Artificial Intelligence: A guide for Thinking Humans by Melanie Mitchell

This is a thought-provoking, accessible book. It includes some technical content but it is not difficult to read, maybe intermediate. It definitely requires your attention. This book has clarified some false impressions I had about Artificial Intelligence. The book includes some chapters explaining the history of AI. It provides a useful classification of AI - Symbolic AI, programmed using symbols (words or phrases) which are understandable by human beings, and Subsymbolic AI, which capture the “sometimes unconscious thought processes” underlying human’s fast perception (recognising faces or words). An example of Symbolic AI is Expert Systems. An example of Subsymbolic AI is Machine Learning, of which the most popular are Deep Learning and Deep Neural Networks. There are chapters dealing with various kinds of AI technologies, like visual recognition (e.g., object recognition), games (e.g., Chess, Go), languages (e.g., speech recognition). In all, Mitchell explains the technologies and their potential, focusing on how machines learn and on enquiring on the ethics and trustworthiness of AI. But most importantly, Mitchell highlights the limitations for each approach, first stating that each technology is just a limited, narrow aspect of human intelligence and providing, sometimes funny, examples of failures. My favourite chapters are at the end. Chapter 14: On Understanding, which looks into the mystery of how humans “understand” things, and Chapter 15: Knowledge, Abstraction and Analogy in AI. The conclusion of both chapters, and I think of the book (which was published in 2018), is that AI is far from achieving human intelligence, let alone superhuman intelligence. To be intelligent machines would need to acquire common sense like we do. We are either born with it or we develop it in infancy. Machines would need learn how to abstract concepts from a few examples, identify and use analogies. Because our mental models, those which form the concepts that we use, are created using abstraction and analogy. We are far from understanding how those concepts are created, emerge and or develop in our human brains, so we are even farther from embedding those in a machine. And to end a quote: “only the right kind of machine – one that is embodied and active in the world – would have human-level intelligence in its reach.”