Totally agree and there are some reasons to be optimistic about logic programming and the like - Bayesian systems like PRISM and answer set programming seem to me to be big steps forward since the last time we went round the block on this!

However I think that the cost of knowledge engineering is still massive and the challenge of this kind of project so sharp that it makes me quail to imagine execution! I think three or four years of full on tool development might make this more practical...