I don't think he's stretching; the way in which Tesla is estimating future FSD version performance was answered by Ashok in the Q1 earnings call. Long story short is that they have measured the scaling rate of FSD performance as compute, data, and architecture scales. Here's the long story long:
Your quoted text has nothing to do with what I am talking about - which is the wording Elon uses is loose.
"We are starting to get to the point where, once known bugs are fixed, it will take over a year of driving to get even one intervention."
"We are starting to get to" - this is entirely vague. How close to the
point is "starting to get to"? It's not
at the point. He's not saying V12.6 is at 1 disengagement per year, it is somewhere less performant than that. We are left to speculate on how much less.
"
once known bugs are fixed" - this is not C code. You can't just go in edit some lines of code and remove this. He is making it seem like these are "easy" to fix, which is speculative. They haven't actually fixed them yet. And "fixing" them means retraining with an updated data set that hopefully removes these errors.
So these errors are still real driving errors, but I imagine ones they are more confident can be fixed by updated the curated data fed in for training.
So Musk sees some report from Ashok that miles per intervention on V12.6 is at say 1,000 MPI. Ashok shows him that 80% of the interventions are ones that they hope they can fixed with better training data. Musk calculates that means they are at 5,000 MPI without these "easy" issues, then estimates they are "close to" the 10,000 MPI mark barrier that is close to the one a year rate for most drivers.
Note, mean and median are different from miles driven per year because of skewed distribution. Musk knows this and I doubt uses the mean when he estimates. Median is closer to 10,000 miles per year.
My point isn't these exact numbers are correct, but we have to put this sort of thought into it because the statements given are purposefully vague.
First step is to admit they are vague.