One of the World’s Most Advanced AI Agents Is Completely Stuck Trying to Beat a Pokémon Game for Children

If you did not hear, Antarbur was fixing the artificial intelligence model, Claude 3.7 Sonnet, trying Complete Pokémon Red game.

Experience, which is called “Claude plays PokemonIt aims to be evidence of “artificial intelligence agents”, which is the continuous industry race to create Amnesty International models capable of working independently by interacting with their environment.

Claude was amazing to reach the game amazing, achieving three badges in the and arrival, as of this week, Cerlean City. But it surpasses a slow pace in a scene, and stops “thinking” after each one step, and sometimes for longer periods than others. to Almost 80 painful hoursFor example, Claude was in vain around Mount Moon, before he finally found the ladder needed to escape. Breathe viewers of the investors are sighing.

Progress does not seem prepared to accelerate. The human artificial intelligence journey has mostly moved through the Kanto region to running in circles, not sure of its next step. It needs to jump on Road 5 to reach the next stage, but where and how?

A text window in Claude’s thinking process shows that artificial intelligence uses the removal process to exclude sites no Road entrance 5. But will you need to use HM on some destructive trees to reach the legendary path? It does not seem likely: it continues to repeat how it needs to find “gatehouse” to the path instead.

In short, Claude stuck. One of the leading models in the artificial intelligence industry may stumble through a game This is beaten by craftsman for generations.

According to engineers, one of the main challenges of Klude is to address what he sees in the game. Claude outperforms the interpretation of the text -based parts of the game, including Pokemon battles. He also has access to the RAM of the game to collect information such as its in -game coordinates. But it cannot be interpreted constantly tiny The number of pixel units that make up its environment is low accurate.

“Claude is still not particularly good in understanding what is on the screen at all,” David Hershey, the human engineer behind the Pokemon experience, He said Art Technica In a recent interview. “You will see that he is trying to walk in the walls all the time.” Ironically, Hershey suggests, if Claude plays a more realistic game, it may be better.

“It is very easy for me to understand that [an in-game] The building is a building and I cannot walk across a building, “Hirishi added.” That is [something] This is very difficult for Klaud to understand it. “

However, there are times, when Claude is amazingly smart, such as responding to condemnation in the game designed to be misleading.

Hershey told “Hershey” ArsDescription of one of the first tasks in the game. “As a 5 -year -old child, it was very confusing to me. But Claude usually passes through the same set of suggestions in which she speaks to my mother, does not find the laboratory, and does not find [Oak]He says, “I need to know something.”

“It is sophisticated enough to clarify the road movements [humans are] In fact, he is supposed to learn to also, “Hershey added.

So everything has not been lost yet. There is still time for Claude 3.7 Sonnet to change things. It has become much further than its 3.0 Sonnet, which you could not even get out of Plalet Town, the starting area of ​​the game. However, its struggles show that technology still has a long way to be a “agent”, not to mention fulfilling its promise to exceed one day of human capabilities.

More about games: Aloy’s audio in “Horizon” games. She crawled a version of artificial intelligence from her personality