Jev, an AI decision model created by TypeSafe AI, has completed a playthrough of Pokémon Red in 37 hours and 40 minutes without using any scripts, cheats, or direct memory writes. According to the project's GitHub repository, the model was guided only by a harness that reads the game's RAM and presents Jev with a set of legal actions accompanied by relevant facts, from which it chooses the next button press.
The image illustrates the gameplay setup described in the article, showing the AI's interaction with the game through the harness that provides legal actions based on memory reads.
The harness supplies contextual information such as the player's location, party status, and on‑screen text, then computes possible moves—ranging from menu selections to battle actions—and adds facts like type matchups, damage estimates, and progress toward the next story milestone. Jev’s decision-making is limited to selecting from these options; it never writes to game memory. To mimic human play, the system enforces house rules: text speed is set to fast, battle animations are disabled, the player is named JEV and the rival BLUE, and each caught Pokémon receives a unique, made‑up nickname spelled letter by letter. Loop protection prevents the model from repeating ineffective actions by tagging tried choices and, if necessary, substituting an alternative or reloading a recent milestone checkpoint.
Over the course of the playthrough Jev made 16,150 decisions, consuming roughly 39.2 million input tokens at a cost of about $0.042 per million tokens, for a total expenditure of approximately $1.65. The median time per decision was around 0.4 seconds. The final team consisted of Charizard (level 83), Graveler (62), Nidoqueen (45), Beedrill (44), Haunter (39) and Primeape (29), with the project noting 16 team wipes (14 occurring during the Elite Four) and 15 attempts to overcome the Elite Four. These figures illustrate both the efficiency of the approach and its reliance on the supplied milestone knowledge and loop‑protection mechanisms.