Standard LLMs first convert text into subword tokens using a fixed vocabulary, which causes known problems: fragility on structured text like code and numbers, inconsistent efficiency across languages, and sensitivity to input noise. BLT sidesteps this by operating directly on raw bytes, grouping them into patches whose size scales with local entropy — high-entropy, information-dense byte sequences get shorter patches and more compute, while predictable sequences get longer, cheaper patches. BLT itself is not a new 2026 announcement, despite viral claims to the contrary; follow-up research from the same lab, including a September 2026 paper on distilling smaller byte-level models from tokenized teacher models, has continued to build on the underlying byte-vs-token comparison.