On Tuesday afternoon, Anthropic launched Claude Performs Pokémon on Twitch, a livestream of Anthropic’s latest AI mannequin, Claude 3.7 Sonnet, enjoying a sport of Pokémon Purple. It’s grow to be an enchanting experiment of types, showcasing the capabilities of as we speak’s AI tech and folks’s reactions to them.
AI researchers have used all types of video video games, from Road Fighter to Pictionary, to check new fashions — usually extra for amusement than utility. However Anthropic mentioned that Pokémon proved to be a helpful benchmark for Claude 3.7 Sonnet, which might successfully “think” via the kinds of puzzles the sport incorporates.
Like OpenAI’s o3-mini and DeepSeek’s R1, Claude 3.7 Sonnet can “reason” its manner via powerful challenges, like enjoying a online game designed for kids. Whereas the mannequin’s non-reasoning predecessor, Claude 3.5 Sonnet, failed the very starting of Pokémon Purple — exiting the participant’s residence in Pallet City — Claude 3.7 Sonnet managed to win three gymnasium chief badges.
The most recent Claude nonetheless runs into hassle, although. Hours into the Twitch stream, the mannequin was deterred by a rock wall, which it couldn’t stroll via regardless of how arduous it tried.
One Twitch consumer summed up the state of affairs this fashion: “who would win, a computer AI with thousands of hours put into programming it, or 1 rock wall?”
Finally, Claude realized that it may navigate across the wall.
On the one hand, it’s irritating to look at Claude traverse Pokémon Purple with the velocity of a Slowpoke, reasoning via each step with excruciating contemplation. But it’s additionally oddly compelling. The left of the stream exhibits Claude’s “thought process,” whereas the precise exhibits real-time gameplay.
At one level, Claude tried to find Professor Oak inside his laboratory, however obtained confused, as a result of there have been different NPCs within the scene.
“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 wrote. “This might be Professor Oak! Let me go down and talk to him.”
Claude then proceeded to mistakenly discuss to an NPC apart from the Processor — an NPC the mannequin had spoken with a number of occasions earlier than. Among the thousand-odd folks within the Twitch chat began to get antsy. Others, notably those that’d been watching the stream for various minutes, have been much less anxious.
“Guys chill,” one individual wrote within the chat. “Before we exited and entered Oak’s lab like 10 times before understanding how to move on.”
For longtime Twitch customers, the format of Anthropic’s stream may really feel nostalgic. Over a decade in the past, tens of millions of individuals tried to play Pokémon Purple directly in a first-of-its-kind on-line social experiment known as Twitch Performs Pokémon. Every consumer may management the participant character by way of Twitch chat, leading to predictably chaotic gameplay.
Some AI researchers have cited Twitch Performs Pokémon as an inspiration for his or her work. In October 2023, Seattle-based software program engineer Peter Whidden revealed a YouTube video detailing how he educated a reinforcement studying algorithm to play Pokémon. His AI spent over 50,000 hours enjoying the sport earlier than it discovered to efficiently navigate it. One problem was that the AI most well-liked to admire the pixelated surroundings as an alternative of really enjoying the sport.
AI-powered “reenactments” of Twitch Performs Pokémon like Whidden’s and Anthropic’s are entertaining, however slightly bittersweet on the similar time. The unique stream was such a pivotal second in Twitch historical past as a result of it introduced folks collectively in an surprising manner. Everybody was on the identical staff, working towards the purpose of getting the participant character to cease operating in circles and truly progress via the sport.
In 2025, it appears we’re now not teammates, however spectators, watching an AI mannequin attempt to play a sport many people obtained the grasp of after we have been 5 years previous. It’s an AI-motivated microcosm of a bigger development: Our experiences on-line are transferring from shared, communal actions to extra solitary ones.
