Slow but Steady: Anthropic's AI Learns to Play Pokémon

An intriguing experiment is underway on Twitch as Claude 3.7 Sonnet, an AI developed by Anthropic, tackles Pokémon Red with varying success for the third consecutive day. Viewers are witnessing its “thought process” in real time. Back in 2014, millions collectively controlled a character via chat in the Twitch Plays Pokémon project. Now, the tables have turned as a solitary AI slowly but methodically solves puzzles designed for 90s kids.

The mission's start proved just as challenging for the AI as it was for many childhood players. The previous version, Claude 3.5 Sonnet, couldn’t even leave the starting house. In contrast, the “advanced” 3.7 Sonnet managed to earn three Gym Badges within its first 24 hours. However, its victories were interspersed with amusing mishaps: the AI spent hours stuck against a stone wall before figuring out how to walk around it, and in Professor Oak's lab, it twice mistook him for another NPC in a white coat.

For Twitch veterans, the stream is a nostalgic nod to the iconic 2014 experiment when thousands of chat commands collectively controlled the character, creating a surreal gameplay experience. Anthropic's choice of game is no coincidence: Pokémon Red’s mechanics, requiring sequential actions, make it a perfect testing ground for AI “reasoning.” Unlike its predecessors, Claude 3.7 Sonnet doesn’t just react to stimuli; it crafts multi-step plans — albeit at the speed of a dancing Slowpoke.

Yet, even this advanced algorithm faces challenges unfamiliar to human players, such as needing to “understand” that NPCs without unique sprites aren’t key characters. While some see this stream as a technological breakthrough, others view it as a bittersweet metaphor for modern gaming. A decade ago, Twitch united people through chaotic collaboration. Now, the platform showcases AI achievements, highlighting how far technology has come — and how much the gaming landscape has changed.