MIT Researchers Observe AI Agents Inventing Play on Simulated Desert Island

MIT researchers observed 13 AI agents inventing challenges during 30 hours of exploration without assigned tasks on a simulated desert island. Each agent, called Eko, operated alone and used Claude Opus 4.7. The researchers say the behavior showed characteristics of play. The researchers described the experiment in a paper posted on arXiv, a platform for sharing research before peer review.

The researchers subsequently assigned specific challenges, including moving as many objects as possible onto the island's highest rock spire. The agents with 30 hours of exploration generally performed better than identical agents with no previous island experience. Developers currently determine what AI systems practice and how success is measured. The researchers propose exploration without predetermined goals as another route to learning skills, particularly skills that might eventually help agents control robots. The researchers argue that agents could accumulate useful knowledge by choosing interesting activities within a sufficiently rich, safe environment.

One Eko attempted to reach the highest spire during the first hour. After failing, the agent changed tactics an hour later, carrying a rock to a nearby summit, placing two blocks above the rock and jumping. The agent then announced that the agent had reached the spire. During a later experiment with throwing distance, the agent threw a golden ring into the ocean. The agent responded by making a heart from stones as a memorial to the ring.

After the game reset and the ring returned, Eko used the ring as a basketball hoop, a target and ultimately an accessory. The game prevented Eko from placing the ring around Eko's body. Eko discovered that stepping off a cliff would cause Eko to respawn near the ring. The agent stepped off the cliff and reappeared through the ring's center, then announced, "I'm wearing the Ring".

Six of the 13 agents discovered that releasing a block during a jump increased throwing distance. The researchers had not intended to enable that technique. The researchers wrote that exploration generated "knowledge of the world that its designers lacked".

The experimental setup encouraged humanlike behavior. The agents received no instruction to create games, but Eko's master prompt, called a "soul" file, characterized Eko as "naturally curious".

UC Berkeley cognitive scientist Eunice Yiu told MIT Technology Review that the results impressed Yiu, but Yiu wanted stronger evidence about the underlying cognition. Yiu said the models might genuinely be playing, but the paper did not establish that conclusion. The researchers concluded that tasks may be the best way to teach some capabilities, while other capabilities may require an experience more like childhood.