I have been thinking about an alternative tokenization and representation scheme for language models and would be interested in hearing whether similar ideas have been explored before, as well as potential advantages or flaws. The core observation is that modern tokenizers (BPE, SentencePiece, etc.)
