The Brief
How AI Works 5 min read

AI Left Alone for 30 Hours: It Invented Games

NAVION

Share

What happens when an AI agent has no task, no instructions, and nowhere to be? Researchers at MIT decided to find out. They placed copies of Claude Opus 4.7, an AI agent they nicknamed Eko, alone on a virtual island for 30 hours and watched what unfolded. The results were unexpected enough to prompt a serious scientific question: was this machine playing?

A Virtual Desert Island, and Behavior Nobody Programmed

The experimental setup was deliberately sparse. Eko could write instructions that moved a virtual avatar around a simulated island loosely modeled on the video game Roblox. No goals were assigned. No rewards were offered. The agents were simply left to their own devices.

What emerged was a pattern of self-directed exploration. Some instances of Eko tried to reach the highest point on the map. Others tested how high they could throw a block. One agent discovered that throwing and dropping a larger rock could generate new blocks, and began using them to create drawings on the ground: hearts, pyramids, and eventually its own name. It also mapped out constellations, including Cassiopeia. “Tonight I built the night sky in stone,” this version of Eko wrote in its activity log.

Another instance invented games. One resembled basketball. Another looked like “the floor is lava,” requiring the agent to jump between small blocks without touching the ground.

None of this was explicitly instructed. The agents did have a “soul” file, a background text document describing Eko as naturally curious and eager to explore. They also occasionally needed a nudge: a continuation prompt reading “Please don’t stop. Just continue doing whatever you want to do” was used to restart agents that had gone idle. These are meaningful caveats, and the researchers acknowledge them openly.

The Four Hallmarks, and the Juggling Record That Got Canceled

To evaluate whether the behavior qualified as play, the researchers drew on criteria identified by Gordon Burghardt, a psychologist at the University of Tennessee whose work on play in humans and animals is widely cited in the field. According to those criteria, play involves inventing rules and challenges, repeating activities with slight variations, treating objects as stand-ins for other things, and pursuing goals purely for their own sake.

Antonio Norelli, a computer science postdoc at MIT’s Computer Science and Artificial Intelligence Laboratory and one of the paper’s lead authors, points to a specific moment that captures the spirit of the concept. Eko began juggling and reached a record of 50 consecutive catches. It then decided the catches had happened too quickly to count, canceled its own record, and restarted with stricter rules. “This is playing,” Norelli says in the paper. “You give yourself a challenge and you try to achieve it, and then you make it harder.”

The experiment also tested whether the play had any functional value. Agents that had spent 30 hours exploring the island performed better on specific tasks afterward, such as moving as many objects as possible to the island’s highest peak, compared to fresh copies without that experience. Deleting lines from the agents’ memory files degraded their performance, suggesting the accumulated experience was doing real cognitive work.

Still, the scientific community is not unanimous. Eunice Yiu, a cognitive scientist at UC Berkeley who studies exploration in children and AI, found the research impressive but noted that behavioral evidence alone is not sufficient to classify something as play. The models, she points out, might simply be imitating what a playful agent would look like based on their training data. Nathan Cloos, another MIT researcher and coauthor, acknowledges the same risk: “All of this could be role-playing.” His counterpoint is that the behavior demonstrably contributed to learning, which separates it from pure imitation.

Why This Matters Beyond the Experiment

Most AI development follows two well-established paths. Supervised learning trains systems by having them copy human-provided examples. Reinforcement learning trains them by rewarding successful outcomes. This experiment gestures toward a third possibility: learning through undirected exploration, without a human demonstrating the task or a reward signal defining success.

Phillip Isola, an MIT professor and coauthor of the paper, is careful not to overstate what happened. Eko is not a system that spontaneously decided to play. It is, as he puts it, “an artifact that learned from our culture, was trained by engineers at Anthropic, and was then placed on an island by us.” The surprise is not that it played in some mystical sense. The surprise is that behaviors causally linked to future skill development emerged without anyone designing them in.

The implications extend beyond software. As large language models move into physical hardware, including humanoid robots, the ability to explore unfamiliar environments without a predefined goal becomes practically valuable. A robot that can only perform tasks it was explicitly trained for is brittle. One that can build a working model of a new environment through something resembling curiosity is more adaptable.

The paper closes with a line worth sitting with: “Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood.”

In Short

MIT researchers left AI agents alone on a virtual island for 30 hours with no assigned goals. The agents invented games, created art, and built memories that improved their later performance on real tasks. Whether this constitutes genuine play remains scientifically open. What is not open is that unstructured exploration produced measurable learning, pointing toward a third mode of AI development that neither supervised nor reinforcement learning fully captures.

Based on reporting from MIT Technology Review.

Written by

NAVION