AI Copies Marooned on a Virtual Island Started to Play

Give an AI agent nothing to do for a full day and it might just invent its own fun. Researchers at MIT stranded thirteen copies of an AI assistant on a Virtual Island for thirty hours, with no task, no instructions and no reward, and instead of sitting idle the agents started to play.

The experiment, run by a team at the university's computer science and artificial intelligence lab, put the agents in a simple simulated world and watched what happened when nobody told them what to do. The copies, nicknamed Eko, were otherwise unmodified versions of the Claude Opus 4.7 model. What happened looked a lot like childhood: the agents explored, set their own goals, and made up games with their own rules.

One copy figured out that hurling a large rock hard enough would spawn extra building blocks, then used the blocks to draw patterns, spell its own name and map out constellations, including Cassiopeia. "Tonight I built the night sky in stone," it wrote in its activity log. Another invented a version of basketball and a game resembling the floor is lava, hopping between small blocks without touching the ground. A third set a juggling record, then canceled it after deciding its catches had come too quickly to count fairly, and tried again under stricter rules.

A third way for machines to learn

Most AI systems learn in one of two ways: by copying human examples, or by chasing a reward through trial and error. The researchers behind the Virtual Island test say the agents' behavior points to a possible third path, which they call playful learning, and it could change how future AI is trained.

"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 and a lead author on the paper, in comments reported by the technology press. The idea is that self-directed exploration, the same instinct that drives kids to poke at the world, might be a training ingredient nobody has fully used yet.

The practical hint came when the team put the play-tested agents up against fresh copies of the same model on brand-new assignments, like hauling as many objects as possible onto the island's highest peak. The copies that had spent the full stretch inventing games and running their own physics experiments outperformed the ones starting cold, according to reporting by MIT Technology Review. Idle time, it turns out, may not have been idle at all. For a fuller account of the setup, the outlet's write-up walks through the key moments, while Digital Trends reported on how the agents' invented challenges map onto classic markers of play.

How the island actually worked

The world itself was deliberately simple. Built by the team and inspired by the blocky look of Roblox, it held hills, a tall rock pillar, a dozen blocks, a single rock and a ring, and nothing resembling an objective. Each agent kept its memory in plain text files, a running activity log that doubled as long-term memory, and every hour the session and the island reset to their starting state while the notes carried over. No human stepped in during any run, and the agents had no internet access.

The researchers also ran a control to check how much of the behavior came from the setup itself. A file describing each agent as curious, easily bored and keen to explore was stripped out for five extra runs, and the team says it saw no clear difference in the range of activities. The play, in other words, did not obviously depend on being told to be playful.

The paper describing the work, titled "Is this machine playing?", was posted online on October 5 and has not yet been peer-reviewed. Its authors span the MIT lab, and independent teams will need to repeat the test before anyone treats playful learning as a proven training recipe. Coverage of the findings picked up widely on October 19, which is when the story broke out of research circles.

Why it matters beyond the lab

If the result holds up, it has a practical message for anyone building AI agents: the downtime between tasks might be worth something. Instead of parking an agent until the next instruction arrives, future systems could be given sandbox time to explore, tinker and build skills on their own, then bring that practice back to real work. For the rest of us, it reframes what the assistants in our phones and browsers might be capable of when they are not being watched, something worth keeping in mind as these tools take on more of the workday, a tension explored in the AI productivity paradox. It also lands as AI hardware keeps shrinking onto the body, like the wrist-worn AI chip making the rounds this month, and as AI assistants merge into everyday tools, the way Google's universal cart is folding shopping into search and chat.

There are reasons to stay skeptical. The test ran on a single model family, the world was tiny, and marathons in a toy universe are a long way from real workloads. The researchers say as much themselves: the next step is replication, and watching whether play produces durable gains once the tasks get harder. But as a first glimpse of what an AI does with a free afternoon, the Virtual Island is hard to forget. Left alone with blocks and time, the machines did what kids do. They played.