Google’s Gemini panicked when playing Pokémon

AI companies fight to dominate the industry, but sometimes fight in Pokémon Sports Halls.

Like Google and Anthropical Both learning how the most recent AI models are managed by the Early Pokémon, can be as fun as they educate the results – and this time there is Google deprmind Written in a report This Gemini 2.5 Pro causes panic when close to the death of Pokémon. This is due to the EU’s activities, according to the report, a quality degradation in the ability to justify the model “can cause to live.

AI Evaluation – or, the process of comparing the performance of different AI models – a Suspicious art often provide Less context for the real opportunities of a given model. But some researchers think that To learn how to play video games of AI models can be useful (or, at least, a funny type).

During the last few months, two developed two developments with Google and Anthropic, set up appropriate leap streams called “Twins are playing Pokémon“And”Claude Pokémon plays“Everyone tries to walk a children’s video game 25 years ago in real time as an AI.

Each flow shows the “reasoning” process of the AI ​​- or shows that how the EU has evaluated and responds to the answer – this models respond to the road.

Although the advancement of these AI models is impressive, they are still not very good in playing Pokémon. It takes hundreds of hours for the gem, through a game where a child can finish in less time in the exponent.

An AI Navigate is interesting to watch a Pokémon game, it’s not so much to complete the time, rather than how much time behaves.

“During the playthrough, the Gemini 2.5 Pro is becoming various situations that cause the model to simulate” panic “,” he said.

This situation of panic can result in the deterioration of the model performance because the EU can suddenly stop using certain means to use certain tools. Even if AI does not think or does not live in emotion, his actions are amazing, but a disturbing answer, which is poor, in the rush decisions when a person is under stress.

“This behavior occurred in a sufficient number of cases that Twitch Chat members actively noticed when the formation occurred,” he said.

Claude also demonstrated some interesting behaviors in their trips along the Canto. In an example, the player will return to the center of “White” and Pokémon, when the EU is exhausted by all Pokemon’s health.

When clinging to the CLOOD MT Cave, it will be taken to the center of Pokémon in the next city in the cave.

But how the game works. When all Pokémon dies, you can return to the Pokémon Center you use the nearest geographically. They watched the audience in horror because AI tried to kill himself in the game.

Despite the shortcomings, there are several ways that EU can be superior to human players. According to the Gemini 2.5 Pro, AI can solve puzzles with effective accuracy.

With some people, the AI ​​Agent, the agent of the agent – in cases created by the Gemini 2.5 Pro in certain tasks – created to solve the boulder puzzles and find effective routes to reach a place.

“Explaining the description of a proposal that describes only one suggestion and how to check a valid path, 2.5 Pro is able to hit some of these complex boulders, which is required to advance through victory,” he said.

Gemini 2.5 Pro It is said that Google can be able to create these tools without human intervention, because it does many things in its establishment. Who knows, maybe the Gemini will affect himself to create a “panic” module.