I wouldn't say the interpretability of word2vec embeddings is suprising - it's just a reflection of words being defined by context/usage, and these embeddings being created based on that assumption.