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.
Original 10.2 safety score beta tester here; got the 12.3 update last night. My short morning commute was pretty solid.
  • FSD finally drives at an appropriate speed in my neighborhood (~15 MPH). On v11, it would try and blaze through at 30-40 MPH.
  • Stop sign/creeping behavior is still kinda clunky. It stopped too early and crept for probably 10 seconds before committing to go. Visibility isn't great out of my neighborhood...not sure any software can fix the camera placement. Regardless, it wasn't a safety problem, it performed the turn well and it didn't stick out into the road before committing like v11 would often do. It just took an eternity :)
  • Odd random slowdowns on country roads. I wouldn't call it phantom braking because it was very subtle; it was more like frequent brief hesitations.
  • The speed stuff that's been mentioned 1,000x already. It's pretty slow to get up to speed when the limit changes, but it seems to mostly stay up to speed after giving it a little juice.
  • Still not slowing down for school zones.
Overall, I'd basically given up testing the last couple iterations of v11 because it was a frustrating mess and felt like it didn't matter anymore. I'm only one short drive in on v12.3, but I'm very encouraged! Hopefully we'll see more steady improvement going forward.
 
If I don’t have 12.3 by 6am I’m going to find a 1967 Dodge Dart in a junkyard, fill it with a combination of 87 octane and kerosine, fill the oil tank to WELL ABOVE the fill line (gasp!), remove the catalytic converter, park near a grove of young trees, and hold the RPMs at 6,000 as the trees get doused in poisonous gas until the tank runs dry.

My terms have been laid out. Tesla, you have less than 8 hours to comply.
You can even be helpful, go do that in Grünheide and help Tesla out with clearing that forest!
 
Original 10.2 safety score beta tester here; got the 12.3 update last night. My short morning commute was pretty solid.
  • FSD finally drives at an appropriate speed in my neighborhood (~15 MPH). On v11, it would try and blaze through at 30-40 MPH.
  • Stop sign/creeping behavior is still kinda clunky. It stopped too early and crept for probably 10 seconds before committing to go. Visibility isn't great out of my neighborhood...not sure any software can fix the camera placement. Regardless, it wasn't a safety problem, it performed the turn well and it didn't stick out into the road before committing like v11 would often do. It just took an eternity :)
  • Odd random slowdowns on country roads. I wouldn't call it phantom braking because it was very subtle; it was more like frequent brief hesitations.
  • The speed stuff that's been mentioned 1,000x already. It's pretty slow to get up to speed when the limit changes, but it seems to mostly stay up to speed after giving it a little juice.
  • Still not slowing down for school zones.
Overall, I'd basically given up testing the last couple iterations of v11 because it was a frustrating mess and felt like it didn't matter anymore. I'm only one short drive in on v12.3, but I'm very encouraged! Hopefully we'll see more steady improvement going forward.
Agree on basically given up on 11. Only time I used it a lot was on the interstate. The times I’d use it in city/rural, I’d leave messages like, “are you fkn’g drunk?? Who fkn’g drives like this??? Stop getting in the fkn’g turn lanes when you know I need to go straight!!! Yeah! Accelerate right up to a bunch of cars at a red light and dynamite your fkn’g brakes you autistic twat!!!”

Pretty much gave up on 11, yeah. V12? I’m staying off the interstate and taking rural roads exclusively because it’s so good.
 
I did my initial drive this morning with little to no traffic and overall it's still pretty rough. Acceleration and braking have a ridiculous step like response. That's not an average driver behavior so not sure how the NN learned it. Something's a miss.

Speed control needs work as others have repeatedly shared.

Still too slow when approaching stop signs. It makes sense to slowly approach crosswalks when people/bicycles are present but otherwise WTF! There needs to be human-like profiles to distinguish when an intersection is clear, no people/etc are present, or the intersection is occluded and we need to get our nose out to see what's coming. Humans don't have one default way of approaching intersection/crosswalks/stop signs so why isn't training data capturing that?

Turns much improved although still jerky steering at times. Steering is laggy and frequently crossing over lane lines. There's too much lane drift. Uncomfortably close to curbs. Glad v12.3 rid itself of turn-right to turn-left. Turning into my neighborhood is finally feels comfortable with appropriate speed in turn and no apex whisky throttle.

Several times v12.3 didn't acknowledge or slow for obvious dips (rain run-off) in the road. There should be training data for dips.

Occasional dumb turn signals and dry wipes.

Overall I was a bit shocked and underwhelmed given all the praise. Maybe it will get better on my next drive. :)
 
I wonder if hard, or at least early, braking is somewhat intentional to verify road conditions while the car can still do something about it.
How would that tie in with training data? Are you suggesting that they found examples of people braking like that and chose them over sane braking? I really wonder if they're still hacking around with heuristics in an effort to clean up whatever it is that the neural net is telling the car to do.

