That's a big problem though, mechanistic interpretability/explainability is really important for the whole "not wrecking everything" part of AI. You could also

That's a big problem though, mechanistic interpretability/explainability is really important for the whole "not wrecking everything" part of AI. You could also just hope AI stalls out or goes back to more hand-tuned Bayesian stuff, but hope is not a plan. Progress past standard analysis has been made, so I'm not sure how important this paper is.