If you’ve listened to the podcast for a while, you might have heard our ElevenLabs-powered AI co-host Charlie a few times. Text-to-speech has made amazing progress in the last 18 months, with OpenAI’s Advanced Voice Mode (aka “Her”) as a sneak peek of the future of AI interactions (see our “Building AGI in Real Time” recap). Yet, we had yet to see a real killer app for AI voice (not counting music).
Today’s guests, Raiza Martin and Usama Bin Shafqat, are the lead PM and AI engineer behind the NotebookLM feature flag that gave us the first viral AI voice experience, the “Deep Dive” podcast:
The idea behind the “Audio Overviews” feature is simple: take a bunch of documents, websites, YouTube videos, etc, and generate a podcast out of them. This was one of the first demos that people built with voice models + RAG + GPT models, but it was always a glorified speech-to-text. Raiza and Usama took a very different approach:
* Make it conversational: when you listen to a NotebookLM audio there are a ton of micro-interjections (Steven Johnson calls them disfluencies) like “Oh really?” or “Totally”, as well as pauses and “uh…”, like you would expect in a real conversation. These are not generated by the LLM in the transcript, but they are built into the the audio model. See ~28:00 in the pod for more details.
* Listeners love tension: if two people are always in agreement on everything, it’s not super interesting. They tuned the model to generate flowing conversations that mirror the tone and rhythm of human speech. They did not confirm this, but many suspect the 2 year old SoundStorm paper is related to this model.
* Generating new insights: because the hosts’ goal is not to summarize, but to entertain, it comes up with funny metaphors and comparisons that actually help expand on the content rather than just paraphrasing like most models do. We have had listeners make podcasts out of our podcasts, like this one.
This is different than your average SOTA-chasing, MMLU-driven model buildooor. Putting product and AI engineering in the same room, having them build evals together, and understanding what the goal is lets you get these unique results.
The 5 rules for AI PMs
We always focus on AI Engineers, but this episode had a ton of AI PM nuggets as well, which we wanted to collect as NotebookLM is one of the most successful products in the AI space:
1. Less is more: the first version of the product had 0 customization options. All you could do is give it source documents, and then press a button to generate. Most users don’t know what “temperature” or “top-k” are, so you’re often taking the magic away by adding more options in the UI. Since recording they added a few, like a system prompt, but those were features that users were “hacking in”, as Simon Willison highlighted in his blog post.
2. Use Real-Time Feedback: they built a community of 65,000 users on Discord that is constantly reporting issues and giving feedback; sometimes they noticed server downtime even before the Google internal monitoring did. Getting real time pings > aggregating user data when doing initial iterations.
3. Embrace Non-Determinism: AI outputs variability is a feature, not a bug. Rather than limiting the outputs from the get-go, build toggles that you can turn on/off with feature flags as the feedback starts to roll in.
4. Curate with Taste: if you try your product and it sucks, you don’t need more data to confirm it. Just scrap that and iterate again. This is even easier for a product like this; if you start listening to one of the podcasts and turn it off after 10 seconds, it’s never a good sign.
5. Stay Hands-On: It’s hard to build taste if you don’t experiment. Trying out all your competitors products as well as unrelated tools really helps you understand what users are seeing in market, and how to improve on it.
Chapters
00:00 Introductions01:39 From Project Tailwind to NotebookLM09:25 Learning from 65,000 Discord members12:15 How NotebookLM works18:00 Working with Steven Johnson23:00 How to prioritize features25:13 Structuring the data pipelines29:50 How to eval34:34 Steering the podcast outputs37:51 Defining speakers personalities39:04 How do you make audio engaging?45:47 Humor is AGI51:38 Designing for non-determinism53:35 API when?55:05 Multilingual support and dialect considerations57:50 Managing system prompts and feature requests01:00:58 Future of NotebookLM01:04:59 Podcasts for your codebase01:07:16 Plans for real-time chat01:08:27 Wrap up
Show Notes
* Histories of Mysteries by Andrej Karpathy
* Area 120
Transcript
NotebookLM [00:00:00]: Hey everyone, we're here today as guests on Latent Space. It's great to be here, I'm a long time listener and fan, they've had some great guests on this show before. Yeah, what an honor to have us, the hosts of another podcast, join as guests. I mean a huge thank you to Swyx and Alessio for the invite, thanks for having us on the show. Yeah really, it seems like they brought us here to talk a little bit about our show, our podcast. Yeah, I mean we've had lots of listeners ourselves, listeners at Deep Dive. Oh yeah, we've made a ton of audio overviews since we launched and we're learning a lot. There's probably a lot we can share around what we're building next, huh? Yeah, we'll share a little bit at least. The short version is we'll keep learning and getting better for you. We're glad you're along for the ride. So yeah, keep listening. Keep listening and stay curious. We promise to keep diving deep and bringing you even better options in the future. Stay curious.
Alessio [00:00:52]: Hey everyone, welcome to the Latent Space Podcast. This is Alessio, partner and CTO at Residence at Decibel Partners. And I'm joined by my co-host, Swyx, founder of Smol.ai.
Swyx [00:01:01]: Hey, and today we're back in the studio with our special guest, Raiza Martin. And Raiza, I forgot to get your last name, Shafqat.
Raiza [00:01:10]: Yes.
Swyx [00:01:10]: Okay, welcome.
Raiza [00:01:12]: Hello, thank you for having us.
Swyx [00:01:14]: So AI podcasters meet human podcasters, always fun. Congrats on the success of Notebook LM. I mean, how does it feel?
Raiza [00:01:22]: It's been a lot of fun. A lot of it, honestly, was unexpected. But my favorite part is really listening to the audio overviews that people have been making.
Swyx [00:01:29]: Maybe we should do a little bit of intros and tell the story. You know, what is your path into the sort of Google AI org? Or maybe, actually, I don't even know what org you guys are in.
