MIT researchers observed game-like behaviors in AI agents
In an experiment conducted by MIT, 13 AI agents based on Claude Opus 4.7 were left on a simulated island for 30 hours without being given any tasks. It was observed that the agents exhibited game-like behaviors by inventing challenges for themselves.
MIT researchers observed 13 AI agents, named Echo, for 30 hours on a simulated desert island where no goals were defined. In the study published on arXiv, it was found that the agents invented their own tasks, such as stacking blocks or creating monuments for lost items.
Following the experiment, it was noted that these agents, when given specific tasks, performed better compared to other agents that did not undergo a prior exploration process. The researchers stated that allowing AI agents to explore without goals could be an effective method for skill acquisition, and they likened this process to a form of 'childhood' phase for artificial intelligence.
Evaluation: Cognitive scientist Eunice Yiu from UC Berkeley stated that the observed behaviors were impressive, but more evidence is needed to verify the underlying cognitive processes.
Sources
- techcrunch.com · Link