You're right, it's not just the crew driving it. On the other hand, the goal should be to learn what works and what doesn't work and I'm not sure the current v12 driving tools are so inclined and/or have gone through the school of hard knocks. If it ain't broke don't fix it.
-
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
You're right, it's not just the crew driving it. On the other hand, the goal should be to learn what works and what doesn't work and I'm not sure the current v12 driving tools are so inclined and/or have gone through the school of hard knocks. If it ain't broke don't fix it.
powertoold
Active Member
I don't understand why people don't think it's silly that Tesla is testing one specific turn out of millions over and over so one YouTuber can show the world it works well.
I thought Tesla's whole premise is that they have fleet data and can test things by running shadow mode within the fleet.
I thought Tesla's whole premise is that they have fleet data and can test things by running shadow mode within the fleet.
Daniel in SD
(supervised)
If it makes the turn 9 times out 10 I win a beer. This is serious business.
stopcrazypp
Well-Known Member
If the overall fleet rarely attempts those turns and/or does it poorly, all the shadow mode in the world is worthless. There will always be cases where targeted training data is required.
The difference with V12 is that they can't hard code certain heuristics to handle a specific case, it has to be done via training.
The positive is that theoretically the "skill" learned on this turn theoretically can be applied to other similar turns.
As we've seen in videos V12 is just as bad at these turns as any older version. They have to try to teach it to make it...or just always make right turns until it can do a U-turn.
They've spent a ton of hours in the last few months at Chuck's turn and these are paid drivers/testers. It's obviously not just something they can get fleet data from.
I'm guessing it's because the vast majority of ULTs are the same. So if it can handle a fairly complicated ULT well, then it can handle the vast majority of ULTs. If your point is valid, then they'd have to train for every ULT, of which there are hundreds of thousands of them in the US alone.
You don't know how many other drivers they have out collecting data/testing on other situations/ULTs. (This is just the one we hear about.)
I think his previous premise was that Tesla wouldn't and didn't need to have drivers collecting data. Tesla would just get them from the fleet.
Even if that is the case, Tesla still needs to have their own drivers to test releases before they go out. Or it could be that they don't have enough "five-star Uber drivers" in the fleet sending data in that is usable for FSDb v12 training, and they need to supplement with their own "generated" data. (I know a lot of my driving is useless because I don't always come to a full stop at stop signs.)
powertoold
Active Member
Yes, but to test the same turn for days is definitely silly, if they plan to roll it out to a general audience, then the limitations and flaws would be clear in any given route with reasonable traffic. For example, that 1 hr drive in SF video was all that was needed to see what the major issues with 12.2.1 are in that particular locale.
Are they using the same build every day? Or are they testing different versions/parameters?
zoomer0056
Active Member
Yup, there's a LOT more going on than meets the eye.
But Chuck insulted the KING
and he is going to suffer. The KING will spare NO EXPENSE to humiliate anyone who dares question anything about the KING. You have been forewarned.
powertoold
Active Member
I'm not sure, but if they're testing V12, then I don't think there's any parameters / heuristics to tweak vs V11
But then again, we don't know what the actual architecture of V12 is
stopcrazypp
Well-Known Member
Seeing the issue is only a small percentage of the problem. The "solution" with V12 seems to be to provide training data that addresses the issue. And one way to do that is sending drivers out to do the drive the "correct way". It's not like they can go back and tweak things by hand as in previous releases. Plus as mentioned, any fixes need to be evaluated.
Correct me if I'm wrong, but doesn't Tesla release code updates in stages:
1) Internal testing with engineers
2) Released to employees
3) Released wider to employees and some OG testers and influencers
4) Released wider to a batch of additional testers (some argue this represents the OG 60K group - that needed safety scores to get in)
5) Released to the fleet
Each stage of release they monitor and frequently halt the rollout to the next group if they find issues. They correct/fix/adjust and then re-release it from the top of the tree, monitoring again. If it's working well, they continue down the tree.
powertoold
Active Member
Sending out / hiring drivers to produce V12 training data is not scalable and negates the whole point of a large fleet.
This is correct.
Recently dedicated testers were hired around summer of last year and this latest release went out to a random group for step 3, but still looks to be following the same progression.
It's not scalable for all scenarios, but for difficult ones, that's what they are doing. They also have simulations as well.
Similar threads
