On Tuesday, Anthropic launched a live Twitch stream featuring its newest AI model, Claude 3.7 Sonnet, playing a game of Pokémon Red. The experiment has become a fascinating showcase of the capabilities of today's AI technology and people's reactions to it. AI researchers have long used video games as a testing ground for new models, often more for amusement than utility. However, Anthropic claims that Pokémon Red proved to be a useful benchmark for Claude 3.7 Sonnet, which can effectively "think" through the sorts of puzzles the game contains.

Claude 3.7 Sonnet is capable of "reasoning" its way through tough challenges, much like OpenAI's o3-mini and DeepSeek's R1. In the Twitch stream, the model managed to win three gym leader badges, a significant improvement over its predecessor, Claude 3.5 Sonnet, which failed to exit the player's home in Pallet Town. However, the model still runs into trouble, struggling to overcome obstacles such as a rock wall that it couldn't walk through. The stream's chat was filled with humorous comments, with one user jokingly asking, "Who would win, a computer AI with thousands of hours put into programming it, or 1 rock wall?"

The stream's format, which shows Claude's "thought process" on the left and real-time gameplay on the right, provides a unique insight into the model's decision-making process. At one point, Claude attempted to locate Professor Oak inside his laboratory but got confused due to the presence of other NPCs in the scene. The model's written "thoughts" were displayed on the stream, revealing its attempts to reason through the situation. While some viewers grew frustrated with Claude's slow progress, others were more understanding, having witnessed similar struggles in the past.

The experiment has drawn comparisons to Twitch Plays Pokémon, a 2014 social experiment in which millions of users controlled the player character via Twitch chat, resulting in chaotic gameplay. Some AI researchers have cited Twitch Plays Pokémon as an inspiration for their work, including Seattle-based software engineer Peter Whidden, who trained a reinforcement learning algorithm to play Pokémon. Whidden's AI spent over 50,000 hours playing the game before it learned to successfully navigate it.

The rise of AI-powered "reenactments" of Twitch Plays Pokémon, such as Anthropic's stream, has sparked a sense of nostalgia among viewers. However, it also highlights a shift in online experiences, from shared, communal activities to more solitary ones. In 2025, it seems that we're no longer teammates working together to achieve a common goal, but rather spectators watching an AI model try to play a game that many of us mastered as children.

Despite this, the experiment remains an entertaining and thought-provoking demonstration of AI capabilities. As AI technology continues to advance, it will be interesting to see how it is applied to various domains, including gaming and beyond.