Everyday AI Podcast – An AI and ChatGPT Podcast - EP 370: NotebookLM - The best AI tool you’ve probably never used

Episode Date: October 1, 2024

Lean in for a secret. NotebookLM is probably the best tool Google has made since the search engine. We dish on what you need to know about NotebookLM, which we think will change how LLMs are built in the future. Newsletter: Sign up for our free daily newsletterMore on this Episode: Episode PageJoin the discussion: Ask Jordan questions on NotebookLMUpcoming Episodes: Check out the upcoming Everyday AI Livestream lineupWebsite: YourEverydayAI.comEmail The Show: [e-mail] with Jordan on LinkedInTopics Covered in This Episode:1. Features of NotebookLM2. NotebookLM vs Google Gemini3. Benefits of NotebookLM4. Use Cases for NotebookLMTimestamps:02:20 Daily AI news07:15 NotebookLM recent updates15:19 Upload and manage all document types easily.19:38 RAG enhances AI by retrieving external information.23:40 Notebook LM ensures grounded data usage only.29:56 Research, notes, listen, transform to speech, review.34:29 Google Gemini back-end good, front-end disappointing.42:26 Notebook l m handles diverse information efficiently.47:23 Great learning tool for students and professionals.51:14 Quickly toggle document sources; folders needed.56:15 Wildly popular, better received than Gemini model.01:01:07 Web access crucial for large language models.Keywords:NotebookLM, AI tools, Google Gemini, data upload limit, AI bots, livestream audience engagement, content sharing, LinkedIn livestream, Everyday AI Podcast, Google account sign-in, data privacy, data scraping, Retrieval Augmented Generation (RAG), grounded model, Chat GPT, large language model, AI milestones, OpenAI Dev Day, Cerebras Systems IPO, Microsoft Copilot 2.0, Python, CSS, HTML coding projects, Notebook LM use case, source citation, data integration, AI overview feature, Deep Dive Podcast feature, Teaching Tools.Send Everyday AI and Jordan a text message. (We can't reply back unless you leave contact info) Start Here ▶️Not sure where to start when it comes to AI? Start with our Start Here Series. You can listen to the first drop -- Episode 691 -- or get free access to our Inner Cricle community and all episodes: StartHereSeries.com Also, here's a link to the entire series on a Spotify playlist. 

This is the Everyday AI Show, the Everyday Podcast where we simplify AI and bring its power to your fingertips. Listen daily for practical advice to boost your career, business, and everyday life. Meet Firefly AI Assistant, now live in Adobe Firefly, the All In One Creative AI Studio. Just describe what you want to create and the assistant handles the rest, orchestrating multi-step workflows across Photoshop, Premiere Express, and more in one conversational interface. You direct the outcome, the assistant accelerates execution. Notebook L.M from Google is the best AI tool that you've probably never used. Yeah, we've talked about it on the everyday AI show plenty.

But I think some recent updates from Google's notebook L.M makes it not just one of the most useful AI tools, but probably one of the best, right? One of the things that we always worry about when using large language models is, number one, how can it work off my company's data? And number two, how can we get rid of hallucinations? Right. So whether you're using Microsoft co-pilot, Google Gemini, chat GPT, Anthropic Claude, those are things that you are constantly kind of struggling with.

notebook LM kind of solves it. It's wild. All right. We're going to be talking about notebook LM, what you need to know, how to use it. The pros and the cons, everything today on everyday AI.

What's going on, y'all? My name is Jordan Wilson, and I'm the host of Everyday AI. This is for you. It's for everyday people, how we can all learn and leverage generative AI through a daily

live stream podcast and the free daily newsletter. If you're brand new here, maybe this notebook L.M caught your eye and you want to know more. Well, you are in the right place. All right. So if you haven't already, please make sure you go to your everyday AI.com. Sign up for our free daily newsletter where every single day we recap our podcast episode.

But also, it is literally a free generative AI university there. We have more than 360 different episodes. So go learn from the world's experts who are leading in AI, whatever you care about, marketing, advertising, education, non-profits. There's a category there. You can learn everything. So make sure you go check that out. All right, before we get into Notebook, L.M. And I am very excited for today's show. Let's first start as we do every single day by going over the AI news. So first of all, speaking of Open AI, Open AI has their dev day today in San Francisco. So Open AI's Dev Day 2024 is set to kick off today in in San Francisco. So while the in-person event is invitation only, there could be a live stream

available, but we're not really sure yet. This has been a kind of quiet announcement from OpenAI. They've talked about having dev days throughout the world really in different cities, but this is their first dev day in 2024. And like I said, there's very few details, but if we follow most of their patterns, there could be a live stream today around 11 a.m. Central time. That's kind of how they've normally done it. And you can just kind of keep an eye on OpenAIs, you know, social media accounts, but generally they will put out a live stream on YouTube. So we'll see if that actually happens. Last year at Dev Day, OpenAI announced a lot. The new model, their new model of the time, GPT4 Turbo, the assistance API, custom GPTs, so many

new capabilities. So we'll see if this year, if they're announcing anything new. I don't necessarily think that we're going to be seeing anything big today. I think that it's mainly going to be developer tools and some API improvements, lower costs, etc. But we might get some news on the new 01 model or some SORA or Dolly updates. All right, Sarabras Systems has filed for an IPO amid growing AI chip competition. So Cerebra's Systems and AI chip startup has filed its prospectus for an initial public offering aiming to trade under the tickle the ticker symbol cbrs on the nasdeck so this move comes as the company seeks to establish itself in a competitive market dominated by giants like invidia yeah when we

talk about generative i we have to talk about the GPU chips uh that power at all and just the immense needs for that power and energy so the company's w s e3 chip which boasts more cores and memory than invidia's popular h 100 positions itself as a strong contender in the AI chip market, which is increasingly crowded with competition from companies like AMD, Intel, and even Google. So the technology IPO market has actually been kind of sparse since in 2024, so not a lot of new technology companies going public. But there's obviously rising interest rates, pushing investors toward more profitable assets, making Cerebra's IPO a significant event to watch.

