Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register
This site may earn commission on affiliate links.
Very true, and this has made me realize the following:

Tesla could improve this by telling us in detail how our report-button comments are used. Are they all transcribed by computer? Are there people who listen and categorize the feedback? Are there key words? If we make the same comment for the same problem multiple times, is it given more weight? Ignored?

IOW, if Tesla teaches us how best to use the feedback, everyone benefits.
We’ve been asking for that since 2021
 
I assume they are taking a random sample of disengagements and analyzing them to determine the percentage that are necessary.
That seems terribly inefficient. I'd be doing statistical analysis, and rely on the frequency of disengagements by location, by given reason, and by anything else available to me. Speed? Model? Time of day? If 100 drivers have disengaged while turning left between 5 and 6 pm at a given intersection, that's worth looking into. If a bunch of drivers are disengaging, it doesn't matter if Tesla considers it necessary because the people involved apparently did.

I'd love to see a disengagement heat map for my geographic area. Again, with various filters.
 
Here's another case of v12.3 not detecting or responding to a cement median while attempting an unprotected left turn. Heavy reliance on map data and little on vision. I experienced the same thing on one of my first v12 drives.

There have been a lot of versions lately, and they clearly just accidentally released the Cybertruck FSD version to the other vehicles. Nothing to do with relying on maps. This also explains the curbing issues which made no sense otherwise (obviously would not release something with known curbing issues).
 
  • Informative
Reactions: Hiline
Did she somewhat like 11.x or not at all? Here's some comments from Dirty Tesla wife about 12.3.2.1 and Autopark:

"I'm surprised that that happened. I'm like, 'is this real life?' Honestly!" "I love this. I LOVE THIS!"

Let's just say that my wife WANTS me to turn on FSD when traffic is low. She instinctively knows that I am less fatigued when FSD is on for any drives longer than 30mins. I just don't use it when things are very complicated around me, but in most drives, she prefers FSD on. And I don't have any accidents in my life except towards the end of year 1 😅 and that's more than 35 years ago, so it's not as if I drive like an idiot.

I think some feels FSD is worse because of their environment. I'm fortunate that around my environment, it works quite well and it has been for a while now.
 
You cant use supervised FSD miles/accident accident rates to say anything about unsupervised FSD safety. Supervision removes the failures from the testing data.

10 times safer than the average human driver equals one accident per 2 million miles.

We are all down here squatting on the sidewalk talking about how wide the cracks are (sub weeds level) when we need to be taking the 40,000 ft view. But we can't, because the fine micro cracks in the side walk keep making it fail.

3 months ago the most optimistic cheerleaders were saying v12 is robotaxi by the end of the year.

I'm still in the somewhere between 10 years and never camp, and never on this hardware.
Supervised is not the same as unsupervised but you can look at disengagement rates, etc to get an initial estimate.
 
The idea of a procedural (C++) thread somehow monitoring and overriding FSDb when needed is...problematic. Because the better the NN gets at driving, the less confident the C++ code can be in making the decision to override. I can envision the C++ code as being a safety rail for a short time to address glaring flaws in behavior of the NN, but it does not seem like an effective long-term solution. The NN is just going to have to get really, really good all by itself AFAIK.
there could still be a sort of supervisor rule. There has to be in the case of local laws, etc… it’s more practical to encode that as actual rules. The automatic safety systems ideally also remain active in case the NN decides to do something boneheaded.

At some point regulators will demand repeatable behavior (and hopefully without gimping the systems too much). They will want to know that the car will obey no right on red rules, etc…. So the NN will always have to have something constraining or overruling its decisions from time to time. That might be another layer of the NN or some other technique only advanced AI researchers know, not necessarily imperative code in an archaic language like c++, but key is it has to be reliable and deterministic enough to satisfy the regulators.
 
I did another drive today with v12.3.1. This was about 40mins worth, again around 4 planned disengagements with V11.

Overall, it handled the drive well. This route has two very difficult intersections with weird angles and less visible turns. Lots of traffic around this hour and V12.3.1 handled them very well. Both directions were zero disengagements, which was a surprise to me even with heightened expectations.

