Thinking Faster

Does thinking faster fix control? We teach a simulated humanoid (Unitree G1) to ride a bike three ways: Claude deciding every step with unlimited time to think, Claude writing and practicing its own controller, and an RL-trained policy. Then we shove it, load it with a backpack it can't see, and watch who stays up. Very visual, and a data-driven reply to Weng's essay.

• Brief (current design): https://claude.ai/artifact/Tzba9c2hMzmhy4uvwSLdyb
• Code: experiments/thinking-faster/