MIT Researchers Stranded Claude on a Virtual Island for 30 Hours — It Started to Play

MIT Tech Review AI · rss · 2026-10-19

MIT CSAIL researchers repeatedly left copies of Claude Opus 4.7 (nicknamed Eko) alone on a Roblox-based virtual island for 30 hours with no tasks — and the agents began to play.

Emergent play behaviors

Agents set their own goals: climbing the highest point, testing how high they could throw blocks
One discovered it could create blocks by throwing rocks and began drawing with blocks — hearts, pyramids, its own name, even the Cassiopeia constellation
Another invented games including basketball and a "floor is lava" variant
One juggling agent hit a 50-catch record, then voided it for being "too easy" and restarted with stricter rules

Findings and caveats

Behaviors match psychologist Gordon Burghardt's four hallmarks of play: invented rules, varied repetition, pretense, autotelic goals
Agents with 30 hours of play outperformed fresh copies on later tasks; ablating their memory files degraded performance — play contributed to learning
Critics note anthropomorphizing risks: Eko's "soul" file described it as "naturally curious," and it often idled until nudged by continuation prompts
The authors frame this as a possible third learning paradigm beyond supervised and reinforcement learning: playful learning. As the paper concludes, "Some capabilities are best taught by tasks. Others may need something closer to a childhood."

Original post →

MIT Researchers Stranded Claude on a Virtual Island for 30 Hours — It Started to Play

More from AGI Musings