And as you can see, if you do work out the bits, you find that cryptonector is wrong, since UTF-8 (as it has been standardized from almost the start of the 21st century, and as codecs in the real world have taken to implementing it since) encodes no more bits than UTF-16 does. It's 21 bits for both.
With regards to the comment then: the range downshifting you mentioned is merely a step in the encoding process though -- the code point is still whatever it was. If you read parent comment, it had claimed that, in a surrogate pair, each of the 2 code units encodes 10 bits of the code point... but that would be missing 1 bit when the code points need 21 bits to be represented. That's all I was saying there. The extra bit indicating that it's in fact a surrogate pair isn't some kind of implicit dummy bit that you can pretend isn't encoding anything -- if it wasn't there then clearly it wouldn't be encoding the code point for a surrogate pair anymore.
> I'm not at all convinced that 2^21 codepoints will be enough, so someday it'd be nice to be able to get past UTF-16 and move to UTF-8
UTF-16 currently uses up to 2 16-bit code units per code point, whereas UTF-8 uses up to 4 8-bit code units per code point, and the latter wastes more bits for continuation than the former. How is "getting past UTF-16 and moving to UTF-8" supposed to increase the number of code points we can represent, as claimed above? If anything, UTF-16 wastes fewer bits in the current maximum number of code units, so it should have more room for expansion without increasing the number of code units.
[1] https://news.ycombinator.com/item?id=17771351