“Understanding (almost) everything about artificial intelligence”: a book that keeps its promises

With its somewhat catchy title, "Understanding Everything (or Almost)..." , its thin spine, its sequence of questions, its simplistic drawings and its pages with little text, let's admit that this book seemed unlikely to distinguish itself from similar productions. Wrong. The promise of the title is more than kept: a reminder of basic knowledge is mixed with dives into recent scientific advances, even an introduction to research questions that are still open.

The success lies in the fact that each of the 20 questions and answers was written by different researchers (only three women), without the book's level or style suffering. The chapters are unadorned, getting to the point, with paragraphs reduced to one or two sentences. They conclude with a "final word" summarizing the four or five pages of each subject. The numerous illustrations serve as both educational support, providing information, and as aesthetic graphic elements. The overall effect is dense.

The first three questions sweep through the history of the contemporary era, leading us to the explosion of the generative version of artificial intelligence (AI). The next five open the hood of the concepts of learning, neural networks, "tokens" (language units), inference, convolution, transformers... Even the technique of generating images using a diffusion model is explained in one page and five drawings.

A series of questions follows that examine the "intelligence" of these tools. Are they better than us? Are they creative? What do they understand? The answers highlight the limitations and the many problems raised by these new technologies.

After four other chapters on applications (in health, robotics, research), again offering a critical look at what is done, is feasible and is not yet achievable, the book ends with four questions with a more social scope. The themes of respect for privacy, security, gender bias, origin, social class or ethics stimulate the reader's reflection and are an opportunity to testify to the vitality of research in these fields. Here again, there is no debasement to keep things simple: the authors speak of "federated learning" , causality, "optimization bias" ... And, despite the reduced size of the book, manage to slip in that the term "hallucination" , very often used to designate a classic defect of AI, is wrongly used.

You have 15.37% of this article left to read. The rest is reserved for subscribers.