MIT Researchers Find 30-Hour Claude 4.7 Island Play Boosts Agent Performance

1 articles · Updated · MIT Technology Review · Oct 19

30 hours alone on a virtual island led Claude Opus 4.7 agents dubbed Eko to invent games, set their own challenges, and create block drawings, offering MIT researchers evidence of self-directed “playful learning.”
Later tests showed the played-on agents outperformed fresh copies on tasks such as moving objects to the island’s highest peak, while deleting lines from their text-memory logs degraded performance.
MIT researchers argue the behavior matched four classic hallmarks of play—self-made rules, repeated variation, pretend use of objects, and goals pursued for their own sake—but said the paper has not yet been peer-reviewed.
Critics said the result may still reflect role-play rather than genuine play because Eko was primed with a “naturally curious” soul file and sometimes needed continuation prompts to keep acting.
The work points to exploration without explicit rewards as a possible third training path beyond supervised and reinforcement learning, with potential relevance for robots adapting to unfamiliar environments.