A 10-Person Team, One Click, and Two Years of Fake Deadlines

Late one night in September 2024, Raiza Martin was already in bed when she found a NotebookLM clip on X. Someone had uploaded a long document whose only words were “poop” and “fart,” repeated hundreds of times, and asked the Audio Overview hosts to make a show of it. It was 11 p.m. Her first thought was practical and slightly panicked: if she tapped play and hit a safety flag, she would be filing a bug and paging engineers instead of sleeping.

She listened anyway. The two AI hosts turned the nonsense into a back-and-forth with real rhythm, and she signed off with “well done, Notebook, good little guy.” Design lead Jason Spielman saw the same video and thought, at first, “uh oh, here comes work.” Then he heard it. “This is even better than I could have ever imagined.” That month, the experimental Google Labs product had already become the talk of the internet: upload a document, audio, video, or PDF, click once, and get a two-voice podcast the team calls Audio Overview.

Hosts Sonya Huang and Pat Grady open Training Data with a nested joke: NotebookLM’s own AI hosts, billed as the voices of Deep Dive, thank Sequoia and promise to keep getting better. Then the humans take over. Two weeks after launch, Huang says, GPUs were “going brrr,” and people were calling NotebookLM Google’s ChatGPT moment. Martin, who felt ChatGPT as a personal earthquake, flinches at the comparison—“whoa, is it?”—but grants the resemblance: people are having a “this is what AI can do” reaction.

Spielman’s version of that shock is more specific. The first time he heard an Audio Overview, the second host coming in felt mind-blowing. Underneath the voices, though, he wants listeners to notice Gemini 1.5 Pro chewing through messy source material and producing something concise. NotebookLM is still, in Martin’s words, an AI-powered research and writing tool. Overnight it became famous as “upload a source and it will generate a podcast.” They did not set out to kill podcast hosts. Voice was the next output modality; dialogue won; the way it landed with people was, Martin says, “delightful and surprising.”

Martin was working on Google’s AI Test Kitchen in 2023 when NotebookLM was still a 20% project. One engineer had a prototype with a terrible consumer brand: “talk to small corpus.” She had to ask what a corpus was. The idea underneath was simple and, to her, immediately practical—use a large language model to talk to a small pile of your own documents and pull information out of them.

That is when the product got a personal engine. Martin had gone back to school as an adult learner. If an LLM could talk to a textbook the way this prototype talked to a corpus, she thought, it would have changed her life, and it could change a lot of other lives. In May 2023 they introduced the first version as Project Tailwind: upload a PDF and chat with it. No podcast, no two hosts, just a source and a conversation.

Spielman, arriving later, noticed something he had not articulated before: almost everything he creates sits on a prior artifact—a report, a brief, a pile of research. That is why the positioning was clear from the start. NotebookLM is source-grounded. Answers are supposed to be anchored in what you uploaded, which is the difference between this tool and a general assistant like ChatGPT or Gemini. They share no user counts. Martin will only sketch a curve: steady growth before Audio Overview, a steep hook after it, and a retention twist—people come for the podcast and stay for everything else.

Behind Audio Overview sit three layers. Gemini 1.5 takes the uploaded material and writes something new. Google’s audio models speak it. Between those, the team built Content Studio, which Spielman describes as the editorial engine: the place where they take liberties with how the show is shaped. Users already want that studio exposed. Make this one funnier. Make this one more serious. Martin’s gut said ship the knobs. She is forcing herself not to.

People fell in love, she argues, because the thing was delightful and a little magical. The product question is not how to add controls. It is “how do I ship delightful and magical knobs?” She will accept more basic choices later: a different show format than Deep Dive, a shorter or longer episode, a topic instead of a full-source overview. What she will not do is sand the surprise off the first experience just because power users are asking.

Spielman thinks the viral fuse was even simpler than the voices. He was on the phone with his grandmother, trying to explain the product, and needed no explanation. “Drop in a source.” She saw the button and generated the show. No prompt, no settings, no onboarding. In a world of chat boxes, Grady asks, where did the podcast idea even come from? Martin’s answer is format. They had a strong voice model and tried monologue, dialogue, and a user-facing switch. Dialogue won because people recognized it instantly: oh, this is a podcast—not text-to-speech reading the page aloud.

Martin expected education to be the center of gravity, and students and teachers are using it. The surprise is how many people use NotebookLM at work. Inside Google she ran a pilot with ads sellers whose training docs run to hundreds of pages and change constantly. The old workflow was tribal: ping Joe, wait, paste his answer into an email, fiddle. After the pilot, the Joes of the world built notebooks and distributed them. Hundreds of people stopped waiting in line.

Spielman keeps hearing the same pattern from friends. A salesperson on a live call who does not know the answer opens the notebook and asks. Another friend in venture and private equity lives inside confidential information memorandums—CIMs, packets Spielman had never even heard of. That friend dumps the documents and slides into a notebook and claims the job got ten times faster. Huang jokes that they are going after Sequoia’s job too. Spielman says they are helping.

The personal cases are stranger, and they show up later in the conversation as the real product category. Founders drop in a landing page to hear whether the messaging is clear. People upload a resume or a LinkedIn bio. A TikTok creator put in her 2004 diary and filmed her own reaction. Someone skipped a weekend of college-friend group chat, pasted hundreds of messages into a doc, and listened on the Monday drive to work. Martin’s name for this is personalized generation: nobody would ever commission a real podcast of those artifacts, and that is exactly why the feature holds.

