Tim Ferriss recently wrote something interesting about the death of prescriptive nonfiction. His argument: LLMs have made information books obsolete. Why read 600 pages when you can ask ChatGPT for a personalized protocol in 15 seconds? His conclusion is to retreat—to focus on storytelling, personality, "transformation." Things AI can't replicate.
It's a smart argument. It's also, I think, wrong in a very specific and important way.
The Surrender
Ferriss's logic goes like this: standard LLMs do a "good enough" job for most people seeking information. Millions believe a free chatbot can synthesize his books and spit out a customized diet plan. Therefore, the information game is over. The only moats left are voice, taste, and deep personal narrative. He's retreating from the 10 million to focus on the 1,000 true fans.
This is a reasonable response if you accept the premise. But the premise is wrong.
The Salad Bowl Problem
Standard LLMs aren't actually good at providing accurate, source-backed information. They're good at sounding like they are. There's an enormous difference.
When you ask a general LLM to summarize a book's argument, it gives you a confident, fluent response that blends training data from reviews, summaries, Wikipedia entries, Reddit discussions, and maybe fragments of the actual text. It's a salad bowl—everything mashed together, sources indistinguishable, provenance lost.
This works fine if you want a rough gist. It fails catastrophically if you need to know exactly what an author argued, compare two specific frameworks, or extract a precise claim without hallucinated additions. The confidence is the pernicious part. The LLM doesn't say "I'm not sure." It presents its blended guess as fact.
Ferriss sees this and surrenders the information market. But you could also see it and recognize an enormous gap.
Running Into the Problem
There are two responses to the observation that AI is eating the information landscape:
- Run away. Focus on what AI can't do (storytelling, personality, curation-as-art).
- Run into it. Build a better tool for the thing AI does badly.
The "Concentrated Knowledge" thesis takes approach two. The insight is that the value isn't in chat-with-a-book (general LLMs already do that poorly but confidently). The value is in structured, source-grounded extraction from curated corpora.
Take 10 specific books on a topic. Keep their semantic boundaries distinct—don't blend them into soup. Let users query across them and get answers that are traceable to specific passages, comparable across sources, free of hallucinated synthesis.
This is what RAG was supposed to deliver, but most implementations are lazy about it. They chunk text, embed it, retrieve the top-k fragments, and let the LLM improvise from there. The "Concentrated Knowledge" approach treats the source material as sacred—pre-vetted, pre-mined, semantically mapped.