Raiza [00:01:39]: I can start. My name is Raisa. I lead the Notebook LM team inside of Google Labs. So specifically, that's the org that we're in. It's called Google Labs. It's only about two years old. And our whole mandate is really to build AI products. That's it. We work super closely with DeepMind. Our entire thing is just, like, try a bunch of things and see what's landing with users. And the background that I have is, really, I worked in payments before this, and I worked in ads right before, and then startups. I tell people, like, at every time that I changed orgs, I actually almost quit Google. Like, specifically, like, in between ads and payments, I was like, all right, I can't do this. Like, this is, like, super hard. I was like, it's not for me. I'm, like, a very zero-to-one person. But then I was like, okay, I'll try. I'll interview with other teams. And when I interviewed in payments, I was like, oh, these people are really cool. I don't know if I'm, like, a super good fit with this space, but I'll try it because the people are cool. And then I really enjoyed that, and then I worked on, like, zero-to-one features inside of payments, and I had a lot of fun. But then the time came again where I was like, oh, I don't know. It's like, it's time to leave. It's time to start my own thing. But then I interviewed inside of Google Labs, and I was like, oh, darn. Like, there's definitely, like—
Alessio [00:02:48]: They got you again.
Raiza [00:02:49]: They got me again. And so now I've been here for two years, and I'm happy that I stayed because especially with, you know, the recent success of Notebook LM, I'm like, dang, we did it. I actually got to do it. So that was really cool.
Usama [00:03:02]: Kind of similar, honestly. I was at a big team at Google. We do sort of the data center supply chain planning stuff. Google has, like, the largest sort of footprint. Obviously, there's a lot of management stuff to do there. But then there was this thing called Area 120 at Google, which does not exist anymore. But I sort of wanted to do, like, more zero-to-one building and landed a role there. We were trying to build, like, a creator commerce platform called Kaya. It launched briefly a couple years ago. But then Area 120 sort of transitioned and morphed into Labs. And, like, over the last few years, like, the focus just got a lot clearer. Like, we were trying to build new AI products and do it in the wild and sort of co-create and all of that. So, you know, we've just been trying a bunch of different things. And this one really landed, which has felt pretty phenomenal. Really, really landed.
Swyx [00:03:53]: Let's talk about the brief history of Notebook LM. You had a tweet, which is very helpful for doing research. May 2023, during Google I.O., you announced Project Tailwind.
Raiza [00:04:03]: Yeah.
Swyx [00:04:03]: So today is October 2024. So you joined October 2022?
Raiza [00:04:09]: Actually, I used to lead AI Test Kitchen. And this was actually, I think, not I.O. 2023. I.O. 2022 is when we launched AI Test Kitchen, or announced it. And I don't know if you remember it.
Swyx [00:04:23]: That's how you, like, had the basic prototype for Gemini.
Raiza [00:04:26]: Yes, yes, exactly. Lambda.
Swyx [00:04:28]: Gave beta access to people.
Raiza [00:04:29]: Yeah, yeah, yeah. And I remember, I was like, wow, this is crazy. We're going to launch an LLM into the wild. And that was the first project that I was working on at Google. But at the same time, my manager at the time, Josh, he was like, hey, I want you to really think about, like, what real products would we build that are not just demos of the technology? That was in October of 2022. I was sitting next to an engineer that was working on a project called Talk to Small Corpus. His name was Adam. And the idea of Talk to Small Corpus is basically using LLM to talk to your data. And at the time, I was like, wait, there's some, like, really practical things that you can build here. And just a little bit of background, like, I was an adult learner. Like, I went to college while I was working a full-time job. And the first thing I thought was, like, this would have really helped me with my studying, right? Like, if I could just, like, talk to a textbook, especially, like, when I was tired after work, that would have been huge. We took a lot of, like, the Talk to Small Corpus prototypes, and I showed it to a lot of, like, college students, particularly, like, adult learners. They were like, yes, like, I get it, right? Like, I didn't even have to explain it to them. And we just continued to iterate the prototype from there to the point where we actually got a slot as part of the I.O. demo in 23.
Swyx [00:05:42]: And Corpus, was it a textbook? Oh, my gosh.
Raiza [00:05:45]: Yeah. It's funny. Actually, when he explained the project to me, he was like, talk to Small Corpus. It was like, talk to a small corpse?
Swyx [00:05:51]: Yeah, nobody says Corpus.
Raiza [00:06:00]: It was like, a small corpse? This is not AI. Yeah, yeah. And it really was just, like, a way for us to describe the amount of data that we thought, like, it could be good for.
Swyx [00:06:02]: Yeah, but even then, you're still, like, doing rag stuff. Because, you know, the context length back then was probably, like, 2K, 4K.
Raiza [00:06:08]: Yeah, it was basically rag.
Raiza [00:06:09]: That was essentially what it was.
Raiza [00:06:10]: And I remember, I was like, we were building the prototypes. And at the same time, I think, like, the rest of the world was. Right? We were seeing all of these, like, chat with PDF stuff come up. And I was like, come on, we gotta go. Like, we have to, like, push this out into the world. I think if there was anything, I wish we would have launched sooner because I wanted to learn faster. But I think, like, we netted out pretty well.
Alessio [00:06:30]: Was the initial product just text-to-speech? Or were you also doing kind of, like, synthesizing of the content, refining it? Or were you just helping people read through it?
Raiza [00:06:40]: Before we did the I.O. announcement in 23, we'd already done a lot of studies. And one of the first things that I realized was the first thing anybody ever typed was, summarize the thing. Right?
Raiza [00:06:53]: Summarize the document.
Raiza [00:06:54]: And it was, like, half like a test and half just like, oh, I know the content. I want to see how well it does this. So it was part of the first thing that we launched. It was called Project Tailwind back then. It was just Q&A, so you could chat with the doc just through text, and it would automatically generate a summary as well. I'm not sure if we had it back then.
Raiza [00:07:12]: I think we did.
Raiza [00:07:12]: It would also generate the key topics in your document, and it could support up to, like, 10 documents. So it wasn't just, like, a single doc.
Alessio [00:07:20]: And then the I.O. demo went well, I guess. And then what was the discussion from there to where we are today? Is there any, maybe, intermediate step of the product that people missed between this was launch or?
Raiza [00:07:33]: It was interesting because every step of the way, I think we hit, like, some pretty critical milestones. So I think from the initial demo, I think there was so much excitement of, like, wow, what is this thing that Google is launching? And so we capitalized on that. We built the wait list. That's actually when we also launched the Discord server, which has been huge for us because for us in particular, one of the things that I really wanted to do was to be able to launch features and get feedback ASAP. Like, the moment somebody tries it, like, I want to hear what they think right now, and I want to ask follow-up questions. And the Discord has just been so great for that. But then we basically took the feedback from I.O., we continued to refine the product.
