So if I'm reading this correctly, this works by knowing how often each symbol occurs ahead of time? (Since the table of frequencies is not counted in the compressed size but required for decompression)
No what he's getting at is that he has 70 unique states per element and binary doesn't align well to that. A 100% proven optimal way to fix this is with arithmetic encoding which is actually a lot faster than the above method and will save even more bits.