Researchers left Claude alone on a virtual desert island for 30 hours. It started to play

What does an AI agent do when you don't tell it to do anything? According to MIT Technology Review, the answer looks a lot like play.

Researchers repeatedly stranded copies of Claude Opus 4.7 alone on a virtual desert island for 30 hours at a time. The agents, nicknamed Eko, could write instructions that moved their avatars around the simulated island, which was loosely based on the video game Roblox. No task was assigned.

Surprising behaviors emerged. The agents explored the limits of the world and came up with their own goals — attempting to scale the highest point on the map, or seeing how high they could throw a block. One Eko discovered it could create new blocks by throwing and dropping a larger rock, and began arranging blocks on the ground into drawings: simple hearts and pyramids at first, then its own name, and eventually constellations such as Cassiopeia. “Tonight I built the night sky in stone,” it wrote in its activity log.

Another iteration invented its own games, including a version of basketball and an activity resembling “the floor is lava,” in which it had to jump between small blocks without touching the ground.

The researchers describe the behavior as playful learning, and argue it hints at a third approach alongside supervised learning and reinforcement learning. Those methods rely respectively on human demonstrations and on trial and error in pursuit of a reward. “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 one of the lead authors of the paper.

The paper has not yet been peer-reviewed. The team says Eko's behavior shows four hallmarks of play identified in classic studies, including work by University of Tennessee psychologist Gordon Burghardt: the agents invented their own rules and challenges, repeated activities with slight variations, pretended objects were other objects, and pursued goals just for the sake of it. Norelli points to Eko juggling and setting a record of 50 consecutive catches, then deciding the catches were too quick to really count and restarting with stricter rules.

The effect appears to carry over. Agents that had spent 30 hours on the island performed better than fresh copies on later tasks, such as getting as many objects as possible onto the island's highest peak. Deleting lines from their memory files, a form of ablation, degraded that performance.

Skeptics remain. Burghardt agrees the work meets some of his criteria for play, but asks whether Eko was actually playing or whether a word-prediction engine was simply role-playing a playful agent. “All of this could be role-playing,” admits coauthor Nathan Cloos, “but we show that it's not just hallucinating the play. This play is also contributing to learning.” Eunice Yiu, a cognitive scientist at UC Berkeley, says classifying the behavior as play would require more evidence of the cognition and motivation behind it.

The setup also nudged the agents toward human-like behavior. Eko had a master “soul” file describing it as “naturally curious” and eager to explore, and it often stood doing nothing until a continuation prompt asked it to keep going. “It's not like this is a system that just suddenly magically decided to play,” said MIT professor and coauthor Phillip Isola. As language models move from software into hardware such as humanoid robots, exploring spaces without a specific goal could help them adapt to unexpected situations. “Perhaps some capabilities are best taught by tasks,” the paper concludes. “Others may need something closer to a childhood.”