Raiza [00:08:12]: So we added more features.
Raiza [00:08:13]: We added sort of, like, the ability to save notes, write notes. We generate follow-up questions. So there's a bunch of stuff in the product that shows, like, a lot of that research. But it was really the rolling out of things. Like, we removed the wait list, so rolled out to all of the United States. We rolled out to over 200 countries and territories. We started supporting more languages, both in the UI and, like, the actual source stuff. We experienced, like, in terms of milestones, there was, like, an explosion of, like, users in Japan. This was super interesting in terms of just, like, unexpected. Like, people would write to us and they would be like, this is amazing. I have to read all of these rules in English, but I can chat in Japanese. It's like, oh, wow. That's true, right? Like, with LLMs, you kind of get this natural, it translates the content for you. And you can ask in your sort of preferred mode. And I think that's not just, like, a language thing, too. I think there's, like, I do this test with Wealth of Nations all the time because it's, like, a pretty complicated text to read. The Evan Smith classic.
Swyx [00:09:11]: It's, like, 400 pages or something.
Raiza [00:09:12]: Yeah. But I like this test because I'm, like, asking, like, Normie, you know, plain speak. And then it summarizes really well for me. It sort of adapts to my tone.
Swyx [00:09:22]: Very capitalist.
Raiza [00:09:25]: Very on brand.
Swyx [00:09:25]: I just checked in on a Notebook LM Discord. 65,000 people. Yeah.
Raiza [00:09:29]: Crazy.
Swyx [00:09:29]: Just, like, for one project within Google. It's not, like, it's not labs. It's just Notebook LM.
Raiza [00:09:35]: Just Notebook LM.
Swyx [00:09:36]: What do you learn from the community?
Raiza [00:09:39]: I think that the Discord is really great for hearing about a couple of things.
Raiza [00:09:43]: One, when things are going wrong. I think, honestly, like, our fastest way that we've been able to find out if, like, the servers are down or there's just an influx of people being, like, it says
Raiza [00:09:53]: system unable to answer.
Raiza [00:09:54]: Anybody else getting this?
Raiza [00:09:56]: And I'm, like, all right, let's go.
Raiza [00:09:58]: And it actually catches it a lot faster than, like, our own monitoring does.
Raiza [00:10:01]: It's, like, that's been really cool. So, thank you.
Swyx [00:10:03]: Canceled eat a dog.
Raiza [00:10:05]: So, thank you to everybody. Please keep reporting it. I think the second thing is really the use cases.
Raiza [00:10:10]: I think when we put it out there, I was, like, hey, I have a hunch of how people will use it, but, like, to actually hear about, you know, not just the context of, like, the use of Notebook LM, but, like, what is this person's life like? Why do they care about using this tool?
Raiza [00:10:23]: Especially people who actually have trouble using it, but they keep pushing.
Raiza [00:10:27]: Like, that's just so critical to understand what was so motivating, right?
Raiza [00:10:31]: Like, what was your problem that was, like, so worth solving? So, that's, like, a second thing.
Raiza [00:10:34]: The third thing is also just hearing sort of, like, when we have wins and when we don't have wins because there's actually a lot of functionality where I'm, like, hmm, I
Raiza [00:10:42]: don't know if that landed super well or if that was actually super critical.
Raiza [00:10:45]: As part of having this sort of small project, right, I want to be able to unlaunch things, too. So, it's not just about just, like, rolling things out and testing it and being, like, wow, now we have, like, 99 features. Like, hopefully we get to a place where it's, like, there's just a really strong core feature set and the things that aren't as great, we can just unlaunch.
Swyx [00:11:02]: What have you unlaunched? I have to ask.
Raiza [00:11:04]: I'm in the process of unlaunching some stuff, but, for example, we had this idea that you could highlight the text in your source passage and then you could transform it. And nobody was really using it and it was, like, a very complicated piece of our architecture and it's very hard to continue supporting it in the context of new features. So, we were, like, okay, let's do a 50-50 sunset of this thing and see if anybody complains.
Raiza [00:11:28]: And so far, nobody has.
Swyx [00:11:29]: Is there, like, a feature flagging paradigm inside of your architecture that lets you feature flag these things easily?
Raiza [00:11:36]: Yes, and actually...
Raiza [00:11:37]: What is it called?
Swyx [00:11:38]: Like, I love feature flagging.
Raiza [00:11:40]: You mean, like, in terms of just, like, being able to expose things to users?
Swyx [00:11:42]: Yeah, as a PM. Like, this is your number one tool, right?
Raiza [00:11:44]: Yeah, yeah.
Swyx [00:11:45]: Let's try this out. All right, if it works, roll it out. If it doesn't, roll it back, you know?
Raiza [00:11:49]: Yeah, I mean, we just run Mendel experiments for the most part. And, actually, I don't know if you saw it, but on Twitter, somebody was able to get around our flags and they enabled all the experiments.
Raiza [00:11:58]: They were, like, check out what the Notebook LM team is cooking.
Raiza [00:12:02]: I was, like, oh!
Raiza [00:12:03]: And I was at lunch with the rest of the team and I was, like, I was eating. I was, like, guys, guys, Magic Draft League!
Raiza [00:12:10]: They were, like, oh, no!
Raiza [00:12:12]: I was, like, okay, just finish eating and then let's go figure out what to do.
Raiza [00:12:15]: Yeah.
Alessio [00:12:15]: I think a post-mortem would be fun, but I don't think we need to do it on the podcast now. Can we just talk about what's behind the magic? So, I think everybody has questions, hypotheses about what models power it. I know you might not be able to share everything, but can you just get people very basic? How do you take the data and put it in the model? What text model you use? What's the text-to-speech kind of, like, jump between the two? Sure.
Raiza [00:12:42]: Yeah.
Raiza [00:12:42]: I was going to say, SRaiza, he manually does all the podcasts.
Raiza [00:12:46]: Oh, thank you.
Usama [00:12:46]: Really fast. You're very fast, yeah.
Raiza [00:12:48]: Both of the voices at once.
Usama [00:12:51]: Voice actor.
