Trust and reputation are always in context, current karma system ignores it. But once you add context visualizing and interpreting karma becomes hard. Take a look of how we're trying to deal with this problem (link in my profile, sick of being accused of spamming, sorry).
Utilizing reputations through clustering of voting histories in order to produce their recommendations is a pretty good description of what's going on there now. Next they're planning on giving you even more control by allowing you to control who you trust directly. This combined system should be the special sauce that really solves this problem, I think.
It is using machine learning approach to estimate what the impact of your feedback should be. That way it is depends on context (only topics / tags right now) and is gaming proof (i.e. you can give as much feedback as you want, that would only improve your reputation / predictions for you by the system).
