Google Labs
Google's experimental product team that incubates early AI products and tests them with real users.
Google Labs runs 20 to 30 experiments at once, shrank project teams from five to seven people to two to three, and uses 50 to 100 day milestones. NotebookLM, Flow, AI Studio and Beam started as small, obsessed groups, and Flow grew out of Whisk. A past team of 30 engineers with no product-market fit ended badly. Lesson: staff small, let teams call kills, and hire slowly.
- 1List the almost possibleRunning list plus 82 predictions
- 2Spot the crossingA model or demo makes it real
- 3Staff two or threePeople who cross PM, code, design
- 4Watch eyes, not DAULean in, behavior shift, willingness to pay
- 5Hit a milestoneWithin 50 to 100 days
From Josh Woodward on Google Labs
What was said about Google Labs
1How NotebookLM and Audio Overviews were built
NotebookLM grew from a tiny side project around one idea: talk to your content. It started around 2022 as a 20% project built on a smaller Labs prototype, Talk to Small Corpus, which let an LLM interact with a single piece of content. The early group was Raiza Martin, one full-time engineer and, later, author Steven Johnson.
Only the engineer was full-time; the others joined because the idea was interesting. Martin says "20% project" undersells it, since the effort was far more than that.
Audio Overviews came from shaping an existing audio model into a familiar podcast form. Another Labs team offered powerful audio models and asked for a good application. The team's hypothesis was that AI outputs were always text, and voice output would change how people interact with and feel about the technology.
Labs often starts from a technology rather than a problem, so it forms a hypothesis about the product's shape before launch to get the most learning. The first test podcast was made by two teammates at 6:30 p.m. on a Tuesday, and when the AI hosts made dry British Parliament debates interesting, the team knew to build on it.
The quality came from heavy tuning, not just the base model. Audio Overviews run on Gemini 1.5 Pro with an audio model on top, and a Content Studio, built mainly by engineer Usama, brings out the best of the audio model. Early attempts were not nearly as good, so the team spent a lot of time listening to get the model to behave this way.
The hosts' sign-offs such as "stay curious" were never scripted; the model picks them, and recurring catchphrases emerged on their own.
A very small team with few processes shipped fast. About eight engineers worked on NotebookLM over the year, and three at the first announcement. PMs, engineers and designers write mocks and requirements in one meeting while engineering is already implementing.
Martin wanted a Discord server from day zero; colleagues asked why not Google Meet or a Google group, then let her go ahead. Her biggest fear was that nobody would join.
Personal, playful uses pulled in a much wider audience than planned. People enjoyed it most on their own documents: resumes, quarterly check-in notes, a father's hospital bio. The Discord reached about 60,000 members and retention rose on daily, weekly and monthly measures.
Early users were mostly educators and learners; the number one use case is still students turning study material into audio. Businesses now adopt it faster than the team can build for them, including one that found staff using personal Gmail accounts and asked to make work use official.
The roadmap aims for any input to become any output. Martin drew the vision two years earlier: take video, audio, emails or posts and turn them into a blog post, tutorial video or chatbot. Mobile is the biggest gap, and plain knobs and sliders were rethought because they did not feel magical.
Most user requests already fit that vision. Format depends on mood, such as audio on a walk and text at work.
2The "almost possible" list and where Labs ideas come from
Keep a list of what AI almost can do and act when one crosses over. Labs tracks experiences that are not quite possible, plus 82 predictions about the future, and puts a small team on one the moment a model or demo makes it real. A demo Josh Woodward saw 48 hours before his interview became a project within those 48 hours.
Woodward compares the crossing to a phase shift. The capability must also match a real user problem, ideally one people would pay to solve. "We talk about what's almost possible. We just have a list actually."
Pick futures to explore and list the tech thresholds each one needs. Labs asks what software, creativity, work and entertainment will look like, then lists thresholds such as model latency, number of languages and modalities. Immersive entertainment is the area watched most closely now.
Behavior is studied well beyond the Bay Area, in other cities and classrooms. Woodward expects to be wrong nine times out of ten but says a team needs a point of view.
Ideas come from obsessed side projects, not scheduled ideation. NotebookLM, Flow, AI Studio and Google Beam each began with a very small, obsessed group hacking over weekends, breaks or in hallways. The lab's job is to back those people and funnel the idea forward.
Flow grew out of Whisk, an experiment that blended images and animated them. "You can't microwave good ideas."
Fund people who live at the frontier and connect them to the core. Every company needs people testing every new model, the ones with "eight $200 a month subscriptions on everything," who know the latest viral thing within 24 hours.
They also need a place to build and a bridge to the main business, or the rest of the company dismisses them as an R&D group making toys.
3Judging early fit and deciding when to kill
At the prototype stage, watch people's eyes, not dashboards. Labs takes prototypes outside the building and looks for eyes lighting up and people leaning in, because DAU, MAU and D7 retention are too thin to mean anything that early. A Gemini feature was pulled before launch after in-person testing showed the reaction was not there.
Excitement alone is a weak bar since dazzling demos arrive hourly, so pair it with real behavior: a group who can't live without it and people willing to pay. The "eyes light up" test has not been validated against cases where it misled a team.
Fall in love with the problem, because pivots are certain. Woodward expects three, four or five pivots before an idea works or dies. Teams attached to a solution suffer through them; teams attached to the problem change the solution and carry on.
When real fit arrives it is obvious, and the team swarms it, as happened with Nano Banana on Gemini, NotebookLM podcasts, Flow and Stitch.
Let the team make the kill call, and thank them publicly. The team usually knows an idea is dead before the leader, from fading passion and having tried everything. When a Gemini PM recommended against launching after early data fell flat, Woodward replied to all to thank her.
The leader's job is to make uncomfortable questions safe, so the team can call out good and bad and pivot without blame.
Name the stages: zero to one, one to ten, ten to a hundred. Each stage needs different metrics and people: reactions at zero to one, waterfalls, CPAs and conversion to magic moments at ten to a hundred.
Naming stages matters because large graduated products, such as Flow now, can crowd out zero-to-one work.
4How Google Labs is set up and run
Give the lab independence, senior air cover and a clear goal. Woodward says a lab cannot sit under an existing business unit, where weird ideas die; some companies need it reporting to the CEO. Decide whether it graduates features into existing products or invents new product lines, and define winning and losing.
Google Labs, about three years old at the NotebookLM interview, started as a graduation lab and expanded to new categories under a VP mandate to "ship AI products and build businesses out of them," with "far, far, far fewer processes, maybe even to a fault." Woodward's claim that big-company labs fizzle in three to four years comes with no data or named examples.
Run many experiments with teams of two to three. Labs runs 20 to 30 experiments at once, and project teams shrank from five to seven people to two to three, maybe four, as AI cut the cost of building.
People who can cross PM, code and design are swapped between projects like musicians playing different songs together.
Plan in rolling six-month windows with 50 to 100 day milestones. Labs plans around the roughly six-month cadence of new model releases and keeps a multi-year vision instead of a multi-year roadmap. The best teams reach a meaningful milestone in 50 to 100 days.
The year runs in named seasons: explosive stretches around big model launches that people can opt in or out of, then hacking periods after events like Google I/O that seed the next season. "Never waste an event or like a date."
Celebrate small user wins and put users ahead of the product. Labs celebrates 10,000-user wins that other Google dashboards would not count, under the mantra users first, Google second, product third.
The same mantra links Labs to the Gemini app, which has over a billion users and shares its zero-to-one people and priorities.
Use the lab to pilot new ways of working. Seeing roles blur, Labs piloted a builder job ladder before it rolled out across Google, which Woodward calls creating "good trouble."
