Layar, Wikitude and Gamaray use a hybrid of the the Heads Up Display and the the Tricorder. My main pet peeve is the use of the Heads Up Display interaction pattern which is represented as a radar like scope with a wedge representing the field of view. Don't get me wrong, I spent many happy hours playing Battle Zone, but when I have to physically rotate the device (and myself) to see objects to my side or behind me, a simple 2D map is a more practical interface.
The Tricorder pattern is equally annoying, since this involves holding up the phone in front of you while scanning the horizon for alien life forms. Doing this made me feel like an alien in Herald Square. While I don't have a problem being an alien, it also made me feel like one of those self-important dorks who wore a blue-tooth headset constantly when they first came out.
Oh the promise!
Gamaray differs from Layar and Wikitude in that it implements the Holochess pattern by allowing users to place 3D object in the field view. This where things start getting interesting because users can interact with the application and reality instead of passively viewing info bubbles. It's not a far jump to the locative art installations as described in William Gibson's book Spook Country. The Mannahatta Project reconstructs the changing ecology of Manhattan back to 1609, and a AR app could overlay the historical view over the current view. AR applications are a natural fit for these visualizations. Temporal overlays could also be used to overlay houses,tax parcels, and demographic data in areas hit by tornados or hurricanes for use by emergency workers and insurance adjusters.
Where's the content?
Granted that AR is in its infancy, but the currently available content detracts from the idea of AR as a useful tool. A number of apps use Wikipedia as a starting point for content, but the entries in Wikipedia are scale inappropriate and are spatially coarse to be of interest. Tobler's law and its corollaries lands with a heavy thunk. Wikitude, Layar, and Gamaray all support user created content but there doesn't appear to be standardization around any particular content model. Crowd sourcing geospatial content worked well for Google, but Google had the advantage of a single data schema in KML as well as a single API.
AR is a natural fit for built environments but geospatial data is rarely collected at the resolution of individual manmade structures and data collection at this scale often means 3D. CAD data does exist for many urban areas but it is frequently locked up in hard drives of various planning agencies and engineering companies; its difficult to imagine that any of that data will be readily available. An alternative is Sketchup and apps are starting emerge, but none on a mobile platform so far.
Not significantly advanced technology
Considering that 18 years ago I carried a GPS that was larger and heavier than my current notebook and 7 years ago my Garmin Etrex was bigger than my circa 2000 Motorola Startac cell phone, you would think that I would be pretty happy with a device that combines GPS, compass, cell phone, camera, internet terminal, and music player into a single package. Actually, I am happy that I own this technological wonder that can begin to implement AR, but where AR apps fail to be indistinguishable from magic is that they don't use all the sensors. Amidst all the trumpeting that the compass was the killer feature on Android, features such as bluetooth, wifi, and of course the camera seemed to have been forgotten.
Makers have been busily hacking at phone cameras to make scanners and there are accessories to turn the phone camera into a microscope. An application that makes use of image recognition to identify objects in the field of view instead of using proximity and bearing to the phone's location would be infinitely more magical the current applications. An AR app can conceivably take a picture and send it to an image recognition service and then provide additional information based on the picture just like the folks at astronomy.net. That would be magic indeed.