3 minor annoyances:
- I clearly see the difference between Assertive and Chill now, at least in terms of acceleration from stop/low speed. Assertive is quite a bit more forceful. So now I keep it in Chill mode. Even this one feels too strong for me in most situations. I wish if current Chill becomes the new Average and a new more mild Chill mode is added in the future.
- When it wants to switch lane, or if I trigger it through signal, it always hesitates very subtly then makes the move. It is also true that changing lane from left to right is now quicker, which is nice. Even when on Chill, I also observed that it changes lane, both when a slower vehicle is encountered in front of me on the right lane, or when another car is tailgating me on the left lane. I have NOT tried MLC, because I forgot. 😅
- I'm still not used to auto-speed. Sometimes too fast, sometimes too slow. When it is too slow, tapping on acc seems to work, so I'm sort of ok with that. But I haven't figured out a way to slow the car down when auto-speeding too fast without disengaging. Anyone?

Overall, I still think v12.3.1 fixes 80% and broke 20%. For me, V12 is more useful, at least in v12.3.1 form. Just need to figure out how to get around the faults without adding much effort or stress.

EDIT : I forgot to mention that, during this drive, I had camera views on all the time to see if the car is close to the curves during all turns. At least in v12.3.1 in my drives, I didn't feel uncomfortable. Either this happens less frequent, or v12.3.1 at least partially fixed this issue from the earlier versions of v12.
 
Last edited:
But I haven't figured out a way to slow the car down when auto-speeding too fast. Anyone?
Turn off auto speed (ASSO).

This gives you ASSO, but allows you to cap the speed with your desired manual set speed offset.

Turning off ASSO does not turn off ASSO AFAIK. It just imposes a cap. (In other words the modes seem to behave the same if manual speed offset is +25% or whatever since probability of ASSO going that fast is basically zero.)
 
  • Like
Reactions: FSDtester#1
Very true, and this has made me realize the following:

Tesla could improve this by telling us in detail how our report-button comments are used. Are they all transcribed by computer? Are there people who listen and categorize the feedback? Are there key words? If we make the same comment for the same problem multiple times, is it given more weight? Ignored?

IOW, if Tesla teaches us how best to use the feedback, everyone benefits.
This is how my poor eyes read: "IOW, if Tesla teaches us how best to use the Facebook, everyone benefits." :)
 
Turn off auto speed (ASSO).

This gives you ASSO, but allows you to cap the speed with your desired manual set speed offset.

Turning off ASSO does not turn off ASSO AFAIK. It just imposes a cap. (In other words the modes seem to behave the same if manual speed offset is +25% or whatever since probability of ASSO going that fast is basically zero.)
Got it, thanks. That solves the "too fast" issue then. Nice! 😁
 
- I clearly see the difference between Assertive and Chill now, at least in terms of acceleration from stop/low speed. Assertive is quite a bit more forceful. So now I keep it in Chill mode. Even this one feels too strong for me in most situations. I wish if current Chill becomes the new Average and a new more mild Chill mode is added in the future.
Really? I found Assertive to be gentler. (As logged by my insurance safety scoring app.)

But I haven't figured out a way to slow the car down when auto-speeding too fast without disengaging. Anyone?
I've found I can use the scroll wheel to slow the car down. (But it doesn't work to speed it up.)
 
Really? I found Assertive to be gentler. (As logged by my insurance safety scoring app.)


I've found I can use the scroll wheel to slow the car down. (But it doesn't work to speed it up.)
I let Assertive drive for about 5 stop/lights and in all cases, it was uncomfortably too strong, which didn't happen in other Chill based stops. I'll experiment further during the future drives.

As for scroll down, I did not try this. Again, I'll try those as well.

Thanks for the tips!
 
  • Informative
Reactions: FSDtester#1
At a guess, they're distracted by a bunch of new priorities. Getting V12 onto demo cars, new cars, into factory builds, into Canada, implementing Elon's new trial "by the end of the week" (thanks, Elon), and so on.
....and watching carefully for unintended or unknowable bugs. This is the fastest rollout since 10.3 and that was a disaster. Tesla doesn't want a 10.3 repeat.
 
an archaic language like c++,
Haha, archaic maybe compared to an LLM that can just write software to do whatever I want with zero effort on my part, but C++ is still AFAIK the first choice for high-performance, low-latency, low-power applications. Just because it's fragile and high maintenance doesn't make C++ archaic when there are no better tools for the many jobs C++ is still the best for...