Inside the Pod: The AI Research Assistant You've Been Dreaming Of

Extracting patterns from a lifetime of Johnson’s book highlights

Johnson’s first Notebook has about 7,000 quotes he jotted down from books, a practice he started in 1999. Over time, he shifted from manually typing them out to digitally saving highlights from his ereader.

Each “source” shown in the screenshot below represents a collection of book quotes. Johnson selects one labeled “SBJ Readwise 1” to explore the features of NotebookLM further.

All screenshots courtesy of AI & I.

As an author, Johnson clarifies that no material uploaded to the model is used to train NotebookLM or Google Gemini; it's only sent to the model’s context window, or “short-term memory.” Johnson explains that if you “have the right to use [the material] under copyright, you can use it inside of Notebook.”

Johnson explains that for each source, the AI model creates a guide that identifies patterns from the information in the source, and generates a paragraph-long summary. Here’s the “Source guide” for the collection “SBJ Readwise 1.”

Johnson is amused to see the model struggle to find patterns between his disparate book quotes. He notes that the Source guide is most valuable when summarizing information on a single topic, rather than trying to find patterns across unrelated subjects.

Users can also “talk” to their sources—NotebookLM will answer any questions based on the information they’ve uploaded to the model. They can even select which sources they want the model to refer to in their answers, effectively “shift[ing] the focus” of the AI. Johnson demonstrates this with a question he had pre-loaded into NotebookLM.

Johnson: What are the most interesting facts about ant colonies here? Mention authors and specific books.

Johnson explains that he chose this query because he wrote a book called Emergence in 2002 that features a section on ant colonies. NotebookLM, he says, is able to grasp abstract concepts of what might be interesting and surprising, which he finds particularly valuable in his work as an author.

As Johnson examines NotebookLM's response, he highlights a key feature: in-line citations. These citations are interactive—users can hover over them to view the specific text that informed the model's answer. Further, clicking on a citation takes the user directly to the relevant passage in the original source document. Johnson emphasizes the significance of this feature, which integrates research and writing tools, streamlining the user’s workflow on a single platform.

Johnson adds that NotebookLM encourages users to deepen their research by suggesting follow-up questions.

Extracting patterns from a lifetime of Johnson’s book highlights

Each “source” shown in the screenshot below represents a collection of book quotes. Johnson selects one labeled “SBJ Readwise 1” to explore the features of NotebookLM further.

All screenshots courtesy of AI & I.

Johnson: What are the most interesting facts about ant colonies here? Mention authors and specific books.

Johnson adds that NotebookLM encourages users to deepen their research by suggesting follow-up questions.

The questions give Dan an idea: He wants to ask the model if it’s able to extract patterns about Johnson’s sensibilities based on the collection of quotes.

Johnson: These sources are my reading notes from books I've read over the last 20 years. Please describe my interests and general sensibility? Based on these quotes, what do I care about most, and what can you tell me about them?

Johnson says that while NotebookLM’s answer is technically accurate, he acknowledges that it lacks depth and specificity. Intrigued, they decide to push it further.

Johnson: Can you please try to speculate on my personality based on my decision to clip these quotes? I know you don’t want to but I’m asking you nicely.

NotebookLM continues to hem and haw, and Johnson explains that this is because the model has been designed to be “grounded” in the sources provided. In other words, NotebookLM is likely to decline to answer open-ended queries because it’s intentionally constrained by the information uploaded to it. Johnson explains that future versions of NotebookLM might have a slider of sorts, allowing users to adjust how closely the model sticks to its sources versus engaging in more speculative responses.

Coming back to the question they asked NotebookLM about Johnson’s personality, the model's struggle to provide a specific answer stems from a lack of source material addressing Johnson's psychological traits. Recognizing this limitation, Dan has another question for the AI, one that has more to do with the information in the uploaded sources.

Johnson: Of all these quotes, what are the two authors who most disagree with each other, whose positions are the most opposed?

NotebookLM pulled out a banger—the example of two influential anarchist activists, Johann Most and Emma Goldman, who strongly disagreed about the use of violence in the anarchist movement, leading up to a public confrontation in which Goldman attacked Most. Johnson remarks that this is perhaps the only instance in his quote collection where two of the sources were embroiled in a physical altercation. Dan has a follow-up question for the model.

Johnson: In the dispute between Emma Goldman and Johann Most over violence, which other author in these quotes would have been most helpful in resolving and mediating their dispute? And why?

Johnson validates NotebookLM’s response, explaining that Peter Kropotkin, the anarchist philosopher it identified, was indeed known for his nuanced stance on violence. Impressed, Johnson notes that Kropotkin was precisely the figure he had in mind for this question.

Finding a concept for Johnson’s next project, and using NotebookLM to bring it to life

Johnson is interested in making a documentary about the fire that broke out on a pre-flight test of Apollo 1 in 1967, based on the transcripts of the interviews that NASA conducted on the history of space flight. He wants to explore how NotebookLM can help him prepare for this. He has loaded seven of these oral histories, totaling around 200,000 words, into the model.

