For example, how to align AI and human objectives, how to incorporate our uncertainties and irrational decisions? In short, we need algorithms that would show that the end goal is beneficial to humans. But our overall mission is to develop the so-called general-purpose AI, which by Russell is “…a method that is applicable across all problem types and works effectively for large and difficult instances while making very few assumptions.” Sounds good, but as he notes in many instances across the book, many precautions need to be taken along the way. As we know, many applications have the AI label from speech recognition, vision systems, translation, etc. Russell takes us back in history to revisit the routes taken in human learning, decision making, and actions taken, as well as how we transferred the learning process onto a machine. The question is, how do we create an AI system so that we don’t need to consider the control issue, or lack of control for that matter, in every scenario of AI application? How do we overcome the problem of the AI system being willing to be shut down. Human Compatible is an honest, eye-opening account of the AI potential and the right way forward, its current shortcomings, and, if we choose the wrong path, our doom. If you are one of the passengers on the high-speed train, please take the blindfold off and read this book to be able to act accordingly before your research goes a step too far. Some researchers already ceased their research efforts because the results and methods are being used in a way that is not beneficial to society. But does it? Are we driving blindly? Most likely.

It seems as if any project somehow sounds and looks better if we add a pinch of AI to it. Do you notice the problem? With many notable achievements in the field, the AI revolution started with many jumping on the high-speed train. The research in AI is growing without the rules or regulations, but with the common goal of achieving the general-purpose AI.

Human compatible book review