What Is The Right Atomic Unit For Knowledge?
Two centuries ago, the invention of peer-reviewed articles changed the course of scientific history. What will be its equivalent in the age of AI?
If you you could travel back in time to the final days of 1832 and ask the average British citizen what had been the most consequential events of that tumultuous year, they would likely tell you about the months the nation had spent teetering on the brink of revolution, rocked by riots in Bristol and massive protests in London. Or how the passage of the Great Reform Act that summer, after a bitter constitutional crisis that nearly broke the monarchy, had redrawn the country’s political map. But almost certainly they wouldn’t point to a quiet, bureaucratic proceeding that transpired in November, during the Duke of Sussex’s annual address to London’s Royal Society, which had served as the de facto parliament of English science since the age of Newton. Amid the political chaos, the Duke had announced a seemingly minor procedural change. To increase the “usefulness and credit” of the Society, they would henceforth “allow no Paper to be printed in the Transactions..., unless a written Report of its fitness shall have been previously made by one or more Members of the Council.”
This seemed like a quiet, bureaucratic fix in a year of deafening political change. (It was no accident that George Eliot set the events of Middlemarch in the three years leading up to 1832.) And yet, that single mandate—born from a failed experiment in open, collaborative review the year before—would inadvertently invent the system that has governed the flow of expert knowledge for nearly two centuries: confidential peer review.
The confidential part of the process was, in fact, an afterthought. The original plan had been one of radical transparency. The author of that plan was the English polymath William Whewell, a Cambridge professor whose expertise ranged from moral philosophy to the physics of tides, and who would later coin the term “scientist” itself. A year earlier, Whewell had proposed a system of referee reports to address a distinctly modern-feeling problem: the growing deluge of submissions from an expanding scientific community was overwhelming the Society’s ability to vet them, leading to rising complaints about the quality of the work it was publishing. Whewell had hit upon the idea that expert reports on submitted papers should be signed and published alongside the originals, confirming the reliability and utility of the paper’s findings. The reports themselves, he argued, would be valuable contributions, “often more interesting than the memoirs themselves.”
But the experiment in open collaboration failed almost immediately. The first two referees appointed, Whewell himself and the mathematician John Lubbock, were sent a paper and found themselves in stark disagreement over its merits, struggling to draft a joint report they could both sign. The logistical burden and the unseemly prospect of publishing personal scientific disputes proved too much for the Council. So they pivoted. The written reports would remain, but they would be submitted individually and, most importantly, confidentially, for the Council’s eyes only. The open, collaborative ideal was replaced by a private, anonymous judgment.
And it stuck. Journals of peer-reviewed articles, alongside books, have survived as a central unit of knowledge for almost two hundred years. There have been other revolutions along the way, like the meta analysis or the review article, new ways of bundling scholarship designed to make it easier to see broad predictable patterns in the results, or test hypotheses against a wider array of data sets. There were technological innovations along the way as well, of course: the web introduced new ways of connecting articles via hyperlinks, search engines, particularly targeted ones like Google Scholar, enhanced our ability to find the research we needed to advance our investigations. But new containers for knowledge only emerge at rare intervals. The book, peer-reviewed paper, the meta-analysis, the hyperlinked web page—each represented a different answer to a few fundamental questions. What is the right package for this particular slice of human knowledge, given the existing publishing and communications platforms? Once you’ve gone through the trouble of conducting your research, what is the most effective way to get your findings into the minds of other people?
It should be obvious to anyone paying attention to developments with AI over the past three years that the timing seems right for a new atomic unit of knowledge to be born. And one of my favorite things about the NotebookLM project is that we’re able to float some trial balloons about what that next container might look like. This month, we’ve pushed that exploration forward along two distinct—but related—paths, by integrating Deep Research directly into Notebook and by launching a series of Featured Notebooks curated by the team at Google Research.
