Making AI better at math tutoring
Khan Academy’s push to improve its AI math tutor, Khanmigo, is prompting debate over whether large language models can meaningfully replicate or scale one‑on‑one tutoring. Commenters weigh the potential for cheap, personalized help and democratized access to math education against concerns about accuracy, pedagogy, missing evidence of effectiveness, legal and cost constraints, and broader risks of offloading core skills to opaque AI systems. Many argue that deliberate practice, good curriculum design, and human guidance remain central, with AI best seen as an adjunct rather than a replacement for teachers.
Motivation and Nonprofit Status
- Some see the blog post as engagement marketing; others stress Khan Academy’s nonprofit status and mission of free, world‑class education.
- Debate over what “nonprofit” really means: revenue can still fund high salaries, but enriching private individuals via “profit” would be illegal.
- Many commenters argue the org’s history and behavior suggest genuine educational aims, not just monetization.
Effectiveness and Evidence
- Several ask for rigorous results specifically on Khanmigo, not just core Khan Academy.
- Links are shared to existing RCTs showing Khan Academy’s impact and to a registered RCT in progress for Khanmigo.
- Some are frustrated that after a full pilot year, Khan Academy hasn’t yet published concrete learning gains for the AI tutor.
Cost, Scalability, and Access
- Strong argument that AI tutors are pursued because they’re cheaper and more scalable than human tutors, especially for global access.
- Others note high-quality human tutoring is unaffordable for many, so AI is a pragmatic democratization tool.
- GPU and API costs vs $4/month pricing are discussed; some think usage patterns and bulk deals make it viable, especially with donor support.
Pedagogy and Learning Models
- Multiple commenters emphasize that effective math learning hinges on deliberate practice, spaced repetition, mastery progression, and frequent assessment.
- One camp is skeptical of conversational AI, arguing pre-authored, carefully sequenced explanations plus adaptive problem sets are more realistic and controllable.
- Others value interactive, always-available explanations, especially for students failed by traditional instruction.
Quality and Limits of Current AI Tutors
- Positive views: AI can flex from elementary concepts to advanced topics, is patient, nonjudgmental, and available anywhere.
- Negative views: current models often hallucinate, struggle with graduate-level material and proofs, repeat the same explanations, and don’t truly model student understanding.
- Some insist AI can never “reason” like humans; others dispute this, leading to a long computability/AI-theory side debate.
Technical and Ethical Concerns
- Noted improvements: using a calculator backend for numeric work, prompt/chain‑of‑thought tuning, and experimenting with newer models (GPT‑4 Turbo, GPT‑4o).
- Some want full transparency on prompts, techniques, and benchmarks, given the nonprofit mission.
- Serious worry about OpenAI’s restrictive terms for student data/use and the lack of clarity on how those apply, seen as potentially undermining student trust.