---
id: "workshop:industry-series-learning-to-use-google-s-ai-powered-research-and-writing-tools-in-your-work"
type: "InnovateUS workshop"
title: "Industry Series: Learning to Use Google’s AI-Powered Research and Writing Tools in Your Work"
url: "https://innovate-us.org/industry-series-learning-to-use-google-s-ai-powered-research-and-writing-tools-in-your-work"
date: "2024-03-12"
duration: "60 minutes"
language: "English"
instructors: ["Steven Johnson"]
recording: "https://www.youtube.com/watch?v=FJLL0Af2JjM"
transcript: "automatic, from the recording"
source: "InnovateUS (innovate-us.org)"
licence: "CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org"
---

# Industry Series: Learning to Use Google’s AI-Powered Research and Writing Tools in Your Work

InnovateUS workshop · 12 March 2024 · 60 minutes

Join us for this 60-minute virtual workshop with Steven Johnson, Editorial Director at Google Labs. In this training, Steven will introduce NotebookLM, Google's new AI-powered research and writing tool, and will explain how NotebookLM's source-grounding approach allows users to navigate and explore complex documents or archives of information. Steven will walk through several use cases that demonstrate using NotebookLM to capture information from source documents, and then transform it into different formats, including briefing documents, slide decks, and thematic summaries. Finally, Steven will discuss the pros and cons of "collaborating" with AI research assistants, and give a preview of new upcoming features.

*This workshop is part of InnovateUS’s new “Artificial Intelligence Industry Series," where speakers from leading technology companies will teach participants how to use specific artificial intelligence tools. We do not promote or endorse any particular products, but we want to offer our audience a chance to learn more about using specific upcoming tools for their everyday work.*

## Learning goals

- Learn about Google's new AI-powered research and writing tool and explain how NotebookLM's source-grounding approach allows users to navigate and explore complex documents or archives of information
- Learn use cases that demonstrate using NotebookLM to capture information from source documents

## Speakers

- **Steven Johnson**, Editorial Director, Google Labs

Recording: https://www.youtube.com/watch?v=FJLL0Af2JjM

## Transcript

Automatic transcript of the recording: names and terms can be misheard, and the times can be a few seconds off. Each time links to that moment of the video. The transcript alone: https://innovate-us.org/industry-series-learning-to-use-google-s-ai-powered-research-and-writing-tools-in-your-work.txt (plain text) and https://innovate-us.org/industry-series-learning-to-use-google-s-ai-powered-research-and-writing-tools-in-your-work.json (every segment with its start and end second).

**Steven Johnson**

[0:00](https://youtu.be/FJLL0Af2JjM?t=0) for me to get to talk to you all about NotebookLM and what we're doing. We love the extended government use case for this product and we're just getting started, so your feedback, if you get a chance to explore NotebookLM, would be incredibly helpful to us. We're very receptive to new ideas and pushing in new directions. I'm going to shift to present. I'm going to give you a little bit of a background around how I got to this product and what we're doing, and then I'm going to just shift into demo mode and do a full walkthrough of a couple of use cases that I think will be relevant to the wide range of uses that I can see from the descriptions of the participants in the chat already. Then we'll leave 25 minutes for questions. I'm just going to share my screen here. This is an intro to NotebookLM. It's technically called an experiment from Google Labs, which is also a new division of Google. I'll get into a little more detail about what we mean by experiment a little bit later, but you might have noticed from Hannah's introduction that I have spent most of my career not building software. I spent most of my career writing books and making television shows. I did a couple of startups when I was younger, but mostly I've been an author. I got into this interesting adventure with Google almost two years ago now. For a very long time of my career, I had both written books about technology and about innovation, but I'd also been really interested in the tools you could use to create books like that. The research tools, the writing tools, basically tools that help you have better ideas in general, not just if you're a writer, but for anyone. How can you use

