In a new post from Google Research, the team introduces S2Vec, a method that learns the "language" of our cities to help map the modern world. The framing is algorithmic: rather than manually encoding urban rules, S2Vec learns representations from the structure of city data itself.
The approach borrows from language modeling, treating elements of a city as a kind of vocabulary whose patterns and relationships can be learned. This allows the model to capture something like the grammar of urban space, which could make mapping systems more flexible and better able to generalize across different places.
Because this is based on a single source, there are no competing claims to compare. The post focuses on the conceptual direction of S2Vec rather than offering a detailed technical breakdown.