On Tuesday afternoon, Anthropic introduced Claude Plays Pokémon, a live Twitch stream featuring their latest AI model, Claude 3.7 Sonnet, as it plays Pokémon Red. This unique experiment highlights the abilities of modern AI technology and the public’s reactions to its performance.
AI developers have traditionally turned to a variety of video games, ranging from Street Fighter to Pictionary, to evaluate new models—often for entertainment rather than practical use. However, Anthropic claims Pokémon serves as a valuable benchmark for Claude 3.7 Sonnet, which is capable of effectively navigating the game’s various puzzles.
Similar to OpenAI’s o3-mini and DeepSeek’s R1, Claude 3.7 Sonnet can “reason” through challenging tasks, like navigating a video game aimed at children. While its predecessor, Claude 3.5 Sonnet, stumbled right at the start of Pokémon Red—failing to leave the player’s home in Pallet Town—Claude 3.7 Sonnet successfully earned three gym leader badges.
Despite its advancements, Claude faces challenges. During the Twitch stream, it became stuck at a rock wall, unable to get past it no matter how hard it tried.
One viewer summarized the moment perfectly: “Who will prevail, an advanced AI programmed for thousands of hours, or a single rock wall?”
Eventually, Claude figured out it could go around the wall.
Watching Claude move through Pokémon Red is a mix of frustration and fascination as it carefully considers each decision, almost like a Slowpoke. The left side of the stream presents Claude’s “thought process,” while the right side shows its gameplay in real-time.
At one point, Claude attempted to locate Professor Oak in his lab but got confused due to the presence of other NPC characters.
“I notice a new character has appeared below me—a character with black hair and what appears to be a white coat at coordinates (2, 10),” Claude noted. “This might be Professor Oak! Let me go down and talk to him.”
Claude then approached an NPC that wasn’t Professor Oak—the same character it had interacted with several times before. This caused some viewers in the Twitch chat to grow anxious. However, many of those who had been watching for a while were far less concerned.
“Guys, chill,” one participant commented. “We exited and re-entered Oak’s lab like ten times before we figured out how to advance.”
For veteran Twitch users, Anthropic’s streaming setup may evoke feelings of nostalgia. Over ten years ago, countless participants tried to control Pokémon Red collectively in a groundbreaking social experiment titled Twitch Plays Pokémon. Each viewer participated by directing the player character through the Twitch chat, resulting in a chaotic yet entertaining experience.
Some AI researchers have drawn inspiration from Twitch Plays Pokémon for their projects. In October 2023, Seattle-based software engineer Peter Whidden created a YouTube video detailing his experience training a reinforcement learning algorithm to play Pokémon. His AI dedicated over 50,000 hours to the game before mastering navigation. One notable hurdle was the AI’s tendency to admire the pixelated graphics rather than engage in gameplay.
AI-driven “reenactments” of Twitch Plays Pokémon, such as Whidden’s and Anthropic’s efforts, are undoubtedly amusing, yet they carry a hint of nostalgia. The original stream represented a significant moment in Twitch’s history, as it unexpectedly united viewers in a common goal: to guide the player character away from aimless wandering and toward meaningful progress in the game.
Fast forward to 2025, and it seems we have transitioned from being active participants to passive observers, witnessing an AI attempt a game many of us learned to master at a young age. This shift serves as a microcosm of a broader trend: our online experiences are evolving from shared, communal activities to increasingly solitary ones.
Compiled by Techarena.au.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
