Pokémon panicked Gemini as a player who is afraid of losing

Deal Score0
Deal Score0

Google Gemini

Google Gemini

Long -awaited, Gemini (Google Bard), the conversational agent fueled by the AI ​​developed by Google, is deployed in the form of a free online service and a mobile application for Android and iPhone devices

  • License:
    Free license
  • Author :
    Google
  • Operating systems:
    Online service, Android, iOS iPhone / iPad
  • Category :
    IA

It is on Twitch, live, that the AI ​​are tested with “Gemini Plays Pokémon” And “Claude Plays Pokémon”. Google Deepmind has documented its conclusions which show the failures of Gemini 2.5 pro with signs of stress when its creatures approach KO according to the firm of Mountain View, the situation causes a “Qualitatively observable degradation of model reasoning capacities”.

Gemini copies the reactions of stressed players

© Shutterstock/Shutterstock/Thrive Studios

Advertising, your content continues below

Faced with this stress, Gemini 2.5 is likely to use certain tools for a long time. If the chatbot is incapable of real emotions, human decisions under stress are reproduced with irrational and disturbing behaviors.

“This behavior occurred in enough separate instances for the members of the Twitch cat actively when it happens”specifies the Google report.

Claude, the AI ​​ofAnthropic who welcomed the co -founder of Netflixalso faces inconsistent behaviors. In one of the caves of PokémonAI thought that making all of its creatures die in a voluntary way would allow you to be teleported to the exit. Except that this strategy brought back to the last Pokémon center visited. Players know it, even children.

Viewers attended this virtual suicide attempt which shows the limits of understanding of current AIs, as chatgpt which struggles in failures in the face of an atari.

On the other hand, AI excels in other areas. Gemini 2.5 Pro perfectly solves the puzzles of Pokémonespecially when you have to move rocks. “With only one prompt describing the physics of the rocks and a description of how to check a valid path, Gemini 2.5 Pro can suddenly resolve some of these puzzles of complex rocks”note the report.

The AI ​​even created specific tools to solve these puzzles with low human assistance. According to Google, Gemini will be able to self-improve in his future versions.

Advertising, your content continues below

Want to save even more? Discover Our promo codes Selected for you.

More Info

We will be happy to hear your thoughts

Leave a reply

Bonplans French
Logo