I wouldn't call it phantom braking because it was very subtle; it was more like frequent brief hesitations.
Did it correlate with darker patches of asphalt? Meaning literal patches of fresher asphalt. The car may be thinking that they are speed humps for a second, so it slows slightly, then realizes they're not. I think I've seen this behavior.

Overall I was a bit shocked and underwhelmed given all the praise. Maybe it will get better on my next drive.
Depending on where "here" is, you may be driving in an environment for which the system isn't well-trained. There's also the driver settling-in period. I've gotten more comfortable with V12 because I've adjusted my expectations and I know what to expect of it.
 
Depending on where "here" is, you may be driving in an environment for which the system isn't well-trained. There's also the driver settling-in period. I've gotten more comfortable with V12 because I've adjusted my expectations and I know what to expect of it.

Who knows but the vehicle control NNs have the feel of being trained with simulated data. Whatever it is hopefully they can clean it up as it will make a huge difference in comfort.

Gotta wonder if training data for stop sign/intersection approach is also partially simulated data as it's too similar to v11.

The big question - how much simulated data is being used to train v12? It could be a way to get similar drive characteristics from those 300k lines of code.
 
Last edited:
I see the Black Tesla youtube channel noticing that FSD 12.3 seems to be trained on model 3 and model y cars. He is driving a larger Model S Plaid and it seems to get too close to the side of the road in some instances, leading him to believe that FSD 12.3 does not have a sense of the size of the car it is driving. It would also show up in the cybertruck when FSD becomes available. If this is true, future FSD versions would need to be able to assess the size of the car they are driving.
 
I see the Black Tesla youtube channel noticing that FSD 12.3 seems to be trained on model 3 and model y cars. He is driving a larger Model S Plaid and it seems to get too close to the side of the road in some instances, leading him to believe that FSD 12.3 does not have a sense of the size of the car it is driving. If this is true, it would also show up in the cybertruck when FSD becomes available. If this is true, fsd would have to be able to assess the size of the car it is driving.

My Model Y also gets very close to curbs and whatnot
 
Did it correlate with darker patches of asphalt? Meaning literal patches of fresher asphalt. The car may be thinking that they are speed humps for a second, so it slows slightly, then realizes they're not. I think I've seen this behavior.
Regarding the braking/hesitations - No, I didn't notice such a correlation. The asphalt is pretty even in that regard on my commute. I DID notice that it sometimes braked for cresting hills, which is something older versions of FSD used to do, presumably being cautious because it can't see what lies ahead. Of course, having driven these roads a billion times, I know I'm not about to drive off a cliff and I also know where the road curves, so...we'll see how v12 handles this with time.
 
  • Like
Reactions: edseloh
How would that tie in with training data? Are you suggesting that they found examples of people braking like that and chose them over sane braking? I really wonder if they're still hacking around with heuristics in an effort to clean up whatever it is that the neural net is telling the car to do.
Good question, I was admittedly still in the human coded realm of thought.
However, it could still be valid. If the NN has a concept of speed vs distance vs surface for braking and some training clips are from adverse surfaces then it may start breaking based on the conservative examples, realize it overshot the deceleration path, and then reduce braking.
Could also be similar with stopping behind the line at stop signs. It doesn't know if the surface will provide sufficient grip, so it targets zero velocity early versus risking sliding through. Like using winter driving habits in summer.
 
On TesalFi 6.5% of the vehicles have FSDb 12.3, while almost 24% still have FSDb 11.4.9.

So ~78% of the FSDb vehicles are still on 11.4.9.
It sucks there's no real pattern. I thought sub ~40-50 pending triggered a wave. Then time based. Turns out it seems to be manual.

At this point I just hope there is a wave today or tomorrow and the wave is big.

I hate that the 12.3.1 weekend usually means like evening Sunday and the roll will be painfully slow but I guess expect the worst and be happy when it's not? I just need to find something to steal my time while I wait. Like a drug addiction or something.
 
  • Funny
Reactions: edseloh and Tronguy
It sucks there's no real pattern. I thought sub ~40-50 pending triggered a wave. Then time based. Turns out it seems to be manual.

At this point I just hope there is a wave today or tomorrow and the wave is big.

It is most likely that they stopped the 12.3 rollout because they decided that 12.3.1 was needed to fix issues.

I hate that the 12.3.1 weekend usually means like evening Sunday and the roll will be painfully slow but I guess expect the worst and be happy when it's not?
And with 12.3.1 the whole process starts over. It will first go to people that already have 12.3, and those of us without it likely won't get it for at least another week... :(