Why are you confident it's not the other way around? People who decide to pursue alignment research may have prior interest or experience in ML engineering that drives them towards mech-interp.
You should also set model.cfg.normalization_type = None afterwards. It's mostly a formality since you're doing it after initialization. ActivationCache.apply_ln_to_stack() is the only function I found which behaves incorrectly if you don't change this.
In my opinion, this is connected with Sturgeon's Law. I'd guess that to expert pianists and piano tuners, 90% of pianos sound out of tune. I know among hardcore software engineers, a common lament is that almost all software sucks. Windows is almost unbearable to me, but I'm sure most desktop users are happy with it. Most desktop users are not programmers.
90% of all things may be crap to the discerning eye, but the world remains ok with that because each person has only a handful of places where they care to discern.
it’s clear that more exploration is the way to go
Just pointing out that maximizing exploration is not always good. I like this post that argues for more exploitation.
I wasn't doing learning theory in 2016, but the cannonical textbook, Shalev-Shwartz & Ben-David (2014), covers both nonuniform learning and PAC Bayes, so I'm a bit confused because both of those approaches were known at the time and sidestep the killer results from Zhang et al. (2016).
In nonuniform learning, you split up your hypothesis class into a union of countably many smaller classes