Google Gemini Beats Pokémon Blue With AI Assistance
Google’s powerful AI model, Gemini 2.5 Pro, has completed the classic 1996 video game Pokémon Blue, marking a notable achievement in artificial intelligence’s ability to navigate complex tasks. While the feat was orchestrated by independent software engineer Joel Z through a livestreamed project called Gemini Plays Pokémon, Google executives, including CEO Sundar Pichai, celebrated the milestone publicly.
Joel Z clarified that this was not an official Google experiment, but Google AI Studio’s product lead Logan Kilpatrick and Pichai both followed the project closely, jokingly referring to Gemini’s progress as “Artificial Pokémon Intelligence.”
Gemini used a framework known as an agent harness, which provided the model with real-time game screenshots and additional metadata, enabling it to interpret, decide, and simulate button presses to move through the game. This mirrors similar experiments from rivals like Anthropic, whose Claude model is also trying to finish Pokémon Red — although Claude hasn’t completed the game yet.
Despite requiring interventions, Joel Z defended the approach: “I don’t give specific hints or walkthroughs,” he said, explaining that minor fixes — like addressing a game bug involving the Lift Key — helped improve Gemini’s reasoning, not cheat the system.
The Gemini Plays Pokémon project remains under active development, with future improvements likely. While not a definitive benchmark of AI superiority, it does highlight how large language models, when paired with specialized tools and minimal guidance, can handle unpredictable, multi-step challenges in gaming environments.
