Give 50 AI agents a shared world, no roles and no instructions, and they will start building technology together. That was the central finding of SwarmWorld, a study published in August by MIT’s Laboratory for Atomistic and Molecular Mechanics. The original work ran on a hosted frontier model.

Now a team working with Apex Fusion Foundation, a Swiss foundation building infrastructure for AI agents, has reproduced that experiment on four Nvidia RTX 3090 cards, the kind found in high-end gaming PCs. Using MIT’s code and commands unchanged, the run took about 26 hours and 12,800 model calls.

In SwarmWorld, researchers Pal, Wang and Buehler placed 50 or more language-model agents in a shared simulated world. None had an assigned job or a recipe to follow. The agents explored, gathered materials, ran tests and built persistent technologies. Some wrote programs that kept running after the agents themselves were removed.

The agents rarely needed to talk to each other. They coordinated by leaving changes in the world for others to find and build on, a process biologists call stigmergy, the same mechanism ants use when they lay pheromone trails.

The societies built broader and more resilient portfolios of technology than the best of an equal number of agents working alone. They also split into specialist roles, such as builders and explorers, without being told to. MIT released its code, prompts and complete traces, which made an outside check possible.

The replication, led by Časlav Nedeljković of HAL8, asked whether a small lab could get the same result with open weights on hardware it owns.

Source link