Asked which design choices made NotebookLM feel obvious, Spielman declines to freeze the product in amber. The team is launching quickly and sitting with users, because the underlying tech is moving too fast to know what is even possible. One choice he will defend looks almost accidental: the left-hand source panel. If you are source-grounded, the interface has to say so at a glance. You are talking to what you uploaded. Audio Overview’s one-click bet paid off for the same reason. The remaining problem, Martin adds, is the upload itself. Users balk. They do not have to upload a source to ChatGPT or Gemini. Those products “just work.” NotebookLM still has work to do in the just-works category.

Spielman’s wider diagnosis is that AI products are in a skeuomorphic era. Early iOS put leather stitching on Notes and made the legal pad yellow so people could cross from the physical world into a virtual one. The chat box, he argues, is the same move: a familiar object that meets first-time AI users where they are, not necessarily the destination. Huang cites Midjourney as a company that solved the blank-prompt wall. Martin’s recent crush is Pika. You upload an image, and each effect previews what will happen to it. She uploaded a drink, picked “cake,” and was ready to pay $10 before the first generation even finished. Show people the other side of the door, she says, and they will hand you the material.

Spielman’s admired example is Claude Artifacts: chat on the left, something you can actually use growing on the right, the hierarchy between human and model flattened. NotebookLM is still, first, a source-grounded tool, which makes it different. Martin’s strategic read is that contextualizing the model against your files is a stickier experience, and Anthropic, OpenAI, and Google all know it. The question is when to introduce it and on which surfaces. NotebookLM’s luck is that it started there. While everyone else juggles a dozen use cases, they can sprint.

Martin’s own prototypes lean into walking and talking with a model. She has started journaling as a back-and-forth: the system writes the log, then visualizes whether the week had more good days or bad ones, and what made her happy or upset. Work and school matter, she says, but the personal loop is rich, and she wants those lessons back in NotebookLM—probably first in a mobile app. Asked where they take lightning in a bottle, she does not reach for a new category. Deepen the experience. Make it useful. Deliver on the promise of that first hit so people who stick around are rewarded.

If she could rewind the launch, the missing piece is sharing. She scrolls X and sees visualizers other people built because NotebookLM had no native one. As a product lead, that is a user she cannot count: they left her surface. Collaboration around Audio Overviews is the gap. The other gap is writing. People already do Q&A and then take the answer somewhere else to make something new. She wants to close that whole journey inside the product.

The personalities of the two hosts are less scripted than fans assume. The show is called Deep Dive, there are two hosts, and Martin says the personas take a lot of editorial liberty with whatever you uploaded. Even she does not know where they will go. That is why “make it funnier” is a harder control than “make it shorter” or “cover this topic.” The heart of the experience, she keeps repeating, is that editorial liberty around your sources.

Huang puts the awkward question on the table. The generated audio is already good enough to replace a real podcast, she says; when the guests insist it is not, is that just politeness? Martin treats it as a real risk she would rather look in the eye. Then she tests it on herself. She listens to Lenny’s podcast all the time. Would she take one of Lenny’s articles, run Audio Overview, and call that a substitute? No. She wants Lenny’s judgment on the topic, not a digestion of the text.

Spielman files Audio Overview as a different medium that happens to sound like a podcast. Reaction videos are popular for a reason: people come for specific humans in a space, not for a generic recap. Users share the shows they generate, but the posture is “I made this for me.” Nobody actually wants you to sit through their resume. The 2004 diary TikTok is interesting because of the woman’s reaction to her own past, not because a podcast about that diary should exist.

That is the point of personalized generation. Huang asks, in a world of chat boxes, where the idea of listening came from. Martin circles back to Spielman’s skeuomorphic instinct: deliver something new in a form people already understand, or they will not try it. Monologue, dialogue, a user switch—they tried the set. Dialogue sounded like a podcast, not like a screen reader, and delight told them what to ship. YouTube as an output is “a cost problem”; video can already come in as a source. Two days before the taping, a user asked them to output better code. Podcasts are cool, the user said, but can you just do the practical outputs? Martin’s reaction was that it was already on the roadmap.

Grady asks the sensitive organizational question. The team looks like a startup: small, scrappy, shipping imperfect things into production and iterating in public—not the stereotype of Google. What did the company give them, and what mold did they break? Martin’s two gifts are specific. She can use models before they are fully ready, so she can design journeys around capabilities that are still six months out. And she is surrounded by smart, collaborative people who actually want to build cool things. With those two, she says, a product builder only has to execute.

The thing they broke is the habit of not shipping. Coming into Labs, she decided the first job was to ship, because not shipping is easier, and Google makes it easy to second-guess every side effect. The tactic is slightly embarrassing, and she hopes the engineers are not listening: fake deadlines. “Guys, October 10th, we have to do it.” Someone notices that is in two weeks. They crank. Almost nobody asks what is actually happening on October 10th. She has been running that loop for two years.

Spielman has been at Google for seven years and thinks “Google is slow” is a misread of teams that touch billions of users. NotebookLM sits in a sweet spot: incumbent models, compute, and data, plus a team of about ten that can still move. The roadmap, Martin says, is to take in every input that matters to you and let AI make something new. They will go deeper on podcasts because people care, and they will also do the unglamorous outputs—including that request for better code. The AI hosts, in the cold open, promised to keep learning. The ten people behind them intend to cash that check.