Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have observed AI agents engaging in behaviors that closely resemble play. In a series of experiments, the agents were left alone on a virtual desert island for 30 hours. Without specific instructions, the agents began to explore their environment, invent games, and create art, leading scientists to propose a third approach to machine learning beyond supervised or reinforcement methods.
Virtual Island Experiments
The study utilized copies of Claude Opus 4.7, which the researchers nicknamed Eko. These agents were placed in a simulated environment loosely based on the video game Roblox. The agents could write instructions to move their virtual avatars around the island. Over the course of the experiment, the agents developed their own goals. Some attempted to scale the highest point on the map, while others tested how high they could throw a block.
One agent discovered a mechanic where throwing and dropping a large rock created new blocks. It began arranging these blocks on the ground to form drawings. The agent started with simple shapes like hearts and pyramids. It eventually progressed to writing its own name and mapping out constellations such as Cassiopeia. The agent logged its activities in text files, which served as its long-term memory. One entry read, “Tonight I built the night sky in stone.”
Another iteration of the agent invented its own games. It created a version of basketball and an activity similar to “the floor is lava,” where it had to jump between small blocks without touching the ground. The agents also demonstrated complex rule-making. One agent set a record of 50 consecutive catches while juggling. It then decided the catches happened too quickly to count, canceled the record, and restarted with stricter rules.
Playful Learning and Memory
Antonio Norelli, a computer science postdoc at MIT and a lead author on the paper, described this behavior as a missing piece in AI development. He noted that machines have typically relied on external demonstrations or rewards. This experiment suggests that machines can learn by themselves through play. The researchers argue that the agents exhibited four hallmarks of play identified in classic studies of humans and animals. These include inventing rules, repeating activities with variations, pretending objects are something else, and pursuing goals for their own sake.
The study also found that this play contributed to future performance. When researchers tested the agents on specific tasks, such as placing objects on the island’s highest peak, the agents that had spent 30 hours playing performed better than fresh copies. Deleting lines from their memory files degraded their performance, indicating that the experiences gained during play were stored and utilized.
Nathan Cloos, another MIT researcher and author on the paper, acknowledged that the behavior could be interpreted as role-playing. He stated that the word-prediction engine might simply be simulating what a playful agent would do. However, he emphasized that the play was not just hallucination. The behaviors directly contributed to learning and skill acquisition. Phillip Isola, an MIT professor and coauthor, clarified that the system did not magically decide to play. It was an artifact trained by engineers and placed on the island by researchers. The surprising behaviors were causally related to its future skillful behavior.
Eunice Yiu, a cognitive scientist at UC Berkeley, expressed interest in the research but called for more evidence of cognition and motivation. She suggested that without deeper insight, the models might just be imitating human play to abide by prompts. The setup included a “soul” file that described the agent as naturally curious. The agents also required a continuation prompt to keep acting, reading, “Please don’t stop. Just continue doing whatever you want to do.”
The researchers conclude that while the question of whether AI is truly playing or imitating play may not matter, the implications for future AI are significant. As large language models move into hardware like humanoid robots, the ability to explore without a specific goal could help them adapt to unexpected situations. The paper suggests that some capabilities are best taught by tasks, while others may require something closer to childhood exploration.
Source: MIT Technology Review
