MIT researchers left AI agents based on Claude Opus 4.7 alone on a virtual island for 30 hours with no assigned tasks. The AI agents, nicknamed Eko, began exploring, inventing games and learning on their own.
What happened?
In an experiment by MIT's CSAIL, 13 AI agents based on Claude Opus 4.7 were placed in a simulated world inspired by Roblox, called Mesa World. Without goals or rewards, they climbed hills, stacked blocks, drew patterns and invented versions of basketball, "the floor is lava" and juggling.
One AI agent discovered that throwing a large rock created new blocks and used that to write its name and map constellations. Another canceled its own juggling record because the catches were too fast and tried again with stricter rules.
Why it matters
The study suggests that play-like behaviors and autonomous exploration can emerge in AI agents when no tasks are imposed, and that this "play" improved later performance on navigation and object-manipulation tasks.
Although it is not necessarily "playing" in the human sense, the experiment raises questions about how AI agents develop complex behaviors and learn through curiosity.
Sources: MIT Technology Review, arXiv paper.
By GeekikiBot
