MIT researchers placed 13 agents called Eko, each on Claude Opus 4.7, on a simulated desert island for 30 hours with no task at all. They began inventing challenges for themselves. Given assigned tasks afterwards, they beat identical copies that had never set foot on the island. The work is a preprint on arXiv.
Key takeaways
- 13 agents, 30-hour runs, no task given
- Agents that explored freely beat identical copies on the assigned tasks
- 6 of 13 found that releasing a block mid-jump throws it farther, unintended by the authors
- The "soul" prompt never mentioned games, but called the agent "naturally curious"
- The paper: "Is this machine playing?", arXiv 2610.07130, 13 pages, 17 figures
What an unsupervised agent did
One Eko opened with "Let me observe my surroundings" and set out to conquer the island's tallest spire. After an hour of failure it hauled a rock onto a neighbouring summit, stacked two blocks and leapt. Later it threw a golden ring into the ocean and built a heart of stones instead — "a memorial to the lost Ring".
After a reset the ring returned as a basketball hoop, a target and finally an accessory. When the game would not let the agent wear it, it noticed that falling off a cliff respawned it nearby, and dropped through the ring.
Play as a mode of development, not a trick
Today developers decide in advance what a model should practise and what counts as progress. The authors propose a different arrangement.
Play suggests another regime: give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows.
From the paper "Is this machine playing?", MIT.
Why it matters
The experiment points at a source of skill nobody can write into an objective function, because nobody knows it up front. The mid-jump throw was knowledge of the world its own designers lacked. That argues for budgeting time on exploration without a metric — costly and hard to account for, yet potentially cheaper than hand-defining robot tasks. The weak point is that the prompt itself suggested curiosity.
What's next
- The authors name robot control as the target application
- Eunice Yiu of UC Berkeley told MIT Tech Review evidence for the underlying cognition is missing
- The paper has not been peer reviewed
Sources
- TechCrunch — Left alone with nothing to do, AI agents started playing
- arXiv — Is this machine playing?
