FSD v12.x (end to end AI)

[TK211X]

Can I not be cool. My mouth is dry and no 12.3 yet.

 

[Ben W]

I am not talking about C++ parameters.
See training vs AI parameters here:

Training computation vs. parameters in notable AI systems, by domain

Computation is measured in total petaFLOP, which is 10¹ⵠfloating-point operations estimated from AI literature, albeit with some uncertainty. Parameters are variables in an AI system whose values are adjusted during training to establish how input data gets transformed into the desired output.

ourworldindata.org

Parameters in notable artificial intelligence systems

Parameters are variables in an AI system whose values are adjusted during training to establish how input data gets transformed into the desired output; for example, the connection weights in an artificial neural network.

ourworldindata.org

These parameters are only modified by retraining the neural net. (There are two semantic definitions of "parameters": the one described here is neural net weights. The other sense is user-facing parameters outside the model, such as assertiveness or auto-speed or "minimal lane changes".) So I'm not quite sure what you're asking?

 

[rlsd]

These parameters are only modified by retraining the neural net. (There are two semantic definitions of "parameters": the one described here is neural net weights. The other sense is user-facing parameters outside the model, such as assertiveness or auto-speed or "minimal lane changes".) So I'm not quite sure what you're asking?

I was thinking of the weights that might be applied to put the car closer to the curbs that I mentioned earlier. I am not AI expert. Just some thought.

 

[PianoAl]

Today I test drove the section from this video 5-6 times with FSD and with enhanced autopilot (AP aka Autosteer beta).

Here is what I found:

1. It did not go over the yellow line today.

2. The yellow line is dirty/worn, but not too bad. I guess with the right sun and angle, it might be hard to see. You be the judge:

3. FSD was often on the plastic line markers or the rumble strip. AP was not (although close). At one point, FSD seemed too close to the side of the road. I didn't try that section with AP.

4. I pressed the accelerator briefly prior to the turn. FSD took the turn too fast (scary fast), AP did not. That is, AP slowed down appropriately.

5. In this general area, FSD drove too slowly. AP did not have that problem. That is, I set the max speed to 58 MPH, and AP was spot on. It drove at 58 MPH on straightaways and moderate curves but slowed down when necessary. Refreshing!

-------------------

What I've learned: Don't press the accelerator to get up to speed, but if you do, watch for curves and be EXTRA prepared to take control.

In trying to give FSD the benefit of the doubt (and restore my confidence): I wonder if it's possible that I had pressed the accelerator and the car was taking the turn too fast, and in my panic (but before it went over the line) I tugged the wheel enough to disengage FSD. Then the car went over the yellow line. I noted that the Report message came up right afterwards, but it would have been there from the inadvertent disengagement.

Arguing against that, I'm very used to driving with FSD and rarely disengage unintentionally. Also, watching the video, it doesn't look like it was going too fast around the curve. Finally, I did not get a lane departure warning (I should have, right?).

In any case, the Tesla engineers will be able to know exactly what happened.

 

[rlsd]

Can I not be cool. My mouth is dry and no 12.3 yet.

Don't you know that you should not drink Elon's Cool Aid?

 

[kabin]

I saw that too but it's not too close. It seems FSD takes the approach: be defensive as much as possible.

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.

 

[TK211X]

Don't you know that you should not drink Elon's Cool Aid?

 

[rlsd]

Today I test drove the section from this video 5-6 times with FSD and with enhanced autopilot (AP aka Autosteer beta).

Here is what I found:

1. It did not go over the yellow line today.

2. The yellow line is dirty/worn, but not too bad. I guess with the right sun and angle, it might be hard to see. You be the judge:

View attachment 1028908View attachment 1028909View attachment 1028910

3. FSD was often on the plastic line markers or the rumble strip. AP was not (although close). At one point, FSD seemed too close to the side of the road. I didn't try that section with AP.

4. I pressed the accelerator briefly prior to the turn. FSD took the turn too fast (scary fast), AP did not. That is, AP slowed down appropriately.

5. In this general area, FSD drove too slowly. AP did not have that problem. That is, I set the max speed to 58 MPH, and AP was spot on. It drove at 58 MPH on straightaways and moderate curves but slowed down when necessary. Refreshing!

-------------------

What I've learned: Don't press the accelerator to get up to speed, but if you do, watch for curves and be prepared to take control.

In trying to give FSD the benefit of the doubt (and restore my confidence): I wonder if it's possible that I had pressed the accelerator and the car was taking the turn too fast, and in my panic (but before it went over the line) I tugged the wheel enough to disengage FSD. Then the car went over the yellow line. I noted that the Report message came up right afterwards, but it would have been there from the inadvertent disengagement.

