One thing that stands out for AI compared to other types of technology is the confusion and myths it has created in public discourse. Too many, AI seems like this magical thing that, variously, can solve all problems or take over the world and doom the rest of us. The more moderate version of this in the field of IP is the notion that somehow AI can independently solve problems and make inventions. In "YOU LOOK LIKE A THING AND I LOVE YOU", Janelle Shane gives short shrift to these notions and debunks the myths. With examples ranging from optimising cockroach farming to a magical sandwich hole, via a quick stop to generate some pick-up lines, Janelle shows how AI (read machine learning) works and what it can and cannot do. While Janelle mentions some specific technologies such as CNN, RNN, and GAN, the stories she tells illustrate the broader point of how computers learn from data guided by an objective function, a narrow process devoid of any common sense or human-like cognition. Understanding this is crucial to appreciate issues such as bias and explainability and the caution that must go into developing AI solutions.
The book's central message (if you ask me) is to debunk the myths and enable us to understand the circumstances under which AI works well and can generate tremendous value, as well as the circumstances in which this is not the case. AI is good at answering narrow, specific questions, like recognising objects in images, completing sequences of words or playing games with set rules. AI is terrible (really terrible) at answering broad questions like should you trust this person as your babysitter, how likely is someone to re-offend, would this person be a good hire? These points are illustrated by silly and amusing, as well as serious and thought-provoking, examples as you compulsively turn the pages. And while Janelle talks about AI in humanising terms - it learns, helps, tries etc., which I think generally contributes to many of our misconceptions, in the context here, it works well and makes the book very readable.
One thing to be cautious about, though, is the pace of innovation in this area. While the book was written when GPT-2 was state of the art, some of the issues like memory constraints and catastrophic forgetting are now less visible as GPT3, BERT, and all kinds of gigantic transformer models produce ever more impressive feats not only in language generation but also generating images (DALL-E, Imagen) and learning a variety of tasks at the same time (GATO). However, while it would be nice to expand the discussion to these huge transformer models, the fundamental points Janelle drives home have not changed.
In short, if you are into AI (I assume you are if you are reading this) and you are looking for something at the same time amusing and enlightening to read on the beach this summer, add this book to your reading list. Alas, having just finished the book, I will need to find something else to read, but I am looking forward to the sequel - Janelle seems to be creating plenty of material for one on her blog.