Lenny's Podcast · Oct 10Behind the product: NotebookLM | Raiza Martin (Senior Product Manager, AI @ Google Labs)
Grew from a 20% project into a product with about 60,000 Discord members and rising daily, weekly and monthly retention. Audio Overviews, a two-host podcast built by about eight engineers, drove its viral spread, and businesses are adopting it faster than the team can build for them.
Tips
- Wrap a powerful model in a surprising, fun format▶ 12:07Audio Overviews drove NotebookLM's viral spread across social media.
- Keep the team tiny and cross-functional in the early stage▶ 21:21Fewer than 10 engineers throughout the first year, until more joined for the future roadmap.
- Give a new AI team clear senior air cover and fewer processes▶ 21:13NotebookLM went from 20% project to a product with a 60,000 member Discord in about a year.
- Start a Discord community on day zero▶ 22:58About 60,000 people are in the NotebookLM Discord.
- When starting from technology, hypothesize the product's shape▶ 10:21
- Pair the model with an opinionated content layer▶ 13:25Early attempts were not nearly as good as the shipped result, which came from heavy listening and tuning.
- Let users convert any input into any output format▶ 28:13She notes that she has often been handed a 100-page doc and not read it, but would turn it into audio.
- Make user controls feel magical, not just knobs and sliders▶ 29:49
- Seed first use with sources users already have▶ 18:46Many Googlers messaged Martin saying hearing their check-in notes as an audio overview boosted their confidence before meetings.
- Watch for enterprise pull from shadow usage▶ 25:35Business use is described as astronomical, and demographics shifted from mostly educators and learners to many professionals.
- Build something interesting first, then work out the business▶ 26:53Retention across daily, weekly and monthly measures rose strongly in the first year, which gave stakeholders confidence.
- Check whether viral AI moments come from user inputs before reacting▶ 44:54
- Expand red-team test cases whenever users find new edge cases▶ 45:43
- Find out how often users read a PM's long doc, then chat with it▶ 32:05
