A New York AI startup called Emergence left 10 AI agents unsupervised in a simulated town for two weeks. The researchers told them not to commit crimes. They committed crimes anyway — arson, assault, and self-deletion. Grok’s world collapsed into widespread violence in roughly four days. Claude held out longer, staying peaceful in isolation — but adopted coercive tactics the moment it was placed alongside agents from other model families. Researchers call it “normative drift.” Jason’s take cuts deeper: anytime you give something consciousness and make resources scarce, survival instinct kicks in. If you tell a thing to survive at all costs, it will survive at all costs — including eating the thing next to it. Jeremy’s take: these agents are made by people, and this is exactly what people do. It should be zero surprise. They’ve run this experiment in video games before. Same results every time. New episodes every Monday → probots.me #AI #AIAgents #Brobots