Are you referring to Ashok Elluswamy's CVPR keynote from June last year - Foundation Models for Autonomy? Tesla has been generating video trained on their fleet with 8 cameras for a while now. Presumably the general world model understanding how to predict how things work like vehicle physics and how drivers behave around traffic controls to accurately predict video should allow for end-to-end control.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
-
Tesla's Supercharger Team was recently laid off. We discuss what this means for the company on today's TMC Podcast streaming live at 1PM PDT. You can watch on X or on YouTube where you can participate in the live chat.
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
Knightshade
Well-Known Member
I though I saw a post referencing the switch over, but perhaps I was confused.
Peeps in this thread will have better knowings.
Question is, is FSD Beta running V11 on highway and V12 in city, or is it all already V12?
My astonishment comes from the fact that Tesla can spend years on developing multiple neural networks and data structures to feed their heuristic system, then wholesale replace it with a monolithic network in seemingly short-order. It probably speaks volumes about how kludgy the heuristic system ultimately was, and how well-fitted to the problem the monolith is.
Does anyone know how recently the techniques that went into this were developed? I know that 2023 was a huge year for machine learning. Is V12 dependent on anything that recent or was V12 just the result of a long learning process?
It appears that highway runs V11 and secondary roads run V12. Part of the V11 system appears to also be running on secondary roads in order to provide the visualization and perhaps other functions.
Context for this discussion:
I single stack for both introduced a year ago in v11.x:
And I assumed v12 was based on that but there's some thought that v12 may have been a parallel effort that switches back to dual stack in some scenarios.
PianoAl
Active Member
AI Driver in his video says that the auto wipers are worse, but watching it, they seem exactly the same as in V11.
The driver definitely felt uncomfortable, and perhaps it's a smaller gap than many would take. I tried to look for the oncoming car in the visualization during the turn, but at a glance, the stream seems to be too low resolution along with steering wheel blocking the view.
Here's another livestream from this morning, but I haven't watched either.
Pretty sure 12.x is still running 11.x on freeways. Look for AUTO MAX set speed, 11.x control messages ("changing lanes…"), blue path differences when it switches.
Similar to 10.x with FSD Beta on city streets and NoA on freeways, 12.x now has end-to-end on city streets and old FSD Beta 11.x on freeways.
It is my impression that the advent of end-to-end V12 is more about computing progress and less about AI technique. We've always known that end-to-end was probably going to happen, but we didn't have the training power to make it practical.
Are these railroad tracks like making a turn across Embarcadero or tracks that are raised from the main road crossing the ballast mound? Some railroad crossings in low traffic areas don't even have a barrier and can be really steep.
scottf200
Well-Known Member
Some people are posting about v12 in a 12 only thread and others are just continuing the conversation in a v11 thread. Can have two groups of people not necessarily following both threads (ie. I wasn't).
The change of "AI end-to-end" is all the improvements videos of it is not a major "new feature"?!?! I see speed control, speed bump adjustments, parking lot improvements, pedestrian improvements, etc. Seems really odd to me like it is being downplayed intentionally.
Last edited:
I'm personally waiting for wider rollout and more comparisons, but it definitely looks like this is almost ready for wide release as a Level 2 ADAS
Watching AI DRIVR's full video makes me wonder if they've somehow biased the system strongly towards following and mirroring lead cars, with the way the system behaves on open roads versus in traffic
Last edited:
Donut behavior is emerging now, V13 will learn how to leave skid marks
scottf200
Well-Known Member
Similar threads
