On Tuesday afternoon, Anthropic debuted Claude Plays Pokémon on Twitch, featuring their newest AI model, Claude 3.7 Sonnet, playing the classic Pokémon Red game. This livestream has become a fascinating experiment, demonstrating the current capabilities of AI technology while capturing the reactions of viewers.
AI researchers have long used video games, from Street Fighter to Pictionary, to test and showcase new models. While these experiments are often more for amusement than practical use, Anthropic chose Pokémon as a benchmark for Claude 3.7 Sonnet, which can successfully “think” through the puzzles and challenges the game presents. Unlike its predecessor, Claude 3.5 Sonnet, which failed early in the game, Claude 3.7 managed to win three gym leader badges, showing a significant improvement in reasoning abilities.
Despite its progress, the model still encounters challenges. Several hours into the stream, Claude found itself stuck at a rock wall, unable to find a way through. One Twitch user humorously summarized the situation: “Who would win, a computer AI with thousands of hours of programming, or 1 rock wall?” After some time, Claude finally realized it could navigate around the obstacle.
Watching Claude play Pokémon Red becomes a strange mix of frustration and fascination. The model reasons through each step with such deliberation that the gameplay is incredibly slow. On the left side of the stream, viewers can see Claude’s “thought process,” while the gameplay unfolds on the right side in real-time. In one instance, Claude tried to find Professor Oak in his lab but became confused by other NPCs. “I notice a new character has appeared below me — a character with black hair and what appears to be a white coat,” Claude noted, mistakenly assuming it was Oak. Eventually, it spoke to the wrong NPC, causing some impatience among the Twitch viewers. However, some long-time viewers assured others, recalling how they had watched similar moments in Twitch Plays Pokémon, where confusion and repeated failures were part of the fun.
The format of Anthropic’s stream is reminiscent of Twitch Plays Pokémon, the viral 2014 experiment where thousands of people controlled a single Pokémon Red game through Twitch chat, resulting in chaotic and unpredictable gameplay. This earlier experiment became a cultural milestone, bringing people together in a collective effort to navigate the game. Now, watching an AI navigate the same game, it feels like a shift from communal activity to solitary observation. As AI models like Claude 3.7 Sonnet take over tasks that humans once managed easily, it highlights a larger trend: online experiences are increasingly becoming less collaborative and more individualistic.
In contrast to the chaotic teamwork of Twitch Plays Pokémon, today’s experience feels more like watching a spectator sport—where the audience observes an AI struggle through challenges that many of us mastered as children. This shift reflects how our online engagement has evolved from shared, community-driven endeavors to more passive, observer-based interactions.
