[dragonwriter]

Polls aren't predictions (some people draw predictions from polls, but those are different from polls.)

They, particularly, don't even in theory sample the universe of "people who are actually going to vote", the sample registered voters or some model of likely voters. Those are known (to anyone who has more than a casual understanding of political polling) sources of nonsampling error when treating polls as a measure of the actual vote, since the universe sampled is different than the universe of interest, and there was particularly a lot of publicly-aired certainty about the utility of backward-looking likely voters models in this election.

I don't think there is a particular problem with the polls so much as with people's (including, unfortunately, many of the people talking about polls in the media) understanding of what polls measure.

To the extent there was an epic fuck up, it was in poll-based predictions that the effects of these type of things are correlated across states, so they gave Trump a near-certainty of winning. (Better predictors, like 538, had Trump an underdog but with a sizable likelihood of winning -- you can't say a 70%/30% chance is wrong given that the result is the one given a 30% chance is the one that materializes.)