-
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
That's how NN training works in general. Keep monitoring the loss function, test against separate data. Update training data. Repeat.
Tesla has FSD modules running NNs in virtual and replay environments as part of the validation.
Can even create a virtual city populated with cars operating off the networks. Or GANs (NN vs NN).
Generative adversarial network - Wikipedia
You need training data and maintain some appropriate balance of that to learn a variety of behaviors. Previously with coded logic, you could be more sure that the car gives 3 feet for bicyclists, but now as neural networks, if this specific behavior isn't a dedicated set of training data, it will "just" learn what humans typically do (assuming training data is from human examples).
But yes there's potential for the data engine to automatically find disengagements from the fleet to autolabel and continue training indefinitely.
Knightshade
Well-Known Member
Which makes more sense?
Over 500 miles they averaged one intervention every less-than-3 miles--- thus they got a lot of SHORT clips over 4 days and spliced them into one video without telling anyone about that (and they needed 4 days because they didn't have enough footage after 2 or 3- Elon specifically announced pushing back the video release as mentioned in the story)
or
Over 500 miles it repeatedly failed hilariously often (177 interventions in 520 total miles), but then somehow MAGICALLY WORKED PERFECT THE WHOLE TIME ONCE?
(Hint it's not that second one)
willow_hiller
Well-Known Member
Last we heard, v11.4.7 is in the works. Tesla FSD Beta: Musk Says Beta 11.4.7 Goes Out This Week; Hypes FSD V12
It's not a matter of which makes more sense to you, it's a matter of what actually happened and what data we have to determine that. The original claim I responded to was: "I mean we know Tesla took 500+ miles of footage over dozens of takes and edited it into a single short video they presented as a single take for the 2016 FSD demo."
How do we *know*? I see some things you are making inferences from, but that's not the same thing.
As to interventions per mile. That's what one could expect when tuning software for a specific route. Run route, find errors, tweak run route, find less errors, tweak... run route, no errors. It's a variation of preHD mapped system other companies use for robotaxis. Heck, maybe they did do it successfully more than once, but that isn't necessary.
Anyway, this is all tangential to the thread.
Knightshade
Well-Known Member
Which is almost always the simpler explaination.
And in this case the two are very very very far apart on simplicity and likelihood.
Because I find the idea that it would fail, over and over again, nearly 200 failures in a few days, and then MAGICALLY THE VERY LAST TIME work flawlessly for the entire drive to be unlikely to the point of insanity.
Your suggestion they can just "fix" a whole route in a couple days in 2016 also seems counter to the fact they STILL haven't fully fixed Chucks UPL months after they had a dedicated team in his specific spot testing for weeks on a much more advanced version of the HW and SW.
On top of that, if it DID work flawlessly why did they remove about half the route from the video with their edits?
Because the actual route is about 2x the length they included in the vid, and cut to avoid a number of trickier areas throughout the video. Why would you do that on a flawless single drive over the whole route? I mean you were speeding it up a ton anyway so it wasn't just for brevity.
It was because they couldn't even get usable footage on those parts after 500+ miles of trying.
That's someone in 2023 driving the actual full route- he points out a bunch of (tougher) spots actually on the route Tesla just edited out/around entirely in the 2016 vid.
Last edited:
Yep, just like my keys are ALWAYS IN THE VERY LAST PLACE I LOOK. (Because I stop looking once I've found them, just like Tesla would stop running the route once they got a clean run.)
Knightshade
Well-Known Member
I'm not saying that's the least relevant analogy I've ever read on the internet.
But it's at least going to be honored to be nominated!
Wait, I just got home from a weekend trip and expected to have 11.4.x ready to download, are you saying that’s not the case?
Is there anyone we can call, maybe it’s some error.
/s
clydeiii
Member
You clearly never drove it on the very early FSDb versions...
Those versions weren’t widely available to the general public. (moderator edit)
Last edited by a moderator:
The relevance is that the older builds were much worse, falling into your "worst yet" classification. Don't sweat it; we look forward to your participation here, but I'd invite you to leave out the personal insults. They don't contribute anything but animosity, and we stay reasonably civil around here.
Last edited by a moderator:
clydeiii
Member
Just pointing out it isn’t the worst. Perhaps the worst for you. But we’ve seen much worse.
Similar threads