[2:01](https://youtu.be/FJLL0Af2JjM?t=121) software as a way of augmenting your creativity, augmenting your memory, augmenting your ability to persuade people of ideas, get things done? I've always been interested in this general category that's sometimes called tools for thought. This is an image of an application that came out in 1987 when I was in college, sophomore year. It was called HyperCard. Those of you old timers, it was an early version of what became the hypertext revolution that was ushered in seven or eight years later with the World Wide Web. I got a taste of with this application using software to help me write my papers and research my papers. It was very crude by today's standards, but it gave me a glimpse of something. I'd spent the next 30 years or so writing both books and writing about these tools. I'd written a long piece for the New York Times Magazine about GPT-3, the predecessor of ChatGPT that came out about two years ago. I got an email from folks at Google saying, hey, could you come in and talk? We've been following your writing for many years. This thing you've been chasing your whole life, this dream of a virtual research system that can make sense of large bodies of information and turn it into something that is actually usable and packageable for your needs, we can finally build that, that thing you've been dreaming about for more than two decades. We can build it because we have these new language models that allow us to do things with words and sentences and meaning that really no computer in the world could do just a few years before. They invited me to come and help build a small team to build basically my dream application. It was initially called Project Tailwind. Then we announced it last year after a big event called Google I-O.

[3:59](https://youtu.be/FJLL0Af2JjM?t=239) It came out as Notebook LM, as in a notebook designed from the ground up with a language model at its core. It officially launched in the US. I'll talk a little bit about the restrictions on it in a few minutes, but officially launched in December. It's still very early. We're still adding a ton of features, but it is now a functional usable product. Hopefully, you'll see it can do a lot of things that would have seemed miraculous just a couple of years ago. Before we get into actually the demo, though, I want to just walk you through a couple of core concepts that really define what we're trying to do with Notebook. Maybe explain a little bit of how it differentiates from traditional just straight chat application like ChatGPT. The most important one of those is this idea of source grounding. From the beginning, we have had this principle with Notebook LM that the whole interaction with the AI is grounded in sources or in documents basically that you define. You don't really have an open-ended chat with the AI. It's technically not possible to even do it right now in the application. I think we may change that over time, but right now, you can't even get started until you have defined a set of documents that are crucial to the project you're working on. I'll get into the details of what they can be, but you can imagine if you're a student writing a paper, it would be the syllabus for your class, the reading list for your class, all the PDFs, textbooks, whatever that you're working on. If you're a lawyer, it would be the relevant meeting notes with your client and the relevant case history. If you're working in a regulatory space, it might be the existing regulations and a project you're working on.

[5:56](https://youtu.be/FJLL0Af2JjM?t=356) Whatever you do, whatever you upload as your sources, from that point on, the AI is grounded in the information inside those sources. The shorthand way of saying this is, and it's a little bit of an overstatement, but not quite as much as you would think, that when you upload sources to NotebookLM, the AI becomes an expert in those documents. From that point on, when you ask questions, when you ask for summaries, when you start writing, everything the AI says will be based on the information in those documents. As you'll see, you can go in and see citations, you can see where the core original text was that the model used to answer questions, but with rare exceptions, you won't be able to ask about information that is outside the boundaries of those documents that you've defined. I'll show you some examples of that, but basically, it creates a knowledge barrier, so the model is really just focused on the sources that you have shared. This is advantageous for a couple of reasons. One, it greatly reduces the hallucination rate. We all know with a lot of these language models, they have a tendency to just make stuff up. If you ask them, for instance, if you go to any of the AI models, Google's and other people's models, and ask for a biography of Stephen Johnson, there will be some accurate things in there, but there will also be a bunch of prizes that I did not win that will be included in my bio. It'll be like, Stephen Johnson is the Pulitzer Prize-winning author, and I'm like, 13 bucks, and it's just, for some reason, the models think that they're writing about the biography of an author, and it feels like they should probably have won a bunch of these prizes, and sadly, I haven't won a Pulitzer Prize, but a number of these models will report that I have.

[7:51](https://youtu.be/FJLL0Af2JjM?t=471) With source grounding, you really reduce that dramatically. If you have somewhere a bio of Stephen Johnson in your sources, it will just almost always stick to the facts in those docs, so it's very important in terms of accuracy and factuality, and it also means that you can personalize. A general, out-of-the-gate model like Gemini or Chachapiti doesn't know anything about the work that you're doing. It has a lot of general knowledge about the world, but it doesn't know the project that you're working on or the meeting transcript that you need to understand, and so this enables you to basically personalize the model that you're working with. There's a lot of advantage to this approach, and it's kind of a core to what notebook.com does. One of the other things that's very important, kind of a technical thing, but I think it's increasingly important for folks who are not particularly on the cutting edge of technology to understand this, and it's important for a couple reasons, is this idea of a context window.

