JupiterMan
Active Member
Don't think so. with forthcoming 10 x in compute and auto labelling, its all about scale.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Elon: FSD Beta tweets
- Thread starter Terminator857
- Start date
AlanSubie4Life
Efficiency Obsessed Member
Not sure why it would be related to any of this. I suspect it was about realizing revenue at a convenient time, and it being safe enough as implemented, based on data (of course we have no idea how safe it is (other than an upper limit), as there is not any data on it, but internally they know).
The challenge will be keeping it safe enough, and performance with the edge cases (which of course there are not nearly enough miles to find yet).
Last edited:
Video or it didn't happen.
Not all of us take videos of every drive.
Same again this morning as well.
EVNow
Well-Known Member
Nope. That has been Elon's and AP teams known and well publicized targets for a long time. Nothing to do with weirdos like O'Dowd.
EVNow
Well-Known Member
sleepydoc
Well-Known Member
And the safety score was arguably not representative of safety score winners' average driving.
clydeiii
Member
Possibly, but I was let in at 99 and I have hover between 99 and 100 for Tesla Insurance.
In any case, prior to the wide roll out, SS was a confounding factor in generalizing the accident rates of FSD when comparing with national averages, arguably a non-representative sample.
Not fsd related but we can all get a laugh out of this tweet:
Classic Elon! Mar 31, what year??
The rest of the world outside of tesla fans is going to learn about Elon time! Very exciting.
Classic Elon! Mar 31, what year??
The rest of the world outside of tesla fans is going to learn about Elon time! Very exciting.
Last edited:
KArnold
Active Member
It looks about the same as a few hours ago, according to Teslascope there were 2 small rollouts and to expect another later today.
Their current info source seems to be right regarding 11.3.2 because they hinted it was coming earlier in the day and it did.
Seems like getting rid of the human-engineered vector space is gonna be the next big barrier that has to come down. If I was forced to drive using only vector space it would be terrifying and bad. It should be vision -> control with no human understandable features in the middle.
willow_hiller
Well-Known Member
I was actually thinking more about this, and I really appreciate Tesla's approach.
If you've got one network for perception, and one network for planning, and one network for control, you can pinpoint where a problem stems from and debug a part of the system without breaking other parts.
How would you debug an end-to-end vision to control network? Throw more training samples at it and pray?
You can only look at the system in a wholistic way. Like you can look at the dashcam footage and reason about why it did what it did, but you can't examine the intermediate feature space because it won't be human-understandable. That's why deep learning systems are so powerful is because they don't have hand-chosen intermediate layers that engineers picked out.
It's like when people used to make vision-based detectors for cars a decade ago, before deep-learning, you'd make a detector for wheels, and doors, and hoods, and other car-like features that humans picked out. And then you'd decide where a car is based on those detected features. That is crazy talk today, no one would do that. And yeah we've given up the ability to debug why a car wasn't detected at the feature-level. We can only look at the dashcam recording and the final output.
Also it's no different than human drivers, you can ask a human after a crash what happened and they will just make up some story after the fact. They can't access their intermediate mental states and tell you. But the dashcam footage tells you all you need to know.
diplomat33
Average guy who loves autonomous vehicles
That's called the "end-to-end" approach. Wayve is trying it. They have some nice demos where the car goes around a parked car or navigates a roundabout. But there is a big difference between training the car to perform certain driving tasks and achieving the reliability where you could drop the car anywhere and let it drive fully driverless. So far, nobody has made it work reliably at scale. It is a very difficult approach for many reasons. For one, it is very difficult to train a single NN to figure out all of planning and control directly from vision alone. Second, it is very hard to troubleshoot since there are no "human understandable" features in the middle that you can pinpoint where the failure happened. It is a black box. So if there is a failure, you have to retrain the entire NN until it fixes the problem and hope you don't introduce a new failure.
Having said, the trend has been towards consolidating NN's into fewer but bigger NN's. So we may eventually get to "end-to-end" some day but right now doing "end-to-end" from the start has not been solved yet.
That's one of the big challenges with the "end-to-end" approach. It is very hard to troubleshoot since it is one big black box. There is no way to pinpoint what caused the failure. So yes, the only thing you can do is retrain the entire NN until it fixes the problem and hope you don't introduce new bugs or failures. That is one reason why all the AV companies have different stacks. It allows them to troubleshoot better. And you can build a NN to do a certain task like perception and test it and validate it until it is reliable. You can build a NN to do another task like behavior prediction and train it and validate it. And then put it together when you know each separate NN is reliable enough.
Last edited:
DanCar
Active Member
EVNow
Well-Known Member
If "one more refinement is needed" ... why is the release going wide ?
j0shm1lls
Member
EVNow
Well-Known Member
Similar threads