Arguing against that, I'm very used to driving with FSD and rarely disengage unintentionally. Also, watching the video, it doesn't look like it was going too fast around the curve. Finally, I did not get a lane departure warning (I should have, right?).

In any case, the Tesla engineers will be able to know exactly what happened.

I agree with not pressing the accelarator if the car is not going on a straight road. There were times I pressed the accelerator for turning because I was impatient but I was wrong: there were other cars coming that I did not see.

I also don't change speed speed limit when the car is making turn to avoid interferring with the car's thinking.

 

[arnolddeleon]

In reviewing the TeslaFi details for subscriber cars that have received 2023.44.30.25/FSDb12.3 I though I noticed in addition to the initial California bias and the current "SunBelt" bias that there may be a higher mileage bias.

Speaking as the owner of two Teslas over a year old with FSDb purchased but less than 12,000 miles on the odometer, I 'm struck that only four TeslaFi subscribers with less than 12,000 kilometers on the odometer have received this update, out of 432 currently installed.

I wonder if they are preferring not just odometer miles but actual miles driven under FSDb?

I seriously doubt that. We got in on our Y which my wife drives primarily and she doesn't use FSDb. Our 3 is still waiting. I use FSDb almost always when I drive but I don't drive that much.

 

[Ben W]

I was thinking of the weights that might be applied to put the car closer to the curbs that I mentioned earlier. I am not AI expert. Just some thought.

This is almost certainly an emergent property of the neural net training. They'll need to retrain the model with fewer examples of being close to the curb (and more examples being further away) to properly address this. Unclear whether there is supervisory C++ logic that explicitly monitors curb closeness and can nudge the car further away (simulating driver input, basically), but they're really trying to get away from that sort of logic I think.

 

[kabin]

Today I test drove the section from this video 5-6 times with FSD and with enhanced autopilot (AP aka Autosteer beta).

Here is what I found:

1. It did not go over the yellow line today.

2. The yellow line is dirty/worn, but not too bad. I guess with the right sun and angle, it might be hard to see. You be the judge:

View attachment 1028908View attachment 1028909View attachment 1028910

3. FSD was often on the plastic line markers or the rumble strip. AP was not (although close). At one point, FSD seemed too close to the side of the road. I didn't try that section with AP.

4. I pressed the accelerator briefly prior to the turn. FSD took the turn too fast (scary fast), AP did not. That is, AP slowed down appropriately.

5. In this general area, FSD drove too slowly. AP did not have that problem. That is, I set the max speed to 58 MPH, and AP was spot on. It drove at 58 MPH on straightaways and moderate curves but slowed down when necessary. Refreshing!

-------------------

What I've learned: Don't press the accelerator to get up to speed, but if you do, watch for curves and be EXTRA prepared to take control.

In trying to give FSD the benefit of the doubt (and restore my confidence): I wonder if it's possible that I had pressed the accelerator and the car was taking the turn too fast, and in my panic (but before it went over the line) I tugged the wheel enough to disengage FSD. Then the car went over the yellow line. I noted that the Report message came up right afterwards, but it would have been there from the inadvertent disengagement.

Arguing against that, I'm very used to driving with FSD and rarely disengage unintentionally. Also, watching the video, it doesn't look like it was going too fast around the curve. Finally, I did not get a lane departure warning (I should have, right?).

In any case, the Tesla engineers will be able to know exactly what happened.

It could be a lot of things. Maybe those tight turns aren't constant radius turns and v12.3 hasn't had enough trained data for that? Sometimes tight blind turns require more braking which requires more steering input. And then road camber makes a difference too.

 

[FSDtester#1]

Beaver tested 10 times. Perfect arc coming in each time.
Almost as if the beaver was removed from the street.
Also performed my first escape from the cul-de-sac at the end of the street, which it always made me take over for almost 3 years.
Unlike Chuck's UPL, was 10 out of 10.
Hallelujah.. my beaver nightmares are over and don't have to disengage driving into my neighborhood. My wife owes me a beer .

 

[Supcom]

Started with auto set speed on - it doesn't work, unless you like to drive around really slowly. As mentioned, it will speed up if you press the accelerator and sort of stick there, but then over time it will decay back down to some ridiculously slow speed.
Halfway through I switched back to old speed method (I like to set 10% over or so). They broke this too. It doesn't get up to the set speed. You can push the accelerator there too and it sort of works for a while then is broken again.
Mostly speed setting is totally broken, though. Had to think about it, which is not what we want.

