I get the impression that most of the 'understand the innards' work isn't scalable - you build out a careful experiment with a specific network, but the work do

I get the impression that most of the 'understand the innards' work isn't scalable - you build out a careful experiment with a specific network, but the work doesn't transfer to new models, fine-tuned models, etc. I'm just an outside observer, though...