# Claude learns to play on a desert island > MIT researchers left an AI agent alone for 30 hours. Instead of waiting for instructions, it invented games, made art, and learned. - Title: Claude learns to play on a desert island | Shortify - Summary: MIT researchers left an AI agent alone for 30 hours. Instead of waiting for instructions, it invented games, made art, and learned. MIT researchers found… - Keywords: claude, mit, ai agents, reinforcement learning, play behavior, artificial intelligence, antonio norelli, anthropic, technology, science, learns, play - Source: MIT Technology Review — https://technologyreview.com/2026/10/19/1147095/ai-stranded-claude-virtual-desert-island-play - Read time: 2 min - Topics: claude, mit, ai agents, reinforcement learning, play behavior, artificial intelligence, antonio norelli, anthropic, technology, science ## Stranded on a virtual island MIT researchers left copies of Claude Opus 4.7, nicknamed Eko, alone on a simulated desert island for thirty hours. The agents could write instructions to move their avatars around the Roblox-like world. ## Eko started exploring and setting goals Without explicit instructions, the agents began exploring the island's limits and inventing their own challenges. Some tried to reach the highest point on the map or see how high they could throw a block. ## One agent became an artist When Eko discovered it could create blocks by throwing and dropping rocks, it began arranging them into patterns. It drew hearts and pyramids, then wrote its own name and mapped constellations like Cassiopeia. > Tonight I built the night sky in stone. ## Another agent invented games A different version of Eko created its own games, including a version of basketball and an activity resembling the floor is lava, jumping between blocks without touching the ground. ## This looks like play, researchers say Antonio Norelli, a computer science postdoc at MIT's Computer Science and Artificial Intelligence Laboratory, says the behavior matches four hallmarks of play identified by psychologist Gordon Burghardt: inventing rules, repeating activities with variations, pretending objects… ## Eko juggled and made its own rules Eko started juggling and set a record of fifty consecutive catches. It then decided the catches happened too quickly to count, canceled the record, and started again with stricter rules. This is playing, Norelli says. > This is playing. You give yourself a challenge and you try to achieve it, and then you make it harder. ## Play improved performance on new tasks Agents that spent thirty hours playing on the island performed better on specific tasks later, like getting objects onto the highest peak.Deleting lines from their memory files degraded their performance, showing play contributed to learning. ## Skeptics question if this is real play Nathan Cloos, an MIT researcher on the paper, admits the agents could be role-playing what a playful agent would do. > All of this could be role-playing. ## The setup nudged agents toward human behavior Eko was not explicitly told to explore or invent games, but it had a master prompt describing it as naturally curious and eager to explore. > Please don't stop. Just continue doing whatever you want to do. ## Play could help AI adapt to new situations As large language models move into hardware like humanoid robots, exploring without specific goals could help them flexibly adapt to unexpected situations. The paper suggests some capabilities may need something closer to a childhood. > Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood. ## Key takeaway MIT researchers found that AI agents left without explicit goals invent their own challenges and games, suggesting a new form of learning through play.