We have to step back and think about how the driver knows when to intervene.
The overwhelming majority of the time there is a visual clue that something isn't right, or the diver loses confidence that FSD knows what it is doing...
The vast majority of those visual errors, seem to be not recognising or correctly categorizing, something the car sees,,
These fall into 2 categories:-
- More training needed against a known label with a predefined course of action.
- New label needed and perhaps additional driving instructions..
There is no obvious indication Tesla is running out of headroom in terms of training with more data, and no obvious indication they are running out of capacity to add more labels.
In terms of the part of FSD that can be solved with vision alone, there is plenty of scope for further improvement, some of that being rapid improvement, some taking months rather than weeks.
The march of 9s is never 100%, any driving situation that requires sensors or information beyond what vision can provide, has a element of risk.
Vector maps, auditory cues and situational memory in training data can help.
But when I look at the videos, most of the time the driver intervened in response to a visual cue, that sort of problem is a NN training problem.