Todd Burch
14-Year Member
Control code likely mostly scrapped. But the gathering of video data as well as the development of the perception/vector stack was a necessary step to get to end-to-end.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD v12.x (end to end AI)
- Thread starter Buckminster
- Start date
-
- Tags
- #fixphantombraking autopilot chill is not chill elon is an ass fix fsd extreme late deceleration fsd does not commit to go fsd does not maintain proper speed fsd hard and late braking fsd jerky driving fsd late deceleration fsd stops too early too quickly fsd stops too far behind line fsd tailgating in city streets
But but but but Papa Elon said there will be console quality gaming in my…. Flagship refresh Model S
Disclaimer: I'm not an ML engineer but I've been trying to make sense of it myself. Basically, I think the answer is that the code developed up through v11 was not at all a waste or retrospectively unnecessary.
The v12 NN is not a formless blank slate that is trained from zero on the video data; rather it has a starting-point form defined by a large number of non-zero interconnection weights, and crucially a very huge number of zero-valued weights - all of this effectively defining special-purpose localized processing sub-networks that interpret the data and make decisions. You can roughly analogize it to various cortexes of the brain, the visual and auditory etc. corresponding to perception centers, and the cerebral cortex, hippocampus for memory and spatial comprehension and so on.
Without the predecessor work developed by code and incremental ML techniques throughout the prior versions, Tesla wouldn't have the foundation networks to train. You can't just start from zero with everything connected to everything else and let it train on a giant formless network, because that has too many parameters to handle even with enormous compute resources, and it wouldn't converge to anything useful on its own. (unless the given problem - unlike driving - can be represented by a small number of well-defined rules, like the often cited Alpha Zero trained from scratch to become the world's best expert Go player).
Of course this doesn't mean that if Tesla started over again, knowing what they know now, that they would take exactly the same path to get to a trainable and capable network starting point. And it doesn't mean that they won't, sooner or later, do it over again to develop an even more capable and efficient trainable network. But for now, I believe the NN "output network" inherited from v11 is the essential basis of the trainable v12 network structure.
Below are some earlier posts I made on the same topic if you're interested, but I tried to restate and summarize the ideas in this post.
FSD v12.x (end to end AI)
HW4 has three camera imputs that are not used on the cars. On S&X, yes; on 3&Y HW4 has no unused inputs. At least that is what I thought green said.
teslamotorsclub.com
FSD v12.x (end to end AI)
It's been 24 hours since the last wave... what are the odds of another drop tonight? No idea. May as well shake a magic 8 ball. We are also past 80% install from last nights drop.
teslamotorsclub.com
FSD v12.x (end to end AI)
What video are you watching? None of what you describe is shown. Yeah even as dissatisfied as I am with FSD and the years of breathless nonsense from the Chief Twit, this UPL video looks…fine (to me).
teslamotorsclub.com
My car was in the second wave to customers for 12.3, so I'm hoping it gets 12.3.1 early as well. Of course, Tesla will do as Tesla does!
I wonder if they'll offer to give me a demo ride when I pick up my car?
There is. It's called v12.3 FSD.But but but but Papa Elon said there will be console quality gaming in my…. Flagship refresh Model S
There are infinite scenarios and levels, and no one has ever won with Game Over. You can be the first.
Yes, in both cases. With flashing lights above and below the sign.
They retrain from scratch often and you can start without any foundation model. And V11 was not needed for any foundation model, according to Ashok their SORA like video generation module is the foundational model. But I think the main point is that it's very hard to do it right from the start. Karpathy said it well, "labelling is an iterative process".Disclaimer: I'm not an ML engineer but I've been trying to make sense of it myself. Basically, I think the answer is that the code developed up through v11 was not at all a waste or retrospectively unnecessary.
The v12 NN is not a formless blank slate that is trained from zero on the video data; rather it has a starting-point form defined by a large number of non-zero interconnection weights, and crucially a very huge number of zero-valued weights - all of this effectively defining special-purpose localized processing sub-networks that interpret the data and make decisions. You can roughly analogize it to various cortexes of the brain, the visual and auditory etc. corresponding to perception centers, and the cerebral cortex, hippocampus for memory and spatial comprehension and so on.
Without the predecessor work developed by code and incremental ML techniques throughout the prior versions, Tesla wouldn't have the foundation networks to train. You can't just start from zero with everything connected to everything else and let it train on a giant formless network, because that has too many parameters to handle even with enormous compute resources, and it wouldn't converge to anything useful on its own. (unless the given problem - unlike driving - can be represented by a small number of well-defined rules, like the often cited Alpha Zero trained from scratch to become the world's best expert Go player).
Of course this doesn't mean that if Tesla started over again, knowing what they know now, that they would take exactly the same path to get to a trainable and capable network starting point. And it doesn't mean that they won't, sooner or later, do it over again to develop an even more capable and efficient trainable network. But for now, I believe the NN "output network" inherited from v11 is the essential basis of the trainable v12 network structure.
Below are some earlier posts I made on the same topic if you're interested, but I tried to restate and summarize the ideas in this post.