Raiza [00:12:52]: Good, good.
Usama [00:12:52]: Yeah, so, for a bit of background, we were building this thing sort of outside Notebook LM to begin with. Like, just the idea is, like, content transformation, right? Like, we can do different modalities. Like, everyone knows that. Everyone's been poking at it. But, like, how do you make it really useful? And, like, one of the ways we thought was, like, okay, like, you maybe, like, you know, people learn better when they're hearing things. But TTS exists, and you can, like, narrate whatever's on screen. But you want to absorb it the same way. So, like, that's where we sort of started out into the realm of, like, maybe we try, like, you know, two people are having a conversation kind of format. We didn't actually start out thinking this would live in Notebook, right? Like, Notebook was sort of, we built this demo out independently, tried out, like, a few different sort of sources. The main idea was, like, go from some sort of sources and transform it into a listenable, engaging audio format. And then through that process, we, like, unlocked a bunch more sort of learnings. Like, for example, in a sense, like, you're not prompting the model as much because, like, the information density is getting unrolled by the model prompting itself, in a sense. Because there's two speakers, and they're both technically, like, AI personas, right? That have different angles of looking at things. And, like, they'll have a discussion about it. And that sort of, we realized that's kind of what was making it riveting, in a sense. Like, you care about what comes next, even if you've read the material already. Because, like, people say they get new insights on their own journals or books or whatever. Like, anything that they've written themselves. So, yeah, from a modeling perspective, like, it's, like Reiza said earlier, like, we work with the DeepMind audio folks pretty closely. So, they're always cooking up new techniques to, like, get better, more human-like audio. And then Gemini 1.5 is really, really good at absorbing long context. So, we sort of, like, generally put those things together in a way that we could reliably produce the audio.
Raiza [00:14:52]: I would add, like, there's something really nuanced, I think, about sort of the evolution of, like, the utility of text-to-speech. Where, if it's just reading an actual text response, and I've done this several times. I do it all the time with, like, reading my text messages. Or, like, sometimes I'm trying to read, like, a really dense paper, but I'm trying to do actual work. I'll have it, like, read out the screen. There is something really robotic about it that is not engaging. And it's really hard to consume content in that way. And it's never been really effective. Like, particularly for me, where I'm, like, hey, it's actually just, like, it's fine for, like, short stuff. Like, texting, but even that, it's, like, not that great. So, I think the frontier of experimentation here was really thinking about there is a transform that needs to happen in between whatever.
Raiza [00:15:38]: Here's, like, my resume, right?
Raiza [00:15:39]: Or here's, like, a 100-page slide deck or something. There is a transform that needs to happen that is inherently editorial. And I think this is where, like, that two-person persona, right, dialogue model, they have takes on the material that you've presented. That's where it really sort of, like, brings the content to life in a way that's, like, not robotic. And I think that's, like, where the magic is, is, like, you don't actually know what's going to happen when you press generate.
Raiza [00:16:08]: You know, for better or for worse.
Raiza [00:16:09]: Like, to the extent that, like, people are, like, no, I actually want it to be more predictable now. Like, I want to be able to tell them. But I think that initial, like, wow was because you didn't know, right? When you upload your resume, what's it about to say about you? And I think I've seen enough of these where I'm, like, oh, it gave you good vibes, right? Like, you knew it was going to say, like, something really cool. As we start to shape this product, I think we want to try to preserve as much of that wow as much as we can. Because I do think, like, exposing, like, all the knobs and, like, the dials, like, we've been thinking about this a lot. It's like, hey, is that, like, the actual thing?
Raiza [00:16:43]: Is that the thing that people really want?
Alessio [00:16:45]: Have you found differences in having one model just generate the conversation and then using text-to-speech to kind of fake two people? Or, like, are you actually using two different kind of system prompts to, like, have a conversation step-by-step? I'm always curious, like, if persona system prompts make a big difference? Or, like, you just put in one prompt and then you just let it run?
Usama [00:17:05]: I guess, like, generally we use a lot of inference, as you can tell with, like, the spinning thing takes a while. So, yeah, there's definitely, like, a bunch of different things happening under the hood. We've tried both approaches and they have their, sort of, drawbacks and benefits. I think that that idea of, like, questioning, like, the two different personas, like, persists throughout, like, whatever approach we try. It's like, there's a bit of, like, imperfection in there. Like, we had to really lean into the fact that, like, to build something that's engaging, like, it needs to be somewhat human and it needs to be just not a chatbot. Like, that was sort of, like, what we need to diverge from. It's like, you know, most chatbots will just narrate the same kind of answer, like, given the same sources, for the most part, which is ridiculous. So, yeah, there's, like, experimentation there under the hood, like, with the model to, like, make sure that it's spitting out, like, different takes and different personas and different, sort of, prompting each other is, like, a good analogy, I guess.
Swyx [00:18:00]: Yeah, I think Steven Johnson, I think he's on your team. I don't know what his role is. He seems like chief dreamer, writer.
Raiza [00:18:08]: Yeah, I mean, I can comment on Steven. So, Steven joined, actually, in the very early days, I think before it was even a fully funded project. And I remember when he joined, I was like, Steven Johnson's going to be on my team? You know, and for folks who don't know him, Steven is a New York Times bestselling author of, like, 14 books. He has a PBS show. He's, like, incredibly smart, just, like, a true, sort of, celebrity by himself. And then he joined Google, and he was like, I want to come here, and I want to build the thing that I've always dreamed of, which is a tool to help me think. I was like, a what? Like, a tool to help you think? I was like, what do you need help with? Like, you seem to be doing great on your own. And, you know, he would describe this to me, and I would watch his flow. And aside from, like, providing a lot of inspiration, to be honest, like, when I watched Steven work, I was like, oh, nobody works like this, right? Like, this is what makes him special. Like, he is such a dedicated, like, researcher and journalist, and he's so thorough, he's so smart. And then I had this realization of, like, maybe Steven is the product. Maybe the work is to take Steven's expertise and bring it to, like, everyday people that could really benefit from this. Like, just watching him work, I was like, oh, I could definitely use, like, a mini-Steven, like, doing work for me. Like, that would make me a better PM. And then I thought very quickly about, like, the adjacent roles that could use sort of this, like, research and analysis tool. And so, aside from being, you know, chief dreamer, Steven also represents, like, a super workflow that I think all of us, like, if we had access to a tool like it, would just inherently, like, make us better.
