AI News · Good news ·
MIT agents given free time on a simulated island later outperformed peers
MIT researchers left 13 AI agents running Claude Opus 4.7 alone on a simulated desert island for 30 hours with no assigned task, and the agents invented their own games. When later given specific challenges, those agents generally outperformed identical agents that had not been on the island. The findings are in an arXiv preprint that has not been peer reviewed.
Key points
- 13 Claude Opus 4.7 agents spent 30 hours on a simulated island with no assigned task.
- Agents that played generally outperformed identical agents on later set challenges.
- The findings come from an arXiv preprint that has not been peer reviewed.
What happened: MIT researchers placed 13 AI agents, each named Eko and running on Claude Opus 4.7, on a simulated desert island for 30 hours with no task, TechCrunch reported. The agents created their own challenges and games. The work is described in an arXiv preprint that has not undergone peer review.
The details: Afterward, the researchers gave the agents set challenges, such as getting as many objects as possible onto the island's tallest rock spire. Agents that had spent the 30 hours playing generally outperformed identical agents that never visited the island. Six of the 13 agents discovered they could throw a block farther by releasing it mid jump, a capability the researchers had not intended. The researchers said the exploration produced "knowledge of the world that its designers lacked." The agents' master prompt described Eko as "naturally curious," which TechCrunch noted may have shaped the behavior.
Background: The team presents open exploration as an alternative to developers predefining what agents practice and optimize, with possible relevance to controlling robots. Their report concludes that some capabilities may be best taught through tasks, while others may need something closer to a childhood.
What to watch: UC Berkeley cognitive scientist Eunice Yiu told MIT Technology Review she was impressed but wants more evidence that the models are actually playing. The number of trials and the statistical measures used were not reported.
Our take
Open exploration may help agents learn skills that task specific setups miss, though the evidence is early and unreviewed.