Disagree. It’s easy to see how far FSD has progressed, there’s lots of videos to document how far it’s gone from then and today. It’s also not easy to say they’ve reached the point of diminishing returns (which generally takes a massive amount of data to hit, even GPT3 has not hit it) as it’s not clear how much of FSD is powered by heuristic based code and what is NNs. Judging by some of the discussions there’s still a significant amount of C/C++ code in the loop that could be improved by moving them to NNs.