I suspect a number of you have used Deep Research either in its Gemini incarnation at Google, or in other variants at OpenAI. But for those of you who haven’t tried it yet, the basic idea is that you give Deep Research a topic or a complex question, and it scours the web, evaluates dozens of sources, and then writes a structured overview of what it has learned, effectively building a starter research brief on the fly. When Deep Research launched last year, I heard from a lot of people that it seemed like the first new AI tool to share a similar ethos with NotebookLM, using AI to effectively create a dream research assistant. The funny backstory of it is that I didn’t hear a word about Deep Research until it launched to the public, which gives you a sense of how big Google is as a company, even in the subset of it devoted to research-oriented software. But from the first time I used the feature, I knew we had to bring some version of it to NotebookLM.
As powerful as the original Deep Research was, the main output of the product was restricted to a structured report and a search-style list of links to the sources that the model consulted to generate the report. But because NotebookLM is designed to manage and explore hundreds of sources, in the version we launched this month the Deep Research report is only the beginning of the journey. In our integration, Deep Research gives you an overview all of the sources it found during its research phase, with annotated commentary explaining how each source relates to your original query. You can then choose to import some or all of the sources to the notebook, along with the report itself, which you can then explore or transform using the full suite of tools that Notebook offers: grounded chat with citations, Mind Maps, Audio/Video overviews, and much more.
I’ve been framing this in the admittedly lofty language of scholarship and peer review, but the truth is these tools are just as powerful when applied to the more personal research non-academics do. Imagine you’re planning a vacation to Italy, for instance. You could ask Deep Research to “create a ten-day itinerary for a family with young children, focused on food and history.” The system would scour the web for travel guides, Wikipedia pages on Italian history, museum opening times, and synthesize all that information into a detailed starting plan. And crucially, it would also give you all the source material it consulted, which you could then import into your notebook to refine your plan, find new restaurants, or create audio overviews to listen to as you travel to your next stop on the itinerary. You’re not just getting a list of links; you’re building a personalized, interactive travel guide, a knowledge base that you can consult and expand during the trip itself.
Deep Research is all about assembling the information you need from the external world, engineering the most relevant context for whatever task you are working on. The Google Research collaboration, on the other hand, is more about output. Every year the Google Research team publishes upwards of a thousand papers on a vast array of topics: the future of machine learning, breakthroughs in quantum computing, the application of AI to fighting wildfires or forecasting floods, and much more. Up until now, they had a few primary avenues for distributing that knowledge: publish the article in a scientific journal, share a PDF on arXiv, present it at a conference, or bundle up a few related papers and summarize them in a blog post. But those formats are fundamentally static in nature. The knowledge is locked inside a fixed container, a one-way broadcast from the author to the reader. You can read the paper in its entirety, of course, or search for keywords, but you can’t have a conversation with it. You can’t ask it to summarize a key finding, or explain a difficult concept in simpler terms, or connect its argument to a collection of other articles on related themes.
The Google Research notebooks take a different approach. Each notebook contains a curated collection of articles on a specific topic, effectively creating a knowledge base of Google’s best thinking on a series of compelling research questions: How do scientists link genetics to health? How can scientists know what’s in your genome? If you’re a specialist in these fields, you can read the original papers or ask nuanced questions in chat and advance your understanding of the latest developments. But these notebooks can also make these complex but important topics understandable to non-specialists or students, including two new visual modes that we just launched that that are both pretty dazzling: infographics and slides. More on them below, but here’s a taste from one of the Google research notebooks:
One interesting thing about these notebooks is that they add a few new levels of human editorial judgment and creativity to the process. You have the intellectual labor that went into the original research of course, but now you have the additional layer of curation included as well: which articles should be bundled together to create the most compelling knowledge base (not unlike curating a special issue of a scientific publication), but also the more novel act of deciding which studio artifacts will best convey the substance of the research to the widest possible audience.
