FSD v12.x (end to end AI)

[LSDTester#1]

Yep. It looked like the status quo. Max set speed appeared to still be gorked. Indecisive at intersection entry and left turn lane selections. And video sections with heavier traffic were edited out.

We really should be calling it v11.

I see @David Wang post and all I think about is Momofuku

 

[LSDTester#1]

I keep refreshing and it's just a sea of 12.3.x

View attachment 1050472

A watched pot

 

[Yelobird]

Stu doesn't like you... I don't like you either..
We are wanted men and have a death sentence in 12 systems....

(Geeks like me will get that reference)


( Most of us for years have been trying to figure out what Stu likes and doesn't like to no avail... it's pot luck. Ask @Ramphex !! )

Hey! Don’t say the “R” word! His mantra,

“I’m like a dog chasing cars, I wouldn’t know what to do if I caught one, you know, I just do…things.”​

 

[Yelobird]

Dude I'm thinking 1 every 10 years is something I want to avoid. You have my sympathy if you are doing annuals.

#FSDTester#1 likes to get a second opinion every visit. Likes to play it safe.

 

[zoomer0056]

I use the yoke top bar as a reference. Top edge down for "right" button. Lower edge up for "left" button. Like a reverse stalk.

Jeeze, that makes it confusing, negative logic . Top down right, bottom up left. And there's still a stalk in the formula.

But if that works then it's ok.

 

[AlanSubie4Life]

Supposed 12.4 leak from an employee:

Well if so they have not addressed speed control. (9:55, 11:15). Or decision wobble (11:40), (16:45). That dog won’t hunt.

Disappointing. IF this is actually 12.4: Looks like it is going to be a few more years.

No rapid progress.

Hopefully 12.4 will be better. And we should ignore 12.5, since that will be about the same, as we know.

SS, DD.

 

[aronth5]

I’m curious if AP is better. And hopefully v12 could do better but does not seem likely, since it clearly seems to have some of the same issues.

Maybe I didn't read your post accurately but no reason to think V12 for the highway won't be a lot better than v11 and AP. When v11 came out there were lots of posts predicting regressions and while there were a few by and large the common feedback was how much improved the v11 highway stack was. I am expecting the same when v12.5 combines highway and city/street stacks. The improvement should also be reflected in Tesla's accident stats.

 

[aronth5]

Well if so they have not addressed speed control. (9:55, 11:15). Or decision wobble (11:40).

Disappointing. IF this is actually 12.4: Looks like it is going to be a few more years.

No rapid progress.

Hopefully 12.4 will be better. And we should ignore 12.5, since that will be about the same, as we know.

To judge v12.4 after 1 video is silly. You should know better. We have no idea how reliable the post is to begin with and we have no idea if the version is one the public will even see. Patience Grasshopper.

 

[Mardak]

The warning "Please pay attention to the road" pops up at 1:05 and again at 1:53, but it goes away without him touching the steering wheel

From 13:37 leaving the parking lot and immediate adjusting the set speed, it seems like there was no nag for 5 minutes before getting onto the freeway. This included waiting at traffic lights where pre-vision monitoring would be especially naggy when resuming from a red light.

It seems like Vision-Based Attention Monitoring might also be in effect for freeways at 18:41 (still using 11.x) as even though the driver does touch the steering wheel, the message actually went away just before touching. Hopefully it won't take long to get rid of old habits to satisfy nag.

It seems like there's still the immediate small red hand "Please pay attention to the road" such as when Autopilot detects something unusual perhaps the construction signage at 10:49, but it seems to also go away just before the driver touches the steering wheel.

 

[AlanSubie4Life]

To judge v12.4 after 1 video is silly. You should know better.

As you know, I am pretty familiar with this stuff. We know how it goes.

We have no idea how reliable the post is to begin with

True, as I made clear.

have no idea if the version is one the public will even see

We know it won’t be. But how is that relevant? They’re all the same basically - we know that!

12.3.x, 12.4, 12.5 - they just seem to be feature adds for the most part.

The improvement should also be reflected in Tesla's accident stats.

Already seems to be improving, in relative terms (obviously cannot compare to unassisted vehicles).

 

[FSDtester#1]

I keep refreshing and it's just a sea of 12.3.x

View attachment 1050472

I read an article or 2 that Tesla is hell bent on getting everyone on 12.3.6.
Teslafi is apparently showing this to be the case.
Unfortunately that means they will likely finish that project before a release of 12.4
All the added people in the 2024.3.25 que over the last few days is disconcerting.
They are probably lining up 3.25 to be the one going to 12.4, but won't do it until they all come over to it and everyone is off V11.

