Small decisions. Interesting possibilities.Submit contentSubmit

Jev beats Brock in Pokémon FireRed

Jev just beat Brock in Pokémon FireRed in less than 5 minutes. That’s 5 minutes of actual real-world time. You can see the in-game time in the video. This is a completely different way of having an LLM play a game. Jev doesn’t see pixels or directly press buttons. The emulator memory is decoded into the current game state, code generates the legal actions, and Jev decides what to do next. Code handles the execution. The game freezes while Jev thinks, then runs as fast as possible between decisio

Added to Jevfast

How it uses Jev

The emulator memory is decoded into game state and code generates legal actions. Jev chooses an action; code executes it. The game pauses while Jev decides and runs between decisions. Inputs are logged for replay.

App workflow

  1. Decode state

    Code reads emulator memory and converts it to the current game state.

  2. Generate legal actions

    Code prepares the actions Jev may choose from.

  3. Choose and execute

    Jev selects an action and code applies it; the game pauses during the decision.

  4. Log and replay

    Inputs are logged frame by frame so the run can be replayed without model calls.

Why it is interesting

The example cleanly separates state decoding, bounded action choice, and execution; a recorded action log supports replay without additional model calls.