EXPERIMENT 04 / ONE MOVE AT A TIME
World 1—1
Run. Jump. Don’t overthink it.
A tiny model meets a familiar
first level.
Play instantly with the keyboard, watch the rules, or download Qwen to try the same level with local AI.
Fast inference. Continuous play. Visible assistance.
Qwen chooses one complete maneuver from typed options. Its instructions stay cached across updates; only the small, changing option table is processed. Each model update scores one field. Jump press, hold and release are executed by the maneuver controller.
The physics guard simulates short trajectories, filters unsafe maneuvers and rechecks an answer when it arrives. It can brake while waiting. This is a hybrid controller: Qwen chooses, local physics handles feasibility and timing. The game and enemies keep moving during inference. There is no stored route or scripted jump fallback.
Raw buttons preserves the original three-field experiment without a physics guard. Decision steps is available only there. Prompt work reused means avoided token evaluations, not a claimed latency multiplier.
This is a playable recreation of the first Super Mario Bros. level, with original code and drawn artwork. Its physics and details are approximate. A valid control can still be a bad move. Controls, SDK and measurements ↗