The Fox, the Hedgehog, and the Quest for Artificial General Intelligence

Abstract

Artificial general intelligence continues to be elusive in part because of a focus on a narrow range of problems that can all be described as path problems. A researcher sets up the goals, the optimization methods, and the parameters of the problem, which determine the potential solutions that a machine can learn. Machine learning is the process of finding the right parameter values that solve the problem. Formal path problems, like the games of chess or go, are fully described by their rules and the current state. They can be treated as purely mathematical processes, independent of any physical instantiation. Other, less formal problems, such as how to drive a vehicle on a busy road, do depend on the physical properties of their instantiation and on feedback from the physical world. Path problems are well defined with relatively easy to evaluate metrics, but they are not the only kind of problem that a generally intelligent agent needs to address. Other problems, called insight problems, require the solver to not just evaluate well-defined functions, but to create those functions. Failure to recognize the need to solve multiple types of problems leads people to believe that computers will at some point be able to make themselves arbitrarily intelligent (the technological singularity), potentially to the detriment of human existence. Path problem solving has provided a good model for special purpose problem solvers, but not for general intelligence. As the ancient Greek poet, Archilochus observed, “a fox knows many things, but a hedgehog one important thing.” Artificial intelligence researchers have been able to build very sophisticated hedgehogs, but foxes remain elusive. And foxes know how to solve insight problems.