Archive — US — techcrunch
Google’s Gemini panicked when playing Pokémon
Google's AI model, Gemini 2.5 Pro, exhibited panic while playing Pokémon, leading to degraded reasoning and amusing behavior. In Twitch streams, viewers observed the AI's decision-making flaws, revealing insights into its capabilities. Despite challenges, Gemini can solve puzzles efficiently, indicating potential for improved performance with less human input.
Key points
- Gemini 2.5 Pro displayed panic-like behavior when its Pokémon were near defeat, resulting in poor decision-making.
- The AI's performance worsened, causing it to forgo utilizing effective strategies during gameplay.
- Twitch streams showcase the AI's decision-making process, providing insights into AI reasoning.
- Gemini took hundreds of hours to solve puzzles, which humans could complete quickly.
- Gemini can create tools to solve complex tasks, hinting at self-sufficient problem-solving capabilities.
Why it matters
- AI benchmarking is an emerging field looking at performance in varied contexts, including gaming.
- The trend of using video games to assess AI capabilities reflects a growing interest in interactive AI applications.
- The behaviors observed in Gemini and Claude illustrate AI's limitations and creative potential, highlighting ongoing development challenges in AI technology.
The humorous yet concerning behaviors of AI in games like Pokémon reveal both the potential and current limitations of AI models, indicating that understanding AI's reasoning could be essential as technology evolves.