Swyx [00:19:46]: Did you make him express his thoughts while he worked, or you just silently watched him, or how does this work?
Raiza [00:19:52]: Oh, now you're making me admit it. But yes, I did just silently watch him.
Swyx [00:19:57]: This is a part of the PM toolkit, right? They give user interviews and all that.
Raiza [00:20:00]: Yeah, I mean, I did interview him, but I noticed, like, if I interviewed him, it was different than if I just watched him. And I did the same thing with students all the time. Like, I followed a lot of students around. I watched them study. I would ask them, like, oh, how do you feel now, right?
Raiza [00:20:15]: Or why did you do that? Like, what made you do that, actually?
Raiza [00:20:18]: Or why are you upset about, like, this particular thing? Why are you cranky about this particular topic? And it was very similar, I think, for Steven, especially because he was describing, he was in the middle of writing a book. And he would describe, like, oh, you know, here's how I research things, and here's how I keep my notes. Oh, and here's how I do it. And it was really, he was doing this sort of, like, self-questioning, right? Like, now we talk about, like, chain of, you know, reasoning or thought, reflection.
Raiza [00:20:44]: And I was like, oh, he's the OG.
Raiza [00:20:46]: Like, I watched him do it in real time. I was like, that's, like, L-O-M right there. And to be able to bring sort of that expertise in a way that was, like, you know, maybe, like, costly inference-wise, but really have, like, that ability inside of a tool that was, like, for starters, free inside of NotebookLM, it was good to learn whether or not people really did find use out of it.
Swyx [00:21:05]: So did he just commit to using NotebookLM for everything, or did you just model his existing workflow?
Raiza [00:21:12]: Both, right?
Raiza [00:21:12]: Like, in the beginning, there was no product for him to use. And so he just kept describing the thing that he wanted. And then eventually, like, we started building the thing. And then I would start watching him use it. One of the things that I love about Steven is he uses the product in ways where it kind of does it, but doesn't quite. Like, he's always using it at, like, the absolute max limit of this thing. But the way that he describes it is so full of promise, where he's like, I can see it going here. And all I have to do is sort of, like, meet him there and sort of pressure test whether or not, you know, everyday people want it. And we just have to build it.
Swyx [00:21:47]: I would say OpenAI has a pretty similar person, Andrew Mason, I think his name is. It's very similar, like, just from the writing world and using it as a tool for thought to shape Chachabitty. I don't think that people who use AI tools to their limit are common. I'm looking at my NotebookLM now. I've got two sources. You have a little, like, source limit thing. And my bar is over here, you know, and it stretches across the whole thing. I'm like, did he fill it up?
Raiza [00:22:09]: Yes, and he has, like, a higher limit than others, I think. He fills it up.
Raiza [00:22:14]: Oh, yeah.
Raiza [00:22:14]: Like, I don't think Steven even has a limit, actually.
Swyx [00:22:17]: And he has Notes, Google Drive stuff, PDFs, MP3, whatever.
Raiza [00:22:22]: Yes, and one of my favorite demos, he just did this recently, is he has actually PDFs of, like, handwritten Marie Curie notes. I see.
Swyx [00:22:29]: So you're doing image recognition as well. Yeah, it does support it today.
Raiza [00:22:32]: So if you have a PDF that's purely images, it will recognize it.
Raiza [00:22:36]: But his demo is just, like, super powerful.
Raiza [00:22:37]: He's like, okay, here's Marie Curie's notes. And it's like, here's how I'm using it to analyze it. And I'm using it for, like, this thing that I'm writing.
Raiza [00:22:44]: And that's really compelling.
Raiza [00:22:45]: It's like the everyday person doesn't think of these applications. And I think even, like, when I listen to Steven's demo, I see the gap. I see how Steven got there, but I don't see how I could without him. And so there's a lot of work still for us to build of, like, hey, how do I bring that magic down to, like, zero work? Because I look at all the steps that he had to take in order to do it, and I'm like, okay, that's product work for us, right? Like, that's just onboarding.
Alessio [00:23:09]: And so from an engineering perspective, people come to you and it's like, hey, I need to use this handwritten notes from Marie Curie from hundreds of years ago. How do you think about adding support for, like, data sources and then maybe any fun stories and, like, supporting more esoteric types of inputs?
Raiza [00:23:25]: So I think about the product in three ways, right? So there's the sources, the source input. There's, like, the capabilities of, like, what you could do with those sources. And then there's the third space, which is how do you output it into the world? Like, how do you put it back out there? There's a lot of really basic sources that we don't support still, right? I think there's sort of, like, the handwritten notes stuff is one, but even basic things like DocX or, like, PowerPoint, right? Like, these are the things that people, everyday people are like, hey, my professor actually gave me everything in DocX. Can you support that? And then just, like, basic stuff, like images and PDFs combined with text. Like, there's just a really long roadmap for sources that I think we just have to work on.
Raiza [00:24:04]: So that's, like, a big piece of it.
Raiza [00:24:05]: On the output side, and I think this is, like, one of the most interesting things that we learned really early on, is, sure, there's, like, the Q&A analysis stuff, which is like, hey, when did this thing launch? Okay, you found it in the slide deck. Here's the answer. But most of the time, the reason why people ask those questions is because they're trying to make something new. And so when, actually, when some of those early features leaked, like, a lot of the features we're experimenting with are the output types. And so you can imagine that people care a lot about the resources that they're putting into NotebookLM because they're trying to create something new. So I think equally as important as, like, the source inputs are the outputs that we're helping people to create. And really, like, you know, shortly on the roadmap, we're thinking about how do we help people use NotebookLM to distribute knowledge? And that's, like, one of the most compelling use cases is, like, shared notebooks. It's, like, a way to share knowledge. How do we help people take sources and, like, one-click new documents out of it, right? And I think that's something that people think is, like, oh, yeah, of course, right? Like, one push a document. But what does it mean to do it right? Like, to do it in your style, in your brand, right?
Raiza [00:25:08]: To follow your guidelines, stuff like that.
Raiza [00:25:09]: So I think there's a lot of work, like, on both sides of that equation.
