THE CRUNCH
A developer has trained a small AI model on a single RTX 3080 Ti to 'play' Pokémon Red. The model, built on LeCun’s LeWorldModel research, uses compressed summaries of 192 numbers to predict screen changes. It was trained on over 42,000 grayscale frames from a save in Professor Oak’s lab, including scripted routes, random button presses, and wandering. The model operates without an initial reward signal, learning goals through planning with 14 button presses and 512 plan samples. After a fine-tuning attempt, the model successfully selected the starter Pokémon Squirtle in 52 out of 100 runs, compared to zero for random button presses.
The model is a world model with about 12.5 million parameters, designed to be trainable on consumer-grade hardware. It learns what each button press does by observing the screen output rather than predicting the entire next frame. The training data included scripted routes, those same routes with random presses mixed in, and random wandering to prevent the model from associating A with dialogue boxes. The developer hypothesises that small errors compound when predictions are made from the model’s own predictions, which may have caused an initial failure to obtain a Pokémon.
The approach follows a March 2026 paper detailing a single JEPA model with about 15 million parameters, trainable on a single GPU. The developer stated that simply making the current model bigger would not likely lead to the game being beaten, suggesting that difficulty scales exponentially with plan length. The project is one of several recent attempts to beat the game using JEPA models, though those use more sophisticated approaches. The code is open source and available on GitHub as lePokeRed, with a CUDA GPU recommended.