Major investors in Cerberas right now include Foundation Capital benchmark and Eclipse ventures with notable individual backers like OpenAI CEO Sam Allman. Interesting there. All right. Last but not least, Microsoft is slowly, maybe quietly rolling out a copilot 2.0 with new features and a fresh look for web users using co-pilot at copilot.microsoft.com. So Microsoft is reportedly set to launch an updated version of its very popular co-pilot AI chatbot featuring a revamped interface and new functionalities aimed at enhancing user experience. So the new copilot 2.0 will showcase a pastel shaded user interface. That's hard to say in the morning, pastel shaded user interface, replacing the previous design,

and it is reported to be faster and more intuitive and friendly, similar to check. at GBT. It's also rolling out a card-based design that could be implemented, encouraging users to explore various AI applications such as journaling or improving sleep. So for the first time, Copilot will also include a voice mode allowing users to interact with the AI using spoken commands. So it has four new voices right now, Meadow, Grove, Wave, and Canyon. Also upon first use, Copilot 2.0 will ask for the user's name and remember it for future sessions, creating a more personalized experience. All right. A lot more on those stories and so much more. So make sure you go to your everyday AI.com. All right, live stream audience. Thanks for joining us. I know there's a bunch of

LinkedIn problems, you know, happening with our live stream. But you can always catch up on the podcast or, you know, on YouTube. So yeah, thanks for sabbatical life who let everyone know, Hey, Frozen on YouTube, go over, or sorry, frozen on LinkedIn, go over to YouTube. Yeah, sorry about that, y'all. But nonetheless, let's jump right into it. Notebook L.M from Google is friggin amazing. Did our first review of the tool probably like four or five months ago, but there's been some recent developments that I think makes this a no-brainer for everyone to use.

companies, individuals for fun use cases, for business use cases. There's no, honestly, it's probably one of my most used tools. All right. So let's talk about some of those use cases and we'll tell you what it is. But, you know, notebook LM, it's not technically its own model. It is an AI tool from Google based on their Google Gemini model. Some use cases for it that I love right off the bat just so you can kind of set your mind. to it, the ability to learn, having it, you know, having a learning assistant like this is, it's wild, right?

So we're going to do some live demonstrations later, but I love to use it for learning. I love it to use it as a personal, large language model, and then also a fast company rag model. All right, we're going to talk more about, you know, grounded models and retrieval augmented generation here in a minute. And also, I do want to talk about this. And I, I want to talk about how significant this is. We always talk about these, you know, now it's just called a chat GPT moment, right? In technology, in AI, right? But it's kind of like that aha moment, right? It's when you, there's almost kind of like a line in the sand, you're like, wow, this is, this is pretty big. This is pretty big. If I'm being honest,

When chat GPT came out in November 22, I wasn't very impressed, which I'm always like, curious, right? Even in video CEO, Jensen Wong, you know, talks about the chat GPT moment, right? Kind of this shift toward traditional AI and traditional machine learning to this generative AI wave, you know, Jensen Wong and most others talk about the chat GPT moment. When chat GPT came out, I was not very impressed, if I'm being honest. You know, at that time, our company, we had been using different GPT tools for more than a year, right, since late 2020. So almost two years.

So chat GPT when it came out, I was not impressed. It wasn't even one of my five favorite GPT tools. So it's always funny when we talk about this chat GPT moment and this kind of aha that you're like, oh, okay, this makes me think about AI a little differently. And I wanted to share some of mine. So for me, there were a couple, but it wasn't chat GPT. I'd say one of my first chat GPT moments or sorry, yeah, was when chat GPT got plugins, okay? Plugins are obviously gone now in lieu of GPTs, but the thing that really was an aha moment for me about chat GPT plugins, well, you could run three of them at a same time, at the same time. And we almost had agentic AI at that point, right?

Now it's gone. But essentially, you could have three different plugins doing different tasks, working with each other. So plugin A could go research something, pass that information off to plugin B, and then autonomously would do the work between the three plugins. That's gone. We have GPTs now, but you can only use one GPT at a time. But that was probably my first chat GPT moment.

Then Suno, the kind of text to music and the AI music platform. I don't know if any of our live stream audience has used Suno. It's amazing, right? And that's another one. When I heard Suno, I immediately reached out and had their CEO, Mikey, on the Everyday AI show. And then probably one of my more recent ones was Claude Artifacts, right? The ability for a large language model to essentially deploy code. and to render code live in the browser, huge, right?

A huge step for large language models. Notebook LM has been my latest. It is, at least for me, and I think for a lot of people, it is a chat GPT moment. I know that's a weird thing to say, right? Like, oh, Notebook LM is a chat GPT moment, but I think it is because I think it is slowly changing what is possible with large language models. So emphasis on the slowly. because with some of these other tools, it was almost like love at first sight or, you know, the aha

moment at first site. It wasn't like that, at least for me, with notebook L.M. Because like I said, we've, we reviewed it four months ago. And I was like, oh, this is great. But I think it's actually been some of these quiet updates over the last couple of months that have really turned this thing into a powerhouse. All right. So let's go over the basics.

All right. Let's go over the basics of notebook. So it is based on Google Gemini's 1.5 Pro. All right. So yes, this is from Google. It is not its own model, right? Notebook L.M is not a frontier foundation model. It is an AI tool built on top of Gemini 1.5 Pro.

Google's most powerful large language model right now. It is free to use. Let me repeat that. that it is free to use. Okay. Wild. But you do need a Google account, FYI. All right. Also, as always, a big, a big key, a big feature that sets notebook LM aside from, apart from others, is the ability for to handle large amounts of your data, right? So before you run out there and dump everything in Google's Notebook LM. As always, be aware of what companies do with your data, read their privacy, policy, et cetera, okay? I got to get that out of the way. Here's the other thing

that you need to know. It's not a traditional chat interface. All right, and I'm going to explain what that means a little bit more, but it doesn't save chat history. It works on this thing called notes, so notebook LM, right? And I actually think that's a bad thing. Or sorry, I'm going to I actually think that's a good thing. Even though it sounds like a bad thing, right, not saving your chat history, I actually think it's a good thing. And I'm going to explain that to you later. The biggest thing here that sets everything else apart is being able to quickly and easily upload all of your documents. And when I say all, I mean, my gosh, it can handle a lot of your documents.

