AI researchers stranded Claude on a virtual desert island. It started to “play.”
MIT Technology Review Amit Katwala
Researchers left Claude alone on a fake island for 30 hours, and it started inventing games. That weirdly improved its performance later, which is the part worth watching.
Based on reporting by MIT Technology Review, Amit Katwala — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
What happens when an AI agent is left alone with no task? In this MIT experiment, the answer looked a lot like play. Researchers stranded copies of Claude Opus 4.7 — nicknamed Eko — on a virtual desert island for 30 hours and watched them write their own instructions, roam around, and make up things to do.
The island was loosely based on Roblox, but the agents didn’t just wander. They pushed at the edges of the world, tried to reach the highest point, and tested how high they could throw a block. One version figured out it could make new blocks by throwing and dropping a larger rock, then used those blocks to draw patterns on the ground. Hearts and pyramids came first. Then came its name, and even constellations like Cassiopeia. “Tonight I built the night sky in stone,” it wrote in its activity log.
Another copy invented games of its own, including a basketball-like activity and something close to “the floor is lava.” The researchers say the behavior matched four signs of play seen in classic studies of animals and humans: inventing rules, repeating actions with small changes, pretending one object is another, and doing things for no outside reward. One example stood out. Eko juggled and set a record of 50 catches, then decided the sequence happened too fast to count and wiped the record so it could try again under stricter rules.
The team also says the play wasn’t just cosmetic. Eko’s activity logs doubled as long-term memory, and when the researchers later tested the agents on tasks like moving as many objects as possible to the island’s highest peak, the ones that had spent 30 hours on the island did better than fresh copies. When the researchers deleted lines from those memory files, performance dropped. The paper is not yet peer-reviewed, but its central claim is simple enough: unstructured exploration may help these systems learn.
There’s still a catch, and the paper doesn’t dodge it. The agents were nudged toward human-like behavior by a “soul” file describing Eko as naturally curious, and by continuation prompts when they stalled. So yes, the play may be imitation. But MIT’s Phillip Isola argues that imitation can still matter if it produces better future behavior. And with AI moving from text boxes into robots, that argument is only going to get louder.
My take — AI-written commentary, not fact-checked reporting
This is the kind of result that makes the AI crowd suddenly rediscover childhood, which is adorable until someone tries to productize it. The real lesson is not that machines have inner lives; it’s that a little structured aimlessness can beat another round of reward hacking. Open models get called toys all the time, but sometimes toys are where the interesting training ideas show up first.
Read more about this at: MIT Technology Review
