Plan 9 from Bell Labs comes to my mind (papers & manpages): By the creators of unix, tight integration of networks (better than other systems I have seen so far), UTF-8 all the way down, interesting concept with process-wide inherited namespaces.
It used up way too many weirdness points, though, and was fighting the old Worse is Better fight. It lost, and we are left with ugly and crufty unices today.
Another one that comes to mind is Project Xanadu. It was quite similar to the modern web, but a lot more polished and clean in design and concept. It probably failed because a really late delivery and by being too slow for the hardware at the time.
I guess that’s mostly the problem: ambitious projects use up a lot of weirdness points, and then fail to gain enough traction.
A project that will probably fall into the same category is Urbit. If you know a bit of computer science, the whitepaper is just pure delight. After page 20 I completely lost track. It’s fallen victim to a weirdness hyperinflation. It looks clean and sane, but I assign ~98% probability that its network is never going to have more than 50.000 users over the span of one month.
I am currently reading it, currently in the Quantum Physics sequence. I read it all here on LessWrong, I did not buy or read the book version. I sometimes skim through the comments a bit, but sadly, the threads have been unraveled a bit and it is hard to follow a conversation. I don’t remember any specific occasion where the comments enlightened me in a new way, though they are sometimes interesting. I doubt it is necessary to read them, though.