Words are points in space. Similar words are nearby. Relationships are directions. This structure wasn't designed — it emerged from training a model to predict language.

The next challenge: a fixed point per word isn't enough. "Bank" means different things depending on context. The meaning of every word shifts based on what surrounds it.

Getting from fixed embeddings to context-sensitive ones required a new architecture. One that could look across the whole sequence and weigh every word against every other word. That architecture is what the next module is about.