Our choice of problems is telling: Small and technical, not large and important. A large, important problem is to work out the semantics of natural language—including all the required commonsense knowledge—so machines can read and understand the Web. Another is to develop robots that understand what they see and hear. Understanding is hard, so AI approximates it with increasingly sophisticated mappings from stimuli to responses: feature vectors to class labels, strings in one language to strings in another, states to states. I once had a robot that learned to map positive and negative translational velocity to the words “forward” and “backward”, but never learned that forward and backward are antonyms. It understood the words superficially. It had the kind of understanding we can measure with ROC curves. Every child does better.
What to do? Form societies defined by and dedicated to solving large, important problems. Hold conferences where the criteria for publication are theoretical and empirical progress on these problems. Discourage sophistication-on-steroids; encourage integrations of simple (preferably extant) algorithms that solve big chunks of important problems. Work together, share knowledge bases, ontologies, algorithms, hardware, test suites, and don’t fight over standards. Don’t wait for government agencies to lead, do it ourselves. If anyone wants to join a society dedicated to the cognitive, perceptual and social development of robot babies, or to learning language sufficient to understanding children’s books at a 5-year-old’s level of competence, please drop me a line.