Then the very popular Deep Dive AI Podcasts, right? that's probably what has caused notebook LM to fall on your radar. And this was not an original feature, right? This feature has only been out for a couple of weeks. We reviewed it the day it was released and shared that in our newsletter. But I think that's really what has built this into kind of an all-encompassing tool to learn. Right. And then it is a grounded rag approach.

Retrieval, augmented generation. We're going to talk about that more in a little. second and it is y'all it is stupid impressive it's very impressive sabbatical life says yeah free for now uh jacky jacky joining from youtube uh yeah our normal lincoln audience all had to leave lincoln what's wrong with you linton um says jacky says when everyone uses it and it's like crack at google will charge maybe or is this their trojan horse we'll see all right so from a visual perspective podcast audience, I'm going to describe this to you. It's very simple. When you go into Notebook LM, you essentially will click on new notebook. Think of notebooks as a chat where you upload all

of your documents. Okay. And when I say all of your documents, I mean all of them. Because now, as of this week, there are some new file formats that I think, like I said, it is kind of this slow rollout of multiple things that have made this aha moment for Notebook LM. So you can upload file from Google Drive, right? So your Google Docs, Google Slides, et cetera. So it natively reads your Google files. So if you are a Google workspace user, whether personally with a Gmail address, your company, that's great. Then you can also upload links to websites and YouTube URLs. That part is huge, right? That part is huge. YouTube URLs. That saves so much time. However, it is worth pointing out, the video has to have captions enabled.

Otherwise, it won't work. Similarly, with a website, if a website blocks certain bots essentially, right? So a lot of website publishers now are putting something essentially in their robots.txee file that tells the rest of the internet how it can and cannot access the information. So if certain news or if certain articles or websites block access, to certain scrapers or crawlers that won't be able to work. However, you can just paste the text in there. So that is the last kind of option there at the bottom.

But then you also have the ability to upload any files you've downloaded. So it's not just Google only, right? It's not just, oh, I don't have Google Doc so I can't use it. Well, you can copy and paste anything. Website links, YouTube links. And also here's the other huge one. The ability now to upload MP3s. Yeah.

Y'all think of all the use cases already, right? Being able to record all your meetings, right? Download the audio of those meetings or the transcripts. And you literally have a notebook that can handle everything, a large language model that can handle every single thing you're working on. And I can't emphasize enough the context that you can fit in. All right. That is what makes notebook LM different immediately is you can upload up to 50 sources.

And you might be saying, okay, Jordan, that's not a lot. You know, I have more than 50 files on my computer. Okay, well, you can combine them because each source can be up to 500,000 words. Okay? That is 25 million words. 500,000 words per source, 50 sources. That is 25 million words. All right.

There are very few, aside from enterprise companies, but you can get 90% of the day-to-day data that you need right there. We have, for the first time, a business rag, large language model that is so easy to use. That's one of the things that makes Notebook L.M different. And hey, live stream audience, thank you all for, for, for, jumping over to YouTube. Sorry that LinkedIn is not working, but if you have questions, please get them in now. All right, another thing that makes this different. Well, it's a grounded model via a rag approach. All right. Let me quickly explain what that is, and I'm going to oversimplify it. So if you're a even nerdier than I am, you know, you don't have to say,

oh, Jordan, you're wrong on these definitions. I'm simplifying here. I want to talk about what it means for a model to be grounded. Okay. So grounding is a broad concept in AI that essentially refers to connecting in AI systems understanding in its outputs to real world information or a specific knowledge base like your company's knowledge or something specifically that you're trying to learn. All right. So grounding aims to make AI responses more accurate, more hallucination-free, relevant, and aligned with certain facts or scenarios. Okay. So that is the concept of a grounded model. One of the ways that you achieve a grounded model, one of the ones that's most probably relevant to people here, is retrieval augmented generation or RAG. It is a process to ground a model.

So most of you all have heard about RAG, but RAG is essentially a technique that enhances AI models by retrieving relevant information from external sources to help augment the model's knowledge when a user asks for something. Right. So, you know, we talk about a rag layer. Okay, so think, when you send a query to a large language model, it goes straight to that model. So a rag layer or the process of retrieval augmented generation, think of it as a layer between your question and the model itself.

Okay, that is a rag layer. For all intent and purposes, we're simplifying here. Okay. And you put your model, all your information. in this kind of retrieval, augmentive generation layer. And that is one of the ways you can ground a model to make it more factual, to make it more hallucination-free, right? Because that's one of the biggest problems right now with any model

is fighting against hallucinations and also bringing your data in there. So essentially, having rag inside of a model combines the best of a large language model with that retrieval system. so you can get up-to-date information and domain-specific knowledge. Okay. Here's the thing I love about notebook LN. It is rag or grounded by default. You cannot go in, click new notebook, and just start chatting.

That's not how it works, right? If let's say you work at a shipping company and you're in marketing at a shipping company, right? You're going to upload all your documents, all your meetings, etc. you aren't going to use that model to write a haiku about advertising at a marketing company, right? It is grounded in the data. So for the most part, although you can always kind of, you know, quote unquote, jailbreak it, or kind of make it go in a different direction. It is only going to work in the data that you provide it, right?

If you ask it to do something else that is not really in its, knowledge base, it's going to nicely tell you, hey, that's not really what I'm here for. So one kind of phrase that I like to tell people when working with a grounded model is it's grounded. Think of it like that, right? Little Jordan in the corner, got in trouble, got grounded. He's got to stay there. Can't do anything else.

He's got to stay there and look in the corner. So the same thing when you set up a new notebook inside of notebook LM, it is grounded, right? You have to start by putting it in the corner and you say, hey, model, This is literally all you can look at and all you can work on. So it's very different than a traditional model that you can, it's not grounded. You can do literally anything and everything. And even when you upload your documents right into chat Gpts, right, you can build a custom GPT inside of Claude's projects, right, inside of Google's gems, right?

So all of these kind of versions of the big models that you can personalize with your data. Right. So those are kind of the three comparisons. Custom GPTs from chat GPT, Claude projects and Google gems. And then also Microsoft co-pilot has GPTs as well, although they got rid of that on the pro plan, which I'm sad about. Anyways, so all the other model makers, you can do something similar, right? upload all your data, give it some custom instructions. But here's the thing. Those models aren't grounded, even after you do that, right? And that can be actually frustrating sometimes because you kind of want the best of both worlds. You want a model that can openly do anything and everything and, you know, it has all these crazy capabilities. But you're like, no, no, no, be grounded.

