Josh Woodward on Building an AI Labs Team

Watch someone's eyes, not their dashboard. That is the strange, almost anti-corporate advice at the center of how Google builds new AI products, and it comes from a man who has shipped more experimental software than almost anyone alive. Josh Woodward runs Google Labs, the Gemini app, and AI Studio, a portfolio that spans brand-new zero-to-one prototypes on one end and a product with more than a billion users on the other (19:39). He has been at Google for sixteen years (01:29), long enough to watch entire product categories rise and die, and his central claim is simple: the pace of change in AI has made every company, whether a five-person startup or a trillion-dollar search engine, need to think and operate like a lab.

That claim could easily curdle into consultant-speak. What saves it is specificity. Woodward does not say 'innovate faster.' He describes, in granular detail, what his team actually does when it has no idea what to build next, how it decides an idea is dying, and why the skills it hires for have quietly changed. The result is less a pep talk than a field manual, and it is worth taking apart piece by piece.

The Metric Is a Pupil

Start with where ideas come from, because Woodward's answer breaks a piece of standard corporate furniture: the design sprint. Years of watching Google Labs produce hits like NotebookLM, a tool that turns documents into AI-generated podcast conversations, and Flow, a video generation product, taught him that good ideas almost never arrive on a calendar invite.

"They very rarely, at least in my experience, have come from design sprints or times where you set aside on a calendar... You can't microwave good ideas." (06:17), Josh Woodward

Instead, they show up when a couple of obsessed engineers can't stop tinkering over a weekend and drag a colleague over to watch a demo at 6:30 on a Tuesday, which is exactly how the first NotebookLM podcast got made (13:27). To give that obsession somewhere to point, Labs keeps a running document of what Woodward calls 'almost possible' technologies (08:17), a watchlist of capabilities that are just out of reach, so that when one crosses the line (faster model latency, a new modality, a longer context window) the team notices immediately, like watching ice turn to water. The document underneath that thinking contains 82 specific predictions about the future, each beginning with the phrase 'we believe the future is' (11:28). Woodward is candid that something like 9 out of 10 of those predictions will be wrong. The point of writing them down isn't accuracy, it's forcing the team to hold an opinion about where things are headed instead of waiting for permission.

Once an idea exists as a rough prototype, the harder problem starts: knowing whether it's working. Here Woodward rejects the standard toolkit almost entirely. Daily active users, retention curves, even the newest benchmark scores, none of it matters in the first weeks of an unproven idea. What matters is watching a person's face while they use the thing.

"When you're showing people early prototypes... you're looking at people's eyes, and that is the metric." (15:52), Josh Woodward

If someone's eyes widen and they lean toward the screen, that is product-market fit in its earliest, truest form. Dashboards, he says, don't even register counts that low. Labs teams will celebrate hitting 10,000 monthly active users (20:00), a number most mature products would consider a rounding error. The lesson scales down well: a founder pitching an idea before a single metric exists has nothing more reliable to go on than whether a stranger's attention visibly sharpens.

Killing an idea runs on the same intuition, but in reverse, and it surfaces a leadership blind spot. Conventional wisdom says the boss decides when to pull the plug. Woodward has found the opposite is almost always true inside Labs.

"The team usually knows before the leader knows, and the signs are like, does the passion start to run out?" (18:19), Josh Woodward

He describes a recent case where a product manager on his team pulled a feature the day before launch because early user tests showed the spark wasn't there, and Woodward's response was to publicly thank her rather than push back (19:01). The mechanism that makes this possible is a stated priority order the whole team repeats: users first, Google second, the specific product third (21:08). Because nobody's job security is tied to a particular feature surviving, admitting failure early costs less than hiding it.

Hiring for the Unlearning Rate

If ideas and metrics explain how Labs finds and judges products, the second half of the conversation explains who can actually execute inside that kind of uncertainty, and the answer has shifted the shape of Woodward's hiring. He has grown skeptical of 'learning rate,' the popular shorthand for growth mindset, and replaced it with something sharper.

"I'm actually looking a lot right now for people, what's their unlearning rate? How fast can they learn something and then walk away from it?" (32:13), Josh Woodward

In a field where the state of the art changes every few weeks, the skill that compounds isn't knowledge, it's the willingness to discard knowledge that just expired. Paired with that is a phrase Woodward picked up from a biography of tennis player Roger Federer: explosive endurance (32:43), the capacity to sprint at full intensity on a launch and then recover, rather than burning out after nine months. Labs formalizes this with named 'seasons,' explicit windows where a team is told it's in sprint mode for a model release, followed by deliberate slack time, like the month after Google's developer conference I/O, when teams are told simply to go hack and rediscover what's interesting (40:15). Treating recovery as part of the plan, rather than an accident, is itself the protocol.

The other structural shift Woodward has made concerns team size. The standard zero-to-one product team used to run five to seven people. Now, with coding and design assistance built into AI tools, Labs builds teams of two to three (34:55). He compares it to musicians who form different small groups for different songs, with the 'songs' being the products themselves. Crucially, Woodward pushes back on the idea that everyone on these small teams becomes an interchangeable generalist. Product managers, he's found, adapt fastest to blurred roles because their job already involves simulating a lawyer, a salesperson, or a designer in a meeting to keep a project moving (36:32). But he insists specialization still matters.

"I actually feel like... people still have their specialty. It's like their major or minor in college." (37:35), Josh Woodward

A product manager who starts vibe-coding prototypes is meant to use that skill in service of becoming a better product manager, not to abandon the discipline altogether.

None of this works, Woodward argues, without the right organizational container. His historical reading is blunt: most corporate labs become ineffective and fizzle out within three to four years, usually because they get nested inside an existing business unit that eventually starves them, or because the parent company can never commercialize what the lab invents (48:19). His prescription has three parts: give the lab independence, often reporting close to the CEO; decide explicitly whether the lab's job is to graduate ideas into existing products or to invent entirely new categories (48:53); and use the lab for more than technology, including piloting new internal processes like a 'builder' job ladder that blurs engineering and product roles before rolling it out company-wide (49:48). Labs even keeps a hiring reference document, nicknamed 'Labs in a Nutshell,' with seventeen or eighteen specific traits new hires are screened against (51:15), everything from compulsive builders to people energized by genuine uncertainty.

It would be easy to read all of this as culture theater, the kind of thing a large company invents to feel scrappy. But the mechanisms underneath it, the almost-possible list, the eyes-not-dashboards rule, the users-first ordering, the unlearning rate, are specific enough to transplant. The uncomfortable implication for most organizations isn't that they need a foosball table and a cool name. It's that they need to tolerate not knowing, reward people for killing their own ideas, and measure success in a dilated pupil before anyone trusts a spreadsheet.