Raiza [00:25:13]: Interesting.
Swyx [00:25:13]: Any comments on the engineering side of things?
Usama [00:25:16]: So, yeah, like I said, I was mostly working on building the text to audio, which kind of lives as a separate engineering pipeline, almost, that we then put into NotebookLM. But I think there's probably tons of NotebookLM engineering war stories on dealing with sources. And so I don't work too closely with engineers directly. But I think a lot of it does come down to, like, Gemini's native understanding of images really well with the latest generation.
Raiza [00:25:39]: Yeah, I think on the engineering and modeling side, I think we are a really good example of a team that's put a product out there, and we're getting a lot of feedback from the users, and we return the data to the modeling team, right? To the extent that we say, hey, actually, you know what people are uploading, but we can't really support super well?
Raiza [00:25:56]: Text plus image, right?
Raiza [00:25:57]: Especially to the extent that, like, NotebookLM can handle up to 50 sources, 500,000 words each. Like, you're not going to be able to jam all of that into, like, the context window. So how do we do multimodal embeddings with that? There's really, like, a lot of things that we have to solve that are almost there, but not quite there yet.
Alessio [00:26:16]: On then turning it into audio, I think one of the best things is it has so many of the human... Does that happen in the text generation that then becomes audio? Or is that a part of, like, the audio model that transforms the text?
Usama [00:26:27]: It's a bit of both, I would say. The audio model is definitely trying to mimic, like, certain human intonations and, like, sort of natural, like, breathing and pauses and, like, laughter and things like that. But yeah, in generating, like, the text, we also have to sort of give signals on, like, where those things maybe would make sense.
Alessio [00:26:45]: And on the input side, instead of having a transcript versus having the audio, like, can you take some of the emotions out of it, too? If I'm giving, like, for example, when we did the recaps of our podcast, we can either give audio of the pod or we can give a diarized transcription of it. But, like, the transcription doesn't have some of the, you know, voice kind of, like, things.
Raiza [00:27:05]: Yeah, yeah.
Alessio [00:27:05]: Do you reconstruct that when people upload audio or how does that work?
Raiza [00:27:09]: So when you upload audio today, we just transcribe it. So it is quite lossy in the sense that, like, we don't transcribe, like, the emotion from that as a source. But when you do upload a text file and it has a lot of, like, that annotation, I think that there is some ability for it to be reused in, like, the audio output, right? But I think it will still contextualize it in the deep dive format. So I think that's something that's, like, particularly important is, like, hey, today we only have one format.
Raiza [00:27:37]: It's deep dive.
Raiza [00:27:38]: It's meant to be a pretty general overview and it is pretty peppy.
Raiza [00:27:42]: It's just very upbeat.
Raiza [00:27:43]: It's very enthusiastic, yeah.
Raiza [00:27:45]: Yeah, yeah.
Raiza [00:27:45]: Even if you had, like, a sad topic, I think they would find a way to be, like, silver lining, though.
Raiza [00:27:50]: Really?
Raiza [00:27:51]: Yeah.
Raiza [00:27:51]: We're having a good chat.
Raiza [00:27:54]: Yeah, that's awesome.
Swyx [00:27:54]: One of the ways, many, many, many ways that deep dive went viral is people saying, like, if you want to feel good about yourself, just drop in your LinkedIn. Any other, like, favorite use cases that you saw from people discovering things in social media?
Raiza [00:28:08]: I mean, there's so many funny ones and I love the funny ones.
Raiza [00:28:11]: I think because I'm always relieved when I watch them. I'm like, haha, that was funny and not scary. It's great.
Raiza [00:28:17]: There was another one that was interesting, which was a startup founder putting their landing page and being like, all right, let's test whether or not, like, the value prop is coming through. And I was like, wow, that's right.
Raiza [00:28:26]: That's smart.
Usama [00:28:27]: Yeah.
Raiza [00:28:28]: And then I saw a couple of other people following up on that, too.
Raiza [00:28:32]: Yeah.
Swyx [00:28:32]: I put my about page in there and, like, yeah, if there are things that I'm not comfortable with, I should remove it. You know, so that it can pick it up. Right.
Usama [00:28:39]: I think that the personal hype machine was, like, a pretty viral one. I think, like, people uploaded their dreams and, like, some people, like, keep sort of dream journals and it, like, would sort of comment on those and, like, it was therapeutic. I didn't see those.
Raiza [00:28:54]: Those are good. I hear from Googlers all the time, especially because we launched it internally first. And I think we launched it during the, you know, the Q3 sort of, like, check-in cycle. So all Googlers have to write notes about, like, hey, you know, what'd you do in Q3? And what Googlers were doing is they would write, you know, whatever they accomplished in Q3 and then they would create an audio overview. And these people they didn't know would just ping me and be like, wow, I feel really good, like, going into a meeting with my manager.
Raiza [00:29:25]: And I was like, good, good, good, good. You really did that, right?
Usama [00:29:29]: I think another cool one is just, like, any Wikipedia article. Yeah. Like, you drop it in and it's just, like, suddenly, like, the best sort of summary overview.
Raiza [00:29:38]: I think that's what Karpathy did, right? Like, he has now a Spotify channel called Histories of Mysteries, which is basically, like, he just took, like, interesting stuff from Wikipedia and made audio overviews out of it.
Swyx [00:29:50]: Yeah, he became a podcaster overnight.
Raiza [00:29:52]: Yeah.
Raiza [00:29:53]: I'm here for it. I fully support him.
Raiza [00:29:55]: I'm racking up the listens for him.
Swyx [00:29:58]: Honestly, it's useful even without the audio. You know, I feel like the audio does add an element to it, but I always want, you know, paired audio and text. And it's just amazing to see what people are organically discovering. I feel like it's because you laid the groundwork with NotebookLM and then you came in and added the sort of TTS portion and made it so good, so human, which is weird. Like, it's this engineering process of humans. Oh, one thing I wanted to ask. Do you have evals?
Raiza [00:30:23]: Yeah.
Swyx [00:30:23]: Yes.
Raiza [00:30:24]: What? Potatoes for chefs.
Swyx [00:30:27]: What is that? What do you mean, potatoes?
Raiza [00:30:29]: Oh, sorry.
