The Promise and Limitations of AI

Almost everyone who talks about Artificial Intelligence, nowadays, means training multi-level neural nets on big data. Developing and using those patterns is a lot like what our right brain hemispheres do; it enables AI's to react quickly and – very often – adequately. But we human beings also make good use of our left brain hemisphere, which reasons more slowly, logically, and causally. I will discuss this "other type of AI" – i.e., left brain AI, which comprises a formal representation language, a "seed" knowledge base with hand-engineered default rules of common sense and good domain-specific expert judgement written in that language, and an inference engine capable of producing hundreds-deep chains of deduction, induction, and abduction on that large knowledge base. I will describe the largest such platform, Cyc, and will demo a few commercial applications that were produced just by educating it as one might teach a new human employee. But it is important to remember that human beings' "super-power" is our ability to harness both types of reasoning, and I believe that the most powerful AI solutions in the coming decade will likewise be hybrids of right-brain-like "thinking fast" and left-brain-like "thinking slow". That is the only path I see by which we will overcome the current dangerous inability of deep-learning AI's to rationalize and explain their decisions, and will make AI's far more trusted and – more importantly – far more trustworthy. Anyone who understood this abstract and found it interesting should find my actual talk similarly accessible – and hopefully interesting!