Format Malleability: Any Input, Any Output
Stop shipping content in one format — give people the power to remix knowledge into the medium they're in the mood for
- Difficulty
- Moderate
- Time to result
- ~months to results
- Steps
- 5
- Confidence
- 90%
Martin's long-held product vision is an AI editor surface that is fully remixable: feed it anything (video, audio, emails, LinkedIn, Twitter, a 100-page doc) and get back anything (a blog post, a tutorial video, a chatbot, a podcast). The strategic claim is that the format of information is currently imposed by the sender, and handing that choice to the receiver changes the relationship between a person and knowledge.
Origin
Raiza Martin's own vision slide, made two years before this 2024 conversation and coloured lime green as a deliberate 'un-Googly' signal. It predates and underlies both NotebookLM and Audio Overviews.
Core principles
- 01Almost every request people make of AI reduces to: take this thing and make it into something new.
- 02Format is not a property of content; it is a preference of the consumer.
- 03Mood is a real input — audio on a walk, text at a desk.
- 04Today the receiver must accept the format the sender chose; that asymmetry is the opportunity.
- 05A format the receiver chose gets consumed; one they didn't often doesn't (the unread 100-page doc).
How to run it
- 1
Audit where format friction kills consumption
Find the artefacts in your world that go unconsumed purely because of their form — the 50-page strategy doc, the dense research paper, the study materials nobody rereads. These are not information problems, they are format problems.
Pro tip Martin's own tell: she was handed a 50-page vision doc by her boss and simply never read it.
- 2
Decouple the source from its rendering
Treat every artefact as a source to be grounded on, not a document to be read. The input set should be as wide as the world people care about: video, audio, emails, LinkedIn, Twitter, PDFs, resumes.
Watch out Grounding matters — outputs that drift from the source destroy the trust that makes remixing usable.
- 3
Give the receiver the output choice
Let the consumer name the medium: make me a blog post, a tutorial video, a chatbot, an audio overview, a study guide. The interface is an editor for shaping, not a viewer for reading.
Pro tip The cheapest place to start is one high-contrast output (audio) against an all-text baseline — the contrast is what makes the value obvious.
- 4
Design for mood-based routing
Assume the same person wants different formats at different moments. Build so that the choice is made per-session, not per-user profile.
Pro tip Test with the walk/desk pair: what should this artefact become if I'm walking, versus if I'm at my desk?
- 5
Use interrogation as a format
The most under-rated output is conversation. Rather than reading a long document, turn its author (or its content) into something you can question — Martin literally chat-botted her boss instead of reading his 50-page doc.
Pro tip Q&A over a source is often faster than any summary because it lets the reader target only what they don't know.
Watch out This works only when the source is genuinely available to interrogate; a summary of a summary compounds error.
In the wild
When Martin joined Labs, Josh Woodward gave her a 50-page document containing his vision. Instead of reading it, she grilled him with questions like a chatbot, until he told her the answers were in the doc.
→ She got the vision faster in the format she wanted, and the anecdote became a live demonstration of the product thesis: chat is easier than reading.
Lenny fed his mother's short autobiography PDF into NotebookLM, generated an Audio Overview, and also sent her the auto-generated study guide from the same source.
→ His mother shared it with friends and organised a dinner around reviewing the study guide of her own biography — the same content consumed in two receiver-chosen formats.
Andrej Karpathy took Wikipedia articles about historical mysteries and used NotebookLM to render them as a ten-episode podcast series published on Spotify.
→ A publishable, listenable product built entirely by re-rendering existing text sources into a chosen medium — which Martin called a great product in its own right.
Common mistakes
Assuming the sender's format is the right one
The 100-page doc goes unread not because the information is wrong but because the medium doesn't fit the moment. Teams keep optimising the document instead of letting the reader change its form.
Building one-way format conversion instead of an editor
A summariser is a fixed pipe. The vision is an editor surface where any input can become any output, and the person doing the choosing is the consumer, not the tool's author.
Ignoring emotional and situational context
Format preference shifts with mood and setting. Locking a user into one modality via their profile settings misses that the same person wants audio on a walk and text at work.
Is it for you?
Best for
Founders and product leaders building knowledge, learning or content tools who want a durable long-term vision rather than a single feature
Not ideal for
Regulated or compliance contexts where the exact artefact and its form are legally load-bearing and must not be re-rendered
From the transcript
“I imagined that in the future you could have an AI editor surface right fully remixable any input any output right”
“for most of the things that I'm hearing people say they want to do it's usually that it's like take something and make it into…”
“I mean that's exactly it and I think even for myself it depends on my mood right if I'm going on a walk yeah I…”
“instead of reading it I just grilled him I was just like q& aing him like a chat bot”
From the episode
Behind the product: NotebookLM
Raiza Martin (Senior Product Manager, AI @ Google Labs)