[8:51](https://youtu.be/FJLL0Af2JjM?t=531) AIs have their initial training phase where they basically are fed an insane amount of information from information on the web and Wikipedia and other sources, and that's where they kind of get their general knowledge, and that's a process that takes months and months and months to train these models, but then when you are interacting with the model, you are putting information into its context window, and a context window is almost like the model's short-term memory. If you think of their training data as their long-term memory, the context window is what they're focused on in the short term, and what has happened over the last couple of years, kind of behind the scenes, but we felt it as a society, actually, is that the context windows for these models have gotten bigger. You can put more information in their short-term memory, and that's one of the reasons why ChatGPT was possible in the first place, because the context window got bigger, and you could put the history of the conversation you'd have with the model in this larger context window, and so it seemed like you were engaging with an entity that had memory, had memory of what you've just been discussing, all that kind of fluidity of conversation. It actually came about because these context windows were getting bigger and bigger, and when you have something in the context window, that is information that the model, again, will be much less likely to hallucinate with. If you show them a bunch of documents, and you can fit them in the context window, they will be very faithful to those documents. They'll be much less likely to make stuff up, and it just makes for a better experience altogether. When I started working at Google about two years ago, a little less than two years ago, the biggest context window we had was something like you could fit something like 2,000 words into that context window.

[10:37](https://youtu.be/FJLL0Af2JjM?t=637) We just released a new model that is coming online in the next couple of weeks that you can fit 700,000 words into the context window, so you can put four or five big 19th-century novels into the context window, and the model will be very fluent in its ability to answer questions and respond to information in that context, so that's been a major change, and in a sense, one of the things you can think about Notebook LM doing is allowing you to kind of shift information effortlessly in and out of the context window so that you can focus the model, you can ground the model on the information that you want to work with. This is also important from a privacy and security point of view, and I know this is essential for anybody working around and inside of government. While it may seem as though we are training an AI on the information you give it, that is not what is happening. We are just taking the information that you share in the form of your sources, and we are briefly putting it into that short-term memory of the model and allowing you to have conversations with the model based on that. Once you finish that conversation, it is wiped from the memory of the model, and it's as if the AI or Google had never interacted with that information at all, so there is no risk of information that you're working with, whether it's quotes from a book you might have read, or whether it's a private personal document like a journal, or whether it's an important government document that has some restricted access. If you feel secure about uploading that document to a service like Google Drive and using it in Google Docs, you should feel confident using it inside of Notebook LM. There is no way that information is going to leak out into the training of the AI and appear in somebody else's interaction with the AI or somebody else's notebook. We are not learning anything long-term from the information you share, so you should feel confident about using it in

[12:41](https://youtu.be/FJLL0Af2JjM?t=761) that context if you feel confident about using any kind of cloud-based service like Google Drive or Dropbox and so on. I thought that was important to share. Finally, just a couple of basic things. It's available at notebooklm.google.com. It is currently US-only. We hope to change that quite soon. It is technically considered an experimental app. We also hope to change that quite soon. That doesn't really have any other implications other than the fact that if you are signing in with a Google Workspace account, the Workspace Administrator has to approve experimental apps. Maybe Hannah can share later. There is a link to a page that walks through how you do that, but you might see something saying you don't have access to experimental apps. If you sign up for it with your normal Gmail address, it will just work. If you are in the US, if you sign up for it in a Workspace account, you have to make sure that this experimental app thing is toggled on. Those are just a few technical things. I can explain more about that when we get to Q&A. Let's jump into an actual demo here. I will share another screen. What I'm going to do, I'm going to show you two things. I'm going to walk you through creating a notebook and give you a couple of examples. Then I actually have a vaguely city government related sample notebook that I just remembered I had. I'm going to use that to show you a couple of things as well. The first thing that happens is you drop in a notebook and you will see more or less a blank screen. You will see some example notebooks down at the bottom that you can explore. There's an introduction to Notebook Notebook that is actually pretty cool. You can use the product and learning how to use

