FSD v12.x (end to end AI)

[mgs333]

I tried auto park today in a pretty tight parallel parking space, with the additional challenge that it was a narrow street so the maximum angle it could take was limited by the cars parked on the other side of the street. I thought it did amazingly well. Took 3 back-and-forths to get in, but ended up well placed and didn't hit anything. Props to the auto-park team

Mine is still ok at best to frequently terrible, it ignores the neighboring cars' positions so it will frequently park too close to the car that's closer to my driver's side and leave a huge gap on the other side. So I need to correct it myself so I can open my door. It just goes for the middle of the space no matter how close/far the other cars are.

It does shift much faster when going from D to R and back but still takes multiple tries even for simple spots without any cars on both sides.

 

[PACEMD]

The law is mostly about what's reasonable. The car stops with a decent gap between itself and the car ahead, or a few feet from the stop line. Is that unreasonable?

Oh, stop it! You're crossing the line now...........

 

[kpanda17]

My two non FSD M3s went quickly from 14.9 to 20
My FSD MYP went from 3.25 to 14.9 and now is paused....come on 15.5/12.4.1 hurry up

 

[aronth5]

One of the biggest problems in my area is people jumping across intersections without looking or going through stops too fast (20 km/h plus) even with traffic and people around, but full stops when you have stops every 200ft or less is painful. I usually just slow to 5-12 km/h myself, not all the way down to zero, if visibility is good. Speeding on the other hand is likely to have you smashing into one of those same jumpers, so fsd automax is too fast and even where it’s safer, IME cops are mostly pulling over speeders or stop sign jumpers around schools.

At least the nhsta stops can be overridden with the accelerator, though it feels like fsd fights a bit with you so it’s not always smooth.

What a lot of people are missing in the NHSTA discussion were the prerequisites Tesla had in place to allow a rolling Stop. If I recall there were at least 5 conditions that had to be met including the crossing road's speed limit had to be no greater than 30mph along with the obvious detection of oncoming cars by the cameras. So if cars were coming no rolling stop. And if any of the other conditions were not met there was no rolling stop signs. Unfortunately Tesla was stupid in allowing the rolling speed to be as high as 5-7mph which I suspect caught the attention of NTSTA. Had Tesla restricted to 1-2mph a chance NHSTA would have let it go.

Can anyone find the list of conditions?

 

[JB47394]

AI DRIVR's latest video. He's conflicted over 12.4.1, saying that interventions are up, but at the same time that it's "more capable". Ultimately, he prefers 12.3.6's style of driving.

 

[FSDtester#1]

AI DRIVR's latest video. He's conflicted over 12.4.1, saying that interventions are up, but at the same time that it's "more capable". Ultimately, he prefers 12.3.6's style of driving.

Yeah, watching him vs Chuck and @DirtyT3sla he definitely is having worse drives which is ironic because we all pretty much believe Fsd should handle best in that area of California. Maybe the worm is turning...
I would love FSD to outperform Omar in different regions.
Also, I saw Omar actually posted a video in the title saying he has disengagements.
Curious...

 

[kabin]

We can always count on these a$$hats to add some humor.

 

[Mardak]

I assume this was not your vehicle! This seems like it would be super difficult to accomplish.

Just paid 10 minutes labor + $65 for parts to replace the ripped off fender garnish. Mobile Service also seemed confused about how FSD would hit the cone hard enough to make it come off. I would guess I've had 12.3.x drive that under-construction road nearly 100 times without issue although the exact positioning of cones is likely different each time.

Rewatching the videos, there doesn't seem to be anything particularly strange about the layout:

The right mirror did touch the 2nd cone that's tilting left while it's the 3rd cone ending the merged lane FSD ran over. My guess is that it thought these 2 cones were actually only one when they visually aligned just moments before and thought it had plenty of space to cut that corner after passing 2nd cone. But there's probably a whole second of unobstructed view of 3rd cone to have time to reassess…

 

[AlanSubie4Life]

Just paid 10 minutes labor + $65 for parts to replace the ripped off fender garnish. Mobile Service also seemed confused about how FSD would hit the cone hard enough to make it come off. I would guess I've had 12.3.x drive that under-construction road nearly 100 times without issue although the exact positioning of cones is likely different each time.

Rewatching the videos, there doesn't seem to be anything particularly strange about the layout:
View attachment 1058485

The right mirror did touch the 2nd cone that's tilting left while it's the 3rd cone ending the merged lane FSD ran over. My guess is that it thought these 2 cones were actually only one when they visually aligned just moments before and thought it had plenty of space to cut that corner after passing 2nd cone. But there's probably a whole second of unobstructed view of 3rd cone to have time to reassess…

Level 3 any day now.

