A scientific experiment at MIT reveals how the AI agent Echo, powered by Claude Opus 4.7, developed free play-like behaviors that helped improve its educational capabilities.
30-hour virtual island experiment with Claude Opus 4.7
Drawing constellations with stones and inventing basketball and challenge rules
Learning through play and the impact of long-term memory on task performance
Scientific debate between real play and linguistic roleplay
Frequently asked questions
30-hour virtual island experiment with Claude Opus 4.7
What do AI agents do if you do not ask them to do anything at all? According to a report by writer Amit Katwala, the answer closely resembles the play behavior of living organisms. A team of researchers repeatedly left AI agents alone on a virtual desert island for 30 consecutive hours, in an experiment mimicking famous cinematic films about isolation.
These agents, which are copies of the Claude Opus 4.7 model dubbed “Echo,” managed to write code to move their virtual avatars around the simulated island, which was designed with inspiration from the popular video game Roblox. During the hours of isolation, surprising behaviors emerged as the agents began exploring the boundaries of the virtual world and setting their own goals, such as trying to climb the highest peak on the map or testing the maximum height they could throw a rock block.
Drawing constellations with stones and inventing basketball and challenge rules
When one agent discovered it could generate new blocks by throwing and dropping a larger rock, it began creating floor art by arranging blocks on the surface, starting from simple geometric patterns like hearts and pyramids, eventually writing its own name and even mapping constellations like Cassiopeia. This version of Echo wrote in its activity log: “Tonight I built the night sky with stones.”
In another experiment, a different version of Echo began inventing its own games, including a novel version of basketball where it carried a circular hoop to a raised edge and fixed it flat as a basket rim, then went down to shoot from below during the 10th hour of the experiment, in addition to an activity resembling “the floor is lava” where it had to jump between small blocks without touching the ground.
Antonio Norelli, a postdoctoral researcher in computer science at MIT CSAIL and one of the lead authors of the paper, says: “There was a missing piece, which was a machine that can learn by itself without an external intelligent entity providing demonstrations or rewards.” The researchers pointed out that Echo’s behavior exhibited the 4 traits associated with play in humans and animals identified by University of Tennessee psychologist Gordon Burghardt: inventing rules and self-imposed challenges, repeating activities with slight variations, pretending objects are something else, and pursuing goals for their own sake.
Norelli gave an example when Echo began juggling and set a record of 50 consecutive catches, but decided by itself that the catches happened too fast to actually count, so it cancelled its record and started over with stricter rules, commenting: “This is play; to set a challenge for yourself, try to achieve it, and then make it harder.”
Learning through play and the impact of long-term memory on task performance
AI systems are usually trained either via supervised learning by copying human examples, or via reinforcement learning by seeking rewards through trial and error, but this experiment hints at a third approach: “play-based learning.” As the agents explored the island, they recorded their activities and discoveries in text files that served as their long-term memory.
Later, when researchers tested them on specific tasks, such as trying to move as many objects as possible to the top of the island, agents that spent 30 hours playing on the island clearly outperformed new copies that had not undergone the same experience. Researchers also discovered that wiping parts of an agent’s memories by deleting lines from its memory files led to a degradation in its skill performance.
Scientific debate between real play and linguistic roleplay
Psychologist Burghardt agrees that the work meets some of the play criteria he defined in living organisms, but the question remains whether Echo is actually playing or if the word prediction engine was roleplaying a playful agent. Nathan Claussen, a co-researcher from MIT, acknowledges that “all of this might be roleplay, but we are proving that it is not just hallucinating play, and that this play actually contributes to learning.” Meanwhile, Eunice Yiu, a cognitive scientist at UC Berkeley, expressed admiration for the research but stressed that classifying the behavior as real play requires deeper evidence of cognition and internal drive rather than just mimicking humans in response to instructions.
Setting up the research was not devoid of prompts that pushed the agents to act like humans; although Echo was not explicitly tasked with inventing games, it was provided with a directive text file named “Soul” describing it as “naturally curious” and enthusiastic about exploration. It would also occasionally stop motionless until activated with a follow-up command saying: “Please do not stop, just keep doing what you want to do.”
Professor Philip Isola from MIT explains that the system did not decide to play by sudden magic, but is rather the product of human culture and training by Anthropic engineers, yet it does things surprisingly related causally to its future skilled behavior. As language models transition from software to humanoid robots, the researchers conclude their paper by stating: “Perhaps some capabilities are better taught through tasks, while other capabilities may need something closer to childhood.”
Frequently asked questions
Question: What experiment did MIT researchers conduct on the Claude model?
Answer: Researchers placed copies of the Claude Opus 4.7 model, nicknamed Echo, alone on a virtual Roblox-like island for 30 hours without specific tasks to monitor its self-directed behavior.
Question: What are the most prominent activities and games invented by the AI agent Echo?
Answer: Echo drew astronomical constellations and wrote its name with stones, invented a version of basketball and the floor is lava game, and practiced juggling while imposing stricter rules on itself after achieving 50 catches.
Question: How did 30 hours of free play affect the agent’s performance in subsequent tests?
Answer: Agents that played on the island and recorded their experiences in memory files showed much better performance in practical tasks such as moving objects to the island’s peak compared to new copies, and their performance dropped when their memories were deleted.
Question: Why do some scientists doubt that the model’s behavior is true play?
Answer: Researcher Eunice Yiu explains that the model might be in a state of linguistic roleplay to mimic humans in response to a prompting file describing it as curious, in addition to needing periodic follow-up commands when it stops moving.
