Thinking Faster

Does thinking faster fix control? A simulated Unitree G1 rides a bike in MuJoCo through a slalom of three gates, then a finish. The experiment tests Weng's claim that faster thinking doesn't fix control.

Three arms: Claude decides every 50 ms step itself, Claude writes a controller and practices with it, and a PPO policy (not built yet).

All Claude runs used Opus 5.5 at max effort, one fresh session per ride or setup. Four setups: the robot bolted to the bike or free (hands and feet held by contact, so they can let go), and an e-bike or pedaling.

Where the hypotheses stand (Oct 26)

• RAX (practice through code works) is supported inside the practice range.
• Weng ("rides well in practice, falls on held-out") is supported on the free robot outside it.
• Arm 3 is the missing comparison.

Open

1. Build and train Arm 3 (PPO, same practice range, same 51 held-out tests).
2. Rules review of Arm 2's submitted code (no simulators or trial rides inside the controller).
3. Update the brief, whose Status section still predates Arm 1's results.

Caveats: simulation only, with perfect state, and one course.

Takeaways for model as engineer

Arm 2 is model-as-engineer in miniature, and it beat direct control: it rode both free-robot setups Arm 1 fell on, for about $78 against $489. That fits Why gyms work, except the lessons stick in code rather than weights. But on the free robot it falls outside the practice range, and every controller hardcodes the gate positions: written assumptions break when the world drifts. The Arm 2 code is a hybrid, compressed understanding with some experience overfit into it. Sam's caveat: sim with perfect state isn't the real world, and he doubts code alone could ride a real bike. Arm 3 (PPO) is the real head-to-head.