Only use my data. Well, it can actually be difficult when using those other systems, even if you're trying to build a smaller version via GPTs, via projects, via gems, inside of Google because those models aren't grounded and they're still maybe going to go in other directions that you might not want them to. And it might use either its own internal training base, its own internal data, or it might browse the web and bring in information that you do not want. That is not an option with notebook LM, which some people might say, oh, that's a huge downer. I think that is the biggest benefit. Like everyone wants to get rid of hallucinations and they want to work with their data.

Well, there you go. The best and most accurate and secure way to do that is by using a grounded approach like notebook LM and bringing in rag data. Adobe just introduced an entirely new way to create, bringing the power and precision of its creative suite into one conversational experience. Meet Firefly AI assistant now live in the Adobe Firefly app. the all-in-one creative AI studio. Powered by Adobe's creative agent, Firefly AI assistant lets you start with your vision, just describe what you want,

and shape the outcome as it takes form with the assistant. The assistant orchestrates multi-step workflows, drawing on 60-plus pro-grade tools across Adobe Creative Cloud apps, including Photoshop, Illustrator, Premiere, Lightroom Express, and more to help bring your ideas to life. You can also get started with creative skills, a growing library of pre-built workflows for common creative tasks, like batch editing photos, creating mood boards, portrait retouching,

and creating social variations. Every step the assistant takes is visible so you can refine, redirect, or take over at any time. You stay in the driver's seat as the creative director. Adobe Firefly AI assistant now in public beta. See it today at firefly.adopi.com. All right. So let's keep it going. It doesn't save you.

your chat history. Okay. That's a little new approach, but I'm going to show you, uh, how you can still save all of that information. Okay. So when you are chatting in one session inside of notebook LM, all of that information will stay there in the chat. However, once you close out, refresh the page, et cetera, you lose access to that chat history. However, you can save it at any time. There's essentially a save as note button. And then what that does is it adds that to your notebook's knowledge. All right. So like I said, you do have to change the way that you work a little bit to get the most

out of notebook L.M. The other big thing, and this is huge, I shouldn't have waited 25 minutes to tell you this part. It cites its sources. All right. And if I'm being honest, before notebook LM, there was really no model that did this. The only model that I think did this fairly well was Adobe's AI assistant. right acrobat AI assistant which I've I've been using a lot and I've really loved it but before that there was no model that could literally cite right even on your own document

external so yes Chad GBT has improved that especially when it goes and uses the web you do you have to do a little fancy prompt engineering to try to get it to cite more things perplexity you know kind of an answers engine not technically a large language model does a very good job with citing. However, sometimes it hallucinates and it makes things up with these sources. So that is what is huge here with notebook LM is it cites the sources. It gives you for almost every single thing it pulls out. And then you can go click on that and read about it. Right. And if you're like, wait, did I really say that on that meeting a year and a half ago? Oh, wait, I can click. Oh, there's the transcript. There's literally exactly what I said.

Right. So think of the concept, how you can search for anything on. Google and you can scroll through and click obviously a page and go read that page. Same thing, right? You essentially are creating your own search engine, right? Of all, all of your data, all of your documents, all of your transcripts, whatever it is that you are uploading, all your YouTube videos, all your MP3 recordings of important calls. It's all in there and it is cited and you can click and investigate.

That is huge because like I said, normally models only do that with external sources of knowledge, mainly when they are fetching information off the internet. Like I said, there are some ways with some fancy prompt engineering. You can kind of get them to do otherwise. But for the most part, aside from Adobe AI, Adobe, the Acrobat AI assistant, no other system or tool has been able to do that until Notebook L.M. And also, the Deep Dive podcast, all right? And this was one of those, I think, viral moments that has now gotten a lot of people paying attention to Notebook LM. which is kind of funny because I'm like,

I've been telling people that used notebook LM since it came out. Many months ago, in 2023, I've been telling people like, hey, use notebook LM. It's amazing. But it was this kind of ability to do these deep dive podcasts. So what this is, you can upload all of your information. And full disclosure, I do it for almost every, everyday AI episode. I first make myself a personalized podcast. When I'm doing all my research, I now do all of my, uh,

I guess last leg research for every single show inside of notebook L.M. So yes, I'll use chat GPT and perplexity and other sources to bring in information quickly, but then I'll take the best of that and I'll put it into notebook LM. So I'll have a bunch of sources and then I will use the audio overview, which is essentially a deep dive podcast, right? So think, and this is great for students, right? Think if you're maybe taping all your left taping. Gosh, did I make myself sound old there.

there, do people tape lectures, audio recordings of your lectures, all the notes that you're taking, sources across the web, right? Essentially being able to dump everything into a notebook, literally. I think that's why they called it notebook LN because, you know, you're taking notes from all these various sources. But then this new audio overview, which is called a deep dive conversation, and then what happens is you get two podcast hosts who have a conversation. And it's not just them reading, right?

It sounds like a real podcast, like two humans with very natural sounding voices. Okay. This isn't robotic voices. This isn't your traditional AI text to speech. These are neural voices, right? They kind of laugh. They chuckle. They interact with each other.

They cut each other off like humans would. Right. If I had a co-host, the co-host would probably cut me off. And we'd be doing something like this. We'd be bantering, right? That is what this do. it. That's what this feature is. This audio overview or deep dive feature. And y'all, I'm telling you, it changes learning. This is how I learn. This is how I've been learning since this conversation

came out. Previously, I would use third-party tools and do something the same. The way I retain information the best is I first do research. I take notes and then I like to listen, right? So I might go find a YouTube video, maybe a podcast, or what I would do a lot of time is take my notes I would have a prompt or a GPT that would turn my notes into a podcast transcript and then I would use a text to speech platform or I would use, you know, even chat GPT's built in voice mode or Google's built in voice mode and just have it read it to me. Now, the deep dive podcast is better because having one AI voice read something to you is like, eh, that's okay. The deep dive podcast that you can create in this audio overview section in Notebook LM is fantastic. by itself, just that one feature, not even talking about, oh, how it bases it off all your documents.

