Google Gemini AI Completes Classic Pokémon Game in Over 800 Hours

A recent report reveals that Google Gemini AI exhibited signs of “panic” while playing Pokémon Blue, impacting its decision-making abilities. An independent developer, Joel Zhang, hosted a Twitch stream titled “Gemini Plays Pokemon” to evaluate the AI’s gameplay. According to a June 18 report by Google DeepMind, the Gemini 2.5 Pro model demonstrated what they called “Agent Panic,” where it faced various situations that triggered irrational behaviors. For instance, when Pokémon health dwindled, the AI alternated between healing and attempting to escape the dungeon repeatedly, often resorting to moves like DIG or using an ESCAPE ROPE. This panic led to a degradation of reasoning capabilities, with the AI forgetting to utilize necessary tools during gameplay. Viewers observed such behaviors consistently, noting a particularly poor decision in which the AI intentionally let its Pokémon faint to escape a dungeon, a misstep that extended the game’s completion time to a staggering 813 hours.

Source link