My wife's M3 with FSD has never had a single interention on V12.
That's because my wife has never activated V12...
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD v12.x (end to end AI)
- Thread starter Buckminster
- Start date
-
- Tags
- #fixphantombraking autopilot chill is not chill elon is an ass fix fsd extreme late deceleration fsd does not commit to go fsd does not maintain proper speed fsd hard and late braking fsd jerky driving fsd late deceleration fsd stops too early too quickly fsd stops too far behind line fsd tailgating in city streets
That's because my wife has never activated V12...
Yup speed control needs improvement, but I just didn't find your videos all that egregious and none that I was uncomfortable with. There are way more important improvements for FSD.
Last edited:
Todd Burch
14-Year Member
There is a theory that sometimes they do A/B testing with different parameters to see which work better, but as far as I can recall there’s never been confirmation of this.
I'm running my second version of V12, specifically V12.3.3 which isn't the latest. My experience is that I find it useful on the highway for making the drive more relaxing as I don't have to mind the speed quite as closely as if my foot were determining that, but V12 offers less real control than v11 did. V12 has more of a mind of its own and my ability to control speed with the thumbwheel on the steering wheel is less. And going into TACC requires you to come to a full stop first, which is less convenient. As to city driving, v12 is certainly improved with fewer nervous breakdowns at intersections, but I live in an old city with a lot of history including driving habits contrary to normal rules and some non-standard intersections plus some laxness in road markings. The result is problems. I also find the navigational routing problematic at times. I was on a suboptimal FSD route the other day when the car decided to do its right turn before we reached the intersection and I had to stop it from entering an A&W!! Riding outside the city the mapping wasn't up to date with construction and the car again wanted to turn right before the intersection under FSD and I had to intervene to keep us from going into a deep ditch. There was a driveway within 50 ft of the intersection that may have been a factor in the confusion. I always fully share my data with Tesla but I have my doubts about the near term future of FSD. To my mind it isn't good enough to get overall statistically much safer than human drivers if the car still does plainly idiotic things that a human driver could avoid. And that's the problem you have with the limited form of AI technology being used to develop FSD right now. There is nothing comparable to the human hypocampus that allows our brains to cope more effectively with something not seen before.
My 22 MS has always performed better on FSD than my 21 MY.
zoomer0056
Active Member
Yes, that matches exactly with the part of my post that was not quoted:
This sounds similar to the reports of FSD hitting curbs more frequently with Model S and X vs 3 and Y. I suspect it's a limitation of NN. Previously it was using logic to determine its actions, and that logic could be genericized across all vehicles. Now it's "photons in, controls out". So each specific version of each car (model, year, HW version, etc.) needs its own training data set, which is not going to be 100% equivalent across the board. It wouldn't surprise me if FSD performs better with Model Y. Since that is the most common Tesla, there will be more training data for it, and it makes sense to focus attention on it.
I suspect there is no more interpretation of the video to identify objects and their trajectories? Previously I imagine the programmers could issue generic commands like "speed down to 20 MPH" in response to a "collision probable" determination, and the underlying code would compute the pedal and wheel controls necessary to make that happen, based on variables such as vehicle power and road conditions. This scales relatively easily to a variety of vehicles and results in a fairly uniform experience across the board. But with pure NN, every vehicle configuration needs its own set of training data. Cybertruck doesn't even have any version of FSD yet. (Then again, Cybertruck doesn't even have Autosteer, which afaik doesn't use pure NN).
I'm new to Tesla, so a question for those who have been around. The v12 updates are rolling out in waves according to model and year. Was it like this with previous versions of FSD, or is it new to the pure NN FSD?
I'm not sure I understand. Do you never switch to an open lane to go around slow moving traffic? (I see humans do that all the time to cut to the front of the line.)
The problem would be where it was trying to cut in, but I can't really tell if it was aiming for the pole or if she just disengaged before it attempted to merge.
PianoAl
Active Member
We have two, both with the same software versions (see sig). I drive both extensively. I haven't noticed any difference in their behaviors.