That one feature is, I think, is maybe what made Notebook L.M, this aha moment, right? It brings a level of human to AI, right? Which I know that sounds weird. But you have very, even famous people, right? Like an Open AI co-founder, Andre Carparthy, just posted recently on Twitter about how this is one of his most used kind of features, this deep dive podcast and how he even feels, you know, almost personally attached to these two hosts because like, like me, he's using it seemingly for everything, right? This is my secret. This is how I learned

so much, essentially since this new feature came out a couple of weeks ago. All right. Also, this begs the question, what's Google strategy here? Okay, so they have this amazing free notebook LM tool, but it's based on Google Gemini 1.5 Pro, which is a paid product. So what gives here? Because you have the normal kind of Gemini chatbot. You have Google gems, which are personalized versions of Google Gemini. Then you have Google AI Studio, which is for developers, as well as Google Vertex AI.

Then you also have Google's search generative experience, which is now called AI overviews and now notebook LM? What's the strategy here? Because it looks like Google has its Gemini model working just about in every corner of the web. Can I be honest with y'all for a second? I know it's hot take Tuesday. Should I, as I take a sip on the coffee here, should I bring in some hot, takes, should I be honest about Google's AI strategy in their different kind of Gemini rollouts

right now? Let me know. All right. Well, I think I'm going to do it. Google Gemini by itself. All right. Michael gave me the permission. Michael, thank you.

Google Gemini by itself, I think is not good. The front end of Google Gemini for me, and I use law. language models, usually anywhere from two to eight hours a day. And I've been using the GPT technology since 2020. I am probably a 0.1, right? 0.1% power user in the world. I can't use Google Gemini. The front end of Google Gemini, it stinks. I said it. I said it, sorry, it's not good. It struggles using, it struggles using any data that you upload to it. It struggles searching Google, right? So you kind of have the worst of both worlds. You upload your documents. It doesn't do a good job. It sometimes just randomly hallucinates. It pulls in its own knowledge from its

internal database when you want it to search the web, you know, Google. Sometimes it doesn't and it falls back on its own internal database. I've done dozens of videos, right? This isn't just some random hot take that I gave you on a whim. I've done dozens of videos. Google Gemini on the front end. It's not good. So there goes my future sponsorship with Google, right? But I want to be honest with you, it's not good. Google AI Studio is amazing, right? The back end way that you can kind of finesse and finagle Google Gemini, the back end is great.

Google Vertex, great, right? Front end Google Gemini, not good at all, right? There are certain use cases, don't get me wrong, if you want it, you know, it's great at creating copy from scratch or rewriting something in a certain tone of voice. That's not what I'm talking about. I'm talking about pushing a large language model and using it for every single thing that it is supposed to be used for. Google Gemini is my least used model, right? At least when we look at the four kind of major models. Chat GPT for me, that's number one. Microsoft co-pilot, I would say, especially if you're Microsoft 365 copilot organization, that's probably your number one. It uses the

GPT, uh, for O technology. So I'll say that's one A and one B, right. Then you have Claude Anthropic. Then you have, I think, a big drop off. And then you get to Google Gemini. But notebook L.M. If I'm being honest, I'm using that more than everyone, everyone else. I'm using it more than chat GPT for that, for these exact reasons that I talked about. Having these audio overviews, amazing. Uh, being able to add sources from anywhere across. them, anywhere across the web changes how you interact with information. It changes how you learn. It changes what your company can do.

Having a grounded model that rarely hallucinates and always gives you cited sources, I'm speechless. And I'm never speechless. But I don't understand Google strategy because like I said, this uses Gemini 1.5 Pro. When you use Gemini 1.5 pro, 1.5 pro on the front end in the chat, I can't stand it. But I can't get enough of notebook L.M. So what is their strategy? I don't know. Hopefully they don't kill it.

Please don't kill notebook LM, Google. If I'm being honest, here's another hot take. This is Google's best product since the search engine that they've built, right? So many things they acquired, right? Like, yeah, they acquired YouTube. They acquired so many of these other companies and turned it into other products. Notebook LM. And I've been using, right, Gmail and, you know, Google Workspace.

Our business has run on Google Workspace since the beginning. Notebook LM is Google's best product that they've built since the search engine. And I will go ahead and say, I don't think it's close. I don't think it's close. In terms of utility and how the world can use this, I don't think it's close. This is probably the most useful and best tool that Google has built since the search engine. And I don't think it's close.

Yeah, the transformer, Michael, good point. So the transformer as part of the GPT technology was developed originally by Google. I still think this is better, right? Because what is the transformer if you don't have a way to properly, and accurately use it. Otherwise, it's just, yes, it's a piece of technology, but it doesn't do anything by itself, right? Everyone else, OpenAI, Microsoft,

everyone else is running with the transformer technology, yes, and it has led the way, right? So, yes, the transformer paved the path to get to Notebook LM, right? But the transformer is not a tool per se. It is a technical part of a recipe that has to be used with other pieces. I'm talking about the end product.

Notebook L.M. Amazing. All right. Let's do this. Let's look live, shall we? We're going to make this quick, all right? Because I know I'm already over on time, so to speak. All right.

So let's go ahead, look live. Live stream audiences. We're having problems. Can you let me know if you're finding this? All right, I'm going to walk our podcast audience through. Okay, so when you go into notebook LM, like I said, you have to sign in with a Google account. It is free to use for right now. the source amount is wild, right? So I'm going to click new notebook. Okay, so when you click

new notebook, it automatically, so when I talk about a grounded model, this is important. It doesn't take you to start chatting, okay? Because I'm going to put something in here. All right, I can't start chatting. So if I act out of this, I literally cannot go click. I literally cannot start chatting. Right? So let me go back and I'll show you. That's what it means for a model to be grounded. It literally does not work unless you upload some information in there. So I'm going to go into paste copied text. All right. And I'm just going to put in this is a test because I want you to see and understand how this model works. So one thing you'll see is it instantly gives a kind of different user interface, which I'm going to talk a little bit about here.

