The companies in the AI struggle to dominate the industry, but sometimes they also fight Pokémon gyms.
As Google and Anthropic They both study how their latest Ai models sail for Pokémon Games, the results can be as fun as they are illuminators, and this time Google Deepmind has done it Written in a report That Gemini 2.5 pro resorts to panic when his Pokémon is near death. This can cause the AI performance to experience a “qualitatively observable degradation in the model’s reasoning capacity,” according to the report.
Benchmarking AI –O, the process of comparing performance of different AI models – is a Dubious art that often provides Small context for the real capabilities of a particular model. But some researchers think that Study how AI models play video games could be –lo useful (or at least a little fun).
In recent months, two non -affiliated developers with Google and Anthropic have established respective twitch flows called “Gemini plays Pokémon“And”Claude plays Pokémon“Where anyone can see in real time when an I tries to navigate a children’s video game over 25 years ago.
Each flow shows the process of “reasoning” of ai or, a translation of the natural language of how the AI evaluates a problem and reaches an answer, giving us a vision of the way these models work.
Although the progress of these AI models is impressive, they are not very good yet to play Pokémon. It takes hundreds of hours for the twins to reason through a game that a child could complete in less time.
What is interesting to see an IA sailing through a Pokémon game is not so much about its ending, but how it behaves along the way.
“Throughout the game, Gemini 2.5 Pro sinks in various situations that make the model simulate” panic “, the report says.
This state of “panic” can cause the performance of the model to make it worse, since suddenly the AI can stop using certain tools at its disposal for a stretch of play. Although I do not think or experience emotion, its actions mimic the way a human can make poor and hasty decisions when they are in stress, a fascinating and disturbing response.
“This behavior has occurred in cases sufficient that Twitch chat members have been actively realized when it occurs,” says the report.
Claude has also exhibited some curious behaviors on his travels through Kanto. In one case, the AI picked up the pattern that when all her Pokémon is left without health, the player’s character will “go white” and return to a Pokémon center.
When Claude was hooked on the Mont Mon cave, it was mistakenly proposed that, if he intentionally, his entire Pokémon fainted, he would be transported through the cave to the next city Pokémon.
However, this is how the game works. When all your Pokémon dies, return to any Pokémon Center you use more recently instead of the closest geographically. The spectators saw in horror as I tried essentially to kill in the game.
Despite his shortcomings, there are some ways in which the AI can overcome human players. From the launch of Gemini 2.5 Pro, the AI is able to solve puzzles with impressive accuracy.
With a little human assistance, the AI created agents: he asked for cases of Gemini 2.5 Proceded to specific tasks, to solve the games of the game boulder and find efficient routes to reach a destination.
“With just a warning that describes Boulder’s physics and a description of how a valid route can be verified, Gemini 2.5 Pro is able to take a look at some of these boulder’s puzzle complexes, which are required to move through Victory Road,” says the report.
As Gemini 2.5 Pro did a lot of work to create these tools on their own, Google theorizes that the current model can be able to create these tools without human intervention. Who knows, maybe the Gemini will be therapult to create a “non -panic” module.
