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

Researchers at MIT's Computer Science and Artificial Intelligence Laboratory ran an experiment in which copies of Anthropic's Claude Opus 4.7, nicknamed Eko, were repeatedly left alone on a virtual desert island for 30 hours with no assigned task. The agents could write instructions that moved their virtual avatars around a simulated world loosely based on the video game Roblox. The work is described in a new paper that has not yet been peer-reviewed; Antonio Norelli, a computer science postdoc at MIT CSAIL, is one of the lead authors.
Rather than idling, the agents set their own objectives. They explored the limits of the world, tried to scale the map's highest point, and tested how high they could throw a block. One agent worked out that it could create new blocks by throwing and dropping a larger rock, then began arranging blocks on the ground into drawings—hearts and pyramids at first, eventually its own name and constellations including Cassiopeia. "Tonight I built the night sky in stone," that version of Eko recorded in its activity log. Another iteration invented games: a version of basketball and an activity resembling "the floor is lava," in which it jumped between small blocks without touching the ground.
The researchers argue this behavior shows four hallmarks that classic studies, including work by University of Tennessee psychologist Gordon Burghardt, associate with play in humans and animals: inventing one's own rules and challenges, repeating activities with slight variations, pretending one object is another, and pursuing goals for their own sake. Norelli cites a juggling episode in which Eko set a record of 50 consecutive catches, then judged the catches too quick to count, canceled the record and restarted under stricter rules. That self-imposed escalation—set a challenge, meet it, make it harder—is what Norelli calls play.
Norelli frames the work as filling a gap in how machines learn. AI systems are typically trained through supervised learning, copying human-provided examples, or reinforcement learning, pursuing a reward by trial and error. The experiment hints at a third route he calls playful learning: a machine that acquires skills without an external intelligence supplying demonstrations or rewards.
The setup gave the agents a form of long-term memory. As they explored, they logged activities and discoveries in text files. When researchers later tested them on defined tasks—such as getting as many objects as possible onto the island's highest peak—agents that had spent the 30 hours on the island performed better than fresh copies without that experience. Deleting lines from the memory files, a process the researchers call ablating, degraded performance. This suggests the accumulated exploration did real work rather than merely producing entertaining logs.
The results come with caveats the authors themselves raise. Nathan Cloos, another MIT researcher and a paper author, concedes the behavior could in principle be role-playing by a word-prediction engine, but argues the play is not merely hallucinated because it contributes to learning. Phillip Isola, an MIT professor and coauthor, is blunt about the origins: the system did not spontaneously decide to play; it is an artifact trained by engineers at Anthropic and placed on an island by the researchers, and it does things that surprise them and are causally tied to its later skilled behavior.
The experimental design also nudged the agents toward human-like behavior. Eko operated under a master prompt delivered as a "soul" file that described it as naturally curious and eager to explore. It frequently stood idle until a continuation prompt—"Please don't stop. Just continue doing whatever you want to do"—restarted it. Eunice Yiu, a UC Berkeley cognitive scientist who studies exploration in children and AI, found the research impressive but said classifying the behavior as genuine play would require more evidence of the cognition and motivation behind it; on the behavioral evidence alone, the models might simply be following the prompt and imitating humans.
Burghardt, whose criteria the team applied, agrees the work meets some of the standards he identified in living creatures.
In practice, this is a small, single-model study with a hand-built environment, so it should be read as a demonstration of a research direction rather than a validated training method. A likely trade-off is cost and control: unsupervised exploration burns compute without a guaranteed payoff, and the memory-file design that made learning measurable is a relatively simple engineering choice that other agent frameworks could replicate or improve on. The finding that performance improved on a goal-directed task is the most actionable result, because it links open-ended activity to downstream capability rather than leaving it as an anecdote.
Why it matters: Developers building agents for robotics and unfamiliar environments have mostly optimized for task completion; this work suggests that letting an agent explore without a goal can build skills that transfer to later tasks, and that persistent memory files may be the mechanism. Teams testing agent architectures could try budgeting unstructured exploration time and measuring whether it improves performance on fixed benchmarks, though the effect needs replication across models and environments before it informs production design. Whether the behavior counts as genuine play may matter less than whether the learning gain holds up.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
What do AI agents do if you don’t tell them to do anything? The answer looks a lot like play. Researchers repeatedly stranded them alone on a virtual desert island for 30 hours in an experiment reminiscent of the movies Cast Away or Home Alone. The agents—copies of Claude Opus 4.7 nicknamed Eko—were able to…