The hardest problem seems less like content generation and more like accurately modeling what a learner actually understands.