I have 2024.3.10 with v12.3.3 FSD. My question is how should I react to FSD errors? If there is a neural network training process involved, does it learn when I just disconnect and take over manually without explaining why? Or should I always explain why I dropped out of FSD every time? Sometimes it is because of something that wasn't the fault of FSD, other times it is definitely when FSD does something stupid? Case in point, there is a fork in the road on my way home that the route planner always designates as the route, but every time it misses that turn, goes down the wrong road and then gets confused and stops. Is there anything I can do to help train the thing, or does it learn from the daily repeated times I have to manually take over?? I guess I'm just not sure WHAT Tesla expects us to do as "beta testers" and whether it's worth bothering with. They haven't been very clear, and I suspect they have just been gaslighting us to make the NHTSB happy. Any thoughts?? Would be nice to hear from some Tesla engineers!
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD Neural Network Training question
- Thread starter viking78
- Start date
EVRider-FL
Active Member
You won’t hear from any Tesla engineers here. My understanding is that the car doesn’t learn directly from your actions, but disengagements are reported to Tesla, and are used to help refine the model. If you include a voice annotation, that can help Tesla understand what happened. How much of that data Tesla engineers actually look at, only Tesla knows.
I don’t do voice reports unless I disengage FSD because it screwed up.
I don’t do voice reports unless I disengage FSD because it screwed up.
In addition to voluntary and forced disengagements, I believe interventions such as pressing the accelerator to the make the car go or go faster may also be flagged for Tesla to review.
I used to think (or hope) that the cars learned individually - it would be very helpful if I could teach my car how to make the same drives that I regularly make every day between home and work, but that is not the case.
FSD 11.4.x reliably repeated the same, poor behavior for the five months I subscribed to it last year. One example was at a couple of 4-way stops on my way to work where the car would either drive straight through the intersection without even attempting to slow down or, occasionally, come to an ABS-controlled stopped in the middle of the intersection! (Because these rural intersections rarely had other traffic, I would regularly let the car do what it wanted to do just to see what happened.) I lost count of the number of times I disengaged and reported what happened, but it never changed.
During my first week with FSD 12, I've noticed that it doesn't always do the same thing. Sometimes, it improves on a previous maneuver, but occasionally will do something wrong that it previously did right. I will say that whether or not my reports were a factor, I reported a few disengagements with 12.3.2.1 that were ALL resolved a few days later in 12.3.3.
I still have to occasionally disengage 12.3.3, but the number of disengagements has dropped by an order of magnitude over 11.4.x. Before, I rarely had a drive where I did not have to disengage. Now, I rarely have a drive where I do have to disengage. The difference is real - at least for me.
I used to think (or hope) that the cars learned individually - it would be very helpful if I could teach my car how to make the same drives that I regularly make every day between home and work, but that is not the case.
FSD 11.4.x reliably repeated the same, poor behavior for the five months I subscribed to it last year. One example was at a couple of 4-way stops on my way to work where the car would either drive straight through the intersection without even attempting to slow down or, occasionally, come to an ABS-controlled stopped in the middle of the intersection! (Because these rural intersections rarely had other traffic, I would regularly let the car do what it wanted to do just to see what happened.) I lost count of the number of times I disengaged and reported what happened, but it never changed.
During my first week with FSD 12, I've noticed that it doesn't always do the same thing. Sometimes, it improves on a previous maneuver, but occasionally will do something wrong that it previously did right. I will say that whether or not my reports were a factor, I reported a few disengagements with 12.3.2.1 that were ALL resolved a few days later in 12.3.3.
I still have to occasionally disengage 12.3.3, but the number of disengagements has dropped by an order of magnitude over 11.4.x. Before, I rarely had a drive where I did not have to disengage. Now, I rarely have a drive where I do have to disengage. The difference is real - at least for me.
Similar threads
