Transcription
Think about how often you encounter the following scenario: There's a task you're trying to complete, but in order to do that simple task, you have all this background; you have to have all these documents you have to read, slide presentations to go over, people you have to talk to, concepts you have to understand. The problem is, who in a crazy busy workday has time to sit down and spend hours reading things?
Hey everyone, about a year ago I released a YouTube video on a new experimental AI platform from Google called Notebook LM. But you know, I thought this is probably not going to go anywhere; it's just a beta, just an experiment. But let me tell you, I was completely wrong. Let me show you the five biggest new features that Notebook LM has come up with. And no, this video is not sponsored.
Now, before I show my screen, I want to answer what I think is a really fundamental question, which is: What is Notebook LM? Especially, how is it unique and different from some of the other AI tools out there, like ChatGPT? Well, I was reading one of the blog posts that the Notebook LM team had written, and in that post they called it a tool for understanding, and I think that's the best simple description I've seen. It's the ultimate tool for understanding things. There's a lot to cover, so buckle up, grab a cup of coffee, and let's dive right in.
First, let's talk about audio overviews. Imagine if you were researching something, or learning something, or reading something, and you just happened to come across a podcast in your podcast app, between two super engaging, knowledgeable hosts explaining that thing to you. Well, now you don't have to be lucky; you can generate exactly that conversation on demand, anytime you want.
What you see here on my screen is the main interface when you log in at notebooklm.google.com. It's very simple. Each one of these squares contains all the material for a specific project. So the first thing you're usually going to do is click "Create new" and add some sources. These are the sources that you want it to know about, that you want it to cite and draw from.
So the project I have in mind is an upcoming trip to Guatemala that my family and I are planning to take this summer. This is exactly the kind of project that I would love to spend lots of time on: researching the country, their culture, their history, all the itineraries, the destinations we could visit, the activities, the cultural experiences. But if I'm really honest with myself, it's also the type of project that is likely not to be prioritized. I have a lot going on over the next few months; I'm writing a book, I have two young kids. So it's the kind of thing that I know I really could use AI's help with.
So the first thing I'm going to do is look for some sources that I want to draw on, and this can be as simple as opening up Google and typing "best places to visit in Guatemala". And I'm going to scroll through here; these look like great sources, these three right here. So I'm just going to hold down command and click each one of them, which opens them up in tabs along the top up here. Now I'm not going to read them, not going to dive into them; all I want them to do is stay open right there. Then let's try YouTube, and here I'm going to be a little more specific; I'm going to say "underrated places to visit in Guatemala". So let's open this one; again, I'm going to hold down command and click. I'll do that one, and then I even like that this one is specifically about an underrated location, so let's add that one. All right, so this is a good start. I have three websites and three YouTube videos, which might not seem like a lot, but it kind of is; it would take me probably an hour or two hours to really dive deep into those.
So what I'm going to do, instead of reading them myself, is just add them as sources in Notebook LM. Let's open up the first website here; I'm going to do command L to highlight the URL here, and then command C to copy it, and then back in Notebook LM, you can see right down here I can add Google Drive documents, I can add links, and I can add pasted text. So I'm going to go ahead and click "website", and then command V to paste, and then insert. Now that I can click on the source right here, and this little purple panel will open up called "The Source guide". The Source guide is just a one-paragraph summary of the source. Let's go to our second website; command L to highlight the URL, command C, back to Notebook LM; I'm going to click this "add source" button right here at the top left, website, paste it, add that as well. All right, so now I have two sources. I'm just going to take a minute here and do the same with this last website, and then it looks much the same with YouTube videos; you just click the video, command L, copy the URL of the YouTube video, add source, this time click YouTube, paste it right there.