Treat headcount growth as a vanity metric. Woodward once grew a team to 30 engineers with zero product-market fit and an inspiring vision, and it ended badly.
Doubling or tripling a team in a quarter says nothing about fit, so he advises living with some organizational messiness before consolidating.
5Who Labs hires
Hire for unlearning rate, explosive endurance and trust building. Woodward looks at how fast someone drops what they learned once testing shows it is wrong, whether they can sprint hard yet pace themselves over years, and how well they build trust with people and agents.
"What's their unlearning rate? Like how fast can they learn something and then walk away from it?"
Favor builders who show what they just built. Strong candidates cannot stop building, read and listen widely, care about people's problems over ego, and are energized by unstructured situations. Showing something they just built gives instant credibility in interviews.
Treat PM as the major and builder skills as the minor. Roles are not collapsing into one universal builder; specialties still matter. PMs adapt fastest in builder pilots because they are used to simulating other functions.
People from strategy, ops, BD and marketing also make strong builders, and early-career PMs should code and prototype in service of becoming excellent PMs. "Don't lose that major."
Why every company now needs to think and operate like a lab team | Josh Woodward (VP Google Labs)
Woodward's team runs 20 to 30 experiments at once, with project teams shrinking from five to seven people to two to three. It celebrates 10,000-user wins, uses 50 to 100 day milestones, and piloted a builder job ladder before a Google-wide rollout. Flow grew out of Whisk, which blended images and animated them. It began with a small obsessed group and is now one of the large products that risk crowding out zero-to-one work. Whisk blended images and animated them, and that experiment led the team to build Flow.
Audio Overviews came from shaping an existing audio model into a familiar podcast form. Another Labs team offered powerful audio models and asked for a good application. The team's hypothesis was that AI outputs were always text, and voice output would change how people interact with and feel about the technology.
Labs often starts from a technology rather than a problem, so it forms a hypothesis about the product's shape before launch to get the most learning. The first test podcast was made by two teammates at 6:30 p.m. on a Tuesday, and when the AI hosts made dry British Parliament debates interesting, the team knew to build on it.
Keep a list of what AI almost can do and act when one crosses over. Labs tracks experiences that are not quite possible, plus 82 predictions about the future, and puts a small team on one the moment a model or demo makes it real. A demo Josh Woodward saw 48 hours before his interview became a project within those 48 hours.
Woodward compares the crossing to a phase shift. The capability must also match a real user problem, ideally one people would pay to solve. "We talk about what's almost possible. We just have a list actually."
Pick futures to explore and list the tech thresholds each one needs. Labs asks what software, creativity, work and entertainment will look like, then lists thresholds such as model latency, number of languages and modalities. Immersive entertainment is the area watched most closely now.
Behavior is studied well beyond the Bay Area, in other cities and classrooms. Woodward expects to be wrong nine times out of ten but says a team needs a point of view.
Ideas come from obsessed side projects, not scheduled ideation. NotebookLM, Flow, AI Studio and Google Beam each began with a very small, obsessed group hacking over weekends, breaks or in hallways. The lab's job is to back those people and funnel the idea forward.
Flow grew out of Whisk, an experiment that blended images and animated them. "You can't microwave good ideas."
Fund people who live at the frontier and connect them to the core. Every company needs people testing every new model, the ones with "eight $200 a month subscriptions on everything," who know the latest viral thing within 24 hours.
They also need a place to build and a bridge to the main business, or the rest of the company dismisses them as an R&D group making toys.
At the prototype stage, watch people's eyes, not dashboards. Labs takes prototypes outside the building and looks for eyes lighting up and people leaning in, because DAU, MAU and D7 retention are too thin to mean anything that early. A Gemini feature was pulled before launch after in-person testing showed the reaction was not there.
Excitement alone is a weak bar since dazzling demos arrive hourly, so pair it with real behavior: a group who can't live without it and people willing to pay. The "eyes light up" test has not been validated against cases where it misled a team.
