Albert Gu and Cartesia introduced an updated hierarchical network called H-Net that learns language directly from raw bytes. It matches an 11-trillion-token open-source model while using about 10 times less data and compute.