Model as engineer

Emerging trend in LLM robot control: instead of the large model controlling the robot directly from its own priors, it writes the simulation environment, the variables, or just a piece of software that performs the action it's trying to learn. The model is the engineer, not the pilot.

Code vs RL policy: at bottom both are the same kind of thing, a function from observations to actions. The difference is where it comes from. Code is written: explicit, readable, the model's hypothesis about how the world works; strong on structure and logic, brittle where reality breaks assumptions (friction, contact, slip). A policy is found: optimised against experience, opaque weights that soak up messy physics. Code is compressed understanding; a policy is compressed experience.