A London pub conversation sparked the creation of WERLD, an open-source artificial life simulation where agents evolve their own civilization from scratch. The project drops 30 agents with blank neural networks onto a graph, giving them 64 sensory channels, continuous motor effectors, and 29 heritable genome traits to navigate an energy-dependent world without any hardcoded behaviors or reward functions.

Built entirely in Python with stdlib only, WERLD uses NEAT neural networks that evolve their own topology through survival and reproduction rather than backpropagation. The simulation features a Next.js dashboard called the "Werld Observatory" that provides live views of population dynamics, brain complexity, species trajectories, and a narrative story generator. Every aspect of the agents' cognitive architecture has a metabolic cost, from neurons to communication channels, forcing evolution to justify complexity.

In initial runs lasting about 12 hours, 30 agents grew to over 7,000 while surviving 20+ population crises and famines. The agents developed basic communication patterns using signal broadcasts like hunger alerts, though structured language hasn't emerged yet. Their neural pathways visibly evolved across generations, with survivors passing on leaner, more efficient topologies. The creator open-sourced the project to see what emerges when you remove the guardrails of human knowledge and society, inviting others to watch this computational version of the Truman Show unfold.