[14:44](https://youtu.be/FJLL0Af2JjM?t=884) the product. In this case, you would just see a new notebook icon. You click that and we're going to call this RFK Transcripts for reasons that will become clear. You give it a title and you open it And now you have the, this is the basic space that you kind of get dropped into when you begin using Notebook On. We have a light mode and a dark mode. If this is dark mode, if you'd like light mode, you can just change that. And the first thing you're gonna wanna do is add some sources, right? So your sources are gonna be here in this left-hand column, and we're gonna load in, if I hit Add Source, I get a bunch of options. I could load Docs for my drive, just Docs for now. We don't support Sheets or Slides yet, but that's coming. You can load a PDF from your desktop. You can load a text file from your desktop. You can paste in copied text that you've just copied and pasted from another app. And eventually you'll be able to upload a website that I have access to that because I'm on the inside team here, but that won't be available to you if you try it today.

[15:51](https://youtu.be/FJLL0Af2JjM?t=951) In this case, we're gonna load PDFs. And so I click that, and what I have here is these are a bunch of interview transcripts that I happen to have downloaded from the Kennedy Library. And they're long oral histories about JFK and RFK. I selected a few about RFK. And so I'm gonna just batch select kind of five of these transcripts. And these are all, I don't know, maybe something like 30, 40,000 words long. Each of your sources can have up to 200,000 words in it, and you can have 20 sources in a notebook at a given time, and then you can have multiple notebooks. So you can effectively have up to 4 million words that are kind of active in a given notebook, 20 sources times 200,000 words per source. Those are the current limits. This is probably gonna go up. So now the sources are kind of loaded here. I'm gonna select all of them.

[16:51](https://youtu.be/FJLL0Af2JjM?t=1011) And by selecting them, that means they're active, like the AI is gonna be focused on the information in these sources. I'm gonna first show you that I can open up any of these sources, and I can read the text of it here. If you are bringing in a PDF that has complicated formatting like multiple columns, that can kind of be a little funky. If you bring in a doc, it will be very nicely formatted, but you can read it generally here. And then we automatically create a source guide for each source. So the AI has gone through here and figured out like what's in this source, and it creates a summary. So the provided text reveals the complexity of Robert and Edward Kennedy's relationship by highlighting their different approaches to politics and staff management. Edward Kennedy possessed a more amiable and diplomatic style, blah, blah, blah. We also suggest some key topics that you can explore as well. So this is the way you can see your sources. You can read your sources if you want, but now they're all loaded up, and you can tell that I've selected these five sources, and then down here by the chat window here, I can see five sources.

[17:58](https://youtu.be/FJLL0Af2JjM?t=1078) So that means I am talking to those five sources. The AI is focused on those five sources. If I deselect one of them, that becomes four sources, right? So I'm talking to all five of my sources. And now I can just ask any question I want. So I can say, tell me about RFK's relationship sorry, relationship to his brother, Edward. Include direct quotes. So I wanna have it go through all these sources that I've uploaded. It's gonna find the most relevant passages. It's gonna summarize what's in those passages about RFK's relationship to Teddy, and it's gonna give some quotes. Although in this case, it didn't give any quotes, but that's okay. So I'm gonna say, we'll do that, so those are both. So here we see this kind of summary. It creates a, like, it will highlight the core ideas in bold, and it often will give you kind of nice organized kind of bullet points.

[19:04](https://youtu.be/FJLL0Af2JjM?t=1144) And what's crucial here is it will also give you citations. So this is the model's kind of summary based on what was in this document. But you also see these citations down here. And if you click on each one, it will show you the original passage from one of these documents that was relevant to this question. And so you're always, like, one click, if I just click on this number, it takes me immediately to the passage. Here you can see it's all about Edward Kennedy, would not make a bubble, but I clearly nod to Edward Kennedy. Robert Kennedy was talking all about their relationship. So I can jump from a summary of the answer I'm trying to get immediately to check the references, basically. And sometimes what you really want is actually the original quote. You don't even want the model summary. You wanna just be able to get to that point in your passages. And imagine that would be just as fast if we had 4 million words loaded in here. It will find the most relevant passages and take you to them almost immediately. Now, the other thing that's really cool here is that we are also suggesting questions as you go.

