From what I hear trouble is your middle name. Hopefully your trouble will continue when 10.69 comes 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 Beta 10.69
- Thread starter Buckminster
- Start date
AlanSubie4Life
Efficiency Obsessed Member
With this bet I can’t lose.From what I hear trouble is your middle name. Hopefully your trouble will continue when 10.69 comes out.
Even a terrible 95% success rate on this easy turn in significant traffic, if it isn’t done with overfitting, would suggest actual significant improvements in underlying long-standing limitations of FSD.
For @Daniel in SD it is lose-lose. He doesn’t have FSD. Sad!
Daniel in SD
(supervised)
Yes, looks like it makes driving so much easier. Thank you to all the beta testers, I'll buy it when it's out of beta.
Hopefully it's not an overfit or Tesla is going to spend the next year fixing all the ULTs near Chuck.
It seems like what they should be focusing on is improving their simulation tools so they don't have to actually test in person. This seems like a very inefficient process.
AlanSubie4Life
Efficiency Obsessed Member
Haha. Hopefully not. There seems to be a fundamental misunderstanding about over fitting above though. People think it might be happening because Elon has good experiences. But they don’t understand that Elon does not have good experiences either.
How are people using that term? For neural network training, overfitting usually means the training data seems to result in lower error but when evaluated with validation data, the error actually increases.
Practically, this means a given trained situation will likely work fine but the neural network fails to generalize instead believing another situation is basically the same as the trained one. If there was overfitting to this scenario, the network could incorrectly predict for a completely different unprotected left turn that there's 3 lanes of cross traffic in each direction with a certain amount of space to creep or median safety region.
Another aspect of neural network overfitting is that it picks up on signals that aren't actually relevant potentially resulting in false positive detections.
Most of the turns like that around here are in lower speed limits - apart from two that are both on 65mph roads. At those speeds that seems like an impossible hurdle for FSD, it can't look far enough down the road to safely make the turn.Yes, looks like it makes driving so much easier. Thank you to all the beta testers, I'll buy it when it's out of beta.
Hopefully it's not an overfit or Tesla is going to spend the next year fixing all the ULTs near Chuck.
It seems like what they should be focusing on is improving their simulation tools so they don't have to actually test in person. This seems like a very inefficient process.
I can see it working ok at 40mph, not sure what the speed limit is on Chucks road though.
AlanSubie4Life
Efficiency Obsessed Member
Who knows, but the evidence proffered appears to be that Elon has good drives, which is where the story falls apart (since this is not the case of course).
AlanSubie4Life
Efficiency Obsessed Member
Says limit is 45, so people travel about 50-55, sometimes faster.
Definitely it needs to be able to see reliably (motorcycles included) about a couple hundred meters. Consistent success on this turn would suggest they’re getting close to that.
It's not a matter of "obsessing" over his jokes. He is who he is. If you look at the original post, it has more to do with what I have come to perceive as trying to use his humor, childish or otherwise, to cover for the lack of forward direction, a well he is going to far too often at this point. He might as well of said "10.13 will be delayed for another few weeks while we try and make it an actual improvement -- Look at the monkey... LOOK AT THE MONKEY!"
I mentioned as well in another post somewhere...but you are spot on.
Even his dollar amount offer for twitter. Why do I think the financial advisors suggested a dollar amount per share lower than what Elon offered. BUT..he was hard pressed to include either a 69 or a 420 in the figure. As a "look at me. see how funny I am" type of thing. So he went with...$54.20
And well...look at where that got him in that deal. Not so funny anymore, is it Elon?
Last edited:
It'll be over 12 weeks since 10.12 was originally released, so presumably there's been 6 internal 2-week sprints that each could have been deployed if strictly following the release plan. Would you have been okay with 10.69 being called 10.18 instead to represent that there has been significant work more than "just" a regular release?
Or are you suggesting that Autopilot team isn't actually making progress in those 2-week sprints, so that's why there hasn't been releases?
