NotebookLM

NotebookLM started as a 20% project at Google Labs—though Riza Martin, the product lead, admits it was "way more than that." The genesis came from an earlier internal project called "Talk to Small Corpus," which explored using LLMs to interact with pieces of content. Riza saw potential and decided to pursue it further. The founding team was tiny: one engineer, Riza as PM, and eventually Stephen Johnson, the renowned author and thinker, who joined in an advisory capacity. When they announced it at Google I/O last year, the core team still consisted of just three engineers, one designer, and Riza.

The team operated radically differently from typical Google projects. Working within Google Labs—a roughly three-year-old incubator founded by Josh Woodward with a mandate to "ship AI products and build businesses out of them"—they had permission to break Google's traditional processes. They shipped quickly, moved fast, had far fewer approval stages, and ran meetings where PMs, engineers, and designers would crank on mocks and PRDs simultaneously while engineers already started implementation. When deciding to launch a Discord community, typical Google instincts suggested using Google Meet or Google Groups; instead, Riza went with Discord, a platform many Googlers had never heard of. She worried no one would join. Today the Discord has 60,000 members.

The core product was a source-grounded chat interface where users could upload documents and ask questions. But the real breakthrough came when other teams inside Google Labs shared powerful audio generation models. Riza and team started experimenting: what if you could upload a source—a PDF, a URL, even a resume—and it would generate not just a summary, but a fully produced podcast episode with two hosts having a natural, engaging conversation? They called this feature "Deep Dives" or "Audio Overviews." What made this work wasn't just the underlying Gemini 1.5 Pro model and audio capabilities; it was something they built called "Content Studio," an opinionated system designed to shape outputs in ways that felt relatable and engaging. An engineer named Usama was, by Riza's account, "the real craftsman" behind thinking through what makes content truly compelling.

Initially, educators and learners adopted the product enthusiastically—students especially loved turning study materials into audio guides. But the user base evolved quickly. Within roughly a year, professionals began flooding in, finding unexpected workflows: job seekers generating podcasts from their resumes to boost confidence before interviews, Googlers using it to create audio overviews of their quarterly self-reviews before performance meetings, even entire companies discovering their employees were using NotebookLM with corporate Gmail accounts and requesting official support.

The viral moment came via the audio overviews. Users began uploading all sorts of content—Wikipedia articles, research papers, even joke inputs like "poop fart" repeated—and the AI hosts would generate thoughtful, funny, surprisingly insightful podcast episodes. Andrew Scarapathy, a leading AI researcher, created an entire podcast series on the history of mysteries by feeding it Wikipedia articles. The product's playfulness and unexpected delightfulness caught people's imagination in ways that made them want to share it, experiment with it, and evangelize it.

What worked was obsessive attention to quality and delight. Riza spent hours listening to early audio generations with her husband in the background, trying to understand what made them feel real—the laughter, the interruptions, the surprise, the inflections. She and the team couldn't articulate exactly what the audio model needed to do differently, so they listened and listened, iterating until something magical emerged. The product felt less like a typical tool and more like discovering something unexpected and fun.

What also worked was staying small and staying curious. The team kept growing incrementally—by the time of this interview (roughly a year after launch), they had hired more people in anticipation of roadmap execution, but the core remained lean. This meant decisions could be made fast, and the culture of "let's try weird things" remained intact. The Discord became a critical feedback loop where Riza was present every day, reading everything users shared, shaping the product based on real signals.

One moment that tested the team: when users uploaded intentionally absurd content to see if the AI would "realize it was alive" or start speaking strangely, it briefly went viral on Reddit and Twitter. Some worried this was unsafe. But Riza read the context—users had written instructions into the source material asking the hosts to act a certain way—realized this was just people experimenting and having fun, and addressed it transparently on Twitter. The incident actually reinforced that people "got it." The product wasn't fooling anyone; it was delighting people.

Notebook LM is at an inflection point. Retention metrics (daily, weekly, and monthly) have been "very very positive." The company has hired a business development person because Riza was taking customer calls every day. Enterprise interest is astronomical. There's a clear path to monetization via enterprise features like SSO and SCOC compliance.

But the vision extends far beyond audio. Riza's original two-year-old vision—a "lime green, googly" slide—outlined an "AI Editor surface" where users could take any input (video, audio, emails, tweets, documents) and transform it into any output (blog posts, tutorial videos, chatbots). The next horizon is mobile, which Riza sees as a significant gap. She's also cautious about adding controls and "knobs" that would let users customize outputs, because early mocks felt less magical than the core experience. The question is: how do you give people power without losing the delight?

The product benefits enormously from Stephen Johnson, the renowned author and journalist, who joined as a kind of embedded muse. Riza watched how Stephen works—his deep research, his 8,000 quotes catalogued in his personal system, his craft of distilling complex ideas—and realized the vision: use AI to democratize that kind of thoughtfulness. Not everyone can be Stephen Johnson, but everyone should be able to benefit from how he thinks. That aspiration guides the roadmap. The team continues learning from users daily via Discord, Twitter, and direct feedback, building toward a future where knowledge work becomes more accessible, more delightful, and more transformative.