GEMINI from Google felt panic when playing Pokémon

Artificial intelligence companies are fighting to control the industry, but sometimes they are also fighting in Pokemon.

like Google and man Both students are studying for them on early Pokémon games, the results can be entertaining as much as you enlighten – this time, at Google DeepMind Written in a report Gemini 2.5 Pro is spreading to panic when Pokemon is close to death. This can cause artificial intelligence experience “a deterioration that can be qualitatively observed in the typical thinking capacity,” according to the report.

AI measurement – or, comparison of the performance of various artificial intelligence models – Doubled art Which often saves Small context The actual capabilities of a specific model. But some researchers believe this Study how artificial intelligence models play video games It could be useful (Or, at least, a kind of funny).

Over the past few months, two developers are not active with Google and Anthropology have created related stagnant flows “”Gemini plays PokemonAnd “and”Claude plays Pokemon“Where anyone can watch it in an actual time as Amnesty International has been trying to move in a video game for more than 25 years.

Each flow of “thinking” from artificial intelligence – or, or translating a natural language for how artificial intelligence evaluates a problem and reaches a response – gives us an insight into the way these models operate.

While these artificial intelligence models offer impressive, they are still not good in Pokemon playing. It takes hundreds of hours to cause Gemini through a game that the child can complete in a large time.

The interesting thing about watching AI to move in the Pokémon game is not related to the time of completion, but how to behave along the way.

The report says: “Throughout the game, Gemini 2.5 Pro enters into different situations that simulate the” panic “model,” the report says.

This state of “panic” can increase the performance of the model, as artificial intelligence may suddenly stop using certain tools at its disposal to extend the game. Although artificial intelligence does not think or test emotion, its actions simulate the way a person may make poor decisions when they are under pressure – a great but annoying response.

The report says: “This behavior happened in sufficient separate situations that Twitch chat members noticed when it occurs,” the report says.

Claude also showed some strange behaviors on her trips on Canto. In one case, artificial intelligence picked up the style that when Bokimon is running out of health, the player’s personality will return to the Bokimon Center.

When Claude stumbled in the Mount Mon Cave, she mistakenly assumed that if she deliberately got every Pokemon, it will be transferred through the cave to the center of Bokimon in the next town.

However, this is not how the game works. When Pokémon dies, you return to any focus that you recently used, instead of the closest geography. Watch the viewers in horror while artificial intelligence tried to kill himself in the game.

Despite the shortcomings, there are some ways that artificial intelligence can excel over human players. As of the launch of the Gemini 2.5 Pro, artificial intelligence is able to resolve the puzzles with great accuracy.

Through some human assistance, artificial intelligence has created Agentic tools – required in front of Gemini 2.5 Pro towards specific tasks – to solve rock puzzles of the game and find effective ways to reach a destination.

“With only a claim describing Bulder’s physics and a description of how to verify a valid path, Gemini 2.5 Pro is able to shot one of these complex rocky puzzles, which must progress through the victory road,” the report says.

Since Gemini 2.5 Pro has made a lot of work to create these tools on its own, Google looks that the current model may be able to create these tools without human intervention. Who knows, Gemini may treat itself in creating a “no panic” unit.