pretty long?
Just wondering. In any case, it's starting to sound like "code" to me. (nothing but nets!!! woo hoo!!!)

 

[Supcom]

Okay. Just remember this is Fight Club. First rule is we never talk about it. So I never want to hear anyone say to a cop when pulled over "but sir, the car was driving, it's not my fault"

If you want FSD to drive like a human, it's going to get your a ticket now and then.

 

[Supcom]

The livestream from MissJilianne yesterday was mostly going straight in LA traffic but there was an interesting segment at 17:27 where 12.2.1 quickly followed a lead van switching lanes revealing stopped/slow traffic ahead and happened again 17:41 back to the original lane:
View attachment 1020875

Has 11.x been able to do these types of maneuvers this quickly? How much does the lead vehicle behavior affect end-to-end in matching that behavior whether it's these lane changes or yielding at stops, etc?

I've had 11.x do many quick lane changes to get past a slow vehicle. It happens so fast, I have to calm down and remember that FSD is looking in all directions all the time so knew that there was a lane opening.

 

[Dewg]

If you want FSD to drive like a human, it's going to get your a ticket now and then.

#facts

 

[kabin]

Sorry, I probably needs some remedial training with linking X broadcasts. It shows 'status' in the link I copy but then I lose it in the post.

At @3:10 or so, watch the sizeable lag (3+ secs) in response to the lead vehicle's brake. You can see FSD even slowly accelerates for a good stretch.

https://twitter.com/x/status/1760328254274318771

 

[Dewg]

At @3:10 or so, watch the sizeable lag in response to the lead vehicle's brake.

https://twitter.com/x/status/1760328254274318771

The question I have - is that a perception issue, a processing lag, or by design? The more "human" it drives, the more rubber banding it may have. In the past people complained the car overreacted on slow downs. Perhaps this is a cushion. Not sure.

 

[heltok]

If you want FSD to drive like a human, it's going to get your a ticket now and then.

This is where autolabel comes in. It will use multiple drives in the same area to generate the ground truth if any laws where broken and use that to filter our most drives where laws were broken. V12 doesn’t drive like humans, it drives like five star humans…

 

[Supcom]

This is where autolabel comes in. It will use multiple drives in the same area to generate the ground truth if any laws where broken and use that to filter our most drives where laws were broken. V12 doesn’t drive like humans, it drives like five star humans…

I wonder why the five star humans can't drive at a constant speed without a pace car. Maybe it's five stars out of ten.

 

[Knightshade]

This is where autolabel comes in. It will use multiple drives in the same area to generate the ground truth if any laws where broken and use that to filter our most drives where laws were broken. V12 doesn’t drive like humans, it drives like five star humans…

How does "autolabel" know if laws were broken, specifically?

 

[scottf200]

This AI DRIVR example was something else!
Comments by Ethan Knight Vision lead @Tesla_AI

https://twitter.com/x/status/1760764608326123657

 

[Daniel in SD]

How does "autolabel" know if laws were broken, specifically?

It seems like the George Hotz quote "All good drivers are good in the same way, but all bad drivers are bad in different ways," might apply. Maybe the law abiding drivers are clustered which would make it possible to filter out the lawbreakers.
The crossing over double yellows to get in line for a left turn seems like a super difficult problem to solve. It would be interesting to see what Waymo does.

 

[Knightshade]

It seems like the George Hotz quote "All good drivers are good in the same way, but all bad drivers are bad in different ways," might apply. Maybe the law abiding drivers are clustered which would make it possible to filter out the lawbreakers.

It seems the myriad folks insisting "everyone" crosses double yellows and rolls stop signs in all the places they drive would directly contradict that

 

[Mardak]

They should call this, "Please don't give me a ticket, officer, everyone else was driving ten over the speed limit!"

Is your experience that AUTO tends to go too fast? Or perhaps on sections of 101 where the speed limit is 55, it stays at 65? Other reports seemed to be more of it being too slow, but that might be more of behavior on residential streets?

 

[Mardak]

I had an interesting experience tonight with V12. Leaving a store, I decided to see if FSD could drive me out of the parking lot.

Just making sure, did you have navigation set to some other destination? At least 11.x can be pretty surprising without a destination where I've experienced it turning into a side street when the road ahead goes up a hill. I could see 12.x being more consistent in driving straight in my scenario, but unless Tesla explicitly trains for that, we might get relatively random behavior like parking the car.