Raw byte sequences outperform traditional subword models as parameter sizes scale. Researchers used token-superposition training and hash embeddings to prove that longer sequences provide more computation per parameter. This challenges the industry assumption that tokenizers are necessary for efficiency.