CSAIL Alliances Lecture Series | Artificial Superintelligence

Abstract:
Artificial Intelligence systems, including machine learning and neural-based AI systems, have been limited in their ability to respond to situations that are different than they have been trained to address (i.e. new input data that does not intersect with training data set). Is it possible to have a machine respond intelligently to situations for which it has not been trained? The challenge lies in specifying and solving the complex interaction of executive control, goal formation, objective specification, constraint definitions and efficient stochastic topological search.

In this talk we will discuss our results constructing an artificial superintelligence based on advances in genetic programming. We will present the interdisciplinary approach used which control the efficiency of the stochastic search as a function of the high-dimensional topology of input data sets and the interacting partial topological covering of the genetic operators. Implemented using these computational mechanisms is a theorem proving system which locates low entropy solutions to the input data. We will also discuss the negative effect of structured data on locating solutions.

This first task provided to the system was to deconstruct the structure and dynamics of the investment markets for which it has postulated new theorems. We will also discuss our experiences with its operation including its psychological dysfunction, the regulatory role of emotions, ethics, morals and values, the ethical dilemma posed by defining its ethical framework, the emergence of consciousness in these systems, and the benefits, risks and challenges that artificial superintelligence poses to society.