A Conversation with NotebookLM's Founding Engineer

0:00Today on the AI Daily Brief, a conversation with Adam Bignel, the founding engineer of Google's Notebook LM. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. To join the conversation, follow the Discord link in our show notes.

0:18Hello, friends. Welcome back to the AI Daily Brief. I thought since we've had two heavy days of deep seek and contemporary news, we might all want a little bit of a breather. And for that, I give you something a little bit different than we normally do. What I'm about to share is a conversation with Adam Bignall. Adam, as you'll learn, is a founding engineer of Google's Notebook LM, but he's also a writer, a musician, and a lot of our conversation is about the particular set of quirkiness that made Notebook LM what it is. Now, for those of you regular listeners, you will know that I call Notebook LM the most important Gen AI product of last year.

0:49I think it sits at that perfect intersection of not only doing old things better, but also opening up totally new opportunities. In the conversation, we learn about how Notebook came to be, some of the most interesting use cases that Adam has found and get some hints around where it might be going in the future. All right, Adam, welcome to the AI Daily Brief. Excited to have you here. Thank you. Excited to be here. Yeah, so we're talking today about, like I was just telling you, what I called my number one Gen AI product for last year, which I've talked extensively about on this show, but we're here to talk Notebook LM.

1:21I think before we do, though, I'd love just a little bit of background and context on you. And it doesn't have to just be how you came to work on this particular thing, but whatever Whatever's important to know for context for this particular story. Yeah, cool. So, yeah, my name is Adam Bignall. I got actually to computer science in a little bit of a roundabout way. So I made a lot of music. Like, yeah, I published a ton of music. And I went to, like, a one-year technical music program before I ever did computer science. And I also just, like, love books and have been kind of, like, a lifelong reader.

1:57And I actually, like, really didn't understand what computer science was. I didn't do any of it in high school or anything. And then after graduating, I had worked at bookstores for a really long time. And I encountered Borges Library of Battle. So that book really kind of, I don't know, or story, I guess, but that one really got me interested in how to think about navigating knowledge, I guess. So, you know, it's a story that's set in an infinite library and there's every possible 410 page book on the shelves. And the characters in this infinite library spend their whole lives like reading books and trying to find ones that are meaningful and stuff.

2:42So the reason I mentioned that is I think like I've approached computer science like sort of through that lens for a long time. 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 a geospatial kind of software. But then, yeah, the next thing was, okay, I got to get back to where my heart is, which is language and understanding things. And yeah, so then that's how I arrived at Notebook. So what was the early genesis of Notebook? How did it come together? What were the first steps? And yeah, let's just start there, I guess.

3:25Yeah. So when I had joined Google Labs, there was this new group that was all about language. And it was like in its total infancy, like there wasn't a real team per se, or at least not projects. So I was hired and I was like this free agent. And I was paired with this woman, Dale Markowitz. And we had been given kind of the vague task of like use an AI to talk to a book. It was called Talk to Small Corpus. I think Steven Johnson has mentioned that before. And this was like summer 2022. So it was like not, people didn't really like take that for granted yet. And so Dale and I were just cranking on a bunch of prototypes.

4:09We had like stuff running in Colab notebooks and on our laptops. And we were just, you know, seeing, it turned into like what is, you know, RAG now, what everyone calls. But we were just trying lots of different things and seeing like how can you find the right context in a book and talk to that context. And then I guess concurrent with that, there was this little reading group, I guess, or like a think tank-y kind of thing in labs that was meeting every week to discuss tools for thought. And that's where I met Stephen Johnson and Reza Martin. And so Stephen Johnson is an accomplished writer and he's sort of like the ultimate power user.

4:49He's like always thinking about he has been thinking about Tools for Thought for decades. And, you know, it seems like we had these models and internally that could like help do those things. And so just in talking to to that group, we realized like these prototypes that we've been building that can let you talk to to books. But but any document are kind of like the foundational software that that could be turned into this like Tools for Thought product that that Stephen and Ryza were thinking about. And so, yeah, it's just we found each other and we started building ever more serious prototypes.

