I don’t know how LLM do that, but I would think base64 is easier than arithmetic. ‘All’ you need (more or less) to do it perfectly is a table with 256 × 256 × 2

I don’t know how LLM do that, but I would think base64 is easier than arithmetic. ‘All’ you need (more or less) to do it perfectly is a table with 256 × 256 × 256 entries containing four-character ascii strings and the ability to chop up byte sequences at every 3rd byte. Also, changing a character in the input only has local changes; it changes at most 2 characters in other encoded byte stream. In arithmetic, on the other hand, a single character change can have effects at an arbitrary long-range.