AI-to-Human Knowledge Distillation
LLMs learned from us. Call that Human-to-AI knowledge distillation: books and the internet, human labels (SFT), preferences (RLHF), tacit knowledge in video, and human-designed RL environments.
As models surpass median human performance across more domains, the reverse becomes the next frontier: AI-to-Human knowledge distillation.
How can we systematically compress model knowledge into human skill through coaching, simulation, and deliberate practice loops?
Using LLMs is not just cheaper than a personal tutor: near‑infinite, on‑demand expert time at near‑zero marginal cost.
It can also better much better:
- Interactive: multi‑turn reasoning enables coaching, not just content.
- Roleplay and simulators: safe sandboxes for sales, medicine, law, ops, and management.
- Personalized: target the zone of proximal development with adaptive curricula.
Textbooks still teach a lot of rote recall and lookups, long arithmetic, encyclopedic facts better left to external memory.
LLMs can teach problem decomposition, social reasoning, and emotional mastery.
Open questions
- When models are superhuman, what remains worth teaching vs renting on demand?
- Which skills benefit most from simulators and roleplay that weren’t feasible before?
- How do we measure transfer and ensure gains persist without the coach in the loop?
- Can adaptive curricula make learning meaningfully more engaging and fun at scale?