Books as prompts

by Zebulun Arendsee

I own far more technical books than I have read. Usually, I’ll only read the first few chapters and then turn to Google for specific questions as they arise. However, the emergence of AI is starting to change my strategy. Now, I can turn to chatGPT for answers to most of my quick queries, whether it’s for specific programming examples, contrasts between topics, or rephrasing of an explanation.

So, what does the future hold for technical books? While they will continue to serve as curated reference material, the real value will be in how AIs can specialize them for individual users. AIs can take the raw material of a book and use it to answer questions about a given topic, along with other training material. The technical books may become inputs to prompts and the AI will generate answers specific to the users questions and background. In this way, technical books serve as carefully written descriptions of systems and expositions of topics. These books may serve as reference material the AI teacher uses as it interacts with students. We will write for machines and learn from machines.

Between now and the coming equilibrium (a topic for another post), there will certainly be many hacks who write “books” in just a few days (or far less). The AI might design a table of contents, the human might edit and rearrange, then the AI generates each chapter, the user edits and rearranges, then the user publishes it as their own. Many authors are outraged by this idea, citing plagiarism and offense that the “book” was made with too little effort. But I have a rather more prosaic response. I am not offended by the existence of these books, but I believe they lack value since author’s work saves the reader little time. In a traditional book, the author synthesizes a vast amount of information. They read many primary sources and draw from deep personal experience. The author carefully curates all this content and presents it in a concise way. This saves the reader time. But if the author instead generates the book with AI, the value added is less clear. The reader may bypass the author entirely and talk directly with the AI.

The designer of the AI book, though, may offer a new form of value, though. As with traditional authors, they organize the presentation and focus the content (through prompts). They may introduce interesting contrasts and questions. They may curate the content and check the correctness of the content with more sophisticated methods. To this end, they may apply an ensemble of technologies, perhaps some computationally expensive. The product, then, is all the description of the generative process and the “book” as a carefully staged snapshot.

This approach may not be limited to technical subjects. We could create fictional universes through writing. Then the AI could generate endless content consistent with this universe. The author could curate the universe by selecting which generated content was “canon” and which was not. The AI could ensure consistency (though at the moment the models are not good at this).