A Mathematical Description of the Improbability of Superintelligence

초록

In this talk I will examine whether the “superintelligence” will ever emerge. In order to do this, I suggest a mathematical framework with which one can see whether artificial intelligence has reached such a level, adopting Nick Bostrom's definition of superintelligence, “intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest” (Bostrom 2014). One can measure the volume that an intelligent system contains in a multidimensional space of cognitive faculties. If the superintelligence in Bostrom's sense is realized, it will assume a volume that encloses the volume of human-level intelligence as a proper sub-volume in that space. However, it will be a hard task for engineers to develop artificial intelligence that comprises all the dimensions of human intelligence. The prospect for superintelligence gets essentially worse, if we consider that human intelligence develops new aspects of intelligence that correspond to new dimensions in the space mentioned above. But there is still possibility that artificial intelligence develops an exotic dimension of intelligence that we human beings neither possess nor understand. Though it is not the superintelligence in Bostrom's sense, it can threaten human controllability over technologies. Even if it is not certain that there is some intelligent agent behind it, it certainly implies practical risk, of which we should be aware.