LLenny's Podcast
← All frameworks
InnovationRaiza Martin (Senior Product Manager, AI @ Google Labs)

Technology-First Discovery with a Shape Hypothesis

When you start from a capability instead of a problem, form a hypothesis about its shape before you ship it

Difficulty
Advanced
Time to result
~months to results
Steps
5
Confidence
92%

Most product work starts from a user problem and searches for a solution. Frontier-technology work inverts that: you hold a raw capability and must find its practical application. Martin's answer is neither pure problem-first nor pure ship-it-and-watch — it is to commit to an explicit hypothesis about the 'shape' the technology should take for humans, ship that specific shape, and learn from a sharp bet rather than an open sandbox.

Origin

Developed by Raiza Martin as product lead at Google Labs, contrasting her prior problem-first PM training (payments, ads) with how Labs actually incubated NotebookLM and its Audio Overview feature out of a 'talk to small corpus' experiment and a sibling team's audio model.

Core principles

  • 01In Labs-style work the technology is the starting input, not the solution you reach for.
  • 02Raw capability is not a product — it has to be shaped and brought closer to people.
  • 03Dumping a tool in the market and studying usage is a valid but low-information approach; a hypothesis yields more learning per release.
  • 04The right shape is recognisable by reaction: people look at it and say 'I get it'.
  • 05Find the modality gap — where the input is rich but the output is still the same old format.

How to run it

  1. 1

    Inventory the capability, not the problem

    Start from what the technology can now do that it could not do 6-12 months ago. For NotebookLM this was Gemini 1.5 Pro's long-context grounding plus another Labs team's powerful audio model. Write down the capability in plain language before imagining any UI.

    Pro tip Actively scan adjacent internal teams for capabilities that pair with yours; the audio model came from a different team inside Labs, not from Martin's own roadmap.

  2. 2

    Find the nugget

    Look for the small kernel inside the capability that is genuinely interesting and keep revving on it until it becomes useful. The nugget for audio was: people could already interact with their sources via text, but every output was still text.

    Pro tip Phrase the nugget as a gap sentence: 'Users can already X, but the output is still always Y.'

    Watch out A nugget is not yet a product. Do not staff it heavily until you have a shape hypothesis.

  3. 3

    Write the shape hypothesis

    Commit, in advance, to a specific form the capability should take in people's hands: what they give it, what comes back, how it should make them feel. The hypothesis for Audio Overviews: give it almost nothing (a URL, a resume), press one button, and get back something surprising, complete and fun.

    Pro tip Include emotional payload in the hypothesis. Martin explicitly designed for surprise, delight and fun, not just utility.

    Watch out Skipping this step is the 'just put the tool out there' path — a fine approach, but it maximises optionality at the cost of learning.

  4. 4

    Ship the shape, not the sandbox

    Ship the one opinionated format rather than a configurable playground. NotebookLM shipped a single 'Deep Dive' podcast format, one-shot, no controls — a bet sharp enough that its reception was interpretable.

    Watch out One-shot opinionated output frustrates power users. Accept that cost deliberately; it is the price of a legible signal.

  5. 5

    Iterate on the shape until people 'get it' instantly

    Treat shape as the iteration variable. Keep reshaping the same underlying tech until a naive observer's reaction is immediate comprehension and delight — the ChatGPT-moment test, where the model already existed and only the medium changed.

    Pro tip Use the reactions of people who have no idea what you do (spouses, parents) as your comprehension test.

In the wild

Audio Overviews from an existing audio model

Another Google Labs team had powerful audio models and asked what a good application would be. Rather than exposing text-to-speech, Martin's team hypothesised a shape: minimal input, one button, and a complete, surprising two-host conversation about your source.

The feature went viral and became the most-discussed AI product of the moment, with no change to the underlying model class — only the shape.

NotebookLM out of 'talk to small corpus'

A small Labs project explored using an LLM to interact with a piece of content. The team identified the nugget, kept revving on it, and shaped it into a source-grounded notebook rather than a generic chatbot.

Grew from a 20% project with one full-time engineer into a product with rising daily, weekly and monthly retention and enough enterprise demand to force a business-development hire.

Common mistakes

Reaching for a problem statement you do not have

Teams handed a new capability often retrofit a fake user problem to justify it. Martin is explicit that in Labs the honest starting point is the technology, and pretending otherwise produces a solution nobody actually shaped for humans.

Shipping the raw capability as a playground

Putting the tool out there and studying how people use it is a fine approach but yields the least learning. Without a shape hypothesis you cannot tell whether a weak response means the tech is wrong or your form was.

Shipping utility with no emotional payload

The shape hypothesis included fun and surprise on purpose. A technically correct output that nobody wants to share fails the 'wow, I get it' test that signals you have found the shape.

Is it for you?

Best for

Product leads and founders inside AI/R&D labs who are handed a new model capability and must decide what to build with it, without a pre-existing user problem to anchor on

Not ideal for

Mature products with a known user problem, a validated funnel, and business-outcome commitments — there, problem-first discovery remains correct

From the transcript

in labs in particular we start with the technology and it's actually a very interesting place to start where you're like okay what are the…

10:00

we had this ability for you to interact with text but people were like the output is still always text right

10:30

a lot of technology I feel like you have to shape it and bring it closer to people and I think it's like such an…

12:00

From the episode

Behind the product: NotebookLM

Raiza Martin (Senior Product Manager, AI @ Google Labs)