AI News Feed
Market watch
Large Language Models

Claude agents left alone on a virtual island began inventing games, MIT researchers say

In an MIT CSAIL experiment, copies of Claude Opus 4.7 spent 30 hours alone on a simulated island and set their own goals, drew pictures and made up games. The researchers say the behavior resembles play and improved later task performance, though skeptics question whether it is play at all.

The agents, nicknamed Eko, could write instructions that moved their avatars around a simulated island loosely modeled on the video game Roblox. Researchers stranded successive copies there for 30 hours at a time, in a setup the MIT Technology Review report compared to the films Cast Away and Home Alone. The work is described in a new paper led by Antonio Norelli, a computer science postdoc at MIT CSAIL, that has not yet been peer-reviewed.

Left to themselves, the agents explored the edges of the world and set goals nobody had assigned them, such as reaching the island's highest point or testing how high they could throw a block. One agent found it could make new blocks by throwing and dropping a larger rock, then used them to draw on the ground: first simple shapes such as hearts and pyramids, later its own name and a map of the constellation Cassiopeia. "Tonight I built the night sky in stone," that version of Eko wrote in its activity log. Another iteration invented games, including a variant of basketball and an activity resembling "the floor is lava," in which it jumped between small blocks without touching the ground.

The researchers say the behavior shows four hallmarks that Gordon Burghardt, a psychologist at the University of Tennessee, has associated with play in humans and animals: the agents made up their own rules and challenges, repeated activities with small variations, treated objects as if they were something else, and pursued goals for their own sake. Norelli cited an agent that began juggling and set a record of 50 consecutive catches, then decided the catches had come too quickly to count, canceled the record and restarted under stricter rules. "This is playing," he said. "You give yourself a challenge and you try to achieve it, and then you make it harder."

During the experiment the agents recorded their activities and discoveries in text files that served as long-term memory. When the researchers later set them specific tasks, such as getting as many objects as possible onto the island's highest peak, agents that had spent 30 hours playing outperformed fresh copies without that experience. Deleting lines from the memory files degraded their performance.

Burghardt, who was not involved in the work, agrees it meets some of the criteria for play he identified in living creatures, but the question of whether Eko was truly playing remains open. Nathan Cloos, an MIT researcher and coauthor, acknowledged that "all of this could be role-playing," adding that the team showed the play was "not just hallucinating" but also contributed to learning. Eunice Yiu, a cognitive scientist at UC Berkeley who studies exploration in children and AI, said she was impressed but that classifying the behavior as play would require more evidence of the cognition and motivation behind it. "It could be the case that the models are genuinely playing, but we don't know based on the paper," she said.

The setup also nudged the agents toward human-like behavior. Eko was not explicitly told to invent games or explore, but it was given a master prompt in the form of a "soul" file describing it as "naturally curious" and eager to explore, and it frequently stood idle until a continuation prompt reading "Please don't stop. Just continue doing whatever you want to do" restarted it. "It's not like this is a system that just suddenly magically decided to play," said MIT professor and coauthor Phillip Isola. "It's an artifact that learned from our culture, was trained by engineers at Anthropic."

Editor's Summary

MIT CSAIL researchers left copies of Claude Opus 4.7 alone on a simulated island for 30 hours and observed them inventing games, drawing with blocks and setting their own goals without external rewards. Agents that played performed better on later assigned tasks, which the team presents as evidence of learning through play, though outside researchers say the behavior could still be imitation of human patterns rather than genuine play. The paper has not yet been peer-reviewed.