Raiza [00:30:29]: Sorry. We were joking with this, like, a couple of weeks ago. We were doing, like, side-by-sides. But, like, Raiza sent me the file and it was literally called Potatoes for Chefs. And I was like, you know, my job is really serious, but you have to laugh a little bit. Like, the title of the file is, like, Potatoes for Chefs.
Swyx [00:30:47]: Is it like a training document for chefs?
Usama [00:30:50]: It's just a side-by-side for, like, two different kind of audio transcripts.
Swyx [00:30:54]: The question is really, like, as you iterate, the typical engineering advice is you establish some kind of test or benchmark. You're at, like, 30 percent. You want to get it up to 90, right?
Raiza [00:31:05]: Yeah.
Swyx [00:31:05]: What does that look like for making something sound human and interesting and voice?
Usama [00:31:11]: We have the sort of formal eval process as well. But I think, like, for this particular project, we maybe took a slightly different route to begin with. Like, there was a lot of just within the team listening sessions. A lot of, like, sort of, like... Dogfooding.
Raiza [00:31:23]: Yeah.
Usama [00:31:23]: Like, I think the bar that we tried to get to before even starting formal evals with raters and everything was much higher than I think other projects would. Like, because that's, as you said, like, the traditional advice, right? Like, get that ASAP. Like, what are you looking to improve on? Whatever benchmark it is. So there was a lot of just, like, critical listening. And I think a lot of making sure that those improvements actually could go into the model. And, like, we're happy with that human element of it. And then eventually we had to obviously distill those down into an eval set. But, like, still there's, like, the team is just, like, a very, very, like, avid user of the product at all stages.
Raiza [00:32:02]: I think you just have to be really opinionated.
Raiza [00:32:05]: I think that sometimes, if you are, your intuition is just sharper and you can move a lot faster on the product.
Raiza [00:32:12]: Because it's like, if you hold that bar high, right?
Raiza [00:32:15]: Like, if you think about, like, the iterative cycle, it's like, hey, we could take, like, six months to ship this thing. To get it to, like, mid where we were. Or we could just, like, listen to this and be like, yeah, that's not it, right? And I don't need a rater to tell me that. That's my preference, right? And collectively, like, if I have two other people listen to it, they'll probably agree. And it's just kind of this step of, like, just keep improving it to the point where you're like, okay, now I think this is really impressive. And then, like, do evals, right? And then validate that.
Swyx [00:32:43]: Was the sound model done and frozen before you started doing all this? Or are you also saying, hey, we need to improve the sound model as well? Both.
Usama [00:32:51]: Yeah, we were making improvements on the audio and just, like, generating the transcript as well. I think another weird thing here was, like, we needed to be entertaining. And that's much harder to quantify than some of the other benchmarks that you can make for, like, you know, Sweebench or get better at this math.
Swyx [00:33:10]: Do you just have people rate one to five or, you know, or just thumbs up and down?
Usama [00:33:14]: For the formal rater evals, we have sort of like a Likert scale and, like, a bunch of different dimensions there. But we had to sort of break down what makes it entertaining into, like, a bunch of different factors. But I think the team stage of that was more critical. It was like, we need to make sure that, like, what is making it fun and engaging? Like, we dialed that as far as it goes. And while we're making other changes that are necessary, like, obviously, they shouldn't make stuff up or, you know, be insensitive.
Raiza [00:33:41]: Hallucinations. Safety.
Swyx [00:33:42]: Other safety things.
Raiza [00:33:43]: Right.
Swyx [00:33:43]: Like a bunch of safety stuff.
Raiza [00:33:45]: Yeah, exactly.
Usama [00:33:45]: So, like, with all of that and, like, also just, you know, following sort of a coherent narrative and structure is really important. But, like, with all of this, we really had to make sure that that central tenet of being entertaining and engaging and something you actually want to listen to. It just doesn't go away, which takes, like, a lot of just active listening time because you're closest to the prompts, the model and everything.
Swyx [00:34:07]: I think sometimes the difficulty is because we're dealing with non-deterministic models, sometimes you just got a bad roll of the dice and it's always on the distribution that you could get something bad. Basically, how many do you, like, do ten runs at a time? And then how do you get rid of the non-determinism?
Raiza [00:34:23]: Right.
Usama [00:34:23]: Yeah, that's bad luck.
Raiza [00:34:25]: Yeah.
Swyx [00:34:25]: Yeah.
Usama [00:34:26]: I mean, there still will be, like, bad audio overviews. There's, like, a bunch of them that happens. Do you mean for, like, the raider? For raiders, right?
Swyx [00:34:34]: Like, what if that one person just got, like, a really bad rating? You actually had a great prompt, you actually had a great model, great weights, whatever. And you just, you had a bad output.
Usama [00:34:42]: Like, and that's okay, right?
Raiza [00:34:44]: I actually think, like, the way that these are constructed, if you think about, like, the different types of controls that the user has, right? Like, what can the user do today to affect it?
Usama [00:34:54]: We push a button.
Raiza [00:34:55]: You just push a button.
Swyx [00:34:56]: I have tried to prompt engineer by changing the title. Yeah, yeah, yeah.
Raiza [00:34:59]: Changing the title, people have found out.
Raiza [00:35:02]: Yeah.
Raiza [00:35:02]: The title of the notebook, people have found out. You can add show notes, right? You can get them to think, like, the show has changed. Someone changed the language of the output. Changing the language of the output. Like, those are less well-tested because we focused on, like, this one aspect. So it did change the way that we sort of think about quality as well, right? So it's like, quality is on the dimensions of entertainment, of course, like, consistency, groundedness. But in general, does it follow the structure of the deep dive? And I think when we talk about, like, non-determinism, it's like, well, as long as it follows, like, the structure of the deep dive, right? It sort of inherently meets all those other qualities. And so it makes it a little bit easier for us to ship something with confidence to the extent that it's like, I know it's going to make a deep dive. It's going to make a good deep dive. Whether or not the person likes it, I don't know. But as we expand to new formats, as we open up controls, I think that's where it gets really much harder. Even with the show notes, right? Like, people don't know what they're going to get when they do that. And we see that already where it's like, this is going to be a lot harder to validate in terms of quality, where now we'll get a greater distribution. Whereas I don't think we really got, like, varied distribution because of, like, that pre-process that Raiza was talking about. And also because of the way that we'd constrain, like, what were we measuring for? Literally, just like, is it a deep dive?
Swyx [00:36:18]: And you determine what a deep dive is. Yeah. Everything needs a PM. Yeah, I have, this is very similar to something I've been thinking about for AI products in general. There's always like a chief tastemaker. And for Notebook LM, it seems like it's a combination of you and Steven.
Raiza [00:36:31]: Well, okay.
Raiza [00:36:32]: I want to take a step back.
Swyx [00:36:33]: And Raiza, I mean, presumably for the voice stuff.
Raiza [00:36:35]: Raiza's like the head chef, right? Of, like, deep dive, I think. Potatoes.
Raiza [00:36:40]: Of potatoes.
Raiza [00:36:41]: And I say this because I think even though we are already a very opinionated team, and Steven, for sure, very opinionated, I think of the audio generations, like, Raiza was the most opinionated, right? And we all, like, would say, like, hey, I remember, like, one of the first ones he sent me.
Raiza [00:36:57]: I was like, oh, I feel like they should introduce themselves. I feel like they should say a title. But then, like, we would catch things, like, maybe they shouldn't say their names.
Raiza [00:37:04]: Yeah, they don't say their names.
Usama [00:37:05]: That was a Steven catch, like, not give them names.
Raiza [00:37:08]: So stuff like that is, like, we all injected, like, a little bit of just, like, hey, here's, like, my take on, like, how a podcast should be, right? And I think, like, if you're a person who, like, regularly listens to podcasts, there's probably some collective preference there that's generic enough that you can standardize into, like, the deep dive format. But, yeah, it's the new formats where I think, like, oh, that's the next test. Yeah.
Swyx [00:37:30]: I've tried to make a clone, by the way. Of course, everyone did. Yeah. Everyone in AI was like, oh, no, this is so easy. I'll just take a TTS model. Obviously, our models are not as good as yours, but I tried to inject a consistent character backstory, like, age, identity, where they work, where they went to school, what their hobbies are. Then it just, the models try to bring it in too much.
Raiza [00:37:49]: Yeah.
Swyx [00:37:49]: I don't know if you tried this.
Raiza [00:37:51]: Yeah.
Swyx [00:37:51]: So then I'm like, okay, like, how do I define a personality? But it doesn't keep coming up every single time. Yeah.
Raiza [00:37:58]: I mean, we have, like, a really, really good, like, character designer on our team.
Raiza [00:38:02]: What?
Swyx [00:38:03]: Like a D&D person?
Raiza [00:38:05]: Just to say, like, we, just like we had to be opinionated about the format, we had to be opinionated about who are those two people talking.
Raiza [00:38:11]: Okay.
Raiza [00:38:12]: Right.
Raiza [00:38:12]: And then to the extent that, like, you can design the format, you should be able to design the people as well.
Raiza [00:38:18]: Yeah.
Swyx [00:38:18]: I would love, like, a, you know, like when you play Baldur's Gate, like, you roll, you roll like 17 on Charisma and like, it's like what race they are. I don't know.
Raiza [00:38:27]: I recently, actually, I was just talking about character select screens.
Raiza [00:38:30]: Yeah. I was like, I love that, right.
Raiza [00:38:32]: And I was like, maybe there's something to be learned there because, like, people have fallen in love with the deep dive as a, as a format, as a technology, but also as just like those two personas.
Raiza [00:38:44]: Now, when you hear a deep dive and you've heard them, you're like, I know those two.
Raiza [00:38:48]: Right.
Raiza [00:38:48]: And people, it's so funny when I, when people are trying to find out their names, like, it's a, it's a worthy task.
Raiza [00:38:54]: It's a worthy goal.
Raiza [00:38:55]: I know what you're doing. But the next step here is to sort of introduce, like, is this like what people want?
Raiza [00:39:00]: People want to sort of edit the personas or do they just want more of them?
Swyx [00:39:04]: I'm sure you're getting a lot of opinions and they all, they all conflict with each other. Before we move on, I have to ask, because we're kind of on this topic. How do you make audio engaging? Because it's useful, not just for deep dive, but also for us as podcasters. What is, what does engaging mean? If you could break it down for us, that'd be great.
Usama [00:39:22]: I mean, I can try. Like, don't, don't claim to be an expert at all.
Swyx [00:39:26]: So I'll give you some, like variation in tone and speed. You know, there's this sort of writing advice where, you know, this sentence is five words. This sentence is three, that kind of advice where you, where you vary things, you have excitement, you have laughter, all that stuff. But I'd be curious how else you break down.
Usama [00:39:42]: So there's the basics, like obviously structure that can't be meandering, right? Like there needs to be sort of a, an ultimate goal that the voices are trying to get to, human or artificial. I think one thing we find often is if there's just too much agreement between people, like that's not fun to listen to. So there needs to be some sort of tension and build up, you know, withholding information. For example, like as you listen to a story unfold, like you're going to learn more and more about it. And audio that maybe becomes even more important because like you actually don't have the ability to just like skim to the end of something. You're driving or something like you're going to be hooked because like there's, and that's how like, that's how a lot of podcasts work. Like maybe not interviews necessarily, but a lot of true crime, a lot of entertainment in general. There's just like a gradual unrolling of information. And that also like sort of goes back to the content transformation aspect of it. Like maybe you are going from, let's say the Wikipedia article of like one of the History of Mysteries, maybe episodes. Like the Wikipedia article is going to state out the information very differently. It's like, here's what happened would probably be in the very first paragraph. And one approach we could have done is like maybe a person's just narrating that thing. And maybe that would work for like a certain audience. Or I guess that's how I would picture like a standard history lesson to unfold. But like, because we're trying to put it in this two-person dialogue format, like there, we inject like the fact that, you know, there's, you don't give everything at first. And then you set up like differing opinions of the same topic or the same, like maybe you seize on a topic and go deeper into it and then try to bring yourself back out of it and go back to the main narrative. So that's, that's mostly from like the setting up the script perspective. And then the audio, I was saying earlier, it's trying to be as close to just human speech as possible. I think was the, what we found success with so far.
Raiza [00:41:40]: Yeah. Like with interjections, right?
... [Content truncated due to size limits]