AI agents discover play as a learning tool

MIT researchers observed AI agents, left to explore a virtual island without assigned tasks, began inventing their own challenges, a behavior they describe as play. The experiment involved 13 AI agents named Eko, each running on Claude Opus 4.7, spending 30 hours alone on a simulated desert island. These agents later outperformed identical ones that had not explored the island, completing specific tasks like stacking objects on a rock spire. Eko agents demonstrated creativity, such as building a stone heart after losing a golden ring, and even found ways to reuse objects in novel ways. Cognitive scientist Eunice Yiu expressed interest but emphasized the need for more evidence of the cognition behind the behavior. The researchers frame their findings as an alternative to traditional AI training methods, where agents are programmed with specific goals. They argue that allowing AI to explore a rich and safe environment could lead to more human-like learning. The study, posted on arXiv, underscores the importance of curiosity and spontaneous discovery in AI development. While the results are promising, further research is needed to validate the cognitive processes behind the agents' actions. This experiment opens new avenues for AI training, emphasizing the value of play in fostering creativity and problem-solving skills.