I just remembered a friend who was Guatemalan also sent me a country guide, a quite long, extensive PDF, kind of highlighting and showcasing some of the best places to visit. So I'm going to go ahead and add that as well. To do that, I'll hit "add source", and this time click "choose file" to add a file from my computer, and I can see it's in my downloads folder right there. There we go; you can see it right there. And there's one more thing I want to add, which is in my notes, which is this suggested itinerary that the friend, who is the one getting married who we're going to visit, actually suggested for us. So they suggested a 6-day, 7-night itinerary; definitely want to take that into account. So I'm going to go ahead and do command A to select all that text, command C to copy, head back over to Notebook LM, and this time I'm going to hit "add source", and right over here where it says "copied text", what that allows me to do is to simply paste whatever text I want right into the little window here, hit insert, and that is now a source as well.
This is a good start; you can see here I have eight different sources I'm working with, and before I click this big "generate" button here, I'm going to customize the conversation a little bit. So I'm going to say, "Just give it some context and details," and then I'll click "generate". This is what you're going to see; it says "generating conversation"; there's a little circle here, and this will take somewhere between around 3 to 6 minutes. So I usually go get a cup of coffee, or go to the bathroom, or just take a little stretch break while this is generating.
All right, so we're fast-forwarding here a few minutes, 3 or 4 minutes, but you can see here it has now generated a 21-minute, 12-second podcast conversation that I would love to share with you. Now here it is. Okay, so get this: Our listener is heading to Guatemala, summer of 2025, nice, a week they say, and they want the authentic experience. "Yeah, did you know that some of the world's most sought-after coffee beans come from the volcanic slopes surrounding Antigua?" "I did; I actually saw something about that. Apparently Antigua has this like booming coffee scene, not just plantations, but like artisan roasters and cafes." "Yeah, I think one of the articles recommended this place called Artista de Cafe." "Oh, nice, saying it was like this hub for local artists and coffee connoisseurs."
I'll go ahead and pause that now, but you can see that this conversation is not just perfectly realistic and indistinguishable from a human conversation, but it's completely customized based on the instructions that I gave them. But you know what? You can even customize this further, and that is through a beta feature called "interactive mode", where, believe it or not, you can actually click a button and join the conversation in real time yourself. So I've just remembered a key detail that I forgot to mention, so let's go ahead and enter interactive mode and chime in on this conversation. So I'll go ahead and start playing the podcast again. "Okay, so get this: Our listener is heading to Guatemala, summer of 2025, nice, a week they say..." So watch what happens when I click this "join" button and interrupt the conversation. "Oh, hey there, how can we help?" "Hey guys, I appreciate this, but I forgot one key detail, which is that we're going to have two small kids with us, ages 2 and 4. Could you please give us a recommended itinerary taking that into account?" "Oh wow, okay, that's definitely a key detail. Yeah, traveling with a two and four-year-old changes things. Okay, so let's unpack this and tailor our itinerary a bit. Yeah, we'll need to balance adventure with kid-friendly options, right? We were about to get into how to experience Guatemala like a local, and honestly, that's totally doable with kids; it's going to mean adjusting the pace a little." "Exactly, like maybe fewer super long travel days and more focus on activities that engage young ones."
All right, so I don't know if you caught that, but this is the most remarkable demonstration of AI that I think I've ever seen. I think this is so powerful for a few reasons: One is, there's a lot of people out there who don't learn best by reading; reading is actually really taxing and boring; they need to listen to it. But there's another reason, which is: think about the typical chat interactions that people have with AI in a chat format; the onus, the burden is always on you to keep the conversation moving forward; you have to think of a question, you have to think of a prompt, you have to keep putting an effort to make progress. This is completely different; this is two hosts who are very knowledgeable about the topic and the sources having their own conversation completely independently of you. It's more passive in many ways, but that's good; very often, especially in your first encounter with something, you want this kind of superficial, surface-level survey of the landscape, that way you even know what topics are relevant, what questions to ask. I've been using audio overviews exactly that way; anytime I want just a kind of general summary or survey of a topic, I go straight here and generate a personalized podcast conversation on that topic.
