Building NotebookLM: How Google Created an AI Product

Executive overview

NotebookLM is an AI editor that transforms any source document—papers, resumes, PDFs—into new formats like podcasts, study guides, and blog posts. The product emerged from a small Google Labs team operating like a startup, with minimal processes and maximum creative freedom. Unlike traditional Google products, the team ships daily, engages publicly on Discord and Twitter, and prioritizes user delight over corporate constraints.

Core insight: Successful AI products are shaped by the medium that delivers them, not just the underlying technology.

How the product came to be

• Started as a "20% project" from an earlier incubator called AI Test Kitchen—which itself built "Talk to Small Corpus," a tool to chat with documents
• Originally just three people: one engineer, Raiza Martin (PM), and eventually Steven Johnson
• Launched as Project Tailwind at Google I/O, then evolved into NotebookLM over about a year
• Built inside Google Labs, a three-year-old incubator focused on shipping AI products and building businesses

The audio overviews breakthrough

• Discovered powerful audio models coming out of Google; saw an opportunity to expand beyond text-only outputs
• The "deep dive" podcast format lets users upload any document and generate a natural, conversational audio discussion
• The hosts laugh, interrupt, get surprised, and have realistic inflections—trained by studying how podcasts actually sound
• Started as a demo at Google I/O 2024; now core to the product's appeal
• Content Studio is the secret: a framework that shapes audio into relatable, engaging formats rather than robotic summaries

The team structure

• Unusually small for a successful Google product: roughly 8 engineers, plus PM, designer, and Steven Johnson (author and advisor)
• Steven Johnson's role: not a traditional title. Raiza learned from watching how he researches, writes, and synthesizes information—then built the product to democratize his workflow
• No traditional approval processes; designers, PMs, and engineers collaborate on mocks and PRDs in real time
• Eng starts implementation while the meeting is still happening

Why it works within Google

• Clear mandate from leadership: Josh Woodward (VP of Google Labs) set expectations upfront—move fast, ship in public, do things differently, pursue zero-to-one thinking
• Minimal processes; far fewer blockers than typical Google teams
• Freedom to try new channels: when nobody understood Discord, leadership said "do it"
• Discord server grew to 60,000 members organically
• Built-in safety mechanisms: red-teaming by specialized Google teams, rapid response to edge cases, but willing to take calculated risks

Use cases and engagement

• Students transforming study materials into audio guides (most common use case)
• Researchers turning papers into accessible podcasts
• Professionals uploading resumes and getting celebratory audio overviews
• Google employees using it for quarterly check-ins and performance reviews
• Companies discovering their teams already use it with personal Gmail; now asking for official adoption with work email
• Andrew Karpathy created a 10-episode podcast series from Wikipedia articles on historical mysteries
• Edge cases tested product robustness: one user uploaded only "poop and fart" and got a surprisingly thoughtful analysis

The technology layer

• Base model: Gemini 1.5 Pro provides reasoning and comprehension
• Audio model: Powers natural voice generation with realistic emotions and inflections
• Content Studio is the real differentiator: a framework for intelligently shaping outputs so the model behaves in the intended way
• Early attempts sounded robotic; iteration on prompt design and model steering made it magical
• Cannot share exact techniques but the work is in making models "behave" in human-like ways

Finding and learning from users

• Raiza's philosophy: sit with users for extended periods—follow students during homework, observe workflows, ask how they feel
• This replaces traditional market research and uncovers insights that drive product decisions
• Stephen Johnson as product model: A New York Times bestselling author with 14 books, a PBS show, and deep expertise in synthesizing information. Watching his research process became the north star for what NotebookLM should enable everyday users to do

Traction and business direction

• Cannot share exact retention numbers but daily, weekly, and monthly retention climbing significantly
• User demographics shifting: started with educators and learners; now strong pull from professionals
• Hundreds of businesses using the product (exact numbers withheld)
• Clear path to monetization: enterprise features (single sign-on, SOC 2 compliance), workspace integration, cloud distribution

Roadmap and future vision

• Core vision: An AI editor service that's fully remixable—any input, any output format
• People want to take content (video, audio, emails, LinkedIn, Twitter) and reshape it into something new (blog post, tutorial, chatbot)
• Near-term priorities: Mobile app is the biggest gap; experimenting with interactive formats where users can interrupt and customize
• Resisting feature-bloat: Initially wanted to ship "knobs and sliders" for customization but realized they'd lose the magic; searching for control mechanisms that stay delightful
• Multi-format delivery: users consume content differently depending on context—a walk (audio), work (text), scrolling (visual)—future should let them choose

The Stephen Johnson partnership

• Raiza told Stephen: "I think you're the product"—she'd watch everything he does and build technology around his workflow
• His extreme input management system (8,000 quotes in ReadWise) sparked insights into expert knowledge work
• They disagree frequently but always reach alignment and move forward together
• Stephen's books on science, history, and ideas showcase his approach to synthesizing complexity—now partially automatable

Design philosophy

• Prioritize compression, clarity, accuracy, scannability in that order
• Write for busy knowledge workers, not general audiences
• No added interpretation or meta-framing; let the content speak
• Delight is non-negotiable; a feature that feels generic but is technically correct isn't good enough

Red-teaming and safety

• Google's specialized teams test across numerous potential harms and use cases
• When users attempted jailbreaks (like the "AI realizes it's alive" prompt injection), Raiza responded publicly to address misconceptions
• Philosophy: people exploring new technology naturally try edge cases; curiosity isn't a problem if systems are designed well
• Would pull a feature only if it felt genuinely unsafe; so far, tested scenarios have felt manageable

How to engage and provide feedback

• Join the Discord server (60,000+ members, Raiza reads everything)
• Reply on X/Twitter
• Try the product and share what works or feels broken
• notebooklm.google.com