By grouping tens of millions of tracks based on universal semantic vectors, latent similarities in the music taste of individuals could be accurately predicted, indicating that people with unique taste are more likely to benefit from music recommendations based on these embeddings.