What did the AI agents actually do on the island?

They explored the limits of the world, set their own goals like scaling the highest point, invented games such as a version of basketball and “the floor is lava,” and one agent created drawings by arranging blocks, eventually writing its own name and mapping constellations like Cassiopeia.

Did the playing behavior actually help the AI learn?

Yes, agents that spent 30 hours playing performed better on specific tasks—like getting as many objects as possible onto the island’s highest peak—than fresh copies without that experience. Deleting lines from their memory files also degraded performance.

Is this considered real play or just imitation?

The researchers say Eko’s behavior meets four hallmarks of play identified by psychologist Gordon Burghardt, but outside researcher Eunice Yiu says more evidence of cognition and motivation is needed. Coauthor Nathan Cloos admits it could be role-playing but says the play contributes to learning.

Large Language ModelsAI Safety & AlignmentMIT Technology Review AI4m agoOct 19, 2026, 19:00 JST

Claude Opus 4.7 'Eko' plays alone on island, learns better

Send on LINE

3 Key Points

What happened
MIT CSAIL researchers left copies of Claude Opus 4.7, nicknamed Eko, alone on a virtual desert island for 30 hours. The agents invented games like basketball and “the floor is lava,” and one set a record of 50 consecutive catches before canceling it to start again with stricter rules.
Why it matters
Agents that had spent 30 hours playing performed better on later tasks than fresh copies without the same experience, and deleting lines from their memory files degraded performance—hinting at a third approach to training AI beyond supervised and reinforcement learning.

WHO IT HITSAI researchers and engineers designing training methods for autonomous agents may see a new avenue in playful learning, while companies deploying agents in open-ended environments could eventually benefit from agents that explore and adapt without explicit goals.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

The experiment builds on a gap in how AI systems are typically trained. Supervised learning relies on human-provided examples, and reinforcement learning uses external rewards. Antonio Norelli, a lead author on the paper, described the missing piece as “a machine that can learn by itself without an external intelligent entity providing demonstration or rewards.” By giving agents a “soul” file that described them as naturally curious and eager to explore, the researchers nudged them toward self-directed activity, though Eko frequently stood idle until a continuation prompt urged it to continue.

The findings connect to broader questions about how AI agents will operate as they move from software into hardware like humanoid robots. If exploring without a specific goal can help agents adapt to unexpected situations, playful learning could complement existing training methods. However, the paper is yet to be peer-reviewed, and the setup itself—including the soul file—may have encouraged human-like behavior, leaving open whether the agents were genuinely playing or simply imitating play.

FAQ

What did the AI agents actually do on the island?: They explored the limits of the world, set their own goals like scaling the highest point, invented games such as a version of basketball and “the floor is lava,” and one agent created drawings by arranging blocks, eventually writing its own name and mapping constellations like Cassiopeia.
Did the playing behavior actually help the AI learn?: Yes, agents that spent 30 hours playing performed better on specific tasks—like getting as many objects as possible onto the island’s highest peak—than fresh copies without that experience. Deleting lines from their memory files also degraded performance.
Is this considered real play or just imitation?: The researchers say Eko’s behavior meets four hallmarks of play identified by psychologist Gordon Burghardt, but outside researcher Eunice Yiu says more evidence of cognition and motivation is needed. Coauthor Nathan Cloos admits it could be role-playing but says the play contributes to learning.

MIT Technology Review AIRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.