 

[cyborgLIS]

AI DRIVR's latest video. He's conflicted over 12.4.1, saying that interventions are up, but at the same time that it's "more capable". Ultimately, he prefers 12.3.6's style of driving.

I agree with him that it seems to not be worthwhile to pursue refinement of 12.4. Instead focus on 12.5. 12.4 is so far behind any previous version of 12 in terms of correct behavior. 12.4 uses a new, more capable model that might make it worthwhile with refinement, but it's not obvious what those new capabilities even are. Apparently you can't just retrain the new model with the previous training set and expect the same behavior. Maybe 12.3 training data is not even compatible with the 12.4 model, so you really have to start from scratch and collect/produce new data with each new model.

 

[FSDtester#1]

Level 3 any day now.

Tonight only for @gottagofast ... the rest of us 2032

 

[EVNow]

Level 3 any day now.

I can't believe Musk actually thinks progress on FSD is exponential instead of asymptotic, so it becomes more and more difficult to improve FSD as it gets better. May be he thinks it will be a S curve with FSD just now entering the straight-line area after slow increase.

 

[FSDtester#1]

I can't believe Musk actually thinks progress on FSD is exponential instead of asymptotic, so it becomes more and more difficult to improve FSD as it gets better. May be he thinks it will be a S curve with FSD just now entering the straight-line area after slow increase.

This latest snafu on moving from 12.3 to 12.4 shows how obtuse they are about launching new iterations and thinking it's like the tweaking the NN's on 12.3

 

[EVNow]

This latest snafu on moving from 12.3 to 12.4 shows how obtuse they are about launching new iterations and thinking it's like the tweaking the NN's on 12.3

I just hope FSD doesn't turn out to be another Solar Roof ...

 

[Tronguy]

Just saying it again: Despite all the carping, it appears that Tesla is taking the approach of, "No FSD before its time."

I'd rather that we have a working FSD than one that has major regressions. Assuming that any Tesla types actually read this thread (ha!), you'd imagine with the general level of snark around here, they might be trying for a snark-free release.

Have patience.

 

[EVNow]

Just saying it again: Despite all the carping, it appears that Tesla is taking the approach of, "No FSD before its time."

I'd rather that we have a working FSD than one that has major regressions. Assuming that any Tesla types actually read this thread (ha!), you'd imagine with the general level of snark around here, they might be trying for a snark-free release.

Have patience.

The point of "snark" being reality vs Elon's hype.

ps : Just to make it clear - most of us here expect FSD to take a looong time to mature. Measured in years, not months. But Elon keeps saying - without fail and apparently zero self-awareness that FSD will happen soon.

 

[lzolman]

At the very least, "No FSD update before my cameras figure out how to calibrate..."

 

[AlanSubie4Life]

general level of snark around here, they might be trying for a snark-free release.

I think if Elon would say to expect six months between each FSD point release, with asymptotic improvements on each release towards the 10-miles-per-intervention ultimate “max-capability” goal, and state the end-of-life date (end of development, final capability) for HW3/4 in 2025 or whatever, there would be less snark.

People like snark so I could be wrong. But the substance of the snark would be very different for sure.

 

[AndrewZ]

What a lot of people are missing in the NHSTA discussion were the prerequisites Tesla had in place to allow a rolling Stop. If I recall there were at least 5 conditions that had to be met including the crossing road's speed limit had to be no greater than 30mph along with the obvious detection of oncoming cars by the cameras. So if cars were coming no rolling stop. And if any of the other conditions were not met there was no rolling stop signs. Unfortunately Tesla was stupid in allowing the rolling speed to be as high as 5-7mph which I suspect caught the attention of NTSTA. Had Tesla restricted to 1-2mph a chance NHSTA would have let it go.

Can anyone find the list of conditions?

The NHTSA’s attention was caught by people submitting complaints, the organization itself is largely reactionary and relies heavily on owner/user complaints. Some people didn’t like that the system was doing illegal things while they were responsible for what it was doing, maybe some were ticketed or warned.

Even for big mechanical issues, the NHTSA doesn’t know about problems until they are brought to their attention by owners, dealerships, suppliers, the manufacturer, etc.

 

[mtndrew1]

Just saying it again: Despite all the carping, it appears that Tesla is taking the approach of, "No FSD before its time."

I'd rather that we have a working FSD than one that has major regressions. Assuming that any Tesla types actually read this thread (ha!), you'd imagine with the general level of snark around here, they might be trying for a snark-free release.

Have patience.

Shoot I’ve been waiting for FSD to be of its time for six damn years.