Not in this case, as FSDb V12.2 is currently limited to HW3 and Omar has HW4.Think the "special group" is Omar.
-
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 12.4.x is coming autopilot chill is not chill elon is an ass elon rocks fix fsd extreme late deceleration fsd does not commit to go fsd does not maintain proper speed fsd doesn't read signs properly fsd hard and late braking fsd has decision wobble fsd jerky driving fsd late deceleration fsd sometimes may not detect train crossings fsd stops too early too quickly fsd stops too far behind line fsd stops when there is zero visibility fsd tailgating in city streets fsd very slow at stops fsd will never enable l3 with hw4
Last edited:
In addition to Tesla employees another 100 owners have V12 according to TeslaScope and they have the best track record of anybody when it comes to Tesla.Think the "special group" is Omar.
Todd Burch
14-Year Member
We should see some mentions on X any moment now if that's the case.
AlanSubie4Life
Efficiency Obsessed Member
He’s been naughty, since he was not realizing the existential risks to civilization, bless his heart, so he may well be excluded.Hopefully Chuck is part of the special groups.
Has it been 2 weeks yet?
I've been mentally tracking interventions and even minor oddness of my own 11.x city street miles ahead of experiencing 12.x. My number has been around 0.2 miles/instance, but this is more for my own reference as other people will count or not count certain things. This includes unnecessary hesitation where the car coasts down 1mph slower or could have made a slip lane yield smoother as well as two counts for my adjusting set speed down and back up for school zones.
I'm expecting the first 12.x I get to exceed 1.0 miles/instance for my tracking methodology.
That's a fair criteria. Any disruption to normal traffic flow, unnatural behaviors, dangerous maneuvers, breaking the law, poor decisions are interventions/disengagements in my book.
I'd bet my bottom dollar that the 100 owners are special cases like close family, high investors and/or friends.
AlanSubie4Life
Efficiency Obsessed Member
Not great news for me. Looks like they are going to try to nail the first 9 and have a “success” on Chuck’s turn for the first time. Would make a v12 believer out of me - though that one from Whole Mars… (nothing quite like proceeding with tons of margin of 5-6 seconds, but instead doddering into the road at 8mph while attempting to get sideswiped by oncoming traffic in the furthest lane) I hope Chuck is paying super close attention and not expecting the same algorithmic approach. It seems like it could bolt at any time and assume vehicles are not changing lanes.
Interesting though - wonder if they have the simulator up and running for this? Not sure how else to do it.
(nothing quite like proceeding with tons of margin of 5-6 seconds, but instead doddering into the road at 8mph while attempting to get sideswiped by oncoming traffic in the furthest lane) I hope Chuck is paying super close attention and not expecting the same algorithmic approach. It seems like it could bolt at any time and assume vehicles are not changing lanes.Interesting though - wonder if they have the simulator up and running for this? Not sure how else to do it.
Last edited:
Omar is not on the latest build since he has HW4 so one has to assume Chuck would be trying his infamous UPL on a newer build.Not great news for me. Looks like they are going to try to nail the first 9 and have a “success” on Chuck’s turn for the first time. Would make a v12 believer out of me - though that one from Whole Mars…
Interesting though - wonder if they have the simulator up and running for this? Not sure how else to do it.
View attachment 1017073
AlanSubie4Life
Efficiency Obsessed Member
Yep it is probably fixed now. Just retrained and now fixed.
Are you suggesting 12.x neural networks wouldn't be able to drive like that due to architectural or technical limitations? Or that even if a shared neural network can drive with many styles to match behaviors around the world, it might be biased to end up driving like it was in California? Or more that any vehicle more strictly following the law will have trouble fitting in? You also seem to be suggesting that the driver/vehicle needs to communicate with others more explicitly that might require different hardware?
OxBrew
Active Member
It's both technical and lack of sensors. Even with 360 view, the Tesla driver couldn't keep pace with the traffic flow, within literally 6 inches of other vehicles and pedestrians on every side, having to modulate speed and predict about 20 or 30 other trajectories in real time, all of them intertwining and affecting each other, and all negotiating and watching others negotiate, know when to piggy back on someone in front creating showdows in traffic you can duck into, and when to yield to "illegal" cross traffic that now has the right of way only because they won that chess move, and your're actually blocked.
It's very much like high density foot traffic, on a NY sidewalk, or the start of a 10k foot race, except everyone's on scooter, tuk tuks (Autos) some cars and busses, and all kinds of cargo trucks.
No rules, but everyone figures it out. You need to learn the language of interaction, which combines hand gestures, constant honking (I'm here!), road position and maneuvers as words in the language expressing your next intent.
The only way Tesla could navigate it that I can see, would be like the old guy pushing the tomato cart slowly up the on-ramp to the flyover, and all traffic going around him like a boulder in a river. You could literally walk there faster than that.
To be clear, I'm not suggesting the NN could not learn to drive in that, it might, with enough cameras (more than it has, needs near perfect 360 views). I'm suggesting that Teslas can't even pull into a tight garage on smart summon without side swiping something. No way could Tesla control the car to that level of precision relative to surroundings, literally inches away at 25 to 35 mph.
Indian drivers have amazing situational awareness at all times, in every direction. That's what's needed.
Last edited:
mgs333
Member
A jar of honey? What an odd gift to give to a stranger. Maybe Chuck grows bees in his backyard?Not great news for me. Looks like they are going to try to nail the first 9 and have a “success” on Chuck’s turn for the first time. Would make a v12 believer out of me - though that one from Whole Mars…
Interesting though - wonder if they have the simulator up and running for this? Not sure how else to do it.
View attachment 1017073
He’s a bee keeper.
Yes no joke, actually Chuck is a dedicated beekeeper, and he does beekeeping videos in addition to his FSD videos:
Regarding the employee UPL testers refusing to interact with him, that's been going on for a long time and I think has nothing to do with recent Elon disagreements on X. It's easy to imagine that if they did get into friendly discussions during the testing, it would foment lots of "Shill" criticism, beyond what's already out there.
powertoold
Active Member
That's funny because they literally texted and called Omar recently to give him V12 with no strings attached.
I think the Tesla AI team tries to avoid communicating with end-users in general. It has nothing to do with shilling. Remember when we sent them emails about our issues? Never a response. I was telling people that it was futile.
Artful Dodger
"Neko no me"
I think that is signifcant, but not because 'Omar', but because he has a HW4 build of v12 which is supposedly lagging by at least 6 mths behind HD3. It's difficult to tell how that translates in terms of driving abilities until we see some videos from HD3 testers (OG or not). Of course, the real test will come in a general rollout, which I do expect by the end of Q1 (pending positive results to that point). Tesla seems to be approaching this very cautiously.
Even more interesting will be the rate of change going forward. With the 10K cluster of H-100 GPUs added to training resources, how fast can they iterate builds now? And when will "Std Autopilot" be ported over to the FSD engine? That's a sea change which will result in a tsunami of 'edge-case' sample data, potentially with a dramatic increase the pace in improvements, so long as training resources can cope with the flood.
It's just the virtuous cycle, in a malström of machine learning...
Cheers!
