My summary of the idea:
There is an awful lot of redundancy and wasted effort that goes into most papers. From introductions that need to be rewritten every time (when linking to a solid introduction would be both better and less time consuming). Each piece of a piece of a full paper (intro, data, analysis, ...) could be peer-reviewed and published individually. A full paper could then be built from these paper-bricks. Anyway, recommend reading the paper as it's well written and clear.
There's also a YouTube video by the author explaining it: http://www.youtube.com/watch?v=4sorEcLjN04.
"Formalize the structure of papers, such that each paper is composed of one or many (clearly marked) of the following "sections" ("bricks"):
symbol description my description
"I" Introduction ("domain intro")
"PS" Problem Statement ("specific problem")
"HLSI" High-Level Solution Idea ("solution vision")
"D" Details ("solution implementation")
"PE" Performance Evaluation ("benchmarks")
-------- >8 ---------- >8 ----------
Some of the advantages:
* no need to rephrase the same "intro" in every domain paper, just reference an existing "I" brick;
* a benchmark (PE) can just reference many "D"s;
* one can easily work "backwards" -- e.g. start with a benchmark (PE) of existing implementations and already publish it, then propose a new implementation (D);
* if someone publishes a similar paper before you, with similar "vision/idea" (HLSI), this doesn't totally destroy your publication, as you can still publish the part with an alternative implementation (D);
* "I+PS+HLSI or I+HLSI: This is what some communities call a "vision paper" [...]"
* & many more listed in the linked arxiv paper http://arxiv.org/abs/1102.3523. Very nice, short and readable one, this.