I think speculative decoding eliminates a lot of the savings people imagine they're getting from making LLMs use strange languages.

Words <> tokens