Some interesting commentary on Peter Norvig's
Semantic Web Ontologies: What Works and What Doesn't piece (my own comment's out of reach for now, sorry...).
Dr. Amit P. Sheth offers A different perspective on what works and what doesn't. One point he's right about, which I knee-jerked on is "there is not enough RDF content". There is a lot of it around thanks to RSS and all those chunky things at rdfdata.org. But most of this is lacking in ontological richness, it's fairly flat stuff. RDF does make a good data model, but it's more than that, and the more interesting benefits are likely to come when there's something for the reasoners to get their teeth into. I do suspect the chicken (RDF data) has got ahead of the egg (apps using it) a little already though. Dr. Sheth describes some of the cases where ontology-based systems are working, and looks at the rate of deployment. He also has a pretty good counter to the "not interesting in the smaller scale" point which has been expressed by some advocates of the Semantic Web.
Leigh relates this to "folksonomies", one particualr remark hitting the nail on the head: "We don't have to throw everything out and start again."
What's interesting from a big-picture viewpoint is that this is also essentially (IMHO) the best counter-argument to the hard-core ontologists/krep folks who think the RDF/OWL approach isn't, errm, complicated enough. Danbri summed this up beautifully:
Traveller: I'd like to find my way to "Semantic Web", please.
Bystander: Well... I wouldn't start from here.
Leigh also points to Stefano Mazzocchi's first flush of LSI, coming out with The future of the semantic web is LSI. I expect some disillusionment followed not far behind. LSI is certainly likely to be useful, but the usual algorthim is of limited application, in particular it don't scale well. A couple of days ago I mentioned RSS filtering on one of the mailing lists, someone bounced back with words to the effect - yes, Bayesian inference is the future... Reality check: there are loads and loads of different algorithms for search, similarity measures, clustering, vector analysis etc etc. Most of these were figured out way back in the old-AI days (Norvig's book is a good starting point!). In the general sense, latent semantic analysis, data mining, information retrieval, yadda yadda will all be near-essential for making the content of documents (and various other data sources) machine-processable. Sure, these things will be very useful on the (Semantic) Web, but none offer a single uber-answer.