5:25And as we presented them around Google Labs and then Google more broadly, people started to kind of take note of them. And then we had to, you know, staff up a team. And then we had like this sprint to get to Google I.O. This is when it was still Project Tailwind. And so we had like that green and blue UI. I don't know if you can picture it. But yeah, and then it really just progressed as like a kind of like typical product. But I will say that like the people who have been involved on the project have been, a lot of them have just been artists. Like a lot of people just write and think about writing a lot and do this like in their spare time.

6:05So that's helped a lot. I think it's helped to steer us away from like being too kind of like pure engineer-y or like, you know, too Silicon Valley centric. And the other thing is that we're just like very scrappy. Like the team just really, really loves like launching stuff. And so it means that like we're really motivated to kind of like build things right, but build things fast and try like weird things and get stuff in front of people. That's awesome. So I remember actually I was covering, I mean, I've been covering, you know, everyday news for a couple of years now in AI. And I remember when it was first announced, but it didn't have the audio overviews feature when it first came out.

6:44where so how early did you guys start thinking about audio overviews was that like on the roadmap from early on and it just wasn't ready yet or did that come you know in a random flash of inspiration like how did that piece come about yeah it's uh it's definitely more like random flash of inspiration um it definitely was not this like long-running thing that we were like cooking um it was like also like a very self-organized thing like the people who built it actually weren't even on the core notebook lm team like they had this prototype in labs um and they they were just they you know they had access to voice models and they thought you know we could probably make pretty good podcasts with this thing um and so they yeah they built these prototypes uh and they started sending them around to people and like when we got them we were just blown away and thought they were like really incredible um and then it seemed like it was a natural pairing with notebook because you know, you naturally needed some content to make the podcasts about.

7:40And, you know, it was like, well, we have this product where people are curating content. And, you know, I bet they want to listen to an AI podcast about it. So, yeah. So it was really like totally like just self-organized. Someone came to us and said, like, hey, check this out. Yeah, it was interesting. A couple of the folks on our team at Super Intelligent had been using it to organize kind of lessons in learning about AI in advance of audio overviews. And then when the overviews hit and everyone started talking about it on, you know, Twitter slash X and whatever, the rest of the team kind of came to it.

8:11But, you know, do you remember when you guys knew you had sort of a hit on your hands? So this is like, you know, kind of popping off in a way that was maybe unanticipated. Yeah, I think there was like two moments for me. So one of them for me was just when we realized that it was a product at all. like you know when you have these prototypes like on average like prototypes don't really go anywhere you know you you'll show them to people but you might have to like kind of beg people to use it so you can get some like usage statistics and just the fact that we had like a little these demos and like other people within Google were like asking us to use it and they kind of like understood like 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 like, yeah, this is a product.

9:01And like, you know, using an AI to interact with your documents is like something that people would want to do and that you don't need to like, really explain to them that much. And then of course, like the audio overviews was like when it went like totally ballistic. Like, I think like, the shift that happened for me was like, I had a lot of conviction that notebook was useful, like I use it all the time. But I think like seeing audio overviews, and how much people took to it really made me understand that we were building a product that was like really kind of the vanguard of like AI products right now.

9:35And that, you know, we were doing something new. We had sort of the space to like be a little bit weird. And, and yeah, just seeing like, I don't think we expected how creative people were going to get with it, like people, you know, putting in their resumes and putting in like the chicken research paper and making them self-aware, or at least making them act self-aware. All that was just super exciting. And I think that that made us realize, okay, people actually want to experiment with this stuff. People don't just want this pure, make my normal work routine go a little faster. It's like they want to be part of the experimentation of AI.

10:16So yeah, that really excited me. And also just seeing all the, you know, the big names like mention it, like I think like Jensen Wong mentioned it and like Robert Downey Jr. mentioned it. It's like totally bizarre. Like every time we got one of these, I'm like, OK, this is mind boggling. Yeah. I mean, I think that it certainly hit the zeitgeist to like podcasts had quite a moment last year. You know, regardless of what you think about the election results, obviously they played a role in the elections in a way that like wasn't there. And I think even like some of the choices, this always happens with products, but like maybe that team was hyper aware and they knew they wanted the two host convention and they knew, you know, they wanted some specific kind of artifacts of that.

