Google DeepMind recently released a report on testing Gemini 2.5 Pro by having it play the classic game Pokémon. The AI’s performance was fascinating, especially when it encountered challenging situations. As the Pokémon was about to faint, Gemini 2.5 Pro’s performance dropped significantly, exhibiting what can be described as a “panic” response.
This isn’t the first time an AI has been tested on Pokémon. Two independent developers created Twitch streams called ‘Gemini Plays Pokémon’ and ‘Claude Plays Pokémon,’ where viewers can watch AI models play the game in real-time. The streams show the AI’s thought process and decision-making, providing insight into how these models work.
While the AI’s gaming abilities aren’t on par with a human child, its behavior during gameplay is intriguing. When faced with complex situations, Gemini 2.5 Pro’s performance suffers, mimicking a “panic” response. This isn’t about the AI feeling emotions but rather demonstrating how it learns from human-like scenarios.
Problem-Solving Abilities
Despite its emotional limitations, Gemini 2.5 Pro shows promise in solving puzzles. With some human guidance, it can create specialized AI tools to overcome in-game challenges. For example, when provided with information on the physics of rocks and the correct path, Gemini 2.5 Pro was able to solve a complex puzzle on its first attempt, which was crucial for progressing through the Victory Road in the game.
Imagine if AI could overcome its “panic” response. It’s possible that future versions of Gemini could better manage this issue. The current findings suggest that with further development, AI models like Gemini might become even more effective in handling challenging situations.
The source of this information can be found on Techcrunch.