Chris Pedregal - Building Granola - [Invest Like the Best, EP.412]
What this covers
Patrick talks with Chris Pedregal, founder of Granola, an AI notepad that transcribes meetings and enhances the notes a user jots down. The conversation opens with history: how writing, notation and charts extended cognition, and how Larry Tesler's removal of editing modes from early text editors made something once unthinkable feel obvious. Pedregal takes that as a hint that AI interaction will change just as much. He then turns to Granola's own path. The first version wrote full notes in real time, but people read and edited them instead of listening, and the team dropped the approach after six months. Heavy users now type only a few private reactions and later ask the chat a specific question.
The range widens from there. On strategy, Pedregal sorts products into a Power Tool Quadrant: low-frequency, good-enough tasks get absorbed by general assistants, while high-frequency work that demands excellent output stays defensible. He describes startups as the R&D wing of big tech and says Granola switches between foundation models as the best one changes. He also asks teams to label work as explore or exploit.
The more contested material is about people. Pedregal predicts that recording will be accepted at work but will meet Google Glass-style backlash in social settings. He compares chat to a terminal era and the steering wheel that AI collaboration still lacks. He warns that picking from AI-generated ideas could homogenize thought. He also covers one-to-one tutoring, taste as a scarce skill, and a briefing delivered before every meeting.
Pedregal argues AI tools should extend human cognition by externalizing memory and delivering the right context at the moment of need, and that app-layer products win through user-experience design and accumulated context rather than through raw model intelligence.
- Tools for thought work by offloading what must be held in working memory, and LLMs can supply dynamically generated context on demand
- Model providers' intelligence is a shared substrate; differentiation comes from product design for a specific high-frequency use case
- Early product decisions (real-time note generation, Mac app, delayed launch) were reversed or protected to find the right interaction pattern
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Early text editors required switching between insertion, deletion, and copy modes, and Larry Tesler's work removed those modes so users could type, delete, cut, and copy fluidly, a change that was unthinkable before it happened and suggests AI interaction will similarly change.
“Larry Tesler basically went on a vendetta to change this. So now it's actually, you should be able to type and delete and cut and copy”
Studies show that a median student with a one-to-one tutor performs at the level of the top 5 to 10 percent of students, making one-to-one tutoring the holy grail of education technology, though business incentives around it are misaligned with social good.
“if you have a one-to-one tutor, the median student actually performs the top 5 or 10% percentile student.”
Modern foundation models need only a small amount of task-specific data, around 50,000 examples, compared with millions required under older machine learning paradigms, so getting data is less of a barrier to building AI products.
“Now it's kind of crazy. We can get away with 50,000 examples.”
Many AI features people use daily were built by startups, which act as the R&D wing for big tech companies, figuring out features that large companies later incorporate into massive user bases.
“Startups are oftentimes the R&D wing of all the big tech companies, and then once something's figured out, they can incorporate that”
If everyone asks AI to generate ideas and selects from them, people will only look at AI-generated ideas, causing a homogenization of thought.
“There's a danger, though, that that's what everyone is doing, and now we're only looking at the ideas that are coming from AI.”
Granola uses many models across different features and switches to whichever model is best on any given day, though building evaluations to determine that is not simple.
“I think eval is not simple. But what you described is exactly what we do. We don't just use one model in one place.”
Teams should explicitly label feature work as exploit mode, where the solution is known and should be executed fast, or explore mode, where the problem is open and needs investigation first, because the two require completely different approaches.
“are we in exploit mode or are we in explore mode? Because we need two completely different approaches to that.”
Products are vulnerable to being absorbed by general AI systems when a use case is low frequency and needs only good-enough quality, while high-frequency use cases where output must be excellent form a defensible power-tool quadrant.
“everything that is low frequency, where you don't need to be great at it, will be eaten up by the general system.”
A specialized product beats a general assistant on a high-frequency use case because the limitation is UI optimization for that use case rather than the underlying model's intelligence.
“it's not a question of intelligence, it's actually how great is the UI optimized for this use case?”
Granola's original design generated full notes in real time during meetings, but people kept reading and editing the notes, lost attention to the speaker, and the approach was abandoned after six months.
“no matter how great the notes we wrote were, if a computer is writing notes for you real time during a meeting, you can't help but read it.”
It is surprising that AI models remain largely unpersonalized, giving nearly identical answers to different users, while Granola personalizes notes per user because what matters to each person differs.
“It is surprising to me how un-personalized any of these models feel today.”
