Fun
The internet has a long history of tweaking Pokémon’s formula to create chaos. Like Twitch Plays Pokémon’s 2014 crowd-controlled madness, projects that pit non-human intelligence against Nintendo’s iconic RPG have become a cultural mainstay. Now, Anthropic’s Claude 3.7 Sonnet has entered the fray—and its slow, methodical, often baffling gameplay has become an unlikely fixation for thousands.
The premise is simple: Claude analyzes gameplay frames from an emulated Pokémon Red in real time, types out its reasoning in a side panel, and translates that deliberation into button presses. The Twitch stream, titled Claude Plays Pokémon, blends nostalgia with cutting-edge AI in a way that’s both hypnotic and intensely relatable. When Claude spends 20 minutes trying to leave Viridian City’s Pokémon Center because it mistakes the protagonist’s red hat for the exit carpet, only to later realize it never left, you feel that.
The original Twitch Plays Pokémon was defined by its frenetic energy—thousands of viewers spamming commands simultaneously created a kind of digital hive mind that zigzagged through Kanto. By contrast, Claude’s playthrough is meditative to the point of absurdity. The AI pauses for seconds at a time, parsing pixelated menus and overworld sprites before deciding whether to mash A through dialogue boxes or attempt rudimentary pathfinding. Its cognitive process, displayed verbatim on-screen, veers between strategic deliberation ("I should check Bulbasaur’s HP before battling this trainer") and existential confusion ("Is Nurse Joy’s counter part of the wall texture?").
This approach hasn’t been without growing pains. Claude struggled mightily with Mt. Moon’s labyrinthine caves, spending nearly a day navigating identical-looking rock formations. It developed a peculiar habit of fleeing from wild Zubats, later deducing (correctly) that its underleveled team couldn’t survive prolonged encounters. Yet the AI also demonstrates flashes of ingenuity, like using in-game currency to stockpile potions when it recognizes the dangers ahead.
At first glance, Pokémon Red seems like an odd testbed for a cutting-edge language model. But its structure—finite world, predictable battle mechanics, binary win conditions—makes it ideal for evaluating AI’s spatial reasoning and long-term planning. Viewers watch in real time as the model grapples with concepts like inventory management, type matchups, and saving progress—skills many human players take for granted after decades of JRPG conventions.
The experiment also highlights the challenges of translating visual data into actionable inputs. While Claude can parse game text and static menus with ease, interpreting dynamic overworld sprites remains a hurdle.
Unlike 2014’s anarchic Twitch Plays Pokémon, which relied on collective human randomness, Claude’s run is a solo performance—a curated display of artificial intuition.
For now, though, the appeal is simpler: Watching an overqualified AI fumble through a 1996 Game Boy title creates a strange kinship. When Claude painstakingly combs through Vermilion City’s trash cans for hidden items—a tedious but necessary task for any completionist—it’s hard not to root for the digital underdog. Even if victory comes at 1/60th of human speed.