FSD does not work well in the US, and Musk has indicated that it cannot work with HW3 or HW4
And past performance is no indication of future performance, especially with Tesla software! Not sure I could agree with an unqualified 'cannot work on HW4', but seems that everything that has happened so far in Tesla's FSD realm has really just identified a likely picture of just what a huge challenge FSD represents, especially given the way Tesla have approached it. Until native HW4 code and data is in public use, we have no idea if that will just carry on evolving from the current HW3 emulation performance or if it will be another start over.
I have no first hand real world experience of training NN's, but it seems to me that training done on low res vision could be pretty much useless when you switch to higher resolution. Compare to people with life long vision impairment where they devise different ways of identifying what they 'see' to work around their impairment. If all you have had is a certain (compromised) view of the world, a sudden switch to high res could be very confusing!
Now they should have no excuses for cameras that don't have complete views at junctions, underestimating variations in road layouts and marking conventions, understanding the huge range of climatic conditions outside SoCal, or making a simple task almost impossible through fixation with one myopic approach to design (like Vision Only).
All the videos on YouTube suggest otherwise
I haven't noticed 'all videos' painting a similar picture. There are improvements apparently, but also regressions and I think the concensus is that while the current offering might still be merit worthy in some regards, it is no where near where it should be after this many iterations (especially given the commentary / predictions from the likes of EM).
Last edited: