From Tokens to Bytes I: Scaling Hierarchical Networks | Cartesia
www.cartesia.ai
A newer generation of H-Nets learns directly from raw bytes, and matches an open-source 3B Transformer with 10x less compute and data.