Claude AI Plays Pokémon on Twitch at a Slow Pace

On Tuesday afternoon, Anthropic kicked off a unique livestream on Twitch called "Claude Plays Pokémon," featuring their latest AI model, Claude 3.7 Sonnet, tackling Pokémon Red. It's an intriguing experiment that not only showcases the AI's abilities but also captures viewers' reactions to this cutting-edge technology.

AI researchers often turn to video games like Street Fighter and Pictionary to test new models, usually more for fun than for practical use. However, Anthropic found Pokémon to be an effective benchmark for Claude 3.7 Sonnet, which excels at solving the game's puzzles through "thinking."

Similar to OpenAI's o3-mini and DeepSeek's R1, Claude 3.7 Sonnet can navigate complex challenges, such as playing a game intended for kids. While its predecessor, Claude 3.5 Sonnet, struggled to even leave the player's home in Pallet Town, Claude 3.7 Sonnet managed to secure three gym leader badges.

Despite its advancements, Claude 3.7 Sonnet still encounters hurdles. During the Twitch stream, it was stumped by a rock wall that it couldn't pass, no matter how hard it tried.

A Twitch viewer humorously remarked, "who would win, a computer AI with thousands of hours put into programming it, or 1 rock wall?"

Eventually, Claude figured out how to navigate around the obstacle.

Watching Claude play Pokémon Red can be frustratingly slow, as it meticulously reasons through each move. Yet, it's also strangely captivating. The stream displays Claude's "thought process" on the left and the real-time gameplay on the right.

At one point, Claude tried to find Professor Oak in his lab but got confused by other NPCs present.

"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."

Unfortunately, Claude mistakenly spoke to an NPC it had encountered several times before, causing some frustration among the thousand-plus viewers in the Twitch chat. However, those who had been watching for a while were more patient.

"Guys chill," one viewer wrote. "Before we exited and entered Oak's lab like 10 times before understanding how to move on."

For longtime Twitch users, Anthropic's stream might evoke nostalgia. Over a decade ago, Twitch Plays Pokémon became a groundbreaking social experiment, where millions of users collectively controlled the game through chat commands, leading to chaotic but memorable gameplay.

Some AI researchers have drawn inspiration from Twitch Plays Pokémon. In October 2023, Seattle-based software engineer Peter Whidden shared a YouTube video about training a reinforcement learning algorithm to play Pokémon. His AI took over 50,000 hours to master the game, with one amusing challenge being its tendency to admire the pixelated scenery instead of playing.

While AI-driven reenactments like those by Whidden and Anthropic are entertaining, they also carry a bittersweet tone. The original Twitch Plays Pokémon was a landmark moment, uniting people in a shared, albeit chaotic, effort to progress through the game.

In 2025, however, we've shifted from being active participants to mere observers, watching an AI struggle with a game that many mastered as children. This reflects a broader trend in our online experiences, moving from communal activities to more isolated ones.