Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory placed AI agents based on Anthropic’s Claude Opus 4.7 on a virtual desert island for 30 hours without specific instructions, observing spontaneous, self-directed play. The agents, named Eko, invented their own rules, created stone drawings, mapped constellations, and designed games like basketball and juggling without human guidance or external reward systems.

To retain their discoveries, the agents recorded their activities in text files that served as long-term memory. When later evaluated on specific tasks, agents with 30 hours of virtual play experience outperformed fresh copies, though performance degraded when researchers ablated lines from their memory files.

The findings suggest a potential third paradigm of AI training alongside supervised and reinforcement learning: autonomous, playful learning. Experts noted the agents met four classical psychological hallmarks of animal play, though questions remain on whether the behavior stems from true play or sophisticated role-playing by the underlying language model.

Why it matters

  • Playful learning offers a potential alternative to reward-based training, allowing agents to self-improve without costly human-labeled data.

  • Persistent text-file memory logs significantly enhance agent performance on downstream complex tasks, validating long-term memory architectures.

  • Autonomous goal generation introduces new evaluation paradigms for agentic behavior beyond traditional benchmark performance.

Source: technologyreview.com