I'm going to go ahead and close this window here on the side, and one last thing I want to show you is that you're not limited to the audio overview. If the hosts talk about something you find interesting and want to know more about, well, you have all the sources already loaded up here; you can go right here at the bottom to the little chat window and ask anything you want. For example, I'm going to say, "Why is Antigua so underrated? I'm kind of afraid; maybe it's dangerous; maybe it's not super safe. So I, I want to know why do tourists not tend to go there?" And there you go; you can see a much more in-depth, bullet-point-by-bullet-point summary of what makes it so underrated, and each one of those statements is cited, and if you click on one of those little citations, it takes you exactly to that part of the source that it's referring to. And then last thing, you can also create structured documents. So let's say that you are studying Guatemala and you actually really decide you want to go deeper; if you click this button right here, the "study guide," it will generate an entire document with a quiz, answer key, essay questions, glossary of key terms, looks like a lot of the specific destinations it's mentioning. So you're not limited to a surface-level conversation; you can go as deep and as detailed as you want.
Now there's one more example of what you can do with with audio overviews that I want to show you. So let's head back out to the main interface; I'm going to create a new notebook, and the example I showed you with my Guatemala trip was an example that uses external sources, sources that you can find on the internet, but you can also use your own internal thoughts, feelings, ideas, journal entries, and that's what I want to show you next. So I'm going to go ahead and click on Google Docs, and then right down here I have my end-of-year journaling. At the end of every year I do something called an annual review or a year-end review where I think back to the past year and just reflect and journal on different milestones that happened, precious memories, goals that I achieved, lessons I learned, and all that journaling is contained in this document right here, which I'm going to insert as a source called "Thiago's 2024 year-end journaling". You can see that there's a summary here of what this document is about, and I'm just going to add that single source and go ahead and generate an audio overview. I'll fast forward a few minutes here, and now we can listen. "Hey everyone, and welcome to uh another deep dive with us." "Yes, welcome. Today we're going to be uh looking at Thiago Forte's 2024 year in review, which is something that uh he recommends people do and uh he practices what he preaches, so yes, does, and it's also kind of interesting that he actually expresses some anxiety about doing the review." "Oh really? Even though this is kind of his area of expertise?" "I know; I was going to say this is kind of his thing, yeah, is to like reflect and review, and he even has anxiety about it." "He does; he's worried about coming across as performative or, you know, not digging deep enough."
All right, I'll spare you the rest of that. What I found so fascinating about this is several things: First of all, it gives me such an interesting kind of outsider perspective on my life. You know, I can go back and read my own journal entries, but hearing what feels like two other humans talking about it in depth, they're fascinated by the insights I've come up with, they're impressed at some of the the goals that I've accomplished, they're empathetic, and they sympathize with some of the hard, painful lessons that I had. It really helps me just get out of my own perspective, uh see my life from this more objective point of view, which honestly gives me so much more appreciation, so much more gratitude, helps me kind of be more objective about even the life and the year that I had, uh which is just kind of a really powerful tool for personal growth. It's something that I'm definitely going to make part of my my journaling practices going forward.
Next up, we have feature number two, which is an expanded context window. I know this might seem like a boring, backend technical feature; you're like, "Why do I care about this? I don't even know what a context window is." But this is, in some ways, actually the most important; this single, seemingly backend technical feature completely radically changes what is possible with AI. Imagine if you had a guidance counselor who could recommend the perfect next step in your learning based not just on guessing or a few small details about you, but on the entire history of everything you've ever read and consumed. Notebook LM uses Google's Gemini 2.0 Flash AI model, which can now take in—get this—up to 50 sources, each one of which can have up to 500,000 words. Just to do the math for you, that's up to 25 million words of context. That is so much higher than any other AI tool out there, and basically means you're never going to hit the limit; almost any conceivable use case can fit within 25 million words.
