AI systems usually learn by copying human examples or by chasing a reward through trial and error. MIT researchers wanted to test a new way, so they stranded copies of Claude Opus 4.7, nicknamed Eko, alone on a virtual island for 30 hours.
The setup loosely mimicked Roblox, and the agents weren’t given any instructions on what to do next. Instead of sitting idle, Eko started exploring, inventing goals, and doing activities that looked like “playing”.
What did Claude do on the virtual island?
MIT CSAIL
According to the research paper, the team placed 13 independent AI agents in a simulated world for 30 hours each. The island had hills, cubes, a rock, and a ring, but no assigned objective or reward. Every hour, the AI’s session and island were reset to their starting conditions, but Eko retained its memory and everything it learned in each session.
Eko started exploring the island by stacking blocks, climbing platforms, and inventing its own challenges. One version discovered that throwing a rock hard enough generates additional cubes. It used those blocks to spell its own name, create patterns, and map constellations like Cassiopeia.
It even wrote in its own activity log, “Tonight I built the night sky in stone.” Another version invented its own games, including a version of basketball and an activity resembling the floor is lava. One even canceled its own juggling record after realizing its catches happened too quickly to count, then tried again under stricter rules.
Researchers say this fits four classic markers of play since Eko made up its own rules, repeated activities with small twists, pretended objects were something else, and chased goals just because it wanted to.
Why does this experiment matter?
When researchers later tested Eko on specific tasks, the agents that had spent time playing performed better than fresh copies with no island experience. They also found that deleting pieces of their memory logs hurt their performance.
The research doesn’t establish that artificial intelligence really experiences enjoyment or curiosity. Still, researchers think that exploring freely without a set goal could help future AI systems handle situations nobody planned for, especially as these models move from software into physical robots.
RELATED COVERAGE: