How do they call the last one a 'fail'? The crow got the treat, he just tried both tubes. It seemed to me that once he realized the blue side wasn't the right one (after 2 tries) he moved back to the red side until he could get the treat...to me that's a win!

They tested with multiple birds and multiple trials per bird. Overall they put about the same number of stones into the connected tube as the unconnected tube, actually getting worse over repeated attempts.

This graph shows their performance for each experiment.

I see. Perhaps they should have used an instance that showed their conclusion better then, since they had many to choose from.

Later the crow flew naked through the streets squawking "Eureka!".

My head hurts too much to read the paper right now, but I suspect that most of this can be attributed to learning, and that a 'causal understanding' is not being demonstrated here.

The important figure to look at here is Fig 3, specifically the first trial in figure 3, before the crows had a change to learn each particular task. Performance is, slightly, better than chance for each task except tasks D and F. But these are NOT naive birds. There are only 6 crows used in this experiment, and before the experiment they were all trained to perform a water displacement task for food reward. The better than chance performance on the first trial for some of the tasks could, and probably is, a result of the similarity between the 'correct' choice in that task and the training condition.
A: A water filled tube is more like the water filled tube in the training condition, than is a sand filled tube.
B: Heavy blocks are more like the heavy stones used in the training condition than are the light blocks.
C: Solid black weights are more like the black stones used in the training condition than are the hollow cages.
E: A nearly full tube is more similar to the training condition (which took only two stones to reach the food) than a nearly empty tube.
Tasks D and F, where the crows did not perform better than chance, have choices much more difficult to distinguish from the training condition (with both choices in F being nearly identical to the training condition).
What's more, the sample size is tiny, with a max of 6 birds for any task, and only 4 of these same (very experienced) birds for tasks E and F. You cannot do meaningful statistics with a sample size of 4.
If someone who has actually read the whole paper wants to show me how I'm wrong please do. Until then, I think the editors at PLoS ONE need to be more careful with their choice of reviewers. The journal, although noble in it's stated mission, is still young, and its reputation is still fragile.
"Our results indicate that New Caledonian crows possess a sophisticated, but incomplete, understanding of the causal properties of displacement, rivalling that of 5–7 year old children." How on earth did they let that pass! They know the press will have a field-day with misleading statements like that.

too soon.