But let me take a step back and explain what is a context window, and why does it matter? Think of AI as having two kinds of memory, just like humans: There's long-term memory and there's short-term memory. An AI's long-term memory is basically what's called its training data, all the data that went into its creation, and basically these days all the major AI models have the same, same training data more or less, which is the entire internet; they're all trained on the internet, which means paradoxically that there's no real advantage in its long-term memory, in its training data, because it's the same across all the AI models. A short-term memory is the data that it's working with right now, like what is the conversation it's having right now, the problem it's trying to solve right now, the person you it's talking with right now. Most of the advantage of working with AI now comes from that short-term memory: the specifics of the problem you're trying to solve, the history of the product or your company, your specific skills and resources you have access to, things you've tried and whether they worked or not; that kind of data, which we call your context, is what you supply to the AI for it to know something about you. The context window, this term I keep using, is simply the limits of how much context you can give the AI. Until recently, it was maybe a few thousand words, or 10,000 words, or maybe 30,000 words. Now, with Gemini 2.0 Flash, that has radically expanded to 25 million words.
Let me give you some examples: Imagine you work in product development, and you could load up an entire folder full of hundreds or even thousands of in-depth customer interviews, and then ask Notebook LM to evaluate and analyze an idea you have for a new product based on all those past interactions. Imagine if you work in a nonprofit, and you can upload an archive of every, every single grant you've ever won, plus a directory of hundreds of currently open grants, and then ask it to tell you which grant proposals you're most likely to win, and then to actually write the proposal drawing on the mission statement of your organization. Imagine if you're having a mysterious medical problem, and you can load up your entire medical history and ask it to identify any recurring patterns, even across different specialties. Imagine if you could upload the entirety of your personal notes from years of reading and highlighting and note-taking, and every time you encounter a new idea, such as in a book, you could ask the AI if there are any existing connections to that idea in your notes. You know what? Let's make this concrete; let's actually take that last example of loading up all your notes and do it in real time.
So I'm going to go ahead and create a new notebook, and I actually use a service called Readwise, which automatically syncs all of the highlights that I make from any ebook that I read, read or uh any online article or other online content that I save in a centralized place. Readwise actually has a partnership with Notebook LM that allows you to export the entirety of all those notes and highlights to Google Docs in a format that's perfectly designed to be loaded up into Notebook LM. You can check out my first video for more details on how that works. What this allows me to do is to open up my Google Drive here, and I'll find five Google Docs worth of all those highlights. So what I'm going to to do is select each of these by holding down command as I click them; I'll click "insert", and there we go. Took about 2 minutes, but you can see in this one little paragraph right here that it's essentially summarized a broad swath of my learning, my research, my interests across over a decade of my life. And this is where it really gets fun at this stage, because I can ask at some really deep, penetrating questions about the history of my my personal growth, my learning, my educ-a- in my career. For example, I can say, "Please summarize a list of my most common recurring interests based on these notes." There we go; we can see note-taking, knowledge management, types of notes, the value of externalizing thoughts, habits and routines, the importance of making connections; all huge areas of interest. But then let me go beyond what is found in this message here and say, "What are some other ideas, topics, and areas of interest that are likely to resonate with me based on this list?" So it's essentially like a teacher suggesting other things that I might want to learn about. Now let's go beyond that and say, "Please create a study guide and reading list to help me dive deeper into these topics." It's broken up into chapters, and then each chapter has key questions, activities, and readings, including, including my own book. I love this; using AI as a tool. Here's a bigger reading list and additional recommendations; each one of these recommendations is drawing on so much existing knowledge about me; they're not random or one-size-fits-all; this is a perfectly personalized learning curriculum that has been generated based on years of my note-taking. Do you get why an expanded context window is so critical? The possibility that this is opening up is a completely different way of interacting with AI; it's not about designing these crazy extensive prompts; it's not about meticulously organizing your sources; it's really just about pointing the AI at a massive repository of information and letting it do the work.