[20:07](https://youtu.be/FJLL0Af2JjM?t=1207) So based on the content of the source material and based on the conversation you've had so far, we're gonna get a bunch of suggested questions down here that you can use to just explore just by clicking on them. So one suggested question was, how did Robert Kennedy's China speech affect Edward Kennedy's staff? So let's just click on that. That sounds interesting. I don't know what the answer to that is. I don't even know what the China speech was, but we're gonna find out about that. And so it will go through, gives it kind of a short answer there, but it says that the staff was upset about the China speech, but Kennedy's speech received a big notice. So let's say, as we're going through this, we find really interesting material that is important for whatever we're doing. And when we see something that we like, that's valuable, we can always just pin it. So you just hit this little pin here in the corner, and that adds that response to this noteboard area that's basically where you save all of your notes.

[21:10](https://youtu.be/FJLL0Af2JjM?t=1270) And so if you go back to the chat and you also wanna capture this small thing about the China speech, boom. Now I've got two saved responses. I can see the original citations. I can open up this and read it in more detail. If I wanna go to one of these citations, let's say this is about the China speech. So I'm gonna click, oh, I can't do it from there. I'm sorry. If I go to here and I go to that point in the text, this is the China speech right here. If I actually want to save, let's say this whole passage, I can just select it in the source and automatically I get a bunch of these options here. So I could add this to a note and it will just take the quote and create a new note with that information straight from the passage. It's actually right here. So this is the actual quote from my passage. So I can like capture information from the sources and I can create these notes here and I can also write my own notes.

[22:13](https://youtu.be/FJLL0Af2JjM?t=1333) So I can add a note up here. I can open up this note. I can give it a title, ideas about RFK and Teddy. And then I can spell Teddy correctly. And then I can say they had a complicated relationship or deep here. And so the green notes are notes that I've written. The purple notes are notes that I've saved either from model response or from the sources themselves. And so I can start to collect ideas that are relevant to the project that I'm working on. I can start to kind of build a kind of a map of information that I'm looking for. And what's really cool is that once I've built out this kind of noteboard space, I can select these notes. And once I selected them, I can do information, I can do transformations of this information, specifically this information. And you can see this because now down here, instead of saying five sources, the AI kind of chat box area is saying four notes.

[23:20](https://youtu.be/FJLL0Af2JjM?t=1400) And that means the AI is now focused on those four notes that I've selected. And so I could summarize those notes. I could find related ideas from the source. I could create a study guide. I could create an outline. I could combine them to one note. In this case, if I create a study guide, it's gonna go through and it's gonna create a set of questions to ask me about the content of the notes. These are central questions to review. And then it creates a glossary of key terms and ideas so that I can understand the material better. But at this point, actually, with these notes selected, I can write anything I want here. I could say, draft a outline for a slide show about RFK and his brother. Include bullet points for each slide. So now based on the information that I have saved as these notes, it's gonna go through and create this new format that I... Oh, well, it didn't work.

[24:20](https://youtu.be/FJLL0Af2JjM?t=1460) I don't know why that happened. That is weird. It should... Let me try it again, actually. That may just be a momentary glitch. Momentary glitch. That is our API not being available, which means it's just getting... So yeah, it's getting slammed with activity. Okay, we won't do that, but you would have seen a lovely little outline there of potential slide ideas that you could create. So the whole model here is to be able to go through all these documents, ask questions, get grounded answers, get grounded citations, grab things that are useful or interesting to you, kind of post them, collect them, save them, and then remix them in new ways. Turn them into an outline for a blog post or an internal briefing document that you wanna share or a summary for your own recollection or a study guide if you're a student. And I wanna show you this other notebook too. And then in about five minutes, we'll switch it over and open it up for questions.

[25:24](https://youtu.be/FJLL0Af2JjM?t=1524) So this is this notebook that we created. It's all kind of made up, but it's imagining a kind of a parks committee in a Midwest town. And they've been put together to preserve this town square called Fire Tower Square. And you can see here that the documents that I've included, the sources that I've included have the board meeting minutes. So transcript of a board meeting. There's a newspaper article about graffiti problems. There's the history of Fire Tower Square. There's a proposal for a new statue that's being built there. And then there's the regulations about public monuments that have been kind of uploaded to those documents. So you've got a mix of source types that are there. And so you can ask all sorts of questions based on these documents. So for instance, you could ask a question like, what are the qualifications for obtaining a permanent permit to erect a public statue?

