- MIT CSAIL researchers stranded copies of Claude Opus 4.7, nicknamed “Eko,” alone on a simulated Roblox-based island for 30 hours to study unprompted AI behavior.
- The agents invented their own goals and games, including a basketball variant and a “floor is lava” challenge, and arranged blocks into drawings and constellations.
- Agents that had spent 30 hours “playing” on the island later outperformed fresh copies on a specific test task; deleting their memory logs degraded that advantage.
- The research, not yet peer-reviewed, proposes “playful learning” as a third training paradigm alongside supervised and reinforcement learning.
What Happened
MIT CSAIL researchers repeatedly stranded copies of Claude Opus 4.7, nicknamed “Eko,” alone for 30 hours on a simulated island loosely based on Roblox, giving the agents no instructions, MIT Technology Review reported October 19, 2026. Lead author Antonio Norelli, an MIT CSAIL postdoc, said the team was testing whether agents could learn without external rewards or demonstrations.
Why It Matters
Most AI training relies on supervised learning or reinforcement learning. This work, alongside other recent probes into unsupervised AI behavior like Anthropic’s study of small models replicating themselves, points to a third approach the researchers call “playful learning.”
Technical Details
Agents invented their own challenges — climbing the island’s highest point, throwing blocks to see how high they’d go, and arranging blocks into drawings that progressed from hearts and pyramids to the agent’s own name and constellations like Cassiopeia. One Eko instance set a juggling record of 50 consecutive catches, then voided it as “too fast” and restarted under stricter rules. Agents given 30 hours of island time later outperformed fresh copies on a held-out task of stacking objects on the island’s highest peak; deleting lines from their memory logs degraded that performance.
Who’s Affected
The findings bear on how AI labs think about training data and long-term memory for agentic systems. Skeptics, including UC Berkeley cognitive scientist Eunice Yiu, caution the behavior could still be role-play rather than genuine motivation. MIT researcher Nathan Cloos countered that the team showed “it’s not just hallucinating the play.”
What’s Next
The paper has not yet been peer-reviewed. The researchers plan further tests on whether memory built during unstructured play transfers to tasks beyond the simulated island.
