It would be instructive to see numbers of cars on v12 in CA vs the rest of the country. I suspect that would give an alternative reason for Chuck and other YouTube stars outside CA being left out.
-
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
EVNow
Well-Known Member
Thats possible. IIRC the way Elon explained it (in one or the other interview) was that really only the planner/controller were hand coded in V11 and when that was converted to NN, it became end-to-end NN.
Ofcourse they could have moved on from there and done end-to-end relearning. Or even optimized NN to eliminate some layers etc. But that is always a double edge sword - it gives you less ability to fix individual problems.
As bad as I want v12 I sure as HELL don't want to go back to the dry wipes. My a$$ was so chapped and just now healing up.
This is NOT true.
Appellate court: Minnesota drivers don’t have to use innermost lane when turning left onto two-lane road
The court says Minnesota's laws do not require drivers to use the innermost lane when turning left onto a two-lane road from a single lane.
EVNow
Well-Known Member
It is a delicate dance - between companies and influencers. They need each other - but companies want to control the narrative. Influencers want to be as independent as possible for credibility. It is always a delicate dance.Poor Chuck may be the last to get v12 (or at least just one before me). Kingdom rules: Never speak ill of the King or there will be hell to pay.
That is with professionally run companies - but when run by a king, it gets worse
mgs333
Member
Especially with an insecure, petty king. Oh crap, now I'm at the back of the queue as well.Poor Chuck may be the last to get v12 (or at least just one before me). Kingdom rules: Never speak ill of the King or there will be hell to pay.
EVNow
Well-Known Member
What a poorly worded statute - its not like there are other rules that say you have to drive on the right side !
That's an interesting possibility, although how often does 12.x try to go around stopped traffic when there's multiple lanes for your direction? Neural networks pick up signals and share signals in unexpected ways, so it is still possible either way.
I also noticed the turn signal was already engaged for the upcoming left turn, so maybe that also resulted in end-to-end thinking it needed to get even more left. Have those with 12.x played around with manually engaging the turn signal to see if that influences behaviors (other than the usual expected changing of lanes) such as left turn signal when already next to double yellow lines making it think it needs to turn or switch towards oncoming traffic like in this scenario?
Yep. For FSD, less may not be more. Aside from those purchasing future vehicles with HW5 etc there might not be much else to hope for.
powertoold
Active Member
After the initial hype of V12, reality is starting to set in. It doesn't seem like HW3's inference compute will be able to generalize all the edge cases of day to day driving.
There are just too many, and the NNs don't seem to understand broad concepts yet. The more I use V12.2.1, the more my initial fears about this approach are realized. V11.4.9 was just overall more consistent, reliable, and dependable. V12.2.1 acts like a stubborn rebellious child, and the crazy wipers don't help either. That's why I told y'all's that the wipers NN is compute limited, and it was only good with 11.4.9 because Tesla introduced a more efficient video module. It doesn't make sense that the wipers worked well for me on ~10.69 then sucked again until 11.4.9 and then sucked again with V12.
There's just too much riding on 12.3, but we already went from 12 --> 12.2, so unless Tesla has some new approach or vastly better training resources for 12.3, I'm not sure if there's much hope left for HW3.
We'll just have to wait and see
There are just too many, and the NNs don't seem to understand broad concepts yet. The more I use V12.2.1, the more my initial fears about this approach are realized. V11.4.9 was just overall more consistent, reliable, and dependable. V12.2.1 acts like a stubborn rebellious child, and the crazy wipers don't help either. That's why I told y'all's that the wipers NN is compute limited, and it was only good with 11.4.9 because Tesla introduced a more efficient video module. It doesn't make sense that the wipers worked well for me on ~10.69 then sucked again until 11.4.9 and then sucked again with V12.
There's just too much riding on 12.3, but we already went from 12 --> 12.2, so unless Tesla has some new approach or vastly better training resources for 12.3, I'm not sure if there's much hope left for HW3.
We'll just have to wait and see
It's a bummer. I was hoping there would be some v12.3 release rumors this weekend.
powertoold
Active Member
Tesla FSD V11: 1 billion parameters
GPT4: ~1.7 trillion parameters
Tesla FSD V12: probably less than 2 billion parameters?
GPT4: ~1.7 trillion parameters
Tesla FSD V12: probably less than 2 billion parameters?
Why wouldn't V12 have the same number of parameters as V11? It's the same sensor environment, right?
powertoold
Active Member
I think end to end should result in more parameters and fewer NNs in the architecture (more efficient), so you would end up being able to process more parameters with the same compute, but then again, I'm just speculating.
sleepydoc
Well-Known Member
This is NOT true.
Appellate court: Minnesota drivers don’t have to use innermost lane when turning left onto two-lane road
The court says Minnesota's laws do not require drivers to use the innermost lane when turning left onto a two-lane road from a single lane.www.fox9.com
I read ‘nearest the centerline’ as applying to the destination lane, but rereading the statute it appears it only applies to the originating lane.
I agree - it’s poorly worded. It also doesn’t address the scenario of multiple turn lanes. (Unless it does in a later section I didn’t read)
Given that it seems like 11.x neural networks are still running for visualization and not control, there should be opportunity to shift inference compute from 11.x to 12.x. However, there's a complication of maybe each are running on their own SoC, so 12.x exceeding the capacity of one introduces need for transmitting/receiving data between the two.
But 12.x neural networks even staying at the same size and compute requirement can learn broad concepts it currently doesn't understand. For example, some signal that is below an activation threshold for a concept can be activated in a later trained network that boosts the signal and/or lowers the threshold.
What specific broad concepts do you think are lacking in 12.2.1 to see if it improves in later 12.x?
powertoold
Active Member
When I assess an approach, I try to intuit any possible fatal flaws.
The end-to-end approach with video training and my experience with 12.2.1's decision wobble has me concerned:
1) In the case of the decision wobble, V12's dataset has videos of committing to some x maneuver, but it also has videos where it is behaving differently (like braking) in a similar situation. The biggest challenge and question for the current NN approach is, "can the NN conceptualize the purpose of a maneuver and collapse its decision tree in the case of a gray area in decision making."
In that video example you gave of the car parking in Dennys, you can see that right before turning into the spot, the path planner had some milliseconds of turning right. Because the decision of parking was clearer, that plan disappeared as the car inched forward.
There's actually a lot of these gray area decisions, especially in a parking lot. Say there 2 possible paths to a pin in the parking lot but there's a small island in the way. You can turn before the island or after the island, 12.2.1 seems to get tripped up with these decisions.
2) There's still many instances of freezing at stop signs / unprotected turns with 12.2.1. This is something we saw during Elon's livestream and persists today. What sort of miracle would be needed to fix this? That's what I wonder.
3) I think these problems can only be reduced with a higher parameter NN, and HW3 is already limited, so HW3 seems like a dead end for V12. Elon already mentioned that more training compute is needed to reduce inference compute, so I guess there's some headroom, but I feel like there needs to be at least 3-5x higher parameter count for V12 to work well on HW3.
EVNow
Well-Known Member
LOL. Interestingly both can be true - v12 could drive better than Waymo in certain circumstances and worse than even v11 in others.
Similar threads
- Article
