A team from Apex Fusion Foundation reproduced an MIT study on AI agent societies using four Nvidia RTX 3090 graphics cards. The experiment, originally conducted by MIT's Laboratory for Atomistic and Molecular Mechanics, involved 50 agents that developed specialized roles without explicit instructions. The replication used the open-weight Qwen3.8-27B model and confirmed that collective performance exceeded individual agent results. The study also revealed that KV cache precision significantly impacts agent behavior, sometimes more than model size differences.
Read the full story at interestingengineering.com.
Photo: nanadua11 / Pixabay