Presumably if they do have simulations of this intersection already, they've probably already evaluated it in many ways and clearly also want real world evaluation. Indeed there most likely can be improvements to minimize the differences, so these tests could also feed into finding out the unknowns or what's lacking in the simulator.
Could they get to not needing real world testing? Maybe? But looking at how Elon Musk runs SpaceX with Starship tests compared to other aerospace companies, all most likely have simulations and formal evaluation of some sort without needing to launch a rocket, but SpaceX seems to prefer real world evaluation more than others in order to reach faster learnings and iterations. There is a cost both money and time in trying to "perfect" the simulator, and there could also be more important things to focus on for the long term needs of FSD.
Also, we don't really know if Tesla is actively testing many other unprotected left turns in other locations. Maybe there are other drivers across the country doing similar tests to evaluate for improvements and regressions potentially also feeding into simulator improvements.
Silicon Desert
Well-Known Member
Well not really. He could always implement 10.xx.xx sub versions. It wouldn't be the first time.Look at it this way. If the next version is 10.69, it really limits future 10.x versions
Like when it thinks it sees a truck in my garage, when what is there is a (Dobsonian) telescope standing on end? See front left rear-view camera below.
It's a surprising lack of situational awareness. It needs to be able to determine that it's parked in a rectangular garage of A*B feet. It doesn't seem to know that. It has drawn no walls nor an open door.
Secondly it has misperceived the objects around it and placed a truck there. From the objects we know it can draw, it could have shown a bike (though the bike is well out of the way), maybe a garbage can. Why would it even consider there is a truck there?
Thirdly it should be aware it can't place a truck because there isn't space anyway.
These are fundamental problems. It doesn't even know where it is or what is around it. (Maybe it does know more than it's displaying, we can't be sure. It should show the scene boundaries and impinging objects)
Last edited:
AlanSubie4Life
Efficiency Obsessed Member
True, however, when they complete this turn safely ~100% of the time like Elon clearly said they are going to, without overfitting, it may as well be the last version 10 anyway! That’s going to mean impressive capability improvements. Won’t be anything left to do but quickly add single stack in version 11 (going to beWell not really. He could always implement 10.xx.xx sub versions. It wouldn't be the first time.
) and then a little cleanup with one or two more iterations, and wide release of completed FSD by the end of the year as promised. Book that revenue!
So when it shows a truck where there is no truck, to what degree is it acting on that wrong information?
Is the display just 'entertainment' and not actually showing what FSD Beta is doing, or is the display the official decision output of FSD Beta? In either case why is it so wrong.
Sometimes we see a truck overlapping the Tesla ego car. If it's in any way accurate shouldn't the car be screeching to a stop.
Can it display objects at arbitrary orientations ie: accurate orientations, or is it restrained. Would it not be better to go back to box placeholders that are oriented accurately rather than fancy, but incorrect, renderings of the objects?
According to Tesla vision I have upwards of 5 trucks and semis in my garage. An empty area on one side (literal blank fence and concrete wall) is rendered as a truck more of then than not, and a neighbor's bicycle causes a compact semi to be detected on the other side. Sometimes they even dance to the music
Yeah, these false positives could be reduced by having labeled training data indicating there's not actually a truck. There's probably multiple predictions that are wrong here such as repeater camera seeing the surface behind the bike as the ribbed wall of a shipping container, and maybe the vertical piece closest to the repeater camera might be a prediction by the pillar camera thinking it's the gap between the truck and semi-trailer, and the unified view piecing together these individual predictions comes up with the final false positive prediction. Training data that reduces the confidence in any of these could be enough to avoid the false positive.
At least right now, it doesn't seem high priority for Tesla to improve predictions for vehicles in a garage as there's also the potential for reducing these false positives to result in false negatives, e.g., switching lanes into a real truck. In some sense, the false positive is "okay" in that it should have FSD Beta not drive into the wall / imaginary truck.
Similar threads
- Article
