The artificial intelligence that plays (and wins) to Pokemon on Twitch

The latest artificial intelligence model developed by Anthropic, Claude 3.7 sonnet is playing live on Twitch at ‘Rosso Pokemon’, obtaining excellent results and winning the admiration of onlookers and fans. The experiment, called “Claude Plays Pokemon”, was launched with a long thread on Twitter. The company told how after more than 24 hours of play, the results obtained are already excellent, especially compared to the previous model, Claude 3.5, which had found many limits in approaching this challenge and struggled even if only to move within the Pokemon universe.

The screen it transmits live, and that tells how several thousand people follow the test,It is divided into two parts: the left side shows the reasoning and thought of the gaming experience while the right side instead is dedicated to gameplay in real time. The comments of the participants are also divided into two large categories of judgment: there are those who complain about excessive slowness and those who, on the other hand, are admired to see the spirit of adaptation of ‘Claude’ in however complicated situations.

Using his learning skills, and in particular his ability to “extended thought“, AI has constantly improved its performance during the progress of the game. In this way she has refined its strategy to ‘win’ against the teachers and conquer the medals of 3 gyms.

Anthropic has decided to use Pokémon to test the reasoning skills of his latest jewel. But that’s not the first time it happens. In the past, other famous video games, such as Street Fighter or Piccationary, They have long been used by researchers to test their ‘artificial creatures’ in relaxed but competitive playful contexts.Even Chatgpt has tried, in the past, to try their hand at Doom.

But it’s not all roses and flowers. Tech Crunch told how after hours of streaming on Twitch, iThe model was stuck in front of a rock wall that could not overcome despite all possible efforts. How did it come out? In the end he understood that he could stop getting incapable and “get around” the obstacle. He also had problems in recognizing Professor Oak, openly publishing his doubt about it. And reading his reasoning process, always reported by the American Tech site, one can remain as fascinated (at least for the spirit and the approach): “I noticed that a new character appeared under me, a character with black hair and what seems like a white coat. It could be Professor Oak! Let me go down and talk to him”. After all of this, did you become curious to see it at work? Well, then you can also follow the live here. At least until the game is over.