Paul Allen's company Vulcan is bankrolling three competing teams that are attempting to create software that does well on the SATs. Early results show that it's possible to get the computer to deal with quantitative questions like "If you mix these two chemicals, what will the resultant pH be?" but, not surprisingly, common sense is hard: "Why is tap water a better conductor than distilled water?" Common sense is the reef on which all AI has run aground. I haven't heard any compelling theories on changing that; I think there's a general hope that processing power will simply chip away at the problem.
In a somewhat related vein, this week's New Scientist has a cover article on consciousness (or "salience") in fruit flies. A fruit fly brain has only 250,000 neurons (the human brain: 100 billion unless you went drinking Friday night). That complexity is within the range of simulation of a modern desktop (it'd be slow, but you could hold the data structure and move data through it). I'm sure an artificial neural network of 250K nodes would produce interesting data.