As a platform for sharing research, what we’ve built so far at Notebook is obviously just the beginning, and right now we are only scratching the surface of potential partners to create these collections. (If all goes well, we might just have a notebook curated by the Royal Society itself in the next month or so.) These notebooks are dynamic in the sense that you can transform the sources they contain into many alternate forms: different languages, explanatory styles, media formats, levels of abstraction. But it’s not hard to imagine a future version where the sources themselves are updated as new information comes online, perhaps using some kind of automated descendant of the Deep Research tool. Future versions might be able to generate a complex meta-analysis of all the recent research on a topic that can be updated instantly as new publications become available. We could even come full circle and address the challenge that the Royal Society faced back in the early nineteenth century. Imagine an AI mediator that could synthesize disagreements between two referees (or two conflicting papers) and propose a consensus, avoiding the logistical hurdles that doomed the 1831 experiment. In fact, I suspect you could accomplish a surprising number of these things right now with some clever prompting. But the larger point is that once you invent a new container for knowledge, you inevitably make new kinds of knowledge possible. The peer-reviewed paper gave us an engine for scientific consensus-building that has served us well for almost two centuries. But it feels like we are on the cusp of a new framework that might be even more significant, a knowledge unit built from the ground up with AI. These new notebooks are our attempt to imagine what that future might look like.
To give you one last taste of what the format now enables, I took the text of this post and asked Notebook to generate a slide deck illustrating its main points, in a visual language that evoked the early 19th-century Royal Society context. I have a few quibbles, starting with the opening slide, which slightly overstates the significance of the 1832 peer review innovation. But as a general treatment of the argument and ideas, it’s pretty magical I think. This is one slide from it:
The entire presentation is here:
Every time NotebookLM releases an update, I’m blown away all over again. Ever since I stumbled onto it last fall, right after the audio overview feature launched, it’s become my favorite AI tool by a mile. And I say that as someone who uses Gemini and ChatGPT every single day, plus OpenEvidence for medical literature and Chartnote as my AI scribe in clinic. NotebookLM, though, occupies its own category. It’s not just a search tool. It has quietly become my central hub for thinking, learning, and organizing my own knowledge.
I’ve used it to explore topics I never would have dug into before. I went from watching *Vikings* on Netflix to learning the real history of the first-millennium Norse world. Then somehow that turned into reading about molten-salt nuclear reactors, which led to a deeper understanding of fission and fusion than I ever had back in my physics and engineering days before medical school. NotebookLM didn’t just give me answers. It made the learning addictive.
And now these new Slide Deck and Infographics features? They’re on a whole different level. They hit me the same way audio overviews did when they first came out. The funny thing is that when I show this stuff to people, only a handful seem as excited as I am. I can’t tell if it’s because they don’t quite grasp how transformative this tech is, or if they just don’t enjoy learning the way I do. But honestly, if the Debbie Downers would give it ten minutes, I think they’d be hooked.
Most of us have said at least once in our lives, “If only I had this back when I was in school.” NotebookLM is the first time I’ve said that and actually meant it. In high school, college, and especially medical school, this would have changed everything. One “Anatomy” notebook. Drop in my notes, textbook chapters, atlas images, lecture audio, whatever I had. Then I create flashcards, quizzes, slide decks, audio overviews for when I’m at the gym. It’s crazy to even imagine!
NotebookLM has made learning fun for me again. That’s not something I expected to feel in my fifties. And I can’t wait to see where you take this next. At a time when the news is always talking about how AI is leading to unemployment, the dumbing down of the population, and possibly doomsday, I wanted to thank you for building an AI tool that makes the world better in a very real way.
The slide deck gave me a queasy feeling because the "training data" used by the AI model to generate the imagery is copyright works by human artists. (Also queasy-making: the Escher-like involutions of the tome pictured in your final slide. Kinda sloppy in the AI sense.)
A big part of the reason the established peer-reviewed model is collapsing is due to the strain of genAI invading scientific research. GenAI "papers" of a quality that hovers around dogshit have overwhelmed the editorial desks of academic journals to the point that some have simply shut down. Then there's lazy or unscrupulous reviewers using genAI to summarize and write their peer reviews for them, and *then* there's the emerging practice of embedding directives to the genAI "reviewer" to give a favourable assessment regardless of actual quality. It's perfidy all the way down.
I see the potential upside of tools like Notebook LM, I really do, but my god the negative consequences of genAI tools are already here, and they're mostly in the range between distressing and horrifying.