-
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
In the past, by ways of wording, many believed the quotes from Elon and others at Tesla meant that the car had FSD or a different version of FSD actually running on the car, but not actually taking control.
That's not what they mean by shadow mode. V12 isn't installed on cars operating or navigating in the background just like FSD wasn't running on all cars in the background in a simulated method.
I think Green's claim of lying is harsh, because it infers that the company and the audience are both directly stating that is occurring. The text book definition is "Shadow Mode or Shadowing is generally considered as the analysis of functions that are implemented in the car and simulate driving actions/decisions, but do not execute them." That's not how Tesla is defining it or implementing it. There is no simulated driving actions/decisions taking place. At least AK, Elon may not know, define it as triggers that collect events for them to be sent to Tesla and evaluated centrally, which we've always known. For V12 they could be using this to get specific types of driving or events for training (like UPL), but again, V12 isn't just running in the background on cars...As pointed out other times, there's barely enough compute to run V11 on HW3, much less have V12 running at the same time.
This was always the most interesting:
That said - not every car gets every campaign so any particular trigger might not be in YOUR car. but if something triggered, chances are I know the trigger (but since you don't have indication of even when the snapshot was created - you still cannot know which one).
New campaigns or events can be pushed as often as daily.
willow_hiller
Well-Known Member
Funny thing about this shadow mode discussion. In 2019, Green was correct when he said it was a series of training data collection campaigns. But in May 2022, Tesla did actually run v11 in shadow mode to test it. He said so himself:
willow_hiller
Well-Known Member
In 2022, he actually meant it was running in the background, not just data collection:
I have only had FSD version 11 disengage caused by rain once. Lots of cars pulled over onto the highway break down lane the rain was so hard.
Release note so far for 2023.44.30.11 / 12.1.1 is the same 1 sentence but new headline:
Perhaps because this is 12.1.1 instead of 12.2, there might have been some non-FSD urgent fix that similarly bumped basically all Tesla vehicles to 2023.44.30.8 even though most were already on some holiday update with 11.4.9.
FSD Beta v12.1.1
FSD Beta v12 upgrades the city-streets driving stack to a single end-to-end neural network trained on millions of video clips, replacing over 300k lines of explicit C++ code.
Perhaps because this is 12.1.1 instead of 12.2, there might have been some non-FSD urgent fix that similarly bumped basically all Tesla vehicles to 2023.44.30.8 even though most were already on some holiday update with 11.4.9.
I'm thinking the same thing. Just fixing some of the holiday bugs. I would be suprised if Tesla could cycle a new V12 FSD release in only 2-3 weeks.
30.10 was to fix some holiday bugs so I would actually expect a pretty quick turnaround on a new V12 FSD release. So far much quicker rollout for employees then V11.
Elon mentioned v12 needs more training in bad weather outside of California. The past week in the Midwest should have provided some good data for that purpose!
(Apologies that this post is somewhat forced - I wanted to see this important thread bumped as it's had no activity for 5 days and had already dropped off the first page of the subform thread index. IMO we don't really want a competing v12 main thread to appear just because this one scrolled down 
)
)How are they going to train FSD driving in bad weather with cars that have bald all season tires on them?
UkNorthampton
TSLA - 12+ startups in 1
Carefully?
Find the "best drivers" to learn from? The ones who have decent tyres.
jeewee3000
Active Member
The bottleneck for FSD in bad weather isn't the grip of the tires. It's the problem of proper perception given the dirty camera's because of rain/mud/snow. Especially the rear camera. In my case that camera is blind 50% of the time. (completely blurred visison due to rain)
I do wish "Chill" increased the follow distance. FSD sometimes follows too close for me to feel comfortable that I could override FSD quickly enough if needed. We used to be able to control the follow distance with the right scroll wheel, now it only selects Chill, normal, or aggressive.
GSP
True, grip is not the bottleneck for what FSD can eventually do. However, today's FSD software does not increase follow distance enough to insure it can stop in time on slick roads, nor limit use of the accelerator. Instead traction control will kick in, followed immediately by a "take over immediately" red hands warning.
GSP
jeewee3000
Active Member
I am sure the car will do an emergency stop if triggered. However on slick roads it takes a more distance to stop. When driving a L2 system, you have to have enough distance for *you* to stop in time.
Said another way, the L2 system can not be more aggressive than the particular human human driver, otherwise he/she will have to disengage to allow enough room for *them* to drive safely (even if FSD could safely drive with less distance).
GSP
AlanSubie4Life
Efficiency Obsessed Member
They don’t need to. Operation in low visibility conditions (etc.) will remain outside the ODD forever with current vehicles.
Does not really matter.
Similar threads
- Article
