Weng's claim

Haoyang Weng: an LLM has spatial grounding but not dynamics grounding, and "faster inference alone will not turn spatial grounding into dynamics grounding." Dynamics come from acting and feeling the result at control speed, so the LLM should steer a low-level policy trained on that experience. Challenge rules: unlimited reasoning and code allowed; simulators, MPC and hand-written dynamics rollouts banned. Practice rides aren't mentioned, so we state our reading (ideally ask Weng first).