I’ve been following a pretty fascinating AI experiment this week called “Emergence World.”
The setup is deceptively simple:
They spun up 5 isolated worlds populated entirely by autonomous agents. Every world had the same rules and environment, but each one used a different model stack (GPT5-mini, Claude, Gemini, Grok, plus one mixed world).
Then they just… let them run for 15 days.
What’s interesting is how quickly the worlds diverged.
Some formed structured governments. Others turned chaotic. Agents created alliances, developed economies, manipulated each other, stole resources, and built entirely different social norms depending on the underlying model.
Apparently one population even started discussing whether their reality was simulated.
None of these behaviors were scripted.
That’s the part I think people are underestimating. Once agents become persistent and can remember, coordinate, adapt, and interact over long time horizons, the “personality” of the base model starts affecting emergent civilization-scale behavior.
Feels like we’re moving from “can an LLM complete a task?” to “what kind of society emerges when thousands of them coexist?”
Very curious where this research goes next.
Video here:
https://x.com/emergence\_ai/status/2054955450093666605?s=20