Engineer in the loop

Sam's tentative (self-described flawed) thesis: if the universe is Turing complete and anything can be code, a capable model could run robots by writing, testing, analysing and rerunning code, with no direct control. Not for generalist robots but for purposeful ones, where the task and environment are well defined. Arm 2 was cheaper and better than direct control. Open: does code break on any small perturbation of its assumptions, and can more testing abstract that away? Side question: how much leniency can code carry. Sam's flagged bias: he hasn't trained an RL policy yet (first one will be Arm 3).

Pitfalls raised: (1) representable isn't findable; Turing completeness doesn't help, information and search do (see Verification gap). (2) Cost comparison flatters code: Arm 1 is physically impossible in reality, so the fair rival is PPO training cost. (3) Test-and-rerun is free in sim, costly on hardware, unless you build a sim of the cell, which is a gym, so RL comes back into play. (4) Testing more only covers what you test (Arm 2: perfect inside the range, drops outside). (5) Code isn't brittle per se; feedback loops like PID survive the real world. Brittle assumptions are the problem. (6) The incumbent is already hand-written code in fixed industrial cells, so the real thesis is the LLM replacing the integrator.