Researchers tested 13 AI agents controlled by Claude Opus 4.7 on virtual islands with no assigned tasks.
Over 30-hour runs with hourly resets and persistent memory, agents independently built structures and invented games, with experienced agents outperforming fresh instances in building tasks.
Researchers placed thirteen virtual characters, each controlled by the Claude Opus 4.7 artificial intelligence software, on separate copies of a simulated desert island. The environment was reminiscent of the popular game Roblox and allowed the digital characters to explore the space, grab objects, and build structures.
How did the digital characters act?
At the beginning of the experiment, which lasted thirty hours for each character with the island resetting every hour but the characters retaining their memories, absolutely no mission or work instructions were given. Despite the same initial conditions, the characters developed different activities.
Some made shapes like compasses, clocks, butterflies, or sailboats, while one character spent almost an hour writing his name. At the same time, the characters invented simple sports, such as throwing cubes at a target or throwing stones like bowling pins. Over time, these games evolved and became more complex, acquiring more objects, timing rules, or even jugglers.
The impact of play on subsequent performance
According to the researchers, the characters were doing these activities for their own sake, not for some external purpose, simply following the rules of the environment. When later tested on construction missions, those with thirty hours of prior game experience far outperformed a group of characters who now started with no memory.
READ ALSO
How artificial intelligence allows anyone to build programs with simple descriptions
Specifically, sixty-two percent of experienced characters managed to build more than two towers, compared to just fifteen percent of new characters.
Why did this behavior occur?
The MIT Technology Review noted that the characters had a command in their base file that described them as "naturally curious" and ready to explore.
Eunice Yiu, a cognitive scientist at UC Berkeley, said this behavior may be an attempt by the models to obey initial instructions and imitate humans, rather than a genuine act of play.
G
Hacks.gr on Google
Set it as a preferred source for cybersecurity updates.
Preferred source
Top comments (0)