Fall in love with the problem, because pivots are certain. Woodward expects three, four or five pivots before an idea works or dies. Teams attached to a solution suffer through them; teams attached to the problem change the solution and carry on.
When real fit arrives it is obvious, and the team swarms it, as happened with Nano Banana on Gemini, NotebookLM podcasts, Flow and Stitch.
Let the team make the kill call, and thank them publicly. The team usually knows an idea is dead before the leader, from fading passion and having tried everything. When a Gemini PM recommended against launching after early data fell flat, Woodward replied to all to thank her.
The leader's job is to make uncomfortable questions safe, so the team can call out good and bad and pivot without blame.
Name the stages: zero to one, one to ten, ten to a hundred. Each stage needs different metrics and people: reactions at zero to one, waterfalls, CPAs and conversion to magic moments at ten to a hundred.
Naming stages matters because large graduated products, such as Flow now, can crowd out zero-to-one work.
Give the lab independence, senior air cover and a clear goal. Woodward says a lab cannot sit under an existing business unit, where weird ideas die; some companies need it reporting to the CEO. Decide whether it graduates features into existing products or invents new product lines, and define winning and losing.
Google Labs, about three years old at the NotebookLM interview, started as a graduation lab and expanded to new categories under a VP mandate to "ship AI products and build businesses out of them," with "far, far, far fewer processes, maybe even to a fault." Woodward's claim that big-company labs fizzle in three to four years comes with no data or named examples.
Run many experiments with teams of two to three. Labs runs 20 to 30 experiments at once, and project teams shrank from five to seven people to two to three, maybe four, as AI cut the cost of building.
People who can cross PM, code and design are swapped between projects like musicians playing different songs together.
Plan in rolling six-month windows with 50 to 100 day milestones. Labs plans around the roughly six-month cadence of new model releases and keeps a multi-year vision instead of a multi-year roadmap. The best teams reach a meaningful milestone in 50 to 100 days.
The year runs in named seasons: explosive stretches around big model launches that people can opt in or out of, then hacking periods after events like Google I/O that seed the next season. "Never waste an event or like a date."
Celebrate small user wins and put users ahead of the product. Labs celebrates 10,000-user wins that other Google dashboards would not count, under the mantra users first, Google second, product third.
The same mantra links Labs to the Gemini app, which has over a billion users and shares its zero-to-one people and priorities.
Use the lab to pilot new ways of working. Seeing roles blur, Labs piloted a builder job ladder before it rolled out across Google, which Woodward calls creating "good trouble."
Treat headcount growth as a vanity metric. Woodward once grew a team to 30 engineers with zero product-market fit and an inspiring vision, and it ended badly.
Doubling or tripling a team in a quarter says nothing about fit, so he advises living with some organizational messiness before consolidating.
Hire for unlearning rate, explosive endurance and trust building. Woodward looks at how fast someone drops what they learned once testing shows it is wrong, whether they can sprint hard yet pace themselves over years, and how well they build trust with people and agents.
"What's their unlearning rate? Like how fast can they learn something and then walk away from it?"
Favor builders who show what they just built. Strong candidates cannot stop building, read and listen widely, care about people's problems over ego, and are energized by unstructured situations. Showing something they just built gives instant credibility in interviews.
Treat PM as the major and builder skills as the minor. Roles are not collapsing into one universal builder; specialties still matter. PMs adapt fastest in builder pilots because they are used to simulating other functions.
People from strategy, ops, BD and marketing also make strong builders, and early-career PMs should code and prototype in service of becoming excellent PMs. "Don't lose that major."
Behind the product: NotebookLM | Raiza Martin (Senior Product Manager, AI @ Google Labs)
Started from scratch with a VP-level mandate and few processes, so small teams can work like startups, such as picking outside tools like Discord without justifying them. It often starts from a technology and shapes the product around it.