 

[sleepydoc]

I can’t believe you made me give you a thumbs down. Colonoscopies are magical!1!! Granted the cleansing takes patience, but the drugs are heavenly and the encouraged methane fart fest after is absolutely glorious. Nowhere else in this world can a man carry on a conversation while simultaneously spewing out audible metric sht tons of air from his sphincter, and have the person you’re conversations with be totally normal like nothing is happening.

I’d do it weekly if they let me

Well, after a colonoscopy they’re actually either CO2 or air insufflated from the scope.

In all seriousness, I’ve seen *many* cases of people getting their colon removed (or worse) for something that could have been diagnosed and treated easily with a routine colonoscopy. The prep isn’t fun but it beats cancer, so please, just get it done!

 

[aronth5]

To judge v12.4 after 1 video is silly. You should know better. We have no idea how reliable the post is to begin with and we have no idea if the version is one the public will even see. Patience Grasshopper.

As you know, I am pretty familiar with this stuff. We know how it goes.

True, as I made clear.

We know it won’t be. But how is that relevant? They’re all the same basically - we know that!

12.3.x, 12.4, 12.5 - they just seem to be feature adds for the most part.

Already seems to be improving, in relative terms (obviously cannot compare to unassisted vehicles).

Your reply just confirmed that judging v12.4 on one video was silly but not surprising.

 

[LSDTester#1]

Your reply just confirmed that judging v12.4 on one video was silly but not surprising.

I think TMC needs a different icon than the borderline racist black thumbs down for disagreeing with a post.

I would respond with so many more icons to Alan’s posts if they had say something like,

… or or or or and this one is the clear winner

 

[gsmith123]

but no reason to think V12 for the highway won't be a lot better than v11 and AP.

Disagree. In my experience V12 has some annoying traits in "near-highway" situations. For instance on a undivided rural two lane road with speed limits in the 45-55 mph range, or even some divided highways with 55mph limits where V12 stays engaged.

It can't hold consistent following distances, especially when a lead car is driving unusually slow (big trucks, for instance or tourists driving slow to take in the sights). V11 performs way better.

It can't consistently stay centered in the lane and unnecessarily drifts right up against the lane lines (for no reason). That said, it does offset in lane well at times (like with adjacent vehicles or in some corners where it's comfortable to have a bit more space.

Overall driving V12 on easy highway driving is not chill for me.

 

[AlanSubie4Life]

I wonder if end2end has any ability to make plans and stick to them (unless other circumstances dictate they can't be stuck to) or whether that is an inherent weakness of the approach? Very interesting. I wonder how many years it will take to develop the right approach which just takes care of this inherently?

 

[gsmith123]

I wonder if end2end has any ability to make plans and stick to them

The consistent lane wobble (typically at turn lane entry) would indicate sticking to a plan is certainly tricky.

I'm not an ML expert and obviously don't know the details of how V12 works, but speculating wildly:

Sticking to a plan means making decisions right now that are consistent with decisions made a few seconds ago. I'd think one way the AI could do that is by inferring what the "plan" is based on current position, speed, steering angle, acceleration, and obviously the whole environment around the car.

For instance, if you're part way through left turn on green arrow it's easy to infer the plan is to complete the turn into the target lane.

This isn't always so easy though. Consider a yellow light at the "dilemma zone" where you either have to hard brake or commit to running the yellow. There's a brief period where both options are roughly tied in terms of being the most probable correct decision. And it turns out we do in fact see V12 being indecisive and wobbling between plans!

Another way the AI could stick to a plan is by having some knowledge of history as an input to its decision making right now. In the yellow light example that means if the car started hard braking 300ms ago then it would remember that and be better able to infert the correct next action is to continue the plan to stop.

 

[JulienW]

Dude, @JulienW is 8hrs away from us

So you'll are on the other side of town?

 

[fasteddie7]

Everyone is a critic and I'm proud to be an everyone.

• Wish they had got out of the car and showed some ASS

I didn’t think actual smart summon was part of 12.4. For some reason I assumed it was 12.5

 

[Dave EV]

Sticking to a plan means making decisions right now that are consistent with decisions made a few seconds ago. I'd think one way the AI could do that is by inferring what the "plan" is based on current position, speed, steering angle, acceleration, and obviously the whole environment around the car.

I don't think this is that hard to do, if they want. You simply need to make sure to include recent data into the dataset that's being used to train the neural net.

The drawback is that the more data you include, the longer the training takes. There's a point of diminishing returns - stuff that happened a minute ago is a lot less relevant that events that happened 1 or even 10 seconds ago.