Even leaving aside the reliability issue (which can be chalked up to the fact that this one is a demo of a non-commercial project that got overloaded), you're comparing two entirely different things.

Check out the "static demo" pages, e.g. http://www.cs.toronto.edu/~nitish/nips2014demo/results/79133...

For this image, the University of Toronto software generates sentences like "a cow is standing in the grass by a car", whereas Rekognition only produces a ranked list of categories. ("sports_car", "car_wheel", etc.)

EDIT: this is an even better example: http://www.cs.toronto.edu/~nitish/nips2014demo/results/89407... I'm cherry-picking the cases where the algorithm does well, of course. But even if it's unreliable, the fact that this works at all is impressive.

The errors are fascinating. "a cow and a car are looking at the camera." "a band plays a group of music [...]". You could almost call them metaphors instead of errors.

what a lovely way of thinking about it.

The demo is clearly designed for the small community of machine learning researchers to play around with it to better evaluate the papers they wrote. They aren't selling a product and probably have a hard time justifying using a lot of computing resources to host the demo. Furthermore, the models are probably optimized for result quality, not speed.