On Tuesday afternoon, Anthrope launched Claude plays Pokémon On Twitch, a live broadcast of the new AI of Anthrope model, Claude 3.7 sonnetPlaying a Pokémon Red game. It has become a kind of fascinating experiment, which shows the capabilities of the current AI technology and people’s reactions.
IA researchers have used all kinds of video gamesof Street fighter to PictionaryTo try new models, often more for fun than utility. But Anthrope said that Pokémon proved to be a useful reference point for the Claude 3.7 sonnet, which can effectively “think” Through the types of puzzles that the game contains.
Like OpenAi O3-mini and Deepseek’s R1Claude 3.7 Sonnet can “reason” through difficult challenges, such as playing a video game designed for children. While the unreasonable predecessor of the model, Sonnet Claude 3.5He failed the beginning of Pokémon Red, leaving the player’s house in Pallet Town, Claude 3.7 Sonnet managed to win three badge of gym leaders.
However, the newest Claude still has problems. Hours in the contraction current, the model was dissuaded by a rock wall, which could not walk no matter how hard he tried.
A Twitch user summed up the situation in this way: “Who would win, an AI of the computer with thousands of hours to program it, or 1 rock wall?”
Finally, Claude realized that he could navigate around the wall.
On the one hand, it is frustrating to see Claude traverse Pokémon Red with the speed of a LazyReasoning through each step with unbearable contemplation. However, it is also strangely convincing. The left of the transmission shows Claude’s “thought process”, while the right shows a real -time game.
At one point, Claude tried to locate Professor Oak inside his laboratory, but he was confused, because there were other NPC on the scene.
“I notice that a new character has appeared under me, a character with black hair and what seems to be a white coat in coordinates (2, 10),” Claude wrote. “This could be Professor Oak! Let me go down and talk to him. “
Claude later proceeded to speak erroneously with an NPC other than the processor, an NPC with which the model had spoken several times before. Some of the thousands of people on the Twitch chat began to get anxious. Others, particularly those who had been seeing the transmission for more than a few minutes, were less worried.
“The boys relax,” a person wrote in the chat. “Before leaving and entering the Oak laboratory as 10 times before understanding how to move on.”
For Twitch users for a long time, the Anthrope transmission format can feel nostalgic. More than a decade ago, millions of people tried to play Pokémon Red at the same time in a first online social experiment of their kind called Twitch plays Pokémon. Each user could control the player’s character through the chat twitch, resulting in a predictably chaotic game.
Some AI researchers have summoned Twitch Plays Pokémon as an inspiration for their work. In October 2023, the Software Engineer based in Seattle Peter Whidden published a YouTube video detailing how he trained a reinforcement learning algorithm to play Pokémon. His AI spent 50,000 hours playing the game Before he learned to sail successfully. A challenge was that the AI preferred to admire the pixelated landscape instead of playing the game.
The “recreations” of Twitch promoted by AI play Pokémon as Whidden’s and Anthrope’s are entertaining, but a bit bittersweet at the same time. The original current was such a crucial moment in the history of Twitch because it brought together people in an unexpected way. All were in the same team, working towards the goal of getting the player’s character to stop running in circles and really progress in the game.
In 2025, it seems that we are no longer teammates, but the spectators, see a model of trying to play a game that many of us get used to when we were five years old. It is a microcosm motivated by a broader trend: our online experiences move from communal activities shared to more solitary.