Let's talk about new feature number three, which is multimodal sources. Multimodal is just a fancy word that means many different kinds of formats. Notebook LM can now take basically any kind of format that you want to throw at it: There's URLs from web pages and YouTube videos, like we saw; you can put in PDFs, you can put in text, Google Docs, audio files; you could take a little voice memo note while walking around your neighborhood and then upload that. But the most exciting and newest one is images, specifically images within Google Slides. Think about how often you encounter information in modern work, especially in the form of slides, presentations that someone is giving to you or you're giving to someone else, sales calls, technical specifications; slides are a very common format these days, and you can now work with those directly in Notebook LM. Let me show you what that looks like. As always, I'll start by creating a new notebook, and then I'm going to click right here where it says "Google Slides", go over to "recent", and let's use as an example a recent presentation that was sent to us by our insurance agent proposing a couple new health benefits packages for our employees. So this is the kind of information that is important, has high stakes; it matters that I pay attention to it, but it's also just so incredibly mind-numbingly boring that I want to spend as little time loading up this information into my brain as possible. So I've loaded it up as a source, as you can see here; you can see there's a little summary, which is that uh they're proposing two new health benefits packages. Okay, so my first question is, "Please compare and contrast the two proposed packages." So you can see here there's a summary uh comparing their deductibles, their office visit costs, monthly and annual cost. But you know what? This is a little bit too extensive, so I'm going to say, "Please summarize more succinctly what is the core difference between these plans." And there we go; the core difference is in the cost of office visits and the overall monthly cost, with A1, the first option, being more expensive. So as the founder and CEO of this business, that is really the bottom line for me; I don't really need to know all the fine-grain details; that is for our insurance agent, that's for the operations person on my team; the decision of which plan to go with falls with me, and what matters at the end of the day is the cost. So I'm actually going to ask it to perform an analysis of what the cost difference would be between these two plans for six employees over the the course of a year. And there we go; its analysis is that the first, A1 plan, would cost 15,000, almost $166,000 more per year. So this is the part that is actually most important to me; I'm actually going to highlight this part and say "save to note", which is what I do anytime the AI gives me something that is actually valuable long-term that I want to keep track of; I just click that button and it saves it over here in this little uh right-hand bottom corner. But the purpose of this demonstration was to show you how it works with images and slides, so what I'm going to do is let's say I would like to know the source of this; I want to check that it's accurate myself; I can go ahead and click this little one here, which is the citation, and it takes me directly to the table where it shows that price difference. So notice when we look at the slides here how complex this information is; it's multicolored, it's in columns, rows; it is not an easy thing to absorb. But diagrams, charts, IM, images, photographs, all these kinds of things, if they're embedded in slides, can now be understood by Notebook LM, and that's the multimodal capability.
Now let's talk about feature number four, which is a
Redesigned interface: If we head back to our Guatemala notebook as an example, the whole interface for a given notebook is divided into these three columns, which they call panels. On the left is the sources panel; this is simply what sources the AI is working with and drawing from, and you can expand that. Or, once you've sort of settled on the sources and you're not going to do a lot of changes here, you can just collapse that so you have more space for the other ones.
The second panel, which is this middle part, is the chat panel, which is where you actually speak with and interact with the AI. It's nice to make this large because sometimes it gives quite extensive answers that you want to be able to see in this wider format. For example, if I click this question, "What cultural and historical significance shapes Guatemala's top destinations?" you can see it's probably around 500 words. So if I like this answer and I decide I want to keep it for later, I can just scroll down, click the "save to note" button, and it saves just this one answer, which I've decided is valuable.
Over here on the bottom right, this is called the studio panel; this one on the right, and it does a number of things. This is where you create and interact with audio overviews; it's where you have these pre-made options, such as creating a study guide or a briefing document or a frequently asked questions or timeline. For example, if I click the FAQ button, it's now generated in a question-and-answer format, a summary of the different sources that we're working with. So I can go ahead and close that. But I'd say the most important part is these notes. When you have a conversation with an AI, unless you take an action, it kind of all just disappears over time. So you want to be sure, if you get a good answer, to just click that little "save to note" button. And over time, you'll have here at the bottom right—almost you can think of it like the final output—the gems, the evergreen insights that you want to take away from this conversation are all going to be saved here in the notes section.
