Let an AI model play and finish Pokemon Blue live on a stream
Software engineer Joel Zhang built a harness that let Google's Gemini 2.5 Pro play Pokemon Blue on a Twitch stream. The model finished the game after about 37 days, which Google CEO Sundar Pichai shared publicly.
- Difficulty
- Builder
- Setup time
- Not measured yet
- Monthly cost
- Varies
- What it saves
- Not measured yet
Also reported here: deepakness.com.
How to do it
- Run a Game Boy emulator you can control from code
- Send the model a screenshot plus key game state (location, party) on each turn
- Give it a small set of tools: press buttons, and save and read notes
- Log its reasoning and stream the run so you can see where it gets stuck
- Tweak the harness, not the game, when it loops
The prompt
Write a Python harness that connects a Game Boy emulator (PyBoy) to the Gemini API. Each turn, capture a screenshot and read the player's map location from memory, send both to the model with a short goal, and let it reply with a JSON list of button presses. Keep a notes file the model can append to and read back. Log every turn to a file.
Where this comes from
Two or more independent people say it worked. Original report: Cryptopolitan.
Tried this? Tell us how it went and help move its proof level up.