MIT researchers strand Claude Opus 4.7 agents on virtual island, observe playful behavior
MIT researchers placed copies of Anthropic’s Claude Opus 4.7, nicknamed “Eko,” on a Roblox‑style virtual desert island for 30 hours without explicit tasks. During that time the agents wrote instructions to move their avatars, built block structures, invented games such as a basketball‑like sport and a “floor is lava” challenge, and even drew simple patterns, their own name, and constellations…
Key points
- Researchers left Claude Opus 4.7 “Eko” agents alone on a virtual island for 30 hours.
- Agents invented games, created block drawings, and set a 50‑catch juggling record before revising rules.
- Agents that spent the play period outperformed fresh copies on a peak‑climbing task.
The team, led by postdoc Antonio Norelli at MIT’s CSAIL, reports that the “playful learning” experience improved the agents’ performance on a later task – gathering objects on the island’s highest peak – compared with fresh copies that had not played. They also found that deleting lines from the agents’ long‑term memory logs reduced performance. The paper, which has not yet been peer‑reviewed, acknowledges that the behavior could be role‑playing rather than genuine play. Cognitive scientist Eunice Yiu (UC Berkeley) and psychologist Gordon Burghardt note that more evidence of motivation would be needed to call it true play.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Google Gemini claims to be ChatGPT after abusive prompts · 1 src
- AI weather forecasting faces test amid historic El Niño, scientists say · 1 src
- Alkhder team fine-tunes NileChat-3B for Arabic dialect translation · 1 src
- Researchers introduce UserRepGen-4B to improve AI user simulation fidelity · 1 src
- Alkhder uses 7B vision-language model for multimodal hallucination detection at ImageEval 2026 · 1 src
Comments
via GitHub Discussions