FSD v12.x (end to end AI)
HW4 has three camera imputs that are not used on the cars. On S&X, yes; on 3&Y HW4 has no unused inputs. At least that is what I thought green said.FSD v12.x (end to end AI)
It's been 24 hours since the last wave... what are the odds of another drop tonight? No idea. May as well shake a magic 8 ball. We are also past 80% install from last nights drop.FSD v12.x (end to end AI)
What video are you watching? None of what you describe is shown. Yeah even as dissatisfied as I am with FSD and the years of breathless nonsense from the Chief Twit, this UPL video looks…fine (to me).
Imo this is what Tesla, SpaceX and Elons companies do really well. The iterative very quickly with lots of embarrassing explosions and silly behaviours along the way until one day you have Starship, FSD V12.3.1 leaving the competition in the dust with the bears still convinced that Tesla is a joke company and Elon is a total fraud who didn't do anything except steal the company from its original founders.
Imo the main useful things from V11 is a huge dataset of complicated situations, of known failures of the previous networks and a way to classify good/bad drivers and drives. Without this doing V12 from scratch would have to do a lot more iterations before it reached V12.3.1 level performance, probably a few years of iterations.
Matches my experience. It's pretty clear V12.3 is not the release where it learns how to drive in winter. That said, the dirt road performance (which is excellent) gives me some hope that the current architecture is amenable to further training and improvement on snow conditions. Dirt and snow are somewhat similar: no road markings, lower grip, bumpy, narrow, etc
100% second being much much more cautious in snow testing.
The visualizations for us humans are from the V10-era. Active safety features (automatic emergency braking and such) are based, at least in part, on the V11 architecture building blocks for perception and are used across the fleet. That alone makes all those years not a waste!
fasteddie7
Active Member
Thank you, here’s something interesting I wonder if anyone else is seeing. In past versions, what I assume was map updates, I would get different behaviors suddenly after experiencing the same mistakes over and over, then one day it would be corrected without an installed update. I made a video my first day with 12.3, but because I used a new camera and some of it was out of focus, I decided not to edit and post it. Fast forward a few days, I took my normal test route again, and lo and behold, a couple of the issue spots drove flawlessly. I ran them again and the same result. Now here’s the kicker, I grabbed our model x and ran the route, and it made the same mistakes my model s made a week prior. What would be insane is if in a few days the model x stopped making the mistakes also. I ended up recording it again and if this happens I’ll def edit a video for it.
sleepydoc
Well-Known Member
Another day with more examples of 12.3 handling situations that 11.9 couldn't.
Finally, a general observation: There's always been variability in people's experiences but there seems to be greater variability in v12.3 than in past versions. Not sure why that is but that's what I've noticed based on reports here.
- Exiting the freeway it had to merge across 3 lanes of traffic in about 500' to make a left turn. No problem
- Reverse route it had to wait for an opening so it could turn right then merge left to turn onto the freeway. No problem.
Finally, a general observation: There's always been variability in people's experiences but there seems to be greater variability in v12.3 than in past versions. Not sure why that is but that's what I've noticed based on reports here.
Knightshade
Well-Known Member
But but but but Papa Elon said there will be console quality gaming in my…. Flagship refresh Model S
There is!
diplomat33
Average guy who loves autonomous vehicles
Yea, there it is, I’m still waiting for Tesla to ship me my unit to put into the MS
I'm now 24 hours into my FSD 12.3 experience. So far seems pretty darned good. If there is one thing I'm seeing a trend about, it's that it doesn't seem very decisive when it comes to picking a lane if there is more than one lane in a curve. Sometimes seems to waver back and forth, and often seems to pick the inside lane, even when it's not a good choice.
2/20/24 updated to 12.2.1
3/15/24 to 12.3
3/25/24 to 12.3.1
My first FSD was:
6/9/22 to 10.12.2
TeslaFi will show you the update history for any recent update. In the list of recent updated cars, click on View.
willow_hiller
Well-Known Member
Definitely a lot of my 10.x/11.x issues are fixed, if not all? Still experiencing wrong lane choices that were fine in 11.4.9 (regression).
Wobbly lane choices suck. Auto speed control seems to work better in more dense traffic conditions. I live in a rural area, and 42mph on 55mph posted roads ain’t gonna cut it. Nor does 17mph on a 40mph posted road work either. Drunk drivers have better speed control on these roads. If a cop was behind me while doing 17 in a 40, I’m getting pulled over with suspicion of being the drunkest guy in 3 zip codes. No one drives like this, no one should drive like this.
Wobbly lane choices suck. Auto speed control seems to work better in more dense traffic conditions. I live in a rural area, and 42mph on 55mph posted roads ain’t gonna cut it. Nor does 17mph on a 40mph posted road work either. Drunk drivers have better speed control on these roads. If a cop was behind me while doing 17 in a 40, I’m getting pulled over with suspicion of being the drunkest guy in 3 zip codes. No one drives like this, no one should drive like this.
That says a one month trial, not a 12 month trial like the suspected fake email says. It also says everyone, not just new purchases. It, also, excluded Model 3s.
Last edited:
PianoAl
Active Member
I'm seeing a lot of that. For example, lane drift at a location, then no lane drift there next time. I'd swear today's drive had less speed stuttering.
Is this expected?
Similar threads
- Article