I agree. I believe that the Max Speed or Auto Max only sets the maximum speed that the car will go. The car will still select the speed it wants to go up to that limit. What we are seeing is that the car usually does not want to go that fast. Most of the time, this has been less than the speed limit for me, which is too slow.

I also found the Max Speed setting borked up when it switches to V11 on highways. I keep my speed offset at +8% and I found FSD setting +10% whenever it switched to V11 mode. Dialing it back down worked as expected, but what happened??

Driving in daylight has been sublime, aside from driving too slowly most of the time. I'm doing zero-intervention drives like it's an easy thing. There is some hard braking way ahead of a red light followed by gentle braking up to the stop. That's a bit annoying, but no deal breaker. No lane drift. Slowing for a lead car turning into a parking lot is so much better than V11. You can see that the latency has been reduced considerably. The car slows appropriately, then speeds back up exactly as it should.

I had one case where an opposing car was waiting to make an unprotected left across my car's path. My car slowed just a fraction before the car turned across my path then sped back up. It was almost imperceptible but, I swear my car sensed that the other car was going to turn in front of me and just gave it a little extra room. This is just what a human might do. V11 would have hit the brakes hard AFTER the other car had made the turn and cleared the lane.

Last night, V12 tried to steer to the left of a median into the oncoming lane. Today, in daylight, it took the correct path at this spot. More testing is required here.

 

[SidetrackedSue]

Um. “.. never using FSDb.” Yet, here you are spewing vitriolic comments about how terrible FSD is and generally implying that people should stay away from it. In a discussion thread dedicated to those who are actually using it and reporting their findings to the crowd.

There are those here who see the glass as empty; those here who see the glass as full; but in all cases the posters on both sides of the fence appear to be reporting based, generally, on actual experience. Bar some thumb-twiddling and beer bets while waiting for the next point release.

And then there’s your generally negative posts based on zero personal experience.

I kind of get it; you regret, I think, plopping down cold, hard cash for FSDb, which is certainly not at present a finished product. But why inflict that angst on us? We’re busy.

Assuming you haven't already ignored me:

1. I don't use it for city driving because it consistently fails to meet our willingness to use in on city streets but it is used for highway.

2. I hang out here (and across TMC) to figure out which updates are worth downloading.

3. I know from posts in this thread there is no point in downloading 12.3+x because we only use FSDb on highways and that's the V11 stack at this point. So we have the decision to make as to whether it is worth it to download (when available) but know that at this time, what's available doesn't change our confidence in using FSDb V11 on highways since, essentially, the same known flaws are still there. Nothing new to learn.

4. If/when highways change over to the V12.3+x stack, this thread is where I'll hear it first.

5. Currently I'm uninterested in updating from where I am due to increased nags AND have learned from this thread that any nags I get won't be able to be dismissed with the scroll wheels which is my preferred method since my hands are already on the wheels. While we don't experience the new camera-triggered AP nags with our version, we do get camera-triggered FSDb nags so are used to them and have a method that works for us to dismiss them. Why would I want to lose that capability on the highway, having learned in this thread that method won't work for us under V12.3+1x. There have been comments saying removal of scroll-wheel nag dismissal was just another inadvertent bug that will be fixed in a future release. In which case, this thread will inform me which of the V12.3+1x updates will have if fixed so I can wait for that and not downgrade my current FSDb highway driving experience by losing my ability to easily dismiss nags.

There are likely other things I've learned in this thread, most of which are more to do with attitudes towards driving reinforcing my thoughts as to why on city streets, I will never trust FSDb (at least V11 but it is to be seen if V12 is any better.)

 

[Todd Burch]

Still no HW4 cars. A couple appear to have gone out but they aren’t actively rolling out at all.

 

[powertoold]

I wonder if Tesla trains videos specific to each car model.

We have some evidence that each model is separate?
1) For Chuck's turn, he saw all kinds of Tesla models.
2) During the Elon livestream, Ashok mentioned some issue specific to Model S.

 

[DanCar]

Drove it around a few miles in my neighborhood. Fixes some issues , has some new issues. Very subjectively I would say it is 5% better than v11.
The bad:
1. Drives slow
2. Stops in new places, but doesn't stop were v11 did stop for no reason.
3. Still jerky, but better compared to v11. I have high expectations for a smooth ride.
4. Accelerates towards red lights. Same as V11.
5. I don't like the aggressive acceleration from stop. Slightly worse than v11?
6. Encountered a pedestrian jay walking. Got too close for my comfort. I would have moved all the way to the left. Instead v12 hung around the middle of the road to avoid the pedestrian that was crossing from left to right.

