I agree with most of what you're saying, and I premise my comment by saying I have some expertise in AI applications, but I would not say I have deep expertise.

What I have read (and seen in my own practical experience) is that:

1) There is a big gain to be made using a large data set and (if I understand what you mean in your first sentence correctly) training on the features;

2) The data set(s) required to improve on those initial gains increase massively, not decrease OR you need human expertise to make the judgement.

In a practical (and current) example from autonomous driving: AI can be trained pretty well to recognize an object in your path ahead. AI can recognize that you can maneuver around that object. AI can recognize other objects moving in the opposite direction to your line of travel but not directly in front of you. AI is not yet able to put those three things together and evaluate in real time and make a correct decision on what is safe:

https://www.youtube.com/shorts/tNf2Np7Ek5E

This is the point where looking at features (i.e. "trends") fails and an understanding of all features as they relate to each other AND the individual data points must be looked at.

The YouTube short above is public, but I have practical engineering examples (that I can't share because of proprietary information) of exactly the same sort. AI is a help up to a point, but then engineering expertise and practical operational experience is required to tell the AI what to do. This will improve the AI in the engineering case that I know personally, but I don't know of the same type of learning capability in AD/FSD.