Google’s Gemini AI Conquers Pokémon Blue (With Some Help)
▼ Summary
– Google’s Gemini 2.5 Pro AI model reportedly beat the 1996 game “Pokémon Blue,” as announced by CEO Sundar Pichai.
– The livestream was created by an independent software engineer, Joel Z, but received public support from Google executives.
– The project was inspired by Anthropic’s Claude AI playing “Pokémon Red,” though Claude hasn’t yet completed the game.
– Both AI models rely on agent harnesses to process game screenshots and make decisions, with Joel Z clarifying that interventions improve reasoning but aren’t direct walkthroughs.
– Joel Z emphasized that the Gemini framework is still evolving and cautioned against direct comparisons between AI models due to differing tools and inputs.
Google’s advanced AI model Gemini has achieved an impressive feat by completing the classic GameBoy title Pokémon Blue, marking a significant milestone in artificial intelligence capabilities. The achievement was celebrated by Google CEO Sundar Pichai, who shared the news on social media, though the project itself was spearheaded by an independent developer named Joel Z.
While not an official Google initiative, the Gemini Plays Pokémon livestream has garnered attention from the company’s leadership. Logan Kilpatrick, Google AI Studio’s product lead, previously noted Gemini’s progress in the game, highlighting its ability to outperform other models in earning badges—a key milestone in the Pokémon universe.
The choice of Pokémon as a benchmark isn’t arbitrary. Earlier this year, rival AI firm Anthropic demonstrated its Claude model playing Pokémon Red, emphasizing how the game serves as a test for AI reasoning and adaptability. Both versions, released in 1996, remain beloved entries in the franchise, making them ideal challenges for AI systems.
However, Joel Z cautions against direct comparisons between Gemini and Claude. “This isn’t a definitive benchmark for AI performance,” he explains. The models operate with different tools and receive distinct inputs, making fair competition difficult. Additionally, neither AI plays the game entirely on its own—both rely on external frameworks that translate game visuals into actionable data, allowing the AI to make decisions.
Human intervention played a role in Gemini’s success, though Joel Z clarifies that these adjustments weren’t shortcuts. Instead, they refined the AI’s reasoning without providing direct solutions. For example, Gemini was informed about a minor game bug requiring repeated interactions with an NPC—a fix later implemented in Pokémon Yellow.
The project remains a work in progress, with ongoing improvements to Gemini’s decision-making framework. As AI continues to evolve, mastering complex, unpredictable tasks like retro gaming showcases its growing versatility—one badge at a time.
(Source: TechCrunch)
