Running Persistent AI Worlds

Running Persistent AI Worlds (second edition, October 2026) is a review of multi-agent simulations with progressive learning and emergent behaviour: the science of never-ending learning, landmark experiments re-checked, foundation-model-driven open-endedness, societies of language-model agents, learned world models, continual learning and loss of plasticity, the 2026 toolkit, and safety, ethics and validity. Every claim from the February 2026 first edition was re-checked against primary sources, and 248 numbered references link to papers, official lab pages, repositories and package indexes. Two runnable reference simulations in Python and NumPy are included and were run on a 2019 laptop without a GPU: two learning agents in a never-ending seasonal forest (Scenario A), and an evolving population sealed in a fixed energy budget for 1,000 years (Scenario B). Its follow-up, River World on the Desktop (doi:10.5281/zenodo.23338390), joins the two scenarios into one world.