A term at least popularised and maybe coined by Douglas Hofstadter to refer to a very specific and constrained realm in which to develop and test AI techniques. The basic idea is that rather than try just to produce artificial intelligences with immediate applications to the real world in all its ridiculous complexity - which is the general gist both of Good Old Fashioned AI (GOFAI) and of the more recent trend towards developing "expert systems" - we should first try to sort out the fundamental problems of intelligence in a toy world testbed. By one analysis, GOFAI was driven down the wrong track by naivety, and expert systems by the lure of quick profits, and neither really do much to advance pure AI research. Sure, we have programmes that can beat the human race at chess, and which can do a decent job of finding the best site to drill for oil next, but sadly we still can't make a computer know, well, anything really. Never mind think. Or abstract. Or make analogies. Or come up with new ideas. Or know just what a good idea is.
Which is why we need to restrict ourselves, for now, to the small-scale. Hofstatder's research group - the Centre for Research into Concepts and Cognition (CRCC) - has done some good work in the fields of analogy and creativity, by working in microdomains such as letter-strings (answer questions like "if abc goes to abd, what does xyz go to?") and letter-forms (trying, given a few sample letters, to complete the typeface "in the same way"), but disappointingly little seems to have been done by others (please /msg me if you know otherwise).