PhaseWhite
Member
To expand on this :
Avg crash rate : 4.2 MM - 99.9996%
6 9s, 4x better : 1.05 MM - 99.9999%
7 9s, 40x better : 0.105 MM - 99.99999%
If we assume Tesla is at 99% - and improving exponentially by one 9 every 6 months - they will be about human avg in 1 1/2 years. 4x better in 2 years. 40x better in 2 1/2 years
Achieving each additional 9 is 10x more difficult than the last, so I don't think it makes sense to assume a linear time scale like every 6 months. Getting from 99% to 99.9% is a 10X improvement or like going from 1 failure in 100 attempts to 1 failure in 1000 attempts.
Getting to 99.9999 is a 10,000X time improvement over 99%.
The only way I could believe in exponential progress like that is if the NN training pipeline was 100% automated based something like using fleet data for re-enforcement learning (I'm not a Machine Learning expert), but from the recent autonomy investor day we know they still rely on human labeling. That doesn't scale to 10,000X very easily so I think a new approach will be needed to get there.
As far as where they are today, I think we'll know when they get to 99% because I think they'd want to start releasing the "FSD" city driving features with a hands on requirement like NOA is today. That will enable them to start gathering the data they need to achieve those extra 9's
If you disagree, please make your case...