This Easter weekend, Andrej Karpathy’s post on LLM-powered knowledge bases went viral — and for good reason. He described a self-maintaining, self-correcting knowledge system built entirely by AI, and then shared the concept as a natural-language “idea file” instead of code, letting anyone’s agent build their own version. For readers who collect highlights and annotations, this is bigger than it sounds: you’re already closer to a living knowledge base than anyone else — you just need the right architecture to unlock it. That architecture has two dimensions: vertical structures that capture each book’s argumentation and logic, and a horizontal layer that connects insights across books, guided by what you’re actually trying to learn or solve.

Books have always done two things for society: they preserve what we know, and they help us create what we don’t yet know. Now AI can produce, search, and recombine text at a scale no human can match. But a closer look at what AI has actually achieved — and what it hasn’t — reveals why books remain irreplaceable.

Any book can be summarized by AI in seconds. So why would anyone still read the whole thing? The answer requires rethinking what books actually are. A book is a mediator — one possible bridge between an insight and your mind. Sometimes it’s the best bridge. Sometimes AI is. And sometimes the shortcut you think saves time quietly destroys the thing you were trying to acquire.

Choosing a book is not like picking something to watch tonight. It is a serious commitment of hours, sometimes days — and the opportunity cost of every other book you won’t be reading during that time. That’s why readers have always needed to trust their sources. But here’s the counterintuitive part: that trust has never really been about books. It’s been about finding people who have already found the books you were looking for.

When anyone can write a book, more people will. More topics, more perspectives, more niche knowledge that was previously locked inside practitioners’ heads. But what about the top — the works that change how we think? AI has had access to virtually all of human knowledge for three years now. It has not produced a single original explanation for any unsolved problem. What does that tell us about the future of books?