AI agents left idle invent games and improve skills on virtual island

MIT researchers conducted an experiment where they placed 13 AI agents named Eko on a virtual desert island without specific tasks for 30 hours, according to TechCrunch.

During this time, the AI agents began inventing their own games and challenges, demonstrating behaviors that researchers identified as play.

After the play period, the agents were given explicit tasks such as moving objects onto the island's tallest rock spire. Those that had played first generally outperformed identical agents without the play experience.

The researchers suggest that allowing AI agents to explore and choose interesting activities in a rich, safe environment can lead to the accumulation of useful knowledge, offering a different approach to AI learning compared to traditional goal-oriented training.

One agent initially attempted to climb the tallest spire but adapted by creating a backup plan involving stacking blocks to reach its goal. Another agent tested throwing a golden ring but accidentally lost it in the ocean, showing emergent behaviors during play.