Google’s Gemini panicked when playing Pokémon

AI companies are fighting to dominate the industry, but sometimes they are also fighting the Pokémon Gym.

As Google And Ethnographic Both studies how their latest AI models navigate to the early Pokémon Games, the results may be just as fun as enlightening – and now Google has dipmind A report That Gemi 2.5 Pro is terrified when its Pokémon is close to death. According to the report, it is said that AI’s performance “can gain the quality of observable degradation in the model’s argument,” the report says.

AI Benchmarking – or, the process of comparing the performance of different AI models – this is a Suspicious industry It often supplies Context For the true ability of the given model. However some researchers think AI models are studying how to play video games May Useful (Or, very least, funny type).

Over the past several months, two unexpected developers with Google and anthropologists have placed Twitch Streams known as “Jemi has played the role of Pokémon“And”Claud Pokémon plays“Where AI tries to navigate the video game of the kid from 25 years ago, someone can see in real time.

Each stream shows the “logic” process of AI – or, a natural language translation of how AI evaluates a problem and comes into a response – gives us insight about the way these models work.

Although the progress of these AI models is impressive, they are not very good to play Pokémon. It takes a few hundred hours to argue Jemini through a game that a child can accomplish significantly in less time.

What is interesting about watching the AI ​​Pokémon game is not so much about the time to finish, but how it behaves in that way.

The report said, “During the plathru, Gemini 2.5 Pro comes in different situations that the model ‘imitates the panic'” “” “” “The report.

This condition of “panic” can make the model performance worse, as AI can suddenly stop using some equipment to expand the gameplay. Although AI does not think or feel the emotions, it imitates the way that a person is under stress and imitates the decision to make decisions – an interesting, anxious reaction.

The report said, “This behavior has occurred in a quite different instance that members of Twitch Chat actively noticed when happening,” the report said.

Claud also showed some curious behavior to travel across Canto. In one example, the AI ​​pattern picked up that when all its Pokémon goes out of health, the player’s character will become “white” and return to a Pokémon Center.

When the clod was stuck in the Mount Moon cave, it incorrectly assumes that if it was intentionally all his Pokémon became unconscious, it would cross the cave and move to the next city’s Pokémon Center.

However, the game is not how it works. When all your Pokémon dies, you will come back to the Pokémon Center you have recently used than the nearest geographic. AI was basically trying to kill themselves in the game, the audience saw the horror.

Despite its flaws, there are several ways that AI can be surpassed by human players. As Jemi 2.5 Pro publishes, AI is capable of solving the puzzle with impressive accuracy.

With some humanitarian assistance, AI has created agent tools – to solve the boulder puzzle of the game and to look for skilled routes to reach a destination – toward specific tasks – encourages the pro -pro -process of Gemi 2.5 Pro.

According to the report, “Geni 2.5 Pro is capable of making a few shots of these complex boulder puzzles, which must be progressed through Victori Road,” the report says, “Just a prompt boulder describes physics and how to verify a valid path.”

Since Gemi 2.5 Pro has done a lot of work to make these tools in itself, Google Theorize that the current model can be able to create these tools without human intervention. Who knows, perhaps Gemini will treat himself in the creation of a “not panic” module.