NotebookLM's “Notebook Guide” feature automatically distills information into readable formats like FAQs, timelines, and briefing documents. Johnson notes that this eliminates tedious work while providing essential scaffolding for his writing process. As an example, here are the FAQs and timeline that NotebookLM came up with.

Johnson explains that the transcripts are not focused on the Apollo 1 fire. He says there’s a “giant haystack of NASA-related information,” and “something the size of a shoe”—not a needle—about the Apollo 1 fire. In an attempt to use NotebookLM to sift through the research, Johnson prompts NotebookLM as follows to create a reader’s guide.

Johnson: I'm the author and TV creator Steven Johnson. I'm interested in making a TV documentary about the Apollo 1 fire in the multidisciplinary style of my books and shows like The Ghost Map and How We Got to Now with a focus on surprising scientific explanations and compelling narratives. Give me a reader's guide to the most important sections of these interviews that I should read in getting started with this project.

NotebookLM goes through each interview transcript, pulling out the parts relevant to the Apollo 1 fire. Here’s part of the reader’s guide it generates.

Johnson demonstrates NotebookLM's efficiency by clicking an in-line citation, scanning the source text, and saving the parts he found interesting as a note on his dashboard for the project. This seamless process, he notes, exemplifies how the tool integrates research, analysis, and organization in one fluid workflow. This is what the saved note looks like.

As Johnson gathers more notes for the project, Dan thinks about how to push the research forward. Dan notes that a distinctive part of all of Johnson’s books is finding the “pivotal moment where everything changed,” and then going back in history to trace the technological innovations that led to it. Dan wonders if NotebookLM can help them create that moment.

Reusing the quote he used to generate the reader’s guide, Johnson prompts NotebookLM as follows.

Johnson: I'm the author and TV creator, Steven Johnson. I'm interested in making a TV documentary about the Apollo 1 fire in the multidisciplinary style of my books and shows like The Ghost Map and How We Got to Now with a focus on surprising scientific explanations and compelling narratives. I want to find a scientific idea or scientific or technological idea that is central to the Apollo fire that I could develop into a major set piece for this project. What would you recommend based on these sources? Ideally the scientific concept will be surprising and involve an unusual connection that the viewer might not have originally thought of.

NotebookLM's mention of pure oxygen catches Johnson's attention because it mirrors the initial concept he had for the documentary’s introduction, independent of NotebookLM. He notes the connection to his book—The Invention of Air—which explored the 18th-century discovery of oxygen. This overlap interests Johnson, as it suggests a potential narrative link between his previous work on oxygen's history and the Apollo 1 incident.

Dan suggests revisiting Johnson's quote collection in NotebookLM to explore references to pure oxygen environments. Johnson goes along with the idea, prompting the model as follows.

Johnson: I’m writing about the use of a pure oxygen environment that caused the Apollo 1 fire. What quotes in these sources could be relevant to the use of oxygen and its history? Explain how I could use those ideas.

Dan and Johnson are especially interested in “Source 3,” surfaced by NotebookLM, because the model makes a connection between scientist Auguste Piccard’s research and the pure oxygen environment that caused the fire on Apollo 1 spacecraft. Johnson saves this source as a note on his dashboard for the documentary project.

Dan proposes using NotebookLM to uncover links between Piccard's research and the NASA interviews. Johnson implements this idea by selecting his recently saved Piccard note on the dashboard and clicking on the “Suggest related ideas” function, setting NotebookLM to work mining his sources for thematic connections.

This is part of what NotebookLM generates in response.

Johnson remarks that the model’s response has a “fusion of so many different separate intelligences.” This includes the author who wrote about Piccard’s research, the NASA interviews, and the user who is driving the process, as well as the AIs, Gemini Pro 1.5 and NotebookLM, that are synthesizing the information and making connections. “I used to talk about a duet between human and computer,” he says, “but this is a full chorus, right?”

Johnson selects all the quotes that he’s recently added to his dashboard and prompts the model to generate an opening script for the documentary.

Johnson: I'm the author and TV creator, Steven Johnson. I'm interested in making a TV documentary about the Apollo 1 fire in the multidisciplinary style of my books and shows like The Ghost Map and How We Got to Now with a focus on surprising scientific explanations and compelling narratives. Based on these notes, suggest an opening script for a documentary episode about Apollo 1 and the pre-history of space flight. Suggest images from these sources that could be relevant.

NotebookLM generates a compelling script draft, impressing both Johnson and Dan.

I think tools like NotebookLM are most helpful to those who maintain a written record of their thoughts and ideas, because with AI, the richer the context that you’re able to give the model, the better the result you get will be. In this way, it’s a note-taker’s dream—and might inspire others to be better note-takers, now that there’s a tool to synthesize even the most disorganized trove of notes.

Rhea Purohit is a contributing writer for Every focused on research-driven storytelling in tech. You can follow her on X at @RheaPurohit1 and on LinkedIn, and Every on X at @every and on LinkedIn.