I may be different from many here because I primarily "use" FSD rather than "test" FSD. I take over for UPLs or any sketchy situation. If the car starts to slow down too much or come too close to a curb, I disengage. I don't wait around for it at four-way stops, I press the accelerator.
willow_hiller
Well-Known Member
This really reinforces @EVNow 's point about the user-collected disengagement data being too subjective to be useful. In this case, FSD:
1. Drove down a legal, empty lane
2. Tried to merge into a turn lane at the last minute
3. Was angled toward a curb at the moment of disengagement, but it was also actively turning
Nothing about that was inherently illegal or dangerous until she intervened. Rude? Sure. But you cannot say with 100% certainty that it was a critical safety disengagement.
In my opinion, it's more likely that FSD would have found a space and continued turning or applied the brakes (FSD V12 behavior we have seen many times) than it driving over the curb and into a pole (FSD V12 behavior we have never seen before). But again, that's just my subjective opinion. Only Tesla would have the ability to simulate the next couple of seconds and figure out what the system would have done had it not been disengaged.
Sensors are the same, but the MS is more confident, doesn't nag as much, doesn't come as close to curbs, stays more centered in the lane.
Your character assassination of FSD not withstanding, why do people think FSD is smarter than it is? It's a driving assist system. It's not contemplating the meaning of the universe and existence. It's driving. You, as a human, can ask the deep questions, and ponder why the cars are all stacked up. You can come to the conclusion that those cars are probably in line waiting for the upcoming turn, but it's really just a guess. I guarantee you that there are times in your past where you sat, patiently, in line for an upcoming exit/turn, and then later, when you start to see lots of people changing lanes, find out there was an accident up ahead, and you just wasted tons of time sitting there like an idiot. But were you an idiot? You made a decision based on the information you had at the time, and didn't know there was an accident up ahead.
My point is that in that situation - DISENGAGE FSD AND DRIVE MANUALLY!! You, as a human with a sophisticated brain that's taken hundreds of thousands of years to evolve, can outthink a computer that's been learning for just a few years.
*ugh*
Individually, sure. But collectively, Tesla can examine all the data coming in for patterns. If every Tesla that ever tried to turn right at that light during rush hour resulted in a disengagement or an automatic reroute, I think that it would show up in data analysis. Combined with the many different ways drivers describe the problem, an analyst would understand the situation before watching any video.
willow_hiller
Well-Known Member
The user-collected data I'm talking about is the Google Form spreadsheet. Where anyone can submit their disengagement data and label it as a critical safety disengagement or not.
I think it's fine to use that data to show there has been a reduction in disengagements from V11 to V12, but people are taking the nominal number of critical safety disengagements from that spreadsheet and comparing it to services like Waymo. That just doesn't make sense for multiple reasons.
Tesla does have quality data on their own disengagements, but they have not shared any of that publicly.
My apologies. I took that as "user disengagement data going to Tesla". Yes, volunteered data is pretty useless. It requires motivation to submit data, and the motivation, whatever it is, usually skews the data pretty heavily. Which is why we have stuff like double blind studies. But you already know that
Sorry for the distraction.
DanCar
Active Member
Tweeter experiences FSD improvements without update. Quote:
... This left turn was jerky 10 out of ten times yesterday with clear blue skies at the 8am hour, no traffic, and failed 7 of 10 by tuning prior to median. Today, same conditions, smooth perfect turn 10 of 10. And the only thing that is different is upload and download of some data that was not a new build.
... This left turn was jerky 10 out of ten times yesterday with clear blue skies at the 8am hour, no traffic, and failed 7 of 10 by tuning prior to median. Today, same conditions, smooth perfect turn 10 of 10. And the only thing that is different is upload and download of some data that was not a new build.
mgs333
Member
Watching a video of it and experiencing it in the driver's seat are very different things. This ridiculous stopping behavior and the need for me to press the gas pedal several times a minute are the 2 biggest failings of FSD for me.
I've had other cars try to pass and cut it front of me when the car starts doing the slow creep.
AlanSubie4Life
Efficiency Obsessed Member
I’ll guess we won’t get followup.
Yes, exactly. Meanwhile even Elon appears to be confused and is frittering away the focus of SELLING CARS for some pie-in-the-sky dream of robotaxis which obviously is not a good use of limited resources. Focus on making FSD good; maybe people will like it and maybe it will improve safety. And make cars - good cheap ones.
But wasting resources on extremely difficult Robotaxis is obviously a dead end at the moment or at least is starting from scratch. So infuriating. Waste.
Last edited:
Similar threads
- Article