[26:28](https://youtu.be/FJLL0Af2JjM?t=1588) I'm hoping our API is not giving us trouble here and this will work. Okay, that's promising that it's, oh, wow, that's really unfortunate. Okay, I don't know what's going on, but it's okay, I have enough stuff saved that I will, that does not normally happen. So for instance, just earlier actually, when I was testing this, I said, what are the limitations on building a statue in the town? And so based on the public ordinance, it summarizes the statues and public monuments in the town of Woodward require a permit from the city manager. The applicant needs to attain all the necessary approvals, blah, blah, blah. And then I can, again, I can click on the citations. I can go directly into the source and I can see right there that it's number D involving public monuments. If I want to, if I'm thinking, okay, this is all, this whole section here is what I'm really interested in. I wanna make sure I can capture that.

[27:28](https://youtu.be/FJLL0Af2JjM?t=1648) I can just select that text. I can say add to note and it creates a new note where it's captured that information. I can give that a title, permit regulations. And so I can save that and then I could remix that later into another document. I get suggested questions here again, describe the different types of conditions that the city manager can impose, discuss the process by which a person can apply for a permit. All of these things are suggested as you go. And the final thing I'll say is that we haven't released yet but we're hoping to do this soon. You can imagine this is a really valuable kind of interface for a help center. So if you wanna kind of upload regulations or frequently asked questions, you can create a bunch of sources that define all the issues. And then you could create, this is what we haven't fully released yet but it's definitely coming.

[28:29](https://youtu.be/FJLL0Af2JjM?t=1709) Basically a public notebook where anybody can go and ask questions and the model will answer based on the sources that you've defined. You can give a little guidance about how you want the answers to be explained. You can include citations if you want, if you want them to be able to go and read the original documentation that's available as well, or you can turn that off. So it can be both a public notebook and a private notebook. We also finally allow you to share. And so if you share a notebook, you can share it in a view mode where someone can only ask questions and get answers based on the sources that you have defined and which you can share individually with a person or a small number of people, or you can create a truly shared notebook where anybody else who you share it with can add their own sources and they can also add their own notes since you kind of have a shared noteboard space. So we have the beginnings of some pretty cool sharing features but we're just getting started in that direction.

[29:31](https://youtu.be/FJLL0Af2JjM?t=1771) So that's the overview. There's a, it's a really, one of the things that I love about this tool, both as a writer and a researcher and as somebody who's developing software is it's very open-ended in what it can do. And we're constantly discovering new things. The AIs understand a lot of different formats. And so, You can go in and, you know, select your sources and say, hey, I would like to have a draft of a software PRD based on these documents with notes on the software I'm trying to build, and it actually will understand what a product review document is or whatever the acronym PRD is for. And you'll just get these like beautifully crafted kind of first drafts of things grounded in the information that you give it, and then you can kind of refine and add, and then you can bring it into whatever document tool you use, whether it's Google Docs or Word or whatever, to create the final product. So I encourage you to go in and just kind of mess around, and one easy way to do it is to just get a sense of what it's capable of. If you are a user of Google Drive, open up a new notebook and just go in and select the last 10 docs that you have in your drive. Don't even think about organizing them, just like select those and import those as sources, and just kind of explore, like look at its summaries, look at the suggested questions, ask questions to see what it picks up on, follow the citations. I think you'll find it's got, you know, quite powerful knowledge of the information that you've shared, again, also in a very secure and private way, and just doing that is a great way to get your kind of feet wet and get a sense of like what the product is capable of, and then you can start thinking about using it in a more focused way for products here and projects you're working on. Okay, that's the intro to Notebook L. I'm happy to turn it over to questions and comments and use cases, and I can switch back. I mean, hopefully it will

[31:37](https://youtu.be/FJLL0Af2JjM?t=1897) start working again, and I can actually show you a few more things if people want to see specific things.

**Unknown speaker**

[31:42](https://youtu.be/FJLL0Af2JjM?t=1902) Thanks so much, Stephen. Let me just pause the recording.

* * *

Original page: https://innovate-us.org/industry-series-learning-to-use-google-s-ai-powered-research-and-writing-tools-in-your-work · Licence: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org