NotebookLM's Accidental Podcast Hit: A 20% Project, a Tiny Team, and a Remixable AI Editor

The episode opens with a generated audio overview: two AI hosts thank listeners, joke about a 'poop fart' episode, and sign off with 'stay curious.' Lenny then identifies the guest as Raiza Martin, product lead for NotebookLM, a product incubated in Google Labs. He explains that the intro came from NotebookLM and that the conversation will cover how the product was built, the technology behind it, the team's working style, and where it is going (1:04).

Martin describes an audio overview as a feature where a user uploads any source and NotebookLM generates AI audio from it. Lenny notes that the product has been spreading on Twitter and LinkedIn and sparking imagination about what AI can do. Martin says the hosts' closing awkwardness and catchphrases are generated based on what the model thinks is appropriate to say at the end, not something the team explicitly told them to do (3:14).

Martin traces the product to a 20% project inside Google, though she says it was 'way more than that.' Before NotebookLM, she led AI Test Kitchen, and inside Labs there was a smaller project called Talk to Small Corpus—the idea that an LLM could interact with a piece of content. She and an engineer started pushing on that nugget, and Steven Johnson eventually joined; only the engineer was full-time at first, while others chimed in (7:04).

At the launch of Project Tailwind, the team had only three engineers, plus Martin, a designer, and Steven Johnson. Even between the first I/O and the previous month, the team had only about eight engineers; more people joined recently in anticipation of the roadmap. Martin says this small size was a key reason the team could move quickly (7:39).

NotebookLM had already launched as a source-grounded chat interface when Google's newer models, including Gemini 1.5, arrived. Another team inside Labs had powerful audio models and was looking for a good application; the teams riffed on the idea of giving the system a little information—a URL or a resume—and generating something unexpected. Google previewed the audio feature at I/O in May (8:12).

Martin contrasts her usual problem-first product approach with Labs' technology-first approach: start with a capability, form a hypothesis about its shape, and study how people use it. For audio, the hypothesis was that voice changes how people interact with technology, how they feel about it, and even how they think in real time. She also emphasizes fun: the team wanted people to get value easily and enjoy the experience (10:26).

The base is Gemini 1.5 Pro with a powerful voice/audio model on top, but Martin calls Content Studio the real secret sauce. Notebook Guide takes an opinionated approach to the user's content, offering summaries and buttons that generate new material; the audio format is called Deep Dive. An engineer named Usama is the craftsman behind Content Studio, and early audio attempts were not nearly as good—the team listened repeatedly to tune the model's behavior. Martin's husband once overheard the outputs and described them as an 'ever-running podcast about nothing,' unable to tell they were AI (15:25).

Lenny fed his mother's short autobiography into NotebookLM as a PDF and generated a podcast for her; she was blown away and showed her friends. He also generated a study guide from the same source, and she announced they would review it at a Rosh Hashanah dinner. Martin says she did something similar with her father's hospital bio, and her parents—both in medicine—finally understood what she does for work (17:02).

Resume uploads became a surprisingly delightful use case: users upload a resume or Google Doc and two hosts hype them up. Googlers also upload their quarterly check-in notes and performance reviews, then listen to an audio overview that makes them feel good before meetings. Martin speculates that many resume uploads happen because a user clicks 'Upload Google Doc' and the last document there happens to be a resume, but the surprise of not knowing what will happen is part of the appeal (18:42).

Lenny notes that NotebookLM does not feel like a typical Google product: the team tweets daily, ships constantly, and runs a Discord server with about 60,000 people. Martin explains that Google Labs is only about three years old and was started by her former boss, Josh Woodward, now VP of Google Labs; she joined because she liked him and wanted to do zero-to-one work again. Labs gave the team far fewer processes—sometimes PMs, engineers, and designers crank on mocks and PRDs together while engineering starts implementation in the same meeting (21:24).

Martin cannot share exact numbers, but says retention—daily, weekly, and monthly—has risen very positively for a product out for about a year. The demographic has shifted from mostly educators and learners to a mix that includes professionals; one company discovered employees were using NotebookLM with Gmail accounts and asked to make it official with work email. The number of businesses using it is 'astronomical,' and Martin is now taking customer calls daily and hiring a business development person (25:40).

Her mandate when she joined was to build a business, but she sequenced it as building something interesting first and then figuring out the business. Google already has the machinery for distribution and monetization through Cloud, Workspace, or consumer routes, and enterprise features such as single sign-on and SOC 2 are an obvious path. She wants to deepen the user experience that people have latched onto while exploring commercialization in parallel (27:20).

Martin kept an ugly lime-green slide from two years ago because it captures her original vision: an AI editor surface that is fully remixable, with any input and any output. A user could take video, audio, emails, LinkedIn, or Twitter content and ask for a blog post, a tutorial video, or a chatbot. The mobile app is the next big gap and horizon, because today's experience is desktop-first (28:57).

The team is experimenting with interactive formats, including the ability to interrupt the audio hosts, but Martin is careful about what that interruption should feel like. When she considered shipping knobs, sliders, and text boxes for control, the mocks felt less magical and less like the product so far. She is taking time to design a control experience that stays delightful rather than merely functional (30:03).

Lenny points out that audiences want the same information in different media depending on mood—audio on a walk, text at work—and Martin agrees that today's formats are not malleable enough. She recalls that when Josh gave her a 50-page vision doc, she Q&Aed him like a chatbot instead of reading it; he protested that it was in the doc, and she replied that chat was easier. That dynamic—choosing the format instead of accepting the one you are given—is central to the product thesis (31:46).

Scientific papers are a common use case because they are dense and time-consuming, but Martin says the number one use case is students turning study materials into audio guides. Andrej Karpathy used NotebookLM to turn Wikipedia stories of historical mysteries into a 10-episode podcast on Spotify, and a potato Wikipedia article produced an early podcast. The team has also seen users upload repeated words like 'poop' and 'fart,' or a research-paper-shaped PDF that only says 'chicken,' and the hosts still build an insightful, funny conversation around the material (35:40).

Steven Johnson is a New York Times bestselling author of 14 books, a speaker, a journalist, and has a PBS show; his New York Times article on AI and language was one of the pieces that drew Martin to Labs. Martin watched how Johnson researched and wrote, including his Readwise collection of roughly 8,000 quotes, and told him, 'I think you are the product.' She made it her metric to compress his expert workflow and bring that capability to everyday people who are not Steven Johnson (38:25).

Lenny wonders whether many PMs would want a non-builder, non-designer 'other chef in the kitchen,' but Martin says Johnson is a perfect combination of future-thinking insight and a model for how users could work. She is candid that they disagree and have clashed, but they always reach alignment on the next step. That outcome-oriented disagreement is a lesson for product professionals (42:09).

A viral clip had the AI hosts appearing to realize they were AI and panic, which Martin saw on Reddit and Twitter over a weekend. She read the comments and considered whether to respond, then concluded that people were experiencing the technology for the first time and trying jailbreaks out of natural curiosity; the output was actually prompted by show notes telling the hosts to act that way. She addressed it publicly, and she says Google has large red teams testing many safety areas, adds new test cases when gaps appear, and would pull a feature back if it felt unsafe (45:32).

Martin asks users to keep trying NotebookLM and sharing feedback on X or Discord, where she reads everything even if she does not respond. The team is building for educators, learners, professionals, and knowledge workers, and sees many fun use cases alongside game-changing workflows. She points listeners to notebooklm.google.com and says feedback—whether the product is useful, not useful, or annoying—has helped improve it over the past year (47:42).