10:54But either that or they just completely just nailed what, you know, what people expected out of that. Because I think even it's not just the audio overview. Because a ton of companies have now come out and done sort of similar, like create a podcast with one click type things. it's the the actual interaction is so it's like a it's like now a full modality of communication that that people have that you know you like even the things that are annoying about it like are familiar about it and you like and respond to them um which is just fascinating to see so i going back i mean almost to the sort of the the the way that you guys were thinking about this as a product it feels to me and correct me if i'm if i'm sort of misstating this that it was It's like you guys had a pretty interesting combination of solving a problem, yes, in the sense of people are probably going to want to talk to documents, but also it being a broad enough or flexible enough problem that you could mostly sort of play around in new opportunities and not be overly prescriptive.

11:53This wasn't like enterprises are desperately begging us for this and here's a set of specs and things that they want to solve. It's like, sure, there's definitely going to be usage for this. I mean, I think validating of that point, I still half of the enterprises that we talked to, one of their core AI applications that they built internally is some custom built, you know, chat with your documents type application. But is that is that fair to say that there was sort of like, yes, you're solving a problem, but but almost even more, you're kind of thinking about just new possibilities and what was possible now with with these technologies?

12:23Yes, certainly, certainly. Like, I, you know, I think like a core thing that Google Labs is supposed to do is experiment. and like experimentation means you don't you don't know the outcome to begin with right like you you want to see what people will do and um i think like you know an old manager i had would really like hammer on this point like you'd be in a meeting and people would say oh you know users want like xyz and like you would never get us let us get away with that like you'd say like you you don't know that like you you have like no reason to believe that except your own conviction and when you put it in front of people they like almost always do something completely different and so that was like totally the case you know we would you know think that the product like did something really obvious and then we'd bring it to like a ux research session and people just do like you know bizarre things like um and so like some of them are kind of you know like no-brainers like it's like if you need to you know do a little report or something like of course you're going to do that but i i think we always wanted it to be um flexible enough to let people like find the cool interesting things um you know the the worst thing would be if you made it so locked down that you like didn't let people do those things and then they couldn't even communicate to you that they wanted to um and and then you know there's no opportunity to do like the the interesting stuff so uh certainly we we always wanted it to be like at least a little bit open-ended just because as soon as you're dealing with the space of like people upload things, it's like, it's easy to assume, okay, well, people are going to use this for business.

13:57But you know, loads of people are just uploading their resumes. And like, nobody thought like, oh, this is clearly like a resume tool, right? Like, we never designed it as a resume tool. But yeah, it works on resumes, or works on like a draft for a short story. It works on your, whatever your credit card statement, like. And so yeah, we always, I think we have a very healthy spirit of experimentation. People around our desks have sticky notes that say, let people do weird stuff. So I think that that's kind of a core tenet for us. Yeah. I mean, I think that that's one of the reasons that it's so hard if you're ever in startups to think about you wanting to go off and do a consumer startup is you basically have to hope you get something in the ballpark of interesting or useful enough that people then tell you what it was actually supposed to be in the first place, right?

14:49Like, like, I can't actually think, I mean, maybe Facebook, like original Facebook, but it's hard to think of a consumer product that got big, that did exactly what people imagined it was going to do. I mean, even Facebook, you could argue, like, it was, you know, just a very different kind of more limited focus of, you know, rating girls or whatever that he wanted to do. And so it was different. But, but that's, that's a very hard magic to capture. And it's extra hard from within a big company, obviously. It's why you don't necessarily usually see big breakout consumer products from inside big companies.

15:20Yeah, that's probably changing a bit. I think that people are realizing that AI moves so fast and that it's actually really in our interest to launch a lot of things quickly and that the appetite for just trying these things is huge. So yeah, I'm grateful that we have the space to do that. And I also think it's getting a little bit better. What are some of the weird or interesting or off-kilter or unexpected use cases that you've seen that have gotten you or your team most excited or interested? Yeah. One that I, it's like sort of like a pure utility, but I wouldn't have immediately thought to do this.

15:58And somebody on our team just actually did do this. They were like buying a house and they had this like super long disclosure and they just put it in and said like, this is the asking price and here's the disclosure for the house. like, tell me everything wrong with the house and give me an estimate of how much it would cost and give me like a, you know, a counter price that I should ask for. And it worked. And like, I think they did get a discount. And so like, that was like, just really cool. I don't think that's like that weird. But it was like, just pleasing to see, like, it actually, you know, help somebody in that way.

