Nils

This is the 100th "Duolingo for anything" I've seen this year. 7 reasons why ad-hoc lesson creation with AI doesn't work (yet): 1. Curriculum The single most important factor in learning is what the learner already knows. These companies are very proud of their knowledge graphs, but there's no prerequisite model underneath. No placement diagnostic. "Tell us your background" self-report is not a tested knowledge state. 2. Mastery The apps may have quizzes, but that doesn't matter if you can miss every question and advance anyway. Exposure != competence. And because knowledge is hierarchical, the errors compound: shaky early means lost later. Worse, the quizzes couldn't gate on mastery even if they wanted to: recognition-based multiple choice measures familiarity, not recall. 3. Forgetting The two most robust findings in a century of learning science are the spacing effect and the testing effect. Without retrieval, you lose most of what you learned within days. These products are architected as short-term crutches: Feel like you learned something today, forget it by tomorrow. (An optional review deck on the side is not retrieval woven into the learning path.) 4. Application Everything is identify-and-recall. Nothing is produce, solve, build, or transfer. The result: 100% "inert knowledge", information you can recite but never apply. This is fatal twice over. Pedagogically, transfer only comes from applying knowledge in varied contexts. Recognizing definitions never generalizes. Motivationally, if the course never touches your actual goal, there is no reason to show up on day 2, let alone day 20. MOOCs have real instructors, curricula, and credentials and still get ~3% completion. An AI-generated course with no stakes, no cohort, no credential, and no artifact at the end will round to ZERO. 5. Feedback & remediation Explaining is the easiest thing a teacher (or LLM) does. The hard thing is pedagogical content knowledge: standard misconceptions, distractors that diagnose them, or choosing the next worked example based on your specific error. This is NOT something that LLMs can do ad-hoc for any domain. 6. Difficulty If learning feels too smooth (easy, fast, frictionless), we produce less of it. Instead, the conditions that feel effortful, like retrieval, spacing, interleaving, generating answers before you're shown them, produce more. This is where the incentives break. These apps are engagement-optimized, and engagement optimization selects against difficulty, because struggle causes churn. "Fits into your busy schedule" gives it away, since no meaningful learning can hppen in five frictionless minutes a day. 7. Motivation These apps capture the weakest motivation that exists: the idle curiosity spike. "I've always wanted to understand LLMs" at 11pm is not what it takes to chew threw hard to understand materials. There's no external need, no cost to quitting, and the product is single-player by design. When nobody notices when you stop showing up, everybody stops showing up. The good news: these are all solvable. Just get someone with actual science of learning expertise on board.