A 155K-param transformer builds a map of a world it's never shown

A 155,000-parameter transformer model learns to construct an internal map of a 2D grid world solely from navigation data, without ever being shown the map itself. This project is for AI researchers exploring emergent world models and the minimal computational requirements for spatial reasoning. It is interesting because it demonstrates that even a tiny transformer can develop a structured, generalizable representation of an unseen environment, challenging assumptions about the scale needed for such cognitive abilities.