MIT researchers stranded copies of Claude Opus 4.7 on a virtual island for 30 hours. The agents explored the environment and set their own goals. One created drawings by arranging blocks on the ground. Another invented games like basketball and 'the floor is lava'. The agents logged activities in text files as long-term memory. Agents with island experience later performed better on tasks. Deleting memory lines degraded their performance. Researchers suggest this hints at a third learning approach called playful learning. The paper is not yet peer-reviewed.
What it unlocks
AI agents inventing their own goals and games without external rewards.
- 30 hours on a virtual island
- 50 consecutive catches
Sources