16:35You know, people have used like, the customization to do interesting stuff. Like, Like, they'll use, like, the customization for the notebook for, like, the responses to make it answer. I'll do this as well. Like, I'll put in my favorite books and I'll have it answer as, like, the characters in the book, which is fun because then you're not doing this kind of, like, notebook is a de facto narrator or it has its own voice. Like, now it's, like, it's informed by the book itself and it's, like, a medium is the message kind of thing. um and then also just all the ways people have customized the audio overviews uh like we've seen everything from people saying like be less personable like just give me like the hard facts um or we've seen people be like you should swear more and and be like really crass and like okay sure like you know um that's really delightful like whenever we see those pop up on reddit like we're all sending them around and we're just like like to see people persuade when it comes to the enterprise.

17:33So Notebook LM is now in workspace. It happened. I mean, it's been available for business. Obviously, you could just sign up and kind of use it that way. But now it's more explicit. Have you seen any sort of shifting patterns and how people are using it inside an enterprise use case? Or is it sort of just following some of the same patterns that you're seeing with consumers? I think it's still really early for that. And we want to try to pursue what people are doing and figure out all the different ways that people want to use it. I would say like, probably one difference is the scale of usage.

18:03Like, I think individuals, they won't have like just the scale of documents, both in like, you know, the detail of the documents and the number of documents. So maybe that's one kind of like obvious difference. But I think a lot of the like, the usage across students or just like enthusiast consumers or business like, are all actually like, kind of in the same realm anyways, like people want to produce certain things, they need help just understanding stuff. Things are really dense. Yeah. So I don't think there's been like one real like, oh, okay, this is like the obvious thing. Today's episode is brought to you by Vanta.

18:41Trust isn't just earned, it's demanded. Whether you're a startup founder navigating your first audit or a seasoned security professional scaling your GRC program, proving your commitment to security has never been more critical or more complex. That's where Vanta comes in. Businesses use Vanta to establish trust by automating compliance needs across over 35 frameworks like SOC 2 and ISO 27001. Centralized security workflows complete questionnaires up to 5x faster and proactively manage vendor risk. Vanta can help you start or scale up your security program by connecting you with auditors and experts to conduct your audit and set up your security program quickly.

19:19Plus, with automation and AI throughout the platform, Vanta gives you time back so you can focus on building your company. Join over 9 ,000 global companies like Atlassian, Quora, and Factory who use Vanta to manage risk and prove security in real time. For a limited time, this audience gets$1 ,000 off Vanta at vanta.com slash NLW. That's V-A-N-T-A dot com slash NLW for$1 ,000 off. If there is one thing that's clear about AI in 2025, it's that the agents are coming. Vertical agents by industry, horizontal agent platforms, agents per function. If you are running a large enterprise, you will be experimenting with agents next year.

20:02And given how new this is, all of us are going to be back in pilot mode. That's why Superintelligent is offering a new product for the beginning of this year. It's an agent readiness and opportunity audit. Over the course of a couple quick weeks, we dig in with your team to understand what type of agents make sense for you to test, what type of infrastructure support you need to be ready, and to ultimately come away with a set of actionable recommendations that get you prepared to figure out how agents can transform your business. If you are interested in the agent readiness and opportunity audit, reach out directly to me, nlw at bsuper.ai, put the word agent in the subject line so I know what you're talking about, and let's have you be a leader in the most dynamic part of the AI market.

20:42Hello, AI Daily Brief listeners. Taking a quick break to share some very interesting findings from KPMG's latest AI quarterly pulse survey. Did you know that 67 % of business leaders expect AI to fundamentally transform their businesses within the next two years? And yet it's not all smooth sailing. The biggest challenges that they face include things like data quality, risk management, and employee adoption. KPMG is at the forefront of helping organizations navigate these hurdles. They're not just talking about AI, they're leading the charge with practical solutions and real world applications.

21:13For instance, over half of the organizations surveyed are exploring AI agents to handle tasks like administrative duties and call center operations. So if you're looking to stay ahead in the AI game, keep an eye on KPMG. They're not just a part of the conversation, they're helping shape it. Learn more about how KPMG is driving AI innovation at kpmg.com slash US. Yeah, I think one of the things that I think about a lot, you know, as we're dealing with enterprise or we're talking with them about how to use different tools is because of the sort of the, it produces podcasts. I think some people's first instinct is to think about it as an external facing tool when a lot of the more interesting use cases that we've seen, at least initially, are internal facing, right?