The good:
1. Bright shiny new toy.
2. I have more confidence in this one, but maybe I shouldn't.
3. Less jerky than v11.
4. Seems faster at stop signs, but still way too slow. Yes I do rolling stops when no one is around.
5. Seems to have an autopark feature. Not useful to me, but seems cool. I told it twice to go home. Wants to park 100 meters from my house, instead of in my garage.

Opinion is the same as V11, I doubt I'll use this version much. Looking forward to V12.4.

 

[Tronguy]

Assuming you haven't already ignored me:

1. I don't use it for city driving because it consistently fails to meet our willingness to use in on city streets but it is used for highway.

2. I hang out here (and across TMC) to figure out which updates are worth downloading.

3. I know from posts in this thread there is no point in downloading 12.3+x because we only use FSDb on highways and that's the V11 stack at this point. So we have the decision to make as to whether it is worth it to download (when available) but know that at this time, what's available doesn't change our confidence in using FSDb V11 on highways since, essentially, the same known flaws are still there. Nothing new to learn.

4. If/when highways change over to the V12.3+x stack, this thread is where I'll hear it first.

5. Currently I'm uninterested in updating from where I am due to increased nags AND have learned from this thread that any nags I get won't be able to be dismissed with the scroll wheels which is my preferred method since my hands are already on the wheels. While we don't experience the new camera-triggered AP nags with our version, we do get camera-triggered FSDb nags so are used to them and have a method that works for us to dismiss them. Why would I want to lose that capability on the highway, having learned in this thread that method won't work for us under V12.3+1x. There have been comments saying removal of scroll-wheel nag dismissal was just another inadvertent bug that will be fixed in a future release. In which case, this thread will inform me which of the V12.3+1x updates will have if fixed so I can wait for that and not downgrade my current FSDb highway driving experience by losing my ability to easily dismiss nags.

There are likely other things I've learned in this thread, most of which are more to do with attitudes towards driving reinforcing my thoughts as to why on city streets, I will never trust FSDb (at least V11 but it is to be seen if V12 is any better.)

You're not on ignore.

I got zero objections about scanning around for people's responses. Heck, I do that blinking all the time. And, when actually using FSD-b, both the complaints and kudos taken in from monitoring these FSD threads are invaluable in Reducing The Amount Of Surprise. Which is a survival-positive thing; people have told tales around campfires and the like since we got speech (and maybe before), the better to reduce the chances of being eaten, eating the wrong thing, or generally Staying Away From That Sudden Cliff. Storytelling about hazardous things is part of what makes us human, and, like I said, is a survival positive thing.

A little imagination in these stories is a good thing; if X caused Y, then XX causing YY isn't a bad thing to talk about. Or to figure out root causes and all that.

But.. there's this old-time nursery story that, I think, every kid in every culture gets introduced to early on: "The Boy Who Called Wolf." It didn't end well for the Boy. And the story is out there because while Calling Wolf is an attention-getting tactic that gets people running around like mad, once people realize that the person in question is doing an attention getting tactic, then people will ignore that person. Even if, later, said person actually does speak the truth.

You've stated that you believe FSD-b to be unusable; I believe you've stated why. You've also stated that you don't like the development churn that's going on with Tesla. You've been answered by people who have been keeping up with new releases, trying out the new UI's, and have generally been stating, "It's not all that bad. We're going to keep on chugging." As near as I can tell, your response has been to double down.

Which is good clean fun on the internet. I suppose. But the crowd in here generally has been keeping up, has been testing the new releases, and have been reporting results. So.. what are you adding?

My last post on the topic.

 

[EVNow]

I have the impression that they twisted some parameters to make the car behaves a little bit differently, not always retraining FSD. For example: v12.3 is different than v12.2.1: in v12.3 the car is closer to the curb in most situations, not in the center of lane lane like v12.2.1. Did they retrain to have this new behavior?

No parameters to tweak in end to end NN training. That is the main issue with NN.

Think of how you would correct a child - except you tell a child a couple of times to correct but with NN its a few thousand times

 

[Dennisis]

My first drive was zero intervention, zero peddle push… perfect. I had a U-turn and had a message, please assist unable to complete maneuver, but it did it like a pro

My U-turn was the same; displayed the message but did it anyway, very nice...

Edit: For me nags have been non-existent so far. In ver. 11 if I looked at the screen for a moment I'd get a warning - was hyper vigilant. Anyone else noticed these are way down? To be clear, I always keep my hands on the wheel and stay alert, got warnings when I'd look to do something on the screen for more than a few seconds but the same behaviors haven't triggered any yet.