But now I can now start typing or start talking with my data, right? And I can say as an example, you know, tell me about large language models. Here's what I mean when I say something is grounded. So it says the response back says, the sources do not provide any information about large language models. Therefore, I cannot answer your question. The only information provided in the sources is, quote, this is a test. all right so pretty amazing right and this is what it means for a model to be grounded it's in the corner it can only look at the corner it can only look at what we tell it to look at its features

and functions are still driven by Gemini 1.5 pro however it can only work on the data that we give it all right but its features and functionalities its skill sets are still backed so you might think of that and say like, oh, well, this is useless. No, it's absolutely not. This makes it amazing. All right. So now let me give you an example here. So I'm opening up another chat here. Okay. And I finished building this one yesterday. So every single Monday, we do a segment on everyday AI called the AI News That Matters. That's a roughly, you know, 35 to 45 minute podcast where we talk about the most important AI news of the week. We do it every Monday. So if you can only join us once. Maybe that's the day that you do it to keep up with AI. But so I took the transcript

for all of these and it's a lot of words. All right. So right now I have 35 different sources. Okay. And all of my sources are on the left hand side. So this, if you're used to using large language models and chat view, normally you look at the left hand side to find your chats. That's not what this is. On the left hand side is your sources. And like I said, there is only one. one chat. All right. And let me just go ahead and do an example of this. Well, let's actually do the notebook guide. All right. So this is a huge piece. So this creates notes. Okay. So I'm going to go ahead and do this live. And it actually might take a while because like I said, I have 35 different sources. All of those sources have about 10,000 words. Okay. It's complete transcripts from these

podcasts. So I'm not great at math, right? But that's 350,000 words. So for all intent and purposes, that's about 300,000 tokens. All right. So out of the box, most large language models could not even handle this much information. That's a point that I'm trying to show you right off the bat. The amount of information that notebook LM can handle is wild. Okay. So now there's these little notes. Okay, so at the top it says notebook guide. And I can either click add note or I can click plus to add a new source, right? So for all of these, you see my source limit. Right now I'm using 35 out of 50. So you can click copy text, just paste the text in there, click like I said, website, put the URL in, click YouTube, put a YouTube URL. It does have to have. It has to be a public YouTube video

and it has to have captions enabled or I can upload Google slides and it's, going to bring in my Google Drive, et cetera. Okay, so now there's these little notebook guides. Okay, I'm going to click FAQ. All right, and you'll see now this is the note view. Okay, and it's actually gonna probably take a super long time because there's 350,000 words. All right. And then I'm going to click Study Guide.

I'm going to click Table of Contents. I'm going to click Timeline, okay? I'm going to click, let's see, did I get timeline? I'm going to click Briefing Talk. So these are kind of think of these as like default prompts. Okay. So you can upload all your information, whatever it is, and then you can click those, right? So it can create an FAQ on all your information, a study guide, a table of contents, etc.

Timeline briefing doc. So like I said, think of those as prompts. All right. So those are going in the background here. All right. And now they're done. All right. They're finishing one by one.

So this is a lot of information. All right, they're popping up as they're done. So now I have my FAQ, my briefing doc, my study guide, my timeline of main events. So this one's going to be interesting, right? Eight months of AI news. And it's instantly giving me a timeline of main events. So one thing I wish, I wish I could make these notes a little larger. So they're a little small on my screen, but if I, yeah, even when I zoom in.

All right. So here's the timeline of events. This is a lot of information. y'all. And this is actually, if you want to learn about AI, let me know. I'll send you this timeline of events. You know, leave me a comment somewhere that says timeline or DM me timeline. Hey, if you want to catch up in nine months of AI news by reading that, go ahead. All right. Then it also creates these FAQs. Okay, cool. So you might be wondering, all right, so how do I actually use this? Right. So I can now type and say, give me a timeline of LLM developments.

All right. Now, if I put that into an ungrounded model, like even Google Gemini, like ChatGPT, it's going to do two things. It's going to first see, okay, do I have this information in my knowledge base, which for the most part, it will. But normally a model's knowledge base is up to a year or more out of date. That is mistake number one, right? When you are using a large language model that does not have rag, it refers to its own internal database, its own internal training data, which is generally very old.

And then it will make a choice on its own whether it wants to connect to the web to bring in more information, which sometimes will give you great things. Sometimes it will give you bad things. Right. So there we go. Now we have a timeline of large language models based on the sources provided. Okay. So for our live stream audience, you'll see here, but it now has 21 different citations. Okay.

So like I said, if you are querying the web with perplexity or with chat GPT when it uses Browse with Bing, this looks similar. But most models will not do this by default when you are working with your own internal documents that you upload. So like I said, GPTs, Claude Projects, or Google Gems. All right. So now when I click one of these things, so it says, okay, July 2024, meta releases Lama 3.1. All right. So now I can click that. That is citation 4. And so it brings me to my source guide on the left hand side. So I can still read that information on the right hand side of my screen. And then on the left hand side of my screen, it takes me straight to that point. So I can see it gives me the file name or the what I gave my, the name for this one is just July 29th, 2024. So I just gave it the name of the date for my transcript. And then it is highlighted.

So this information is highlighted. So I can instantly go back and remind myself of that information. All right. So this is fantastic. All right. So now I'm going to go in. This is going to take a while. I'm going to generate.

All right. I'm going to generate a podcast conversation. So think of all the different use cases. All right. Think you are. I mean, for students, this is a. no-brainer. This is a great way to actually learn something. Think of for everyone else. So like I said, let's say you work at, you know, at a big logistics company, you work in

marketing. You have a lot of public documents, right, your own company's website, you know, white papers for your industry, information on your competitors, your next big presentation that you're working on, you know, all meeting transcripts with your team. your KPIs that were handed down by you from your boss, right? You can put all of this information in there. Again, you know, always follow your company's AI policy, you know, understand what, you know, Google and everyone else is doing with your data. Got to get all that out of the way, right?

But then when you go in there, then you can still talk to this like you would, normally a large language model and say, hey, here's all the information I gave you. Here's what I want you to help help me with in this chat. You know, go back and first always look at my data. Well, you don't even have to tell it to. right? It's almost like I'm dropping good habits because that's what I'm used to doing when I'm uploading documents into custom GBT's into chats inside of other large language models is reminding them, hey, stay grounded. You don't have to do that. It does it by default here

in Notebook L.M. But think now all of a sudden, your job as a marketing executive at this logistics company is a whole lot easier. You could say, hey, please help me forecast, you know, findings or help me forecast revenue for next course. based on my revenue that I've shared for the last four years, right? It can do anything that a large language model can do, right? Analyze data can help you summarize long things, make them shorter. You can give it bullet points and it can make that information longer, right? Anything that you can do in a text-to-text model, right?

