Notebook LM started out, really, as a note-taking tool – but now, it’s representing some of the most interesting genAI processes around, the transformation of documents into “live” presentations from entities that sound very human.
In many ways, people would say that it’s a transformational technology. It’s certainly something that large audiences are paying attention to right now in the world of AI – Notebook LM has made enormous waves in the industry by allowing users to take a piece of text, and make it into an AI-generated podcast, replete with “characters” that are fully non-human. You can even talk to these hosts, which, for many users, is a sort of eerie navigation of the uncanny valley.
But that process of making Notebook LM more than just a “notebook” has been a process of development and curation. That’s really interesting in its own right. So when Adam Bignell appeared on the AI Daily Brief podcast, it was an interesting window into how Notebook LM got created, and what it’s doing in the context of the modern race toward more powerful LLM systems.
Early on, AI Daily Brief host Nathaniel Whittemore asks Bignell about his background, and Bignell notes that he was a musician prior to becoming a computer scientist. He also likes to read, and as a major influence, he pointed to Jose Borges and the “Library of Babel,” a mid-century short story in which the author posits an endless library of books.
“I’ve approached computer science through that lens for a long time,” he added, “and I really wanted to work on stuff that was about language. So I’ve done other jobs in computer science. I had worked on latency analysis tools. I worked on (geospatial software). But then, yeah, the next thing was, okay, I got to get back to where my heart is, which … language and understanding things, and … that’s how I arrived at Notebook.”
Bignell talked about a long-standing goal to “create a bunch of prototypes” and the idea of pioneering what he called “tools for thought,” along with the idea of using AI to talk to a book, which he was thinking of in 2022.
“We were just trying lots of different things, and seeing: how can you find the right context in a book and talk to that context?” he said.
Eventually, he explained, he began collaborating with Stephen Johnson and Raiza Martin, and the early idea of Notebook LM was born.
“We started building ever more serious prototypes, and as we presented them around Google Apps, and then (around) Google more broadly, people started to kind of take note of them. And then we had to staff up a team…”
Here’s how he described the creative process as it formed, and how the teams behind Notebook LM, to him, are a little different than what you would typically see around Silicon Valley:
“The people who have been involved on the project – a lot of them have just been artists. A lot of (them) just write, and think about writing a lot, and (write) in their spare time. So that’s helped a lot. I think it’s helped to steer us away from being too “pure engineering,” or too Silicon-Valley-centric. The other thing is that we’re just … very scrappy – the team just really, really loves launching stuff. And so it means that we’re really motivated to build things – but build things fast – try weird things, and, yeah, get stuff in front of people.”
Podcasts, as Whittemore points out, are part of the zeitgeist these past few years, and Bignell said that was a sort of organic addition to the concept, as notebook LM moved forward. He called it a “natural pairing:” Notebook LM’s power of curation, and the idea of creating AI podcasts that are self-organized.
Explaining the project’s takeoff, Bignell also noted that early feedback propelled Notebook LM to high priority status.
“On average, prototypes don’t really go anywhere,” he said. “You’ll show them to people, but you might have to kind of beg people to use it, so you can get some usage statistics. And just the fact that we had these demos, and other people within Google were asking us to use it, and they kind of understood what it was for and what you could do with it … and they could kind of imagine the future that we were imagining …I think that was good, and gave us a lot of conviction that, yeah, this is a product.”
Later on, it was the user feedback that kept building the team’s confidence.
“Seeing audio overviews and how much people took to it really made me understand that we were building a product that was really kind of the vanguard of AI products right now,” he said. We had the space to be a little bit weird and… I don’t think we expected how creative people were going to get with it. People wanted to be part of the experimentation.”
He pointed to a UX research session where people were doing all kinds of things, like uploading their own resumes.
“We always wanted it to be flexible enough to let people find the cool, interesting things,” he said. “You know, the worst thing would be if you made it so locked down that you didn’t let people do those things, and then they couldn’t even communicate to you that they want to do, and then, you know, there’s no opportunity to do the interesting stuff. So certainly, we always wanted it to be at least a little bit open ended.”