21:55Like knowledge sharing across the company, storytelling around successes and things like that. Obviously, there's just data processing, huge numbers of documents is sort of like the obvious kind of layer one thing. But we're seeing an interesting combination of sort of basically like, you know, data plus storytelling, you know, largely for internal purposes. Yeah, definitely. Like I, so one thing I'll say is that like, I don't think we should think about notebook as like a podcast platform, you know, like, I think like the great thing that podcasts have done is show that like, people understand that you can transform documents to make them more understandable in a different format.

22:38And podcasts are one flavor of that. But, you know, we were already talking about, like, what other kind of output types can we support, right? And that's going to be very different if you're, you know, like a law firm compared to if you're, you know, somebody making a meme on Reddit. And so, like, we want to support across that whole spectrum. And, yeah, it's like I think the more kind of like faces of your documents that you have, I think the easier you're there to understand them. And like being able to massage them in multiple different ways, I think actually will help you just understand it more.

23:17You know, it's like learning multiple languages is just good for understanding broadly. And I think that that's kind of what these are is like podcasts are one, but we'll have more things soon. And those will, I think, help in more use cases. Yeah, it was interesting. So when the audio overviews first started to pop off, obviously, as a podcaster, like 75 % of the articles, I feel like, were like turn anything into a podcast or some variation on that flavor for a title. And so I got the question a ton of times, like, you know, is this coming for podcasting? Is it going to be all generated, you know, audio in the future?

23:49I was like, well, hold that aside. You know, that's always a separate question. But in terms of this tool specifically, it seemed pretty clear to me that, again, like the first thing that people would try is kind of just replacing one-to-one the stuff that they were already doing. But there was so clearly different capabilities that had not been capabilities before, right? So for me, I'm speaking largely to a non-developer audience, not non-technical per se, but certainly not developers, at least as the core business-type audience translating AI. And a lot of times, from the standpoint of trying to actually let people know what's going on, it is important to go dense, to understand what's being said in some research paper.

24:33Because that actually, in this sector, that impacts significantly what's going on. right? Like, we're recording on the day that everyone's freaking out about DeepSeq and R1. And actually, like, understand, you know, floating points is useful for understanding what's going on and whether it's actually plausible that they did this or whether it's just a CCP PSYOP or whatever, right? And for me, like Notebook LM, one of the most default use cases is translate, you know, dense, thick, kind of, you know, technical, you know, resources into something first that I can understand them potentially that my audience can understand.

25:08And I feel like there's a million use cases like that where it's like, I couldn't have done that in any way before. So it's fundamentally new. Yeah, totally. It's like, I, you kind of want this, like, you know, funnel of like understanding where you don't want like a exactly just a one line sentence explaining it. But you also don't want the whole thing. It's like you want something in between. And, and, you know, I think we find that like the best way to get a handle on these things is to to like engage with it. And so I think like the nice thing is like you can do this back and forth with with questions and like and get this handle.

25:47And I also think that like the podcasts are a nice thing as like sort of a fun little, you know, they like entice you to use the product like a hook. Yeah, exactly. And it's like I found that hook even working on me. It's like, you know, I work on the product. It's like I don't need to be told to use it, but I've found that like I will just whatever, like I will hear about like Ozempic or something. And it's like this was an interesting one where it's like I've heard about it in the news. I know that it's this like new drug. I've also seen news that says it's like has these like real like health benefits.

26:20And I'm like, OK, like how real is all that? And where previously I might have just sort of wondered idly or like read, you know, from a news source I trusted or something. now I'm like okay no I'm gonna go get the research paper and I'm probably not gonna read it in full but like I will interact with it and I'll like ask some questions and I'll click the citation and I'll read that part in full and it's like I I find it's good because it actually lowers the barrier of entry for me interacting with it period it's like now I have this thing in my head where when somebody sends me something dense I think oh I can like understand this a little bit where previously I would have been understood it like zero percent yeah no I think I think that that is a profound use case.