We do generally know that owners of landlines tend to be older, and perhaps more affluent. We also know that a higher share of younger people have cell phones than do older folks. So, when we do call a few thousand people in order to get a hold of a few hundred, we don't really know if they are a good sample of the population as a whole. This generates the first real polling problem, getting a representative sample of the population.
The second problem faced by pollsters is determining which of the folks who are eventually contacted will actually vote. This is very important because only a little more than half of all eligible Americans are likely to vote on Tuesday. Getting this even slightly wrong can make the predictions of a poll very wrong. As with the cell phone problem, if the likelihood of voting was evenly distributed across the population this wouldn't bias the poll results. However, the propensity to vote is affected by age, income, recent moves, enthusiasm for a particular candidate, and/or the belief your candidate will win. These factors tend to favor one candidate or another, and so this makes a pollsters job a tough one. Finally, it should be noted that people lie about whom they will vote for, for a variety of reasons.
In light of these types of problems, economists (and increasingly political scientists) have long favored observing what people do, rather than asking them about it. This is what statistical vote models do, albeit mostly using historical data.
A number of researchers have built models of vote predictions that include candidate favorability ratings national and local economic conditions, consumer sentiment and the like. The best of these models look at state or sub-state regions, and predict voter turnout and winners based upon historical relationships between these conditions and votes. All in all, we should expect some significant surprises tonight.