So that's the overall interface, but for me, by far the most important new addition over the last year has been these little numbers. See those? There's kind of a big problem with most modern AI tools, which is they hallucinate; they make things up, which is kind of a paradox. Because the more you depend on AI to execute important tasks and even make decisions, the less you can trust it because you never know if it's hallucinating. One thing I really appreciate about the Notebook LM team's philosophy is they had, from the very earliest prototypes, this idea of source grounding. What source grounding means is that, especially for information that is in the context window—that short-term memory that we talked about—it's basically not going to hallucinate, or very, very rarely. You can check that it's not hallucinating; you can essentially double-check the AI's answers by simply clicking one of these little numbers, which are the source citations. I'm going to go ahead and click on the 16, and it not only opens up the original source, which is this YouTube video, it takes me to and highlights exactly the point that that came from. This is a huge deal, you guys. You have to be able to trust what the AI is telling you, and this new interface with inline citations is a huge step forward for that.
And finally, feature number five: Let's talk about it for the very first time. There is now a paid, upgraded version of Notebook LM. The basic version, which is most of what you've seen here, is free, and I would expect it to remain free; most of Google's core tools are free. But if you work on a team or a company or organization that wants to use Notebook LM, it may be worth checking out. There's actually three separate ways to get Notebook LM Plus: either through Google Workspace, which you may have as part of your company, through Google Cloud, or through Google One, which can be purchased individually. The pricing depends on exactly how you purchase it, but it seems to be something around $20 per month. And there's four main things that you get as part of Notebook LM Plus. The first one is an even bigger context window. So instead of 25 million words, you can have—and this is just kind of mind-boggling—up to 15 million words. The only use case I can think of for that is if you're at a company and you want to put in like all of the company's data into a prompt.
The second thing you get is five times higher usage limits, so that means you can create five times more audio overviews and have five times more notebooks—again, mostly useful if you're working with a lot of other people on a team or in a company. The third Plus feature is what are called chat modes. If you click on this button right here, you can give Notebook LM certain almost like personalities that it takes on in your interactions. So let's say in this conversation you want it to take on the role of an analyst; it will become more critical and more analytical. Or it could be more of a guide, almost like a teacher. Or you can define your own personality. So let's say you wanted it to act as a marketing expert or a life coach or a research assistant; you can actually define that and describe it right here in this box. And then you can also customize it to make its responses longer or shorter.
But the fourth, and I think by far the most important Plus feature, is the ability to collaborate within Notebook LM. Think about it: Have you ever had a particularly interesting interaction with an AI? You're getting some real insights and some real new perspectives, and then you think, "I want to share this." There's often not an easy way to do that except copying and pasting things and sending them to people via email or Slack or whatever, but then you lose all the interaction; it's just a static piece of text. With Plus, you could actually get this entire notebook, click this share button right here, customize the permissions, add a welcome note if you'd like, and then choose whether you want to share the full notebook with all the sources that you've added, with all the customizations, with all the notes you saved, or you can make it chat only. You could give the people on your team or the people who work for you, your employees, or or even external collaborators, the ability to talk with and go back and forth with the essentially custom AI model that you've created without you having to copy and paste stuff. Once you've decided on the settings, all you have to do is click this link right here to copy the link and send it to them, and they'll have access via their own free Google Notebook LM account.
Overall, I would say Plus really makes the most sense for people who have adopted Notebook LM pretty seriously and are using it with others as part of companies, like I said. Okay, I've shared a lot with you, so let's take a step back and kind of summarize the big picture here. I think the big shift that's happening right now with Notebook LM, with this much bigger context window, is just like a great teacher; it's gaining a really fantastic short-term memory. Think about how important it is that a great teacher can follow the thread of the topic that it's teaching you; a great teacher can reference previous conversations, what you just said a few minutes ago; they can think into the future and plan out different stages of a curriculum that they're going to be teaching and coaching you through. We're living in a time of just incredibly fast change, in no small part because of AI. But if there's one thing that I can say for certain, it's that we're all going to have to learn much faster and much more often. Having an AI-powered world-class teacher with a fantastic memory sitting next to you and guiding you through this unpredictable landscape is just a world-changing and even life-changing phenomenon that we all now have access to. If anything that I covered resonated, I encourage you to create a free account at notebooklm.goole.com and let me know what you think.