I don't believe the OP meant individual cars. But I'm pretty sure a disengagement shows in the logs which can be parsed and eventually followed up on.
As the post you quote mentions- those reports contain very little of use for "fixing" the specific disengagement-
On the current wide-release versions of firmware, when you disengage only a very very tiny report goes to Tesla- No pictures, no video, just location is coordinates, how did you disengage (brake, wheel, stalk), speed, heading and time. That's it. Under 1kb of data in the report.
It could be useful to tell Tesla "Many cars have trouble at this one GPS location" or even "Many cars have trouble at this location if they're going more than X speed" but nothing is "learned" about correcting whatever the issue is simply by disengagement.
If Tesla decides that spot has a real problem it can then actively send out a campaign to collect useful data there- but it can collect that data (photos, video, etc) REGARDLESS of if the driver has FSD or not, and regardless of if the driver manually disengages at that spot or not.
It's part of why folks thinking they need to make FSD super cheap so Tesla can collect more data are misguided- the actual "train the NN" data they need can be collected from a car regardless of if you paid for FSD- it's one of the huge fleet advantages Tesla has.
No, the illusion fooling people is that "no intervention ever" is the only acceptable state of affairs.
Because no driver ever has needed any outside help on the road. Ever, not a single one.
Truth is this is pure FUD..
And yet in the last couple pages we have folks telling us how once Tesla solves FSD somehow BOTH of these will happen:
Folks will stop wanting to personally own cars since RTs can take them everwhere
and
Folks will spend 100k on owning a Tesla since it has FSD.
I'm a bit surprised at all the shock and amazement at what the beta FSD release can do.
I'm not at all surprised by the system working pretty well most of the time.
I AM surprised by folks who think working 90% of the time means it'll easily work 99.9999% of the time Real Soon Now with Just A Little Training. Or people who think software can magically fix blinded-by-dirt-or-weather hardware.
Because that's now how any of this stuff works.
Even Elon, who's always exceedingly time optimistic, has said it's likely at least another year before it does some things "well" that it "can" do now.
I'
By the way, I've seen a couple of videos which indicate that speed bumps are already solved. This leaves statements like "Speed bumps will need to be learned also." looking less than brilliant.
Since there's also videos showing it going right over them without slowing down as well, it's nonsense to claim they're "solved"
They are, at best, "working on it"
I never meant the individual car. I still believe that once the training labels road debris as a non-drivable area and it's uploaded to the fleet the problem will be solved.
There's that "solved" word again- I do not think it means what you folks think it means.
1) There's an almost unlimited # of things that could be road debris. It's not a single thing to train against. (Here fleet size can help a lot to bring back many examples, though it'll still be incomplete for a long while)
2) What do you DO once the car can see the lane it's in as suddenly undriveable a very short distance ahead? THAT needs to be programed- and for a variety of scenarios... Do you swerve into another lane? Which one (if there's a choice)? Do you narrowly (but pretty sure you can make it) cut someone off to do it? Do you brake instead? Do you drive onto the shoulder to avoid it?
It's certainly addressable, and I'd 100% expect them to address it... but it's very very far from a simple or easy fix.
Clarification: car performance can improve due to changes other than the NN.
Agree that the Neural Net does not self update on a per car basis.
However: the camera mapping/ merging parameters must update/ calibrate on a per car basis. If a driving issue was due to poor blending, re-driving an area that exposes the mismatch could improve the blend and drive performance. (Tight curves/ intersections/ roundabouts could fall into this category)
Not quite clear what you're saying here? other than camera calibration (because cameras might physically be very very very slightly different positions from one car to another) nothing is car-specific
Clar
Potentially, the map/ hint data set could self update locally as part of the fleet learning algorithm or a 'things I've seen' cache, but that's theoretical.
Nothing updates locally/uniquely like that- maps are pushed from the mothership just like regular firmware (though not updated as often to this point- wouldn't shock me if we start seeing faster map updates going forward though.)
