I locked AIs in a prehistoric world

I spent several months building a small prehistoric world in my browser, with fires to maintain, fruit to gather, rabbits to hunt, wolves to avoid and tribes of digital humans capable of talking to one another. Then I let AI agents loose in it to see what they would do without a script written in advance. They protected their children, shared food, panicked at tracks they did not understand, developed group habits and sometimes reacted to my interventions as if I were a voice from above.

That is an agentic simulation. An AI receives something other than a question. It is given a world, memory, needs, relationships and the ability to choose an action. A Game Master then translates its intentions into events in the environment. This kind of experiment lets us observe models in action as well as in language. The video tells the story of this first version of the world, its strangest moments, its useful bugs and what it says about the AI agents coming to our tools.