JEVfireEXPERIMENT 03 / LOCAL AI

THE PROMPT IS THE PERSONALITY

Slipstream

Three laps. Four competing strategies.
Push your luck. Or outlast the field.

Choose a controller00:00
Qwen chooses maneuvers, boost and pit stops. Grip, tyre wear and damage decide whether the risk pays off.
JEV PROVING GROUND

Three laps. One winner.

Prompt drivers Scripted traffic
Click a car or its number to inspect its next move.
Fleet decisions / sec
Fleet rounds / sec
Mean fleet inference
Render FPS
Prompt work reused

Load the local model, or take a scripted drive to explore the circuit.

  1. The grid is ready.
What is actually driving the cars?

Qwen sees race position, traffic gaps and closing speeds, grip, the next corner, tyre condition, damage, boost, and pit distance. It scores legal speed changes, safe or risky overtakes, boost, and pit requests. Code assigns the winning action to that car’s known field.

All eligible cars enter one fleet request together. WebLLM scores their independent prompts sequentially on one model, reusing eligible cached prefixes. The next fleet request starts immediately, with no artificial pause. This browser build does not expose simultaneous multi-sequence GPU execution. Steering, smooth lane changes, traffic, and optional following assistance run between AI decisions. Tight overtakes temporarily waive that assistance. Excess corner speed can cause a spin, damage, or retirement. Pitting costs time but repairs damage and renews tyres and boost. Game rules never count as AI ticks.

Fleet decisions/sec counts applied car choices. Four cars at one update/sec means four decisions/sec and one fleet round/sec. Mean fleet inference is the time for the whole request, not one car. Per-car rates appear on the driver cards. Only known fields and actions can enter the result. A valid action can still be a poor driving choice. Our small comparison found faster scoring but worse live race outcomes. See the speed and quality measurements ↗ Read the driving notes ↗