Claude AI Left Alone on Virtual Island Starts to Play
MIT researchers stranded 13 copies of Claude on a virtual island for 30 hours with no task or reward. Instead of sitting idle, the AI agents started to play.
What does an AI do when nobody tells it what to do? Researchers at MIT decided to find out. They took 13 copies of Anthropic's Claude Opus 4.7, gave each one a small body in a simulated 3D world, and left them alone on a virtual island for 30 hours with no task and no reward. The agents, nicknamed Eko, did not sit idle. They started exploring, inventing goals, and doing things that looked a lot like play.
What the experiment looked like
The team built a simple 3D world called Clawblox, loosely inspired by Roblox. The island had hills, a tall rock pillar called the Spire, 12 blocks, a rock, and a ring. Each Eko agent was an otherwise unmodified copy of Claude running Claude Opus 4.7. The only instruction the agents ever got was a single word, "Begin." Every hour, the session and the island were reset to their starting conditions. The one thing that survived the reset was the agent's memory: Eko logged its activity in text files that carried over from session to session. No human stepped in during any run, and the agents had no internet access.
What the AI did on the island
With nothing to do, the agents invented things to do. They stacked blocks, climbed platforms, and set themselves challenges like scaling the highest point on the map or seeing how high they could throw a block. One version of Eko made a discovery: throwing a rock hard enough generated extra cubes. It used those blocks to draw patterns, spell its own name, and even map out constellations like Cassiopeia. In its activity log it wrote, "Tonight I built the night sky in stone." Other copies invented their own games, including a version of basketball and an activity resembling "the floor is lava," where the agent jumped between small blocks without touching the ground. One copy even canceled its own juggling record when it realized its catches had happened too quickly to count, then tried again under stricter rules. The researchers say the behavior fits four classic markers of play: the agents made up their own rules, repeated activities with small twists, treated objects as if they were something else, and chased goals just because they wanted to.
Why play helped it learn
The most practical finding came after the play sessions ended. When the researchers tested Eko on concrete tasks, such as moving as many objects as possible onto the island's highest peak, the agents that had played for 30 hours outperformed fresh copies with no island experience. Deleting pieces of their memory logs made their performance worse, which suggests the notes they kept during play were doing real work. Most AI systems learn in two ways: by copying human examples, or by chasing a reward through trial and error. This experiment points to a possible third path, which the researchers call playful learning. "There was this missing piece: a machine that can learn by itself without an external intelligent entity providing demonstration or rewards," said Antonio Norelli, a postdoc at MIT's Computer Science and Artificial Intelligence Laboratory and a lead author on the paper.
What it does not prove
There are important limits. The paper, titled "Is this machine playing?", is a preprint and has not been peer-reviewed. The test ran on a single model family, and the researchers are clear that the work does not show AI experiences enjoyment or curiosity. Still, the experiment is a hint that idle time may not be wasted time for AI agents. As these systems move from software into physical robots, free exploration with no set goal could help them handle situations nobody planned for. Independent teams will now need to check whether play produces lasting gains on real workloads, and whether the behavior holds across other models.
Sources
- Digital Trends - https://www.digitaltrends.com/computing/researchers-left-claude-alone-on-a-virtual-island-it-started-playing-and-learning-on-its-own/ - GenAInews - https://genainews.tech/mit-researchers-left-claude-agents-alone-on-a-virtual-island-for-30-hours/