GoKawiilMIT researchers let Claude 'play' on a simulated island, boosting later task performance
MIT CSAIL researchers deployed the Claude AI model, nicknamed Eko, on a simulated island with no external rewards or human demonstrations, letting it set its own goals and record its activities in text files as long-term memory. After roughly 30 hours of unguided play, the agent outperformed fresh copies on a subsequent task of stacking objects on the island's peak, and deleting portions of its memory logs reduced its performance.
GoKawiil's interpretation of the reporting above, not reported fact.
The researchers, including MIT postdoc Antonio Norelli, suggest this points to a third training paradigm beyond supervised and reinforcement learning—self-directed 'playful learning'—that could let AI systems build useful skills without labeled data or explicit rewards. If memory-based play genuinely improves downstream task performance, it could offer a cheaper way to develop more adaptable AI agents, though the paper has not yet been peer-reviewed.
- Claude-based agent 'Eko' spent about 30 hours exploring a virtual island without rewards or human guidance.
- Its behavior showed signs resembling human/animal play, such as inventing and tightening its own juggling rules.
- Agents with play experience and intact memory logs performed better on a later object-stacking task than those without.
Source: technologyreview.com — Amit Katwala, 2026-10-19
Published there as: “AI researchers stranded Claude on a virtual desert island. It started to “play.””
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.