NotebookLM grew from a tiny side project around one idea: talk to your content. It started around 2022 as a 20% project built on a smaller Labs prototype, Talk to Small Corpus, which let an LLM interact with a single piece of content. The early group was Raiza Martin, one full-time engineer and, later, author Steven Johnson.
Only the engineer was full-time; the others joined because the idea was interesting. Martin says "20% project" undersells it, since the effort was far more than that.
Audio Overviews came from shaping an existing audio model into a familiar podcast form. Another Labs team offered powerful audio models and asked for a good application. The team's hypothesis was that AI outputs were always text, and voice output would change how people interact with and feel about the technology.
Labs often starts from a technology rather than a problem, so it forms a hypothesis about the product's shape before launch to get the most learning. The first test podcast was made by two teammates at 6:30 p.m. on a Tuesday, and when the AI hosts made dry British Parliament debates interesting, the team knew to build on it.
The quality came from heavy tuning, not just the base model. Audio Overviews run on Gemini 1.5 Pro with an audio model on top, and a Content Studio, built mainly by engineer Usama, brings out the best of the audio model. Early attempts were not nearly as good, so the team spent a lot of time listening to get the model to behave this way.
The hosts' sign-offs such as "stay curious" were never scripted; the model picks them, and recurring catchphrases emerged on their own.
A very small team with few processes shipped fast. About eight engineers worked on NotebookLM over the year, and three at the first announcement. PMs, engineers and designers write mocks and requirements in one meeting while engineering is already implementing.
Martin wanted a Discord server from day zero; colleagues asked why not Google Meet or a Google group, then let her go ahead. Her biggest fear was that nobody would join.
Personal, playful uses pulled in a much wider audience than planned. People enjoyed it most on their own documents: resumes, quarterly check-in notes, a father's hospital bio. The Discord reached about 60,000 members and retention rose on daily, weekly and monthly measures.
Early users were mostly educators and learners; the number one use case is still students turning study material into audio. Businesses now adopt it faster than the team can build for them, including one that found staff using personal Gmail accounts and asked to make work use official.
The roadmap aims for any input to become any output. Martin drew the vision two years earlier: take video, audio, emails or posts and turn them into a blog post, tutorial video or chatbot. Mobile is the biggest gap, and plain knobs and sliders were rethought because they did not feel magical.
Most user requests already fit that vision. Format depends on mood, such as audio on a walk and text at work.
Give the lab independence, senior air cover and a clear goal. Woodward says a lab cannot sit under an existing business unit, where weird ideas die; some companies need it reporting to the CEO. Decide whether it graduates features into existing products or invents new product lines, and define winning and losing.
Google Labs, about three years old at the NotebookLM interview, started as a graduation lab and expanded to new categories under a VP mandate to "ship AI products and build businesses out of them," with "far, far, far fewer processes, maybe even to a fault." Woodward's claim that big-company labs fizzle in three to four years comes with no data or named examples.
More in Photo & Video
Photo & VideoCanva
An online design platform for presentations, social graphics, docs, whiteboards, video and more.
13 sources · 46 tipsOh Snap
A camera app launched in 2014 as an alternative to Apple's default camera.
4 sources · 8 tipsReelshot
A simple video app that was featured by Apple on the App Store.
2 sources · 11 tips
Photo & VideoGlam Up
A consumer app whose founders grew it through short-form UGC videos.
2 sources · 10 tips
Photo & VideoPhotoRoom
An AI photo editing app that removes and generates backgrounds for online sellers.
4 sources · 17 tips
Photo & VideoCaptions
An AI app for generating and editing talking videos, aimed at people who could not make video before.
2 sources · 14 tips
Photo & VideoMojo
A mobile-first app for creating animated social media content.
2 sources · 14 tips
Photo & VideoHalo AI
An AI photo editor where users upload a photo, write a prompt and get an edited image.
1 source · 16 tips