Over the past few decades, a vast literature has accumulated on what machines can and can’t do with respect to “intelligence”. Since Deep Blue vs. Kasparov, a lot of people have developed a reflex action – they respond to any mention of that body of work by saying, “…but that sort of algorithmic search has nothing to do with intelligence and it would be easy to make machines seem like they are intelligent by following that road, but we want the real thing”. Over the past six months, I have heard this general line from at least two relatively smart researchers.
Today, as I was watching my two-year old daughter play with her v-tech interactive train station toy, two things occurred to me: (a) even assuming that automated problem solving is one of the “easier” problems of intelligence, we don’t really have the ability to create an “interactive education-cum-entertainment toy” that can engage a grown-up scientist in quite the same way that the simpler one keeps my daughter engaged (e.g., consider the number of people who go here in search of some brainy fun); (b) while we are waiting for a fully functional general artificial intelligence (and for detractors to be convinced about truth in advertising issues), variants of this toy could play an immensely useful role in the scientific discovery pipeline.
There are a lot of things that are already in place for this purpose, but there are also crucial components missing. With the advent of a powerful arsenal of statistical learning algorithms, we really know how to arrive at reliable models of all sorts of complex processes… but this arsenal is seriously lacking in the ability to efficiently explore (on a hunch, as it were) a complex abstract space. It is hard to “guess” or “conjecture” truly interesting things that would engage me in a meaningful way, or in another role, assist me by reminding me of a direction I may have overlooked. Moreover, whatever exploration is possible seems hard to connect with the way humans might deal with the same problems. GOFAI style state space search also doesn’t seem capable of truly open ended exploration in abstract spaces of the sort I have in mind… this whole affair calls out for a more serious intellectual exploration!
PS: Having spent a few years in industry working on algorithmic toolkits, I am convinced that there is a viable market for a truly intelligent problem solving buddy – and one of my friends is already involved in this effort in the form of a startup company. However, in addition to commercial viability, I believe that this direction is a genuinely interesting, worthwhile and surprisingly open (i.e., with not many people seriously focused on it) scientific challenge.