Comment by sidibe
3 years ago
> I'm guessing that after running the route many times for practice
Why practice? They did the run until it worked well for the video. It wasn't a demo with a live non-Tesla audience. In this case not working 9/10 times is good enough rather than 1/10000
They did practice, though?
The odds I'm talking about are not "it doesn't work as well as the video" but instead "things go horribly wrong in a public way".
There is no reason to think it was better than today's FSD even with those "crutches" (as I believe Elon himself refers to them when talking about other SDC companies), and the chance unsupervised FSD does something horribly wrong in a public way on real streets even in 3.4 miles is far higher than 10000:1 today. After all the car did hit a barrier during the making of this video. I doubt they spent much effort improving the odds when they had infinite takes.
Looking around I see something like 6mi per disengagement [1] which, yes, is unacceptably high. The video is probably about 2mi, which naively would give 1:3. But disengagements have a large amount of precaution, so this isn't an accident rate. And some accidents wouldn't be ones they'd need to be public about, like the one in testing which only came out years later.
[1] https://electrek.co/2022/12/14/tesla-full-self-driving-data-...