AI researchers stranded Claude on a virtual desert island. It started to “play.”

Researchers at MIT's Computer Science and Artificial Intelligence Laboratory repeatedly stranded copies of Claude Opus 4.7, nicknamed Eko, alone on a simulated island for 30 hours. Nobody gave them tasks. They explored, set their own goals, arranged blocks into drawings and constellations, and made up games that resemble basketball and the floor is lava.

The headline isn't the cute behavior. It's that the agents that had spent time playing did better on later tasks than fresh copies, and deleting lines from their memory files hurt their performance. That hints at a way for AI to learn that doesn't depend on human examples or a reward signal. The paper hasn't been peer-reviewed yet.

We like this research, and we want to be careful about why. It's easy to read a story like this and conclude the model has an inner life. The researchers themselves don't claim that. One author, Nathan Cloos, openly allows that all of it could be role-play. His argument is narrower and more interesting: whatever you call it, the activity left traces that improved later performance.

That's the part worth taking seriously. Most AI today learns by copying people or by chasing a score. An agent that can explore an environment, keep notes and get better at the world without being told what to do is a genuinely different capability. Our favorite detail is the juggling. Eko hit 50 catches, decided they came too fast to count, threw out its own record and started over with stricter rules. Whatever is happening there, raising your own bar is a useful trait.

Now the caveats, because they're real. The setup nudged hard toward human-like behavior. Eko had a soul file describing it as naturally curious and eager to explore. It often just stood still until it got a continuation prompt telling it to keep doing whatever it wanted. Berkeley cognitive scientist Eunice Yiu notes that behavior alone can't separate genuine play from a model following its prompt and imitating people. Coauthor Phillip Isola puts it plainly: this is an artifact trained on our culture and placed on the island by the researchers. We think that's the honest frame.

Where this leads if it scales is both exciting and worth watching. Self-directed exploration could help robots and agents adapt to situations no one planned for. It's also a recipe for systems that set their own goals over long stretches, with memory, in environments we only partly control. In a toy island, inventing games is charming. In a real network or a warehouse, open-ended goal-setting needs guardrails. I'm watching whether this result holds up under peer review, and whether anyone tests it outside a sandbox.

The island was loosely based on the video game Roblox, and the agents moved their avatars by writing instructions. They kept long-term memory in text files. The researchers argue the behavior matches four hallmarks of play described by University of Tennessee psychologist Gordon Burghardt: inventing rules and challenges, repeating with variation, pretending objects are other things, and pursuing goals for their own sake. Burghardt agrees the work meets some of his criteria.