Idea: PPO-train a model to generate lessons for you, graded by another model's tests
cephaloform · x · 2026-10-14
A fun thought experiment: can you PPO-train a model to generate personalized lessons for yourself, with another model creating tests — using test performance as the reward signal? Applying RL to a self-learning loop where the AI optimizes how it teaches you.