« Reflections on the Development of Implementable, Winning Strategies | Main | Is Design Thinking the “New Liberal Arts”? »

October 24, 2016


Jerry Leichter

"The AI100 Study Panel offers a circular, operational answer: AI is defined by what AI researchers do." In any academic field, if you drill down far enough, you will come to exactly this answer. What do mathematicians study? Mathematics. What's mathematics? What mathematicians study. Who are mathematicians? Those who other mathematicians recognize as doing mathematics.

Sure, you can find all kinds of secondary characteristics. Someone whose work involves the analysis of the plays of Ibsen is almost certainly not a mathematician - at least when doing that work.

Mathematics, in some broad sense, involves the study of the formal consequences of sets of axioms. But are logicians mathematicians? How about string theorists? You can start endless, pointless debates on these questions.

There's AI as a field of study - defined by its practitioners, and very, very heavily influenced by its funders; there's AI as defined by marketers - anything that will make a product look cutting-edge, spiffy, new, and worth extra money; and there's AI as defined by the "gut feel" of most people - anything that makes a computer act, in some recognizable way and to some reasonable degree, as people expect other people to act. Advances in the field defined by the first of these have lead to an explosion of the second (which will last until the next buzzword takes over) and a slow but steady growth in the third. Exactly how far this will go is impossible to guess. Deep learning, statistical techniques, and "smart" brute force are proving to be much more capable than anyone could reasonably have expected. (Most experts familiar with both Go and deep learning didn't expect that a program would beat a strong professional player for at least a decade more.) Machine translation using these techniques works way better than anyone had expected. And yet ... the results, useful are they are, have a long way to go, and it's not yet clear whether the current crop of techniques can get there.

Predicting the effects of broader employment of techniques in ways that we already know work can be difficult, but it at least starts with a reasonable base of "what we already know". Predicting either what else (that hasn't been shown to work yet) will become accessible to known techniques; or what other techniques might emerge; is speculation on top of speculation. "If we can land a man on the moon, why can't we ...."
-- Jerry

The comments to this entry are closed.