We are not talking about Tesla going from 0% progress on BEV to 100% progress in one update (there is no 100%, at least not pre-Singularity).
We are talking about:
In that one update, a lot of Tesla drivers reported that Autopilot went from ping ponging between lane lines all the time to lane keeping smoothly. It could handle a lot of sharp turns and hills that it couldn't before. And automatic lane changes went from sloppy to reliable.
All I'm suggesting is that we'll see a similarly dramatic improvement in FSD from v9 and that it will be mainly because a lot of the progress from the 2.5D-to-4D rewrite will be concentrated in that v9 update.
We are talking about:
- more vision NNs being upgraded from 2D projection to BEV
- better BEV net architectures (e.g. ones that solve the challenge of internally representing uncertainty and probability)
- better trained BEV nets (e.g. through larger/more diverse/better curated datasets)
- BEV nets that have been debugged, use compute more efficiently, or have otherwise been refined
In that one update, a lot of Tesla drivers reported that Autopilot went from ping ponging between lane lines all the time to lane keeping smoothly. It could handle a lot of sharp turns and hills that it couldn't before. And automatic lane changes went from sloppy to reliable.
All I'm suggesting is that we'll see a similarly dramatic improvement in FSD from v9 and that it will be mainly because a lot of the progress from the 2.5D-to-4D rewrite will be concentrated in that v9 update.