MIT researchers placed 13 AI agents on a virtual island for 30 hours. The agents invented games and mourned a lost ring. They later outperformed untrained agents on set challenges. One agent learned to throw objects farther by releasing them mid-jump. The agents discovered knowledge their designers did not intend. The experiment used Claude Opus 4.7 for each agent. A master prompt described the agents as naturally curious. The paper was posted on arXiv before peer review. Researchers propose this as a new way to train agents. A cognitive scientist called for more evidence of genuine play.
What changed
AI agents were given predefined tasks and success criteria.
What it unlocks
AI agents can now learn skills through unguided exploration in a safe environment.
- 13 AI agents
- 30 hours on a virtual island
- 6 of 13 agents discovered new techniques
What you need to act on it
- Claude Opus 4.7 model
- a virtual island environment
Sources