AI has been the buzz for the past few years and so it is no surprise that there are AI books aimed at a general audience now. One such book is You Look Like a Thing and I Love You written by Janelle Shane. She apparently is behind the AI Weirdness website, which I do not follow. This book is written for the non-tech layman, explaining the basics of what is today’s AI, how it is created and deployed, and what it can and cannot do.
Folks who actually work with deep learning know how far it is from human-level Artificial General Intelligence (AGI). But how do you explain that to the layman? This book can be of help. It lays out 5 principles of AI weirdness: it is not smart enough, has the brains of a worm, doesn’t understand general problems, does exactly what it is ordered and takes the path of least resistance.
The introductory chapters show how AI is all pervasive today: search engines, text-to-speech, text correction/prediction, enhancing photos and pretty much every task we perform with our phones or computers. A chapter on learning uses a simple feedforward network to illustrate DL training. It also glosses over other types of ML like Markov Chains, random forests, evolutionary algorithms, GANs.
The middle chapters are the meat of the book, where the author repeatedly goes over examples of how AI can only solve a narrow problem it is trained for with its training data. So, we are introduced to all the myriad ways in which bad data, unbalanced data, missing data, and reward functions affect the AI created by it. We see how AI always take the shortcut during training, usually in ways unintended by the creating human.
One of the chapters compares AI to the human brain and shows in several ways how it is forgetful, easy to trip, easy to hack and subvert. These issues are already big problems for AIs and are bound to get messier as AIs take over more functions that affect humans (insurance, loan, law, policing).
The final chapters look at how many companies cheat by using humans behind what they call as AI. It also looks at how many products use a hybrid AI+human approaches, where AI handles the boring/easy tasks and calls in a human when it finds the situation difficult. (Think of the chatbots now used by banks and other online services.)
All in all, this is an easy and funny book to read, both for techies and non-techies. Not much value for ML/DL folks, decent value for general techies and great value for a general audience. As a person who is fairly involved in the DL field, I got this general feel that the book was a “could-have-been-a-blog” cobbled together from popular DL Twitter news and memes of the past few years. Another shortcoming was the overuse of text models (recipe creation for example) in several chapters for illustration. The author seems to be a cartoonist, her fun visual illustrations are peppered on every page and they form a huge part of the narrative - a huge plus that makes all the concepts easy to understand through visual aids. This book goes a long way in bringing some sanity to the AI hype that the press is constantly trying to whip up to grab eyeballs.
Rating: 3/4