We solve the learning dynamics of (a close approximation of) word2vec in closed form, revealing what semantic features are learned.