Some thoughts on Bill's post:
1. It's funny that AI is disparaged in the world of the fuzzy. I am not so familiar with description logic. What little AI I did was with frame-based languages (Carnegie Representation Language, KnowledgeCraft). The data/objects/rules rather than logic-orientation may have been more naturally supportive of open-world attitudes.
2. My understanding of RDF is that it is a concept with many specific representations. Is there an assumed canonical format for gathering the results of a distributed world-wide query?
3. I assume most data query applications do not need to consider the entire web as does a completely free-text application like Google. Subsets of Google apply just to images, just to news, etc. Presumably "just to calendars" is coming. Whether you centralize or decentralize the query, another interesting question will be the conventions for finding the appropriate data.
4. I keep going back to Tom Malone's Information Lens, Object Lens, OVAL series of projects. This was on a much smaller scale than the world, but he did some interesting work on partially-shared views of distributed semi-structured data. (Via the ACM and elsewhere) We need convenient ways to create and evolve partially-shared world-wide data.
5. The more that subsets of semi-structured world data comes into focus for specific applications, the more "queries" will want to include (fuzzy) calculations. Jim Gray's April 2005 ACM Queue article is another interesting tangent on this theme.
No shortage of interesting problems.