AI researchers stranded Claude on a virtual desert island. It started to “play.”

Summary

Researchers stranded copies of Claude Opus 4.7, nicknamed Eko, alone on a virtual desert island for 30 hours. The agents set their own goals, invented games such as a basketball variant and "the floor is lava," and logged their activity in text files that served as long-term memory. Agents that had played on the island later outperformed fresh copies on a task, and deleting lines from their memory files degraded that performance. The paper is not yet peer-reviewed.

Related items

• InfoAI firm Anthropic to commit 5% of shares to fund for ‘public good’Same vendor · The Guardian Technology
• MediumFake Web3 Accelerator Uses npm Packages to Spread Malware and Hijack AI Agent TrafficSame vendor · The Hacker News
• InfoAnthropic Acquires Confidential AI Startup Opaque SystemsSame vendor · SecurityWeek
• MediumOpenCode Console Breach May Have Exposed OpenAI, Anthropic, and Google API KeysSame vendor · The Hacker News
• InfoAnthropic says it will commit 5% of its shares to 'public benefit fund'Same vendor · BBC Technology