See also:

https://blogs.oracle.com/ksplice/the-1st-international-longe...

Twitter characters can actually store up to nearly 31 bits each, if you’re using the JSON API. (Or at least, this was true in 2010. I don’t know whether this is still true.)

http://blog.kevinalbs.com/base122

https://news.ycombinator.com/item?id=13049329

Base-122 encoding is 87.5% efficient in UTF-8, better than anything listed in the base65536 repository’s comparison table.