I’m going to start acting demented and wander aimlessly in S.F. Asking people when V12 is coming out like a junkie.
-
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 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 tailgating in city streets
I totally agree with the problem and the solution too. It should use the FSD power of the cars that are plugged in and charging. In addition, limit the drain to 10% or less, as long as your battery is charged over 50%. Every available computing source especially the ones Tesla can reach out to easily should be reached out.
Sorry guys but this is not a practical solution. The main problem with the training phase is not performing the calculation, it is moving the right data into the pipeline for each of the processors so that many operations can be executed in parallel. This is discussed in the Dojo presentations. In addition to this HW3/HW4 are built for inference, not for training. This is why Tesla created Dojo rather than just putting a couple of hundred thousand HW4 systems into a data centre.
My head hurts from reading all the fan-fiction in this thread. Use the customers' the cars for training is the dumbest idea.
Also, Dojo is dead. DOA. Or did it even arrive? Most of the team leads quit, and the thing doesn't scale. Turns out you can't build a super computer architecture that scales using power point. Musk has gone from pushing it as the solution to everything, to calling it a "long shot" in the latest quarterly call and ordering more stuff from Nvidia. I expect Optimus, the teleoperated demo robot, to go down the same path.
Funny how software can be harder than shooting rockets into space.
Also, Dojo is dead. DOA. Or did it even arrive? Most of the team leads quit, and the thing doesn't scale. Turns out you can't build a super computer architecture that scales using power point. Musk has gone from pushing it as the solution to everything, to calling it a "long shot" in the latest quarterly call and ordering more stuff from Nvidia. I expect Optimus, the teleoperated demo robot, to go down the same path.
Funny how software can be harder than shooting rockets into space.
Last edited:
I disagree. Robotaxi can't work on HW3 or HW4 because neither is capable of cleaning the camera lens--except for the in-windshield cameras. I can't count the number of times FSD has given up due to "occluded cameras" while driving on wet, dirty roads. Unless Robotaxi is only going to operate on dry roads.
A "long shot" is not "dead." Even if it is, what's your point? If Tesla builds it's supercomputer from Nvidia chips instead of in-house Dojo chips, how does that affect End to End AI FSD? Perhaps delay it a bit, which I'm sure we all expect anyway... Either FSD will eventually work at an L3+ level or it won't; I don't see how training on Nvidia GPUs as opposed to Dojo chips really matters.
sleepydoc
Well-Known Member
Hardly dumb - There are other examples of distributed computing even if it doesn’t work in this specific case.
New Drive from Whole Mars Catalog
00:06 Engaging FSD from parking space.
00:45 Unprotected left
01:32 Respecting KEEP CLEAR marking on road
02:33 Hairpin right at start of highway on ramp.
04:36 Approaching toll booth. Excessive slowing on approach, accelerator-intervention at booth, speed-setting intervention on exit.
10:50 Disengagement upon encountering a bus lane on an exit ramp. (language warning)
15:22 Slowing on approach to green light intersection (22 to 8 mph)
15:48 Slowing slightly for car coming off curb. Car is signaling and angled to enter Tesla's lane.
15:54 End of drive, move to left side of lane because Tesla cannot get to curb. Left turn signal used.
00:06 Engaging FSD from parking space.
00:45 Unprotected left
01:32 Respecting KEEP CLEAR marking on road
02:33 Hairpin right at start of highway on ramp.
04:36 Approaching toll booth. Excessive slowing on approach, accelerator-intervention at booth, speed-setting intervention on exit.
10:50 Disengagement upon encountering a bus lane on an exit ramp. (language warning)
15:22 Slowing on approach to green light intersection (22 to 8 mph)
15:48 Slowing slightly for car coming off curb. Car is signaling and angled to enter Tesla's lane.
15:54 End of drive, move to left side of lane because Tesla cannot get to curb. Left turn signal used.
At AI Day 2022, Ashok Elluswamy said they've accumulated hundreds of thousands of past failure examples accumulated over many years, and these work more as negative assertions rather than positive adherence. This set is useful in making sure FSD Beta doesn't do something known to be bad whereas there could be multiple good ways to say make a lane change, e.g., change now or wait longer or slow down. These should still be useful in evaluating whether 12.x is doing something it shouldn't be doing.
As networks are trained, they're already being evaluated on how accurate they are predicting the training data to adjust behavior to minimize error, and a similar thing happens for a set of examples excluded from training but evaluated in the same way. Additionally, Tesla can leverage the inference compute available from the whole fleet in deploying "minor fixes" that could include updated neural networks to see how they make predictions in real-world scenarios without changing the driving behavior.
Even with end-to-end, this fleet evaluation can happen when people are driving without Autopilot active so FSD Computer can run the neural networks and compare with human driver behavior, and if Tesla is looking for specific mismatched behaviors, that can result in video being sent back and/or statistics of how often they match or differ.
I don't know about the rest of you, but I really did LOL at that one. Hillarious.
It's the same mistake that 11.x made because it's still running 11.x on freeways (if you're mainly interested in 12.x behavior).
I'm sure at first, Tesla robotaxis will indeed operate only in good weather. But over time it will get more and more capable of peforming its robotaxi duties in worse and worse conditions.
We still don't know if HW3 or HW4 will be able to handle it though.
Interesting that it's still 12.1.2. I saw that they've been rolling 12.2 out...
stopcrazypp
Well-Known Member
Pick summer time to do the drive. You are telling me it wouldn't be possible for there to be a day in the year where you can map a coast to coast route that avoids rain? Nothing in the definition requires it to be "practical". And as pointed out, it can stop until the rain clears and continue.
Rain is just an example, it can also be fog, which is much more rare. If the car can't travel in fog that a human can, it's automatically disqualified from SAE L5, but plenty of people would still consider it L5.
No one asked him the details, people just assumed because he says L5 he means the same as SAE. As pointed out above, the definition for SAE has a ton of nuance that would disqualify a car from a certain level. And his reference to reliability levels also is inconsistent with SAE definitions (that's not a criteria in SAE's definition).
Todd Burch
14-Year Member
Artful Dodger
"Neko no me"
I understand that v12.2 is going to HW3 cars. Omar has HW4.
As in what to look for? Easiest is probably AUTO MAX set speed with 12.x vs a number with 11.x. Other 11.x control visualizations (e.g., blue highlights) and messages (e.g., changing lanes to follow route) appear on the freeway too. If you pay attention to where it switches between 12.x and 11.x, you can sometimes find hand-off issues, e.g., 12.x performing a lane change getting onto the freeway then 11.x decides it wants to return to the original lane (to then show the message that it needs to follow route before changing lanes).
