Superintelligence reading group - Section 1: Past developments and present capabilities

Abstract

This week, we’ll talk about the history of AI and where the field might be headed. AI research has been through several cycles of excitement (often accompanied by over-optimism) followed by disappointment and a “winter” when funding and interest dwindles. By around the 1990s, “Good Old-Fashioned AI” (GOFAI) techniques based on symbol manipulation gave way to new methods such as artificial neural networks and genetic algorithms. These are widely considered more promising, in part because they are less brittle and can learn from experience more flexibly. AI is very good at playing board games and has recently achieved human-level performance at tasks such as board games and image recognition. AI is used in many applications today and (e.g., hearing aids, route-finders, recommender systems, medical decision support systems, machine translation, face recognition, scheduling, the financial market) and in general, tasks we thought were intellectually demanding (e.g., board games) have turned out to be easy to do with AI, while tasks which seem easy to us (e.g., identifying objects) have turned out to be hard. An “optimality notion” is the combination of a rule for learning, and a rule for making decisions. It allows AI systems to improve their behavior over time. Most AI algorithms known today are hill-climbing algorithms–they can find a better decision than the current one, but may get stuck at a local optimum. An “optimality notion” is a mechanism to extract morality from a set of human minds. – AI-generated abstract.