Elon
: "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."
In the United States, the average person drives about 13,500 miles per year.
I assume this is more a figurative number than an exact number. Maybe he means more like 5000miles, or maybe he means >>13500miles.
Previous customers(not Elon level tolerance testing) had about 305miles/intervention (
FSD Community Tracker )
View attachment 1053901
The big question is what his definition of a known bugs is. Is it lack of data of a specific type of situation ie unprotected left turn? Or is it more software bugs?
But guess the good news is that Tesla feels that they know enough bugs that once those are solved they will improve the intervention rate from 300miles/intervention to 5000-15000 miles/intervention. And they know how to do it, just gather tons of high quality data, train on massive compute and validate. So by the end of the year if not a lot sooner we should have a system which is that good.
Also do we believe the rumours of FSD China in 2 weeks that are going around on x?
I'm sorry this is so absurd on multiple levels. One, you cherry picked a moment in time that people were reporting 300 miles per critical disengagement.
Now, with more data, it's around 150.
The amount of caveats in Elon's statement is important.
"Once known bugs are fixed". What is considered a "bug" and how is it fixed? They are operating end to end neural nets, fixing is retraining with an updated data set and hoping it fixes the issue. This is not some "go edit a few lines of code" to fix the "bugs". Just because they are "known" does not mean the solution is trivial.
"We are starting to get to the point" - LOL. So V12.6 (if you exclude "known bugs" might get close to 1 intervention per year. But how much driving does he assume?
Does he assume national average, or is he thinking for people who drive less?
All these assumptions matter, because when they add up, they can change the actual outcome by an order of magnitude.
If you realize "starting to get to the point" might be a 2x or 3x off of actually being there.
If you realize "known bugs" are actually pervasive, critical errors than aren't removed suddenly.
If you realize a year of driving for some people is 5-10k miles, not 15k...
Then the actual state of V12.6 might be 1000 miles per critical disengagement, not 10,000.
And while this is a big improvement, again, it is far off from being viable for robotaxis.