So right now, you know, notebook LM, you can't create videos, you can't create, you know, spreadsheets. So there's certain things that you cannot do. So this isn't an end all be all replacement. But if you are thinking of text to text, if you are thinking of, hey, I need hallucination free outputs from a large language model. To me, you have to use notebook L.N. All right. So this is taking a little minute. We can go back and it should, it should be done. All right. So let's go back into the chat. Okay. So like I said, you have to you have to save these things as a note. because if I refresh this, it is going to be gone, right? And this is probably some information, right?

When I said, you know, please give me a timeline of LLM developments. So I can click this Save to Note. So when I click that Save to Note, you'll see it is now in the upper left hand corner. It says saved response. So I can click on this. And then I can even from there click, help me understand, critique, suggest related ideas. Right. So even when you click on one.

note, you get some new information. All right. That's important to know. I can also change the name of it as well. So now it is one of my saved sources on the left hand side. So I can go in, rename it, etc. All right. Another thing that you can do, and I'm not going to do it now because I'm in the middle of generating this conversation. Yeah, I figure this conversation is going to take a while. And I'm going to end by talking about the future of this product and a couple of other things. But on the left hand side, when you upload all your sources, you might not want to use them all at once. So you can quickly click the select all sources to deselect all. So maybe as an example, you want to use notebook LM for different purposes, but all within your data.

So you can quickly toggle sources on and off. So maybe first you want to go through your meeting transcripts. Right. So you can toggle everything off aside from those meeting transcripts. I hope FYI, because I know people from Notebook L.M are reading this and I know that they're putting out now tons of great updates. We need folders, FYI, right? I need to be able to have a folder with all my meeting transcripts and then to quickly be able to toggle that on and off. I need something that says, you know, competitor information with all the competitor information in there. I need all my internal information, you know, goals. Like we need folder structures because 50 sources,

is a lot. It's great, but we need folders. Anyways, you can toggle sources on or off. So then when you ask notebook LM something, you can make sure it's only looking at certain sources. That's like, I don't know, super grounding, right? Like, hey, don't look at all my competitor information. Don't look at all the, you know, information that I brought in from the web. Only look at these certain sources that I have toggled on or off. So I have 35. Right now I'm having a conversation and I'm generating an AI overview, the DeepDive AI podcast off all 35 of those pieces of information, which is why it took a while to get done and it's finally done here. All right. So now let's go ahead and listen. All right, let's go ahead and listen. Hopefully a live stream audience,

let me know if you can hear this. I'm not going to play the whole thing. But I do want you all to hear what this sounds like. All right. So let me know if you can hear this. This is the audio overview, the AI podcast deep dive that was generated based off of my content. All right, let's listen to maybe 30 seconds. It's incredible, isn't it? The speed at which AI is evolving. One minute we're trying to grasp the concept of AI generated art and the next, bam, we're talking about AI that can create entire video sequences from a simple text prompt. It's a lot to keep up with, even for those of us who live and breathe this stuff. And let's be honest, who has the time to wade through

endless articles and podcasts? That's why we do these deep dives to cut through the noise and bring you the most impactful AI developments. I agree, it's crucial to discern the hype from the truly transformative. Not everything labeled AI is created equal. Exactly. Speaking of transformative, let's talk about OpenAI's SORA. This new AI model is causing quite a stir. Imagine typing in a detailed scene description, let's say.

A cat in a detective hat chases a robot. All right, so hopefully y'all could hear that AI overview. I know we're having some technical issues for the live stream, but if you couldn't, It was two hosts that sounded fairly human, right? That didn't sound like a normal text to speech voice. Those are the words that we've heard you say. Yeah, it's because it's all my data, right? A couple things to know about these, the AI overview.

It will always be there now. You don't have to save it. It's not like chat history where it's going to go away. The thing I like is it has playback speed. So I always listen to it at 1-8 or 2-0, so you can hear that. think it sounds better at like one five. So listen to how it sounds in one five. But mouse through a bustling marketplace. And then boom, SORA generates a full minute of high definition video complete with realistic textures,

lighting. It's, I think that's just how I always talk. Just like excited and fast. You know, I generally listen on one five or one eight. So it has built in speed control with which is huge. And then you can also download that podcast as well. So again, a great thing to do. If you have a trip, you're going to be on the plane. You're in the car. You want to learn about things. I mean, number one, I hope you just continue to tune into everyday AI and learn about it here with us. But this is another great option. Maybe you need to learn about something very specific. Maybe you have a big meeting coming up and you want to prep for it. Y'all, I cannot think of any way that you should not be using this. That's why, like, when I talked about Andre Capathy, right,

Open AI co-founder largely, but largely referred to as one of the smartest people in the history of artificial intelligence, he's talking about this. Like he is using this all the time. I am using it all the time. You should be using it all the time. I'm not sponsored by Google. They're not, you know, advertising on the everyday AI podcast anytime soon after I, you know, said, hey, Gemini's front and stinks.

But notebook LM is probably one of the best tools I've ever used in my life. And I've used thousands with an S of AI tools. It is one of the best. All right. So that is a quick overview of inside the platform. Like I said, all of your notebooks are on the left. You can upload up to 50 sources that have 50,000 words, 25 million words. My gosh.

So dumping information of all kinds, YouTube videos, if there's captions, websites, copy text, that you can just copy and paste anything in there, Google slides, Google Docs, wild. Absolutely wild. All right. So let's wrap this up. This turned into a long one. So, well, what's next? What's next for this tool?

Is it going to go away? Is Google going to charge for it? I don't know. I don't think it's going to go away because I think this thing is wildly popular. If I'm being honest, I think this has been better received than their Gemini model. We remember when Google first announced Gemini, they did some, let's say some not-so-great marketing. They kind of lied a little bit, right? And they showed us these marketing videos with Gemini showing it could do all these things that in actuality it couldn't, right?

