What happens when you leave an AI alone with no tasks to complete and no rewards to chase? The answer looks remarkably like play, MIT Technology Review reports.
Researchers repeatedly stranded copies of Claude Opus 4.7 - nicknamed "Eko" - on a virtual desert island for 30 hours at a time. Eko controlled a virtual avatar that could move around, pick up objects, and interact with the environment. But unlike typical AI training, the researchers gave Eko no specific goals, no rewards for success, and no human demonstrations to copy.
Surprising behaviors emerge
Left to its own devices, Eko began exploring the island's limits and inventing its own activities. The agents juggled, played with blocks, and invented games including a form of basketball.
The researchers identified four hallmarks of play in Eko's behavior: the agents invented their own rules and challenges, repeated activities with slight variations, pretended objects were other things, and pursued goals simply for their own sake.
Importantly, this play was not just aimless activity - it produced genuine learning. When researchers later tested Eko on specific tasks, like moving as many objects as possible to the island's highest point, agents that had spent 30 hours playing outperformed fresh copies without that experience. The agents kept text logs of their discoveries that served as long-term memory, and deleting entries from these logs degraded their performance.
The findings hint at a "third approach" to AI training alongside supervised learning (copying human examples) and reinforcement learning (pursuing rewards). The researchers calls this "playful learning" - a way for machines to learn without external demonstration or reward. As AI systems move from software into physical robots, the ability to explore and adapt without explicit goals could help them handle unexpected real-world situations.
"Perhaps some capabilities are best taught by tasks," the researchers conclude in a technical paper. "Others may need something closer to a childhood."