While this is an interesting thought experiment, what problem are
> you trying to solve with this proposal?
(asked privately but it seems worthwhile answering publicly, bcc'd,
you can id yourself if you like.)
Look, as I said in the original message I was asked to speak to a
group of young "hackers" at the HackerSpace in Singapore.
I wanted to be interesting and thought-provoking, make them think
through how this stuff works for an hour or two, encourage them to
poke holes in it, etc. It was one of the audience who pointed out the
potential MTU problem.
What problem does it solve, potentially?
0. Despite fears expressed herein I am not single-handedly planning to
convert the worldwide internet to this over the weekend. I'm going to
need some help 
1. It eliminates the need for DNS in its generally used form.
Sure, we've overloaded DNS with other functions from SPF -- in fact it
was Meng Weng Wong, inventor of SPF, who graciously invited me to
speak -- to whatever. But that's begging the point, there's nothing
interesting here about distributed, lightweight databases other than
eliminating one. Keep the DNS protocol per se for those things if you
like.
But given this you won't need to translate between host names and
addresses which is really what DNS was invented to do.
2. It makes "addresses" more transparent to humans, particularly when
you consider ipv6 addresses as typically displayed (hex.) Is this an
important goal? Not sure, but it's certainly true.
3. It's a transfinite space.
That just means that like Dewey Decimal etc it can be arbitrarily
expanded, you can add more levels or even stick levels in between plus
or minus some rules regarding SLDs/TLDs, and other rules which might
or might not be imposed (see #4).
But its total address space is as large as you allow a payload, there
is nothing inherent in the scheme that limits the addressing other
than the permutation of all acceptable Unicode glyphs I guess. But
since one can also have numeric parts and the set of integers is
infinite (that's tongue-in-cheek, somewhat.)
4. Also, because it's transfinite it's arbitrarily segmentable.
Again, that just means you can impose any meaning you like on any
substring or set of substrings. So for example host.gTLD is generally
taken to be something of some significance, or host.co.ccTLD, and that
sort of idea can be applied as needed, or not at all.
5. Bits is bits.
I don't know how to say that more clearly.
An ipv6 address is a string of 128 bits with some segmentation
implications (net part, host part.)
A host name is a string of bits of varying length. But it's still just
ones and zeros, an integer, however you want to read it.
The discussion I was responding to on NANOG involved how we got here
and where might we be going.
I brought up an idea I'd worked out somewhat and have even presented
in a small but public forum as being a possible future to consider
further.
Now you can go back to your regularly scheduled Jim Fleming guffawing.