The reference you give isn't for computer vision by the way. The ImageNet project isn't for image recognition. It's for a curated database of images which researchers can use to train algorithms or do other image based research. The error rate shown in the graph on that story (which you submitted) does not show computer vision image recognition error rate and %. It shows the number of errors in their own database which is human curated. In other words, right now if they have 1,000,000 images of cheesecake, and someone searches for cheesecake, and looks at the top 5 results they get, ~96.5% of the time the image will actually be cheesecake.

Having worked on a rather large, and well funded, computer vision project for one of the largest companies in the valley, I can safely say a few things. First, ImageNet database is very good and useful for image research. If you are going to be training an algorithm and need millions of well curate images, they are one nice place to go. Second, we are no where near approaching a 0% error margin on computer vision. Third, AI is certainly a thing and should be watched. It's neat and computer vision is one very neat and highly useful aspect of it. But please do not for a second believe that in 2016 a computer will be able to look at any random image of anything and actually know what it is.