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.
Obviously. My question was if you were a passenger would you notice the speed issues you mentioned? It's not unreasonable to think that most of use who adjust speeds would care less it we weren't driving. Especially if we were in the back seat. I adjust speeds frequently and I'm beginning to wonder whether it's because I want FSD to drive just like I do when maybe that's not so important. The majority of V12 negative feedback seems to be related to speed so just wondering.
I think Tesla is growing up. The wild west days are gone. For FSD to go main stream out of beta, it has to obey all laws. Even though recalls can be done over the air, but why do it in the first place. So don't expect Tesla to relax the not over speed limit implementation and stops (0 speed) at the sign posts. User speeding up using the accelerator is necessary for safety. That would be a definitive user action and can be recorded and probably satisfy the agency requirements.
 
It seems like 12.x is quite impressionable/suggestible. There's the common example of pressing the accelerator to speed up, and it'll generally maintain that new speed at least for a little bit.

I was playing around with 12.3 today and ended up with navigation wanting a left turn where left turns are disallowed with signage and right-turn-only painted road marking. Initially it turned on the left turn signal and waited for cross traffic to make the left turn, but when I canceled the turn signal, it correctly made a right turn. I tried again later without canceling the left signal, and I needed to disengage before it completed the left turn.

Additionally testing at various other stop signs in small residential streets, I was able to get 12.3 to turn right even though navigation and original FSD-Beta-engaged turn signal wanted to turn left. Similarly, even when navigation wanted to go straight, I was sometimes able to trick it to turn by engaging the turn signal. Although it seems like trying to force a turn by manually engaging the turn signal "too early" will result in FSD Beta realizing it doesn't need to turn and turning off the signal.
 
If the 1.1 mile 3rd test drive I did from downtown to home were any example for a STUDENT DRIVER test, the proctor would have failed them THREE times.. should that be some sort of way to judge FSDb at this point… over under vs. a student driver? I know a few YEARS ago, I said it was more like a student driver OR an inebriated driver more than anything else. We REALLY have to get past at least that minimum bar to start to believe that this is achievable.

So, will you continue using FSDb all the time, putting in the hours in order to make it good and keeping your posting privileges here?
 
Last edited:
So, will you continue using FSDb all the time, putting in the hours in order to make it good and keeping your posting privileges here?
Funny how there are so many different opinions. I’ve had it one day. I did four drives and so far zero interventions. I basically never had a zero intervention drive prior to this unless it was extremely short. It’s not perfect by any means but it’s a big jump in the right direction. I’ve had fifty Uber rides that were worse. I see drivers everyday drive far worse than 12.3.
 
Sunday nights are usually the big drops historically..
I wouldn't be surprised to see it
For those waiting on V12, do not make the mistake of installing v11 on a 2024 branch, or you will be waiting for probably a month or so.
I'm not optimistic for 2 reasons. First, I live in the Northeast where only a handful of owners have V12. Second, I have HW4.
But I do expect a bigger drop overnight.
 
At stop signs it seems to fully stop and then have a false start (moves but stops again) more often than the prior version.
After a couple of drives today, I think the stop sign behavior is improved in the way it approaches the creep line. 11.4.9 would often lurch aggressively towards the street edge, causing me to disengage because I couldn't trust that it was only moving up and not really starting to turn in front of an approaching cross-traffic car. In 12.3, it's much more comfortable, but maybe overcorrected to be too slow now. No disengagement needed, but if anyone were behind me I might have had to goose the pedal.

Regarding the "NHTSA Stop" at the sign before that (which we understand is clearly achieved by a training intervention to satisfy the Feds): previously I felt it was stopping with the bumper right at the stop sign, now I think it's stopping a few feet before the sign.
Still ignores "No Turn on Red" signs.
So far, 12.3 has honored No Turn on Red at one particular large T-intersection (@Dennisis eastbound Orange Grove turning onto Skyline Drive). Last night it stuttered just a little but held its place.

There are at least four fixed NTOR signs, multiple red arrows and a couple of special large dot-matrix No Right Turn lighted signs - the overkill is kind of hilarious, yet it took at least a couple of years for the rule to be honored by most drivers - and it's still frequently violated. 11.4.9 wouldn't always honor it, especially not if a lead car took the turn.
1710727574992.png
 
Got the update overnight. Interestingly, FSDb wasn't available yesterday. I had to calibrate the cameras, re-accept terms, and reset the associated the options.