He also mentioned other cases where people used Notebook LM with a seller’s disclosure to get a better price on a house, or made the LLM voice a character from a book, something he said he also personally experimented with.
However, he also boiled the main utility of the application down into one overarching idea:
“Usage, across students or just like enthusiasts, consumers or business (users) are all actually kind of in the same realm anyway,” he said. “People want to produce certain things. They need help just understanding stuff. Things are really dense.”
Enterprise Ramifications
In “part two” of the podcast, after the ads, Bignell and Whittemore went further afield in thinking about how Notebook LM will change things. In talking about some of the most popular applications of the technology, Bignell suggested that although there are lots of external uses, pretty many companies might use this tool internally as well, for things like knowledge-based storytelling.
And then there’s the podcast thing.
“I don’t think we should think about (Notebook LM) as a podcast platform,” he said. “I think the great thing that podcasts have done is show that people understand that you can transform documents to make them more understandable in a different format. And podcasts are one flavor of that. But, you know, we were already talking about what other kind of output types we can support.”
Meanwhile, audio overviews, he suggested, will help in talking to non-developer audiences, and he repeated his suggestion that users should “go dense” to get details out of Notebook LM‘s capabilities.
“One of the most default use cases, (is to make dense resources) into something that I can understand, and that potentially, my audience can understand,” he said. “And I feel like there’s a million use cases like that where I couldn’t have done that in any way before. It’s fundamentally new.”
Talking about a “funnel of understanding,” he described how Notebook LM can help a user to drill down.
“You don’t want just a one-line sentence explaining it, but you also don’t want the whole thing,” he said. “You want something in-between, and the best way to get a handle on this thing is to engage with it. … podcasts are a nice thing – they entice you to use the product. Like, a hook, and I found that hook even working on me.”
He gave the example of researching a new drug, like Ozempic.
“I’ve heard about it in the news. I know that it’s this new drug. I’ve also seen news that says it has real health benefits. … where previously, I might have just sort of wondered idly, or read from a news source, … now I’m going to go get the research paper, and I’m probably not going to read it in full, but I’ll interact with it, ask some questions, I’ll click the citation, and I’ll read that part in full. I find (this method) good because it actually lowers the barrier of entry for me interacting with it … I think that that is a profound use case.”
Summary and Synopsis
The universal summary, the two mused, is a difference between a middle ground between a very short synopsis, and The whole work itself. As Bignell was describing, Notebook LM can bridge that gap, bringing the user something in the middle.
He gave the example of putting one of your favorite books into the platform, and seeing what comes back: beautiful paragraphs and insights curated by the AI.
“It’s a new way to read,” he said. “It makes reading non-linear.”
And again, he pointed out how the technology can make things simpler for human readers, too.
“I find reading extremely dense, long books fun, but lots of people don’t,” he said. “How can we make that fun? If that was fun, then people might do it more. I have some conviction that that would be a good thing. But in general, I think if ever you take these productivity tasks – these things that are strongly associated with work – if they’re fun …having fun on the job is what everyone wants to do, right?”
At the end of an illuminating and wide-ranging interview, Bignell closed with an idea: what he would like to see a super-user do with Notebook LM.
“I would love to see somebody write a really Long Form thing, you know, a novel, a book, …. with Notebook (strongly) in the loop; even just critiquing the outputs. Maybe it’s helping set the stage, or giving these little aha moments …. and then you could imagine: suppose that we have all the output types I would like to have – you could do kind of the full multimedia thing with this, you could publish the podcast alongside it, and whatever else. … I would love to see somebody really use it as a creative prosthetic, (as) essential to the act of creating something that’s really important to that creator, and hopefully, to other people.”
This was such an eye-opener into one of the most popular LLM products of our times! I got an enormous response to a prior blog post breaking down what it looks like to actually chat and interact with virtual AI podcast hosts created on the platform. We’ll see a lot more of this come to life as those intrepid users experiment.
