Mental Simulation: Architectural Bet for Educational AGI or Another Recommender System?

Last week I chatted with a friend working on K12 products. His team spent over a year building a "smart mistake notebook" using transformers, which recommends similar questions based on students' historical errors. The results looked promising—an 85% accuracy rate—but teachers concluded: "This is no different from a standard question bank; students still can't solve what they couldn't before." Essentially, it just applied collaborative filtering to question features and student answer sequences, without touching the cognitive layer of why mistakes happen.