Post

I'm only 30% of the way through or so, but one thing jumps to my mind: there wasn't one single breakthrough here. Instead, we got substantively better at: - tracking training dynamics - nn architectures - interpretability protocols (for lack of a better name) These interact positively together.

This paragraph is interesting. It seems that "bigger, better, newer models train better *and* are easier to keep monosemantic" is a better explanation than "polysemanticity was never in InceptionV1". It's good news either way, but I wonder.

Ok, done reading Roughly, "if we examine modern nns using modern tools we have to build nns, it gets easier". That's good news, but (shock!) the oai guy is overselling it. Minimal counter: someone gives you only a pile of weights, says "this is aligned". "good news mechinterp" won't help you