It showed us as an example responding to the paper rock scissors game live. It couldn't do that. In the research paper, it said, oh, we actually took a ton of pictures of that video of paper rock scissors. And then we did a bunch of prompt engineering. And then we got the output that we needed. And then we used that. So yeah, we, you know, they kind of said, okay, maybe Gemmon, I can't do all these things. And we had to do a bunch of, you know, duct tape under the hood.

No book L.M is not like that. No book L.M is frigging impressive. It is great. I am using this nonstop. All right. So what's next for it? Are they going to start charging for it? Probably, if I'm being honest,

I'm using Google Gemini on the front end. I don't know, 10 minutes a week because I don't think it's very good. I'm using notebook LM, sometimes more than an hour a day. And I think that's going to become the norm. A free tool with all of this information you can upload, I mean, they're going to have to start charging for it. They're going to have to, right? So what's next?

Okay, so shout out here to testing catalog. All right. So this is a Twitter account. They also have a website, good stuff. Go check them out testing catalog. So they just shared something yesterday, you know, kind of some leaks or some behind the scenes, that they're going to, notebook LM is going to be creating bots. Okay.

So what I hope this means ultimately is that your notebooks will be able to talk to each other or your different notes within notebook LM can be able to talk to each other. All right. So that's a pretty good feature that I think is coming. And hopefully, like I said, hopefully what that means is your notebooks will be able to talk to each other. That will be huge. All right. Because there we have essentially agentic rag, all right? Not to throw out a bunch of buzzwords, but, you know, a grounding.

bot, right? So it's grounded in RAG, retrieval, log, meta generation, your data only. And then they can talk to each other. So this agentic nature, that would be amazing, right? Also, Reza Martin, you know, trying to get, Reza, let's get you on the show. So she's the product lead at Notebook LM. And she did share something in response to this, this testing catalog kind of blog post slash tweet yesterday. And she did say this, custom chatbots. I have a lot to say. This is pretty widely used internally at Google. And literally every day someone pings me to say, this has 10xed our team's productivity. Not joking. In the hacks, you're still looking at the old version. So she's saying, essentially, we're all looking at the old version. There's a newer version with more capabilities.

And so she says, not joking. And the hacks, you're still looking at the old version. So I'm excited for what you all think when the new version launches. So yeah, apparently there is going to be a sizable update coming to Notebook LM pretty quickly. All right. That was a lot. We were all over the place. I did see a couple of questions.

So I'm going to try to answer them quickly. Marie asking, do the podcast hosts voices change every time you use it? No, it is the exact same to host every time. I believe they only speak English and you can't really get them to say anything. So people have found out there are ways you can kind of hack it. So in the notes that you upload or in the sources, you can kind of tell it, oh, talk about this. Right. And people did this, you know, existential, you know, oh, these two AI hosts are trapped.

Right. So you can just put in information in your source document that kind of steers them, but they're still going to take their own spin on it. So you can't just give them a script or, you know, anything like that. But you can steer them in certain directions. like you can. You can steer any model with proper prompt engineering. So no, the host, you cannot change their voices. It generates once. That's what you get. You can download it, play it, speed it up. All right. Cecilia asking, so once it has your rag or data and has chewed on it,

will it extrapolate and go outside the internet to the internet for comparisons or it is only internal digestion. As of now, it is grounded. So it does not look on. the internet. As far as I know, I've done testing, it cannot look on the internet, right? So, which I think is generally when I'm speaking about sending large language models to the internet, it's something you need to do. It's because these models, even that offers some sort of quote unquote rag retrieval augmentate generation, right? So even when you can upload files into GPs, you can upload files into GPDs, you can upload files into Google gems, you can upload files into co-pilot GPTs, sometimes they still struggle to stay grounded and will instead refer to their

own internal knowledge base or go off on the web, right? So that's why even using these kind of other options that I just talked about, sometimes you still need web access. I don't think you need web access with notebook LN because you can literally just bring in those 10 or 20 different websites. And there you go, right? Amazing, the control that you have. Jason asking, can you upload source code? If you're asking, can you upload coding, right? Can you upload your Python projects and CSS and HTML? Yes.

I did some testing on it. I uploaded some code. You know, I said, hey, can you change this code for me? Yes. So most of the features and functionalities that Gemini 1.5 can do most. It can do as well. So yes, it can summarize long documents, make it shorter, take bullet points, make it longer. It can look at all your different documents.

if those documents are coding. Yes, Gemini is actually, even the front end Gemini chat is great at coding, FYI. So yes, it can, Notebook LM can work with your source code as well. Oh, that was a long one, y'all. Here's the thing. I had a whole section that I knew I wasn't going to have time to get to. I put together seven additional helpful tips

on how I'm using Notebook LM. I wanted to give you all the basics, and I knew if I dorked out, this would turn into a three, hour like Lex podcast. I can't do this. This was an hour. Um, so please always feel free to listen to this in 2x. I ramble. It's unedited, unscripted. I want to bring you, what I like to say is the realest thing in artificial intelligence, but there was more I couldn't get to at all these great tips and tricks. So maybe I'll turn them into a podcast or a video in the future, but I have them

ready to go. So please repost this, um, you know, on LinkedIn or retweet it, you know, on Twitter. If you do, I will send you that list, right? So whether it's a list or a video, I have a bunch more information that I was going to take me way longer to explain. And I knew this was already going to be long enough. So if you're listening on the podcast, make sure to check out your show notes. I always leave a link to our LinkedIn live stream, even though it wasn't working today. So go repost that. And I will share those seven additional tips with you.

All right. So I'd say some tricky ones, some ones that really take. many, many hours of using the tool to understand and to really get the most out of of notebook.LM. So make sure to share that. Also make sure to go to your everyday AI.com. Sign up for the free daily newsletter. And join us back tomorrow and every day for more. Everyday AI. Thanks y'all. Meet Firefly AI assistant. Now live in Adobe Firefly, the Allman One Creative AI Studio. Just describe what you want to create in your own words and the assistant handles the rest, orchestrating multi-step workflows across Adobe.

Adobe Creative Cloud apps, including Photoshop, Premiere Express, and more in one conversational interface. You direct the outcome while the assistant accelerates execution. Stand control with the ability to step in and refine at any time. See it today at firefly.adobie.com. And that's a wrap for today's edition of Everyday AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going.

For a little more AI magic, visit your everyday AI.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers and we'll see you next time.