I think that when you're not expected to publish any papers to rationalize what you're doing, you're free to use any possible ugly hack to improve your results, (using a "kitchen sink" approach where you just combine the results of lots of unrelated techniques, extracting words from the URL, using the URL to actually fetch some related textual content on the website, etc). This gives private companies a competitive advantage over research institutions - their only purpose is to "make things work", not to introduce new techniques and have interesting insight about them.
Lots of companies and teams are exploring deep neural network with all kinds of application. Rekognition API is the only one I found that provide open API service right now. You could train classifier using your own images. But you need to create an account and upload your images using their web application.
Rekognition:
7.55% fruit; 0.92% dinner; 0.88% produce; 0.87% alcohol; 0.84% sliced
Toronto:
50% American lobster, Northern lobster; 12% plate; 7% crayfish, crawfish, crawdad; 7% Dungeness crab, Cancer magister; 4% king crab, Alaska crab; 4% butcher shop, meat market; 4% grocery store, grocery; 4% pomegranate
I find this interesting because I thought Hinton's group had state of the art tech. Who are these people and how do they do it?