The way neural networks work is a fascinating subject. The more I learn, the more I realize how little everyone knows about "AI". I have seen many entrepreneurs and even highly educated scientists make many ignorant statements regarding "AI".
The reason I specifically brought up Mechanistic Interpretability is its ability to peak into the inner workings of an "AI". I used to think of them as black boxes with extremely complicated inner workings. It turns out this complexity can be somewhat managed. Currently these techniques work on a single layer. Eventually we could see a multi layered analysis that could unlock previously unexpected features out of the "AI" we build.