I've read both Paxos papers from the parent and grandparent comments. Comparatively, the linked article is succinct, well-written, and (relatively) easy to understand.
(As he hasn't responded:) I'm guessing he means the article linked by this entire article thread, as opposed to any of the alternatives mentioned by the comments that he responded to.
I can really recommend implementing (and testing) Paxos (and the failure detector Omega mentioned at the end of the linked page) if you're interested in learning distributed systems, message passing-based communication and juggling state machines.
It may look tricky, but once you penetrate the formalities and understand it, you'll love it. It's a nice kick when you get it working.
Edit: Although you may not want to do it in C, as I had to. At the time, I'd have given my left hand for some tuples, pattern matching and other stuff. :)
Funny you mention this -- this was almost exactly what I and a friend did for a final project last semester. we did it in Go, which was probably much more enjoyable than C :)
Robbert is indeed one of (if not the) smartest men you'll ever meet in relation to fault tolerance in distributed systems. However, if you're looking to build a fault tolerant consensus algorithm, you owe it to yourself to check out randomized consensus (which predates Paxos, and really is the same algorithm - just simpler)
I think Apache Zookeeper is a Paxos implementation. I've heard very good things about it. Various pieces of Hadoop use it, and I know of people who use it for medium-sized webapps that need a canonical, fault-tolerant configuration system.
