21-07-2025
A tool that can make sense of a screen full of documents and tabs? Sign me up
I am writing this column in a text editor on my laptop. I also have a browser (Firefox) running that has – pauses to count – 18 tabs open. Each one represents a web page containing information that I searched for when planning the column: websites , YouTube videos , a list of relevant podcasts to which I should have listened , pdfs of relevant reports , notes that I've made when reading through the sources I've consulted , lists of links that are conceivably relevant – etc, etc.
And, somehow, I have to weave a coherent narrative from all the stuff in those tabs. Cue violins?
Save your sympathy: I'm just an ordinary Joe facing what confronts millions of 'knowledge workers' every day. As Steven Johnson, one of the world's best science writers, puts it : 'You find yourself in these situations where the job you were trying to do involves synthesi sing information that is scattered across 15 open tabs and a bunch of documents sitting on your drive, and, you know, wherever that is, all over the place.
'And whatever you're doing, whether you're writing a book or working on a marketing plan, whether you're conducting user research and you've got 20 different transcripts, or you're working on a documentary and you got all these interviews – it's just very hard to manage.'
And switching from tab to tab, trying to find your way to the information that you're looking for, just doesn't work any more.
We're long past 'information overload' and have moved onto something much worse: a phase of cognitive fragmentation, when information is not just over abundant but scattered, disorgani sed and contextually disconnected, making it nearly impossible to synthesise or see relationships between pieces of knowledge.
What we need, Johnson thought, is a new kind of computing tool that would enable us to spend less time trying to organise all these disparate sources of information. That could be done by providing a single virtual space into which we could dump all those pesky tabs and then get an AI that's 'read' your documents to help you make sense of them.
As it happened, some researchers at Google Labs had been thinking along similar lines and so Johnson hooked up with them – and NotebookLM was born.
It's a remarkable tool that, to date, seems to have flown under the radar of public attention, perhaps because people thought it was just some esoteric tool for academics. The basic version is free ( there is, as ever, a premium version) , it's easy to use and it runs in your web browser. To engage with it, you have to identify a number of sources of information on a topic in which you're interested and upload them to the tool.
Each notebook you create can have up to 50 sources in various formats: audio files, text (both pasted into NotebookLM and in common file types), Google Docs, Google Slides, Markdown (a simple markup language for creating formatted text) and pdf files, web URLs and YouTube videos.
Once the sources are in place, the fun begins. NotebookLM allows you to explore the sources in a variety of ways: by providing summaries or analyses ; picking out common themes or divergences; preparing study guides or a briefing document; creating a visual mind map that may be useful ( for example, in a PowerPoint presentation); creating timelines or notes on particular aspects of the sources in response to a query – and so on.
Intriguingly, it can also generate what Google calls an 'audio overview' of the material you've uploaded in the form of an AI-generated podcast-type conversation. This last feature has now gone viral in a modest way after the New York Times tech journalist Kevin Roose uploaded his latest credit card statement and then had to listen to a pair of AI-generated strangers discussing his profligacy.
Of course, underpinning all this is a large language model (LLM) – in this case, Google's Gemini 1.5 Pro – but here it plays a different role than in its usual form. What Google has done is to use a powerful general-purpose LLM to create a personali sed and private AI tool that has expertise in the information the user gives it. So rather than being a general-purpose chatbot, it was designed to be deeply knowledg able about your specific documents and research materials.
This is achieved using a fancy technique called retrieval- augmented generation (RAG). Within NotebookLM, RAG improves the performance and reliability of Gemini by incorporating external, verified sources of knowledge (the sources uploaded by the user) into its text-generation process. So rather than relying solely on its internal training, it actively retrieves and references the sources as a way of keeping it 'grounded'.
I've been using NotebookLM for a couple of years and have found it genuinely useful for both academic and journalistic work. It is a good example of technology being used for the augmentation of human capabilities rather than as a means of undermining or displacing them.
And it's a pretty good way of controlling browser-tab bloat.
What I'm reading
Admission of gilt
Trump's Gilded Design Style May Be Gaudy. But Don't Call It 'Rococo' is a nice Bloomberg essay by Feargus O'Sullivan .
Brought to book
Regina Munch's essay Encounters with Reality in the Point magazine offers a critical appraisal of Christine Rosen's interesting book The Extinction of Experience.
Town crier
Why I Moved Out of Temple Bar After 25 Years has Frank McDonald writing in the Irish Times about how the coolest part of Dublin became uninhabitable.
