I see half-and-half solutions as more possible: slightly smarter markup and much smarter web spiders would reap about as much benefit as a fully semantic web. I mean, just marking headers properly in <h\d> tags rather than <font size="\d"> and quotes in <quote> is a pretty decent hint for sniffing out semantically important text, and having grammar parsers for named entities like proper names and addresses and dates and such that are as forgiving with that text as BeautifulSoup is with the terrible HTML out there is a good amount of work in the right direction. I worked on a research project doing just that for two years and you can get a LOT of useful information out of a lot of modern HTML pages just by applying some simple statistical/AI magic to it. Anything more sophisticated would be near impossible without a breakthrough bordering on miracle.