Shareholder here as well and dont think we will see that big step in the next two weeks, or before the end of the trial. But does it matter? I think we need to see the major steps ahead of 8/8/2024 to make a big impact on the stock price. If SFSD becomes wall street believable and their attitude changes as Tesla from an auto maker to an AI maker, then the stock still jump.
-
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
FSDtester#1
Craves Electrolytes
Thanks, I will go check mine. MS... service mode, software, autopilot
Figured it out, both banks 2.17 GB... what does this tell us?
What version are you on?
It’s more useful in a comparative sense to check if anything changed. Rather than someone changing the version number or actually updating the bank with a new version.
FSDtester#1
Craves Electrolytes
Scary story. Cool that FSD hit the brakes in time in spite of the poor traction.
We had a scenario wich had me wishing FSD would get a closer to the car ahead: We were a few cars back from a red light in our direction, and FSD stopped a bit far back from the car ahead. Then, a car coming the opposite direction stops and turns on his left blinker, wanting to cross our lane. I look to my right and see the driveway he must want to turn into, and judge that if I pull up to a normal distance behind the car in front of us the the opposite direction car can turn behind us just fine. So I nudge the go pedal to close the gap. At the same moment, the opposite direction car turns right in front of us to use the gap. Speeds are low, but I slam on the brakes, and so does he, no collision. He then backs into his own lane, we creep forward, our light goes green, end of event.
But by stopping so far back from the car we were following, FSD tempted the opposite direction driver to try to sneak through. If he had turned his blinker on sooner, and had I been driving, I would have either closed or open the gap to remove the ambiguity which FSD created.
The good news is that in neither Kpanda's case nor in mine did a collision occur. The bad news is that there are always many paths from now to the future, so decisions are often both right and wrong at the same time. Our job is to survive the decisions, second guessing them is optional.
But all you really need is the one ingredient that trumps (now really need a synonym for that word...) all the rest: too much free time on your hands.
[EDIT: Which all of us here clearly have...]
hey8you
Member
Before I get totally wasted and erase a terrible work week with the magic of alcohol. Im drinking now so I guess I better hurry.... lol. I have mixed feeling about v12.3.4. Exiting my complex with the gate partially closed, V12.3.4 handled it great. Older versions couldn't decide to wait or try another route. While trying to enter today, v12.3.4 tried to pass a stopped car at the gate so I disengaged. It's still stopping short at stop signs and stop lights. Its cutting over to the wrong lane on 2 different intersections and one of them it can't tell when the light goes green. Others similar intersections it handles them fine. Yesterday it followed a car driving in the parking lanes(visual aid below) after it went through an intersection.
Red line car took. green line car should have taken
Red line car took. green line car should have taken
FSDtester#1
Craves Electrolytes
Why do you think there's an A and B for other things like Infotainment? Why do you think the size reported would need to change for Autopilot behavior changes? Why do you think one bank is selected over the other?
It seems like these are separate storage for handling updates so that it can revert to the previous bank if the installation on the other bank fails for some reason. I do think a size change would indicate something has changed, but even the same reported size could have very different behavior for end-to-end, e.g., it takes the same amount of space to represent the number 0 and 1.
mgs333
Member
Isn't it possible that 2 subsequent versions, especially for a tiny minor update, could round to show the same size?
NHTSA requires the slowing and stopping before lights and signs. We hear that might be removed in future code.
Gates are tough and think on the path to being improved with new videos for AI to learn.
12.3.4 was a compatibility merge, not a step improvement. We have only seen one big step, 12.3.X. We are readying for the next big step which will be 12.4. Elon has already teased it. Be patient, supervise, skip gates with FSD for now.
12.3.2.1, 12.3.3, 12.3.4 are all the same just merges in for unique vehicle hardware, hold for the big one 12.4
"Elon Musk has said that Tesla FSD V12.4 is a significant step up in capabilities, with training compute limitations almost gone."
Todd Burch
14-Year Member
The size of two data banks being the same means nothing.
As an example, two 1000x1000 matrices of 16-bit weight values can be completely different and still have the exact same data size.
As an example, two 1000x1000 matrices of 16-bit weight values can be completely different and still have the exact same data size.
I don't know what to say, I just know that it is different. What I don't know is why. Are the NNs different? Is it different "guardrails"? Tuning parameters? So something in that 2.17GB is different. (If it is really camera feed in and control out of the NNs, then they have changed.)
For me it now stops like my grandpa, but still takes off like my dad. (My grandpa was very efficiency minded, and would start slowing down very early for stops, coasting in with the hope that the light would change before he got there. And probably pissing people off behind him. He, also, said my dad "drive like a frustrated race car driver.")
Some proof, here is my State Farm report for the same ~13 mile drive I normally make:
Not a single hard braking event. (I don't think this has ever happened on this route before with FSD driving.) It's great. Now if they can just tone down the take-offs. (I'm sure Alan would have an absolute fit with how soon it starts slowing down and how long it takes to get to the stop line and actually stop.)
Possible but based on some more data I found it would be an edge case. There appears to be always be a greater than 10mb difference in each FSDV12 version change.
E2E can be inherently unstable. You can have exact simultaneous scenarios but something benign in one can affect it. Like imagine if everything was exactly the same except one tester had clouds in the sky and one didn’t. It can cause the model to spit out different actions.
Nvidia had an entire video workshop on Ai, hallucinations, bias in training set, etc.
Last edited:
I can tell you that if the data bank is exactly the same size in the next update that the car isn’t going to stop curbing and destroying wheels. They need to go into the training set and find whatever idiot drivers are sucking through a green straw while driving onto curbs. Retrain the data set. I promise there will be enough difference to change the data bank size.
I understand that, but 12.3.4 seems significantly different than 12.3.3, at least for me in regards to stopping. It will be interesting to see if the stopping remains like this on my future drives.
Similar threads
- Article
