Fifteen years on, I'd have thought my AI complementarity principle would be widely understood...
The idea is that before you can accurately identify (eg) boners in an image, you must also be able to deduce the lightsources and body position and skincolor in that image... and the first (comparatively easy) step in doing this is to gather algorithms that can generate a visual representation of every possible combination of these.
The target image can then be matched to multiple hypothetical models, and the best match selected.
A simplified approach might look just for the cylindrical boner, while deducing the lightsource.
AI startingpoint: 3D model of boner plus stomach plus thighs, from every angle, at every distance, with every lighting, and every skincolor.
Once you've got that working so it can reconstruct, from just the visual output, every permutation it can generate, then you can start addressing the exceptions like clothing, hands, softies...
They claim they won't use Web data, but there's no way they can compile enough databases on their own to handle Jeopardy's general knowledge. Awards, lyrics, plots, characters... the list goes on and on and on.
WolframAlpha is a recent disappointment that's spent years collecting databases and delivers almost nothing useful yet.
I'd suggest celebrity blind-items as a fun test domain that might be manageable, eg:
"Which female rocker, who was once married to a famous old-school singer, now has a penchant for young girls?" [cite]
Even textbooks on this topic don't usually spell out the very simple dependence between atmospheric depth and surface temperature: when you warm the Earth, air molecules 'bounce' higher, so the atmosphere gets deeper. When you cool it, they bounce less high. The higher they fly, the slower they move, unintuitively termed 'adiabatic cooling'.
A small percentage of the highest bouncers can be reheated by the Sun near the top of their bounces, and I assume the reported lower ionosphere is more due to a decline in this factor than to any global cooling.
According to the latest official figures, 43% of all statistics are totally worthless.