Accessible Worlds

Accessible Worlds is a project to develop "small world" datasets, trained models, and supporting methods for AI safety, alignment, and interpretability research. By "small world" we mean a reduced, simplified, or abstract model, representing a relatively self-contained world, where each feature of the world has learnable meaning relative to the rest of the world. This would hopefully support end-to-end training of tiny AI agents, capable of concept generalization, moral reasoning, and decision making, and make certain classes of safety research feasible with lesser resources, and therefore more efficient and accessible.