Initial observations from today's test driving:
  • @AlanSubie4Life, that right turn out of the former Sorrento Valley post office business park onto Vista Sorrento Parkway is much improved. It was downright reasonable and appropriately cautious with this version! No more wanting to blindly blast out into traffic! I guess I won't have to wear my $hi11in' pants anymore for that turn.
  • U-turns are a welcome addition! It completed most OK, although slowly. Even though, as noted by someone else, it showed a message indicating it would assistance, it completed the turns by itself. However in one case, I had to take over to speed things up due to oncoming traffic.
  • It handled most cul-de-sacs well by completing a u-turn and continuing on toward its destination. In one case, a parked car didn't allow enough room for the turn to be completed without a 3-point turn. The car stopped after some wheel jerking, but no message. It's not able to reverse in this scenario -- yet.
  • The auto speed setting results in driving that's too slow -- particularly on residential streets with zero traffic. Gentle goosing of the power pedal seems to increase the speed to a higher range.
  • Speed bumps are handled better, but not rendered on the screen -- which is fine with me.
  • Still ignores "No Turn on Red" signs.
  • Usually maintains centered position in lane, but sometimes is too slow to move between lanes or stays way too far to the right within the lane. Sometimes drives outside the right lane marker for no reason.
  • Parking lot driving is much improved. Smoother and less indecision and resulting wheel jerk.
  • I started FSDb on the top level of a parking deck (mostly empty) with a destination set that was a mile away. This test seemed to confuse it; maybe it thought it was on a lower level or in an open regular parking lot. It would just drive in a circle through that level and then park in a spot -- apparently giving up. The parking was good, and a nice way for it give up. However, there was no message.
  • Immediately after turning into right into a strip mall parking lot, the car issued the red hands of death forcing me to take over. It displayed an error message that was something like "Internal Autopilot Error". The prior version never had a problem with this particular turn.
  • At stop signs it seems to fully stop and then have a false start (moves but stops again) more often than the prior version.
  • Improved approach to red lights with earlier slowing.
As always, two steps forward and one step backward!
  • Still ignores "No Turn on Red" signs.
I saw one opposite situation: it stopped at No Turn on Red but did not see "Monday - Friday" and continued to wait for green light.
 
  • Funny
Reactions: Pdubs and KArnold
It seems like 12.x is quite impressionable/suggestible. There's the common example of pressing the accelerator to speed up, and it'll generally maintain that new speed at least for a little bit.

Does this imply that the auto speed neural nets have some memory of previous speeds? I guess that makes perfect sense. I wonder how long it's lookback period is and how this actually works. Is historical velocity an input to the neural net?
 
Does this imply that the auto speed neural nets have some memory of previous speeds?
Potentially the neural networks are predicting how much acceleration as opposed to a target speed, so the common case prediction would have enough acceleration to maintain the current speed. Unclear if there's explicit memory of previous speeds, but presumably there's enough memory / previous video / internal state to figure out speeds of other objects including static ones like the road.

I haven't tested enough of regular set speed, but it seems like 12.x tries to drive at the speed it wants to unless traditional control code prevents exceeding the set speed (or 50% above detected speed limit). AUTO max set speed effectively is the Autopilot 85mph limit, so the main benefit of numeric set speed seems to be to go slower, e.g., for school zones if end-to-end doesn't understand.
 
400 miles on 2023.44.30.25 so far. 2021 MY (HW3). Auto max speed blows on state highways. Tries to match speed when clusters of speeders pass, otherwise hangs out at 2-3 above the limit.

Folks are reporting v12.3 is cutting some turns short/close and it isn't perfect. No worries about shutting down the release as long as drivers watch closely.
Drove on the wrong side of the road today after taking a protected left. Stayed left of the double yellow on the 300' or so into a SC.
No notable improvement overall vs. v11. Very very different. If they fix the obvious problems quickly it will be quite a bit better than v11 though. The fundamentals of driving behavior seem a bit more natural which is why I think this. But we’ll see if they can fix the obvious problems - in the past this has taken years.
Subjective for sure. 12.3 is wife approved so far. Significant improvement. On 10.5 it couldn't circle the block around our house without intervention. Today was several tens of miles without an intervention. Definitely a step forward but I respect your interpretation.

Biggest thing for me so far was the wrong lane selection above and then stopping too far behind stop lines today and then going without creeping (I know most would celebrate but it was uncomfortable).

At the risk of sounding like Omar - it's just way more smooth... I hope they keep rolling it! Definitely better than 11.4.9 for my 120 mile commute.
 
The majority of V12 negative feedback seems to be related to speed so just wondering.
My complaints about speed control related to facilitating the flow of traffic and to my personal comfort. I'm always going to be sensitive to the latter. Facilitating the flow of traffic is the number one priority for any driver, but as a passenger I'll be less inclined to worry about it unless my attention is drawn to it. For example, honking horns, or even just my noticing strange happenings because I'm occupying my eyes by watching traffic.

If I had reasons of less consequence then I'm sure I'd set them aside as a passenger.
 
Based on my understanding of the v12 E2E architecture, neither general nor specific sign reading would necessarily become an "emergent" capability from the training clips.

Regular stop signs are ubiquitous and are obviously understood by the system today, in almost all cases. But more to the point, even though they advertise that they don't explicitly label those, I believe that the v11 stop sign recognition sub-network and the training starting-point weights were carried over from v11, and then subsequently and continually reinforced by the training.

Other kinds of traffic rule signs, including NToR and school zone directives etc, were only sporadically (if at all) recognized in v11. Again, I don't think these capabilities would emerge easily in a training process that simply added some random set of parameters. In theory it could but at a very high cost in training data and time.)

But if they developed an initial heuristic module tuned for sign text recognition, perhaps one that almost worked but never well enough for deployment, then grafting those tensor elements and starting-point coefficients onto the v12 architecture would be the enabler for much-improved understanding of posted signs.
 
  • Informative
Reactions: FSDtester#1