The Claude Opus 4.7 copies set their own goals and invented games, hinting at learning driven by play rather than rewards.
Researchers at MIT stranded copies of Claude Opus 4.7 on a simulated desert island for 30 hours with no instructions, and the agents began to play. The copies, nicknamed Eko, wrote commands that moved avatars around a world loosely modeled on Roblox. They set their own goals, invented games, and built drawings out of blocks.
AI systems typically learn through supervised training on human examples or reinforcement learning that rewards trial and error. The Eko experiment points to a third path the authors call playful learning. Antonio Norelli, a postdoc at MIT’s Computer Science and Artificial Intelligence Laboratory and a lead author, frames the goal as a machine that learns without demonstrations or rewards.
For builders, the result suggests idle agent time may not be wasted. When researchers later tested the agents on concrete tasks, such as moving as many objects as possible onto the island’s highest peak, the copies that had played for 30 hours outperformed fresh ones. Agents also logged activity to text files that served as long-term memory, a pattern teams can copy today.
The work has not passed peer review, and the researchers ran a narrow test on a single model family. Independent teams still need to check whether play produces durable gains on real workloads, and whether the behavior holds across other models. Watch for replication attempts and for training recipes that reward self-directed exploration.
What matters
- Agents nicknamed Eko mapped constellations, built stone drawings, and juggled with self-imposed rules.
- Builders gain a third training signal, since agents that played alone later solved set tasks better.
- The paper is not yet peer-reviewed, so watch for replication on other models and longer runs.
Why it matters
The paper is not yet peer-reviewed, so watch for replication on other models and longer runs.
This GenAI News article was prepared in original wording using reporting and materials published by MIT Technology Review AI. Source reference: https://www.technologyreview.com/2026/10/19/1147095/ai-stranded-claude-virtual-desert-island-play/.
Drafted by the GenAI News review pipeline.