The real unlock of LLMs is that they can bring extremely relevant context to a person at the moment they need it, and that context can be dynamically generated to fit the needs of that specific moment, making people more capable in real time.
“the real unlock here is that you can use LLMs to bring extremely relevant context to the person in the moment they need it.”
Recording norms will be accepted in workplace meetings because the AI value is high for users, but in social settings AI recording will be a divisive issue, with backlash similar to that against Google Glass.
“I see a pretty stark distinction between in the workplace setting, most people want all these things captured”
As cameras became ubiquitous, everyone became a photographer, and this suggests that as software becomes easy to build, a premium will shift toward taste, so exceptional taste may make skilled builders more valuable.
“At the same time, I feel like there's a premium on taste now.”
Not everyone will build their own software because shared consciousness, social proof, and the perceived quality of popular products make people prefer common apps, similar to how popularity and labels shape how music and wine are experienced.
“I think there's some shared consciousness, shared experience thing that matters for how good something is.”
A new class of software, not yet named, will become where knowledge workers in meetings, communication, and projects spend their day, similar to how developers live in IDEs like Visual Studio or Cursor, and Granola aims to build that workspace from day one.
“We think that there's going to be a new class of software, it doesn't have a name yet, where people like you and I will spend all day in.”
AI tools should unlock human creativity and judgment; the danger is outsourcing judgment along with rote work, because if people outsource thinking such as writing, they lose the value of their own thinking.
“I think there's a fine line where you want to outsource all the rote work, all the boring stuff, the mindless stuff, but you really don't want to outsource the judgment.”
Current chat-based AI interaction is a bottleneck that is akin to a terminal era, and chat will persist but will feel archaic as users have little real control over outputs.
“I really think we're in the terminal era with multiple computers where you type in a command and then the computer would literally spit back a command.”
Early cars with a stick for steering were fine at low speed but dangerous at high speed, and the steering wheel gave fine-grained control; AI collaboration still lacks an equivalent fine-grained control, relying on coarse turn-taking.
“And a steering wheel is a UI that gives you so much fine-grained control when you're trying to turn.”
When a computer writes notes for users in real time during a meeting, people cannot help but read them, which leads them to edit notes and stop paying attention to the speaker, defeating the purpose of being present.
“if a computer is writing notes for you real time during a meeting, you can't help but read it.”
Heavy Granola users write only a few notes per meeting, usually internal reactions like concerns about a person, and rely on the AI transcript; afterward they ask the chat a specific question rather than reading through notes.
“they only write a couple notes per meeting and those notes are usually the internal thoughts that they had.”
A divide similar to the early Google search divide will emerge between AI natives, who intuitively know what context to give AI, and others who default to older habits.
“AI natives will just understand what context they need to give AI and how to work with AI.”
Collecting all of a person's data into an AI is becoming easy, but the hard problem is determining which context is relevant to the task at hand, which may be a technical or UI problem.
“Gathering the context – just getting all the data is not that hard.”
In the future everyone will receive a dossier-like briefing in real time before any meeting, similar to a diplomat's high-stakes negotiation briefing, and the open question is which context (previous meeting, all email, or all information) should feed it.
“I think we're going to live in a world where everyone's getting those in real time whenever they go into any meeting.”
Companies whose customer experience departments were created after 2025 will look very different from earlier ones, likely with far fewer people and different tooling, while older departments will struggle to change because changing existing structures is harder than building new ones.
“I'm pretty convinced that you're going to be able to look at a company and say, 'Was their customer experience department created before or after 2025'”
Pulling themes across all meetings with a given person or on a given topic makes context that would otherwise be lost in forgotten notebooks immediately accessible when making relevant decisions.
“You wrote it down somewhere, but you don't know where that notebook is.”
Granola transcribes audio in real time without storing any audio, which gives up some value such as tone of voice but makes it far less invasive than AI meeting bots that record and retain audio and video indefinitely.
“we made a very conscious decision not to record and store any audio.”
Building as an app-layer product on top of LLMs lets a small team deliver what would otherwise require a massive team, because the technical advances happen at the model layer.
“If we weren't building on top of this foundational technical layer like LLMs, we'd need a massive team to be able to do what we're doing today.”
Great product design can be guided by repeatedly asking how a product makes you feel, because within about 500 milliseconds a viewer registers many emotional reactions, and these reveal what needs fixing, such as clutter or insecurity.
“when you look at a product, you feel like 10 things. And oftentimes those things tell you ex. Oh, it's too complicated, it's too cluttered.”
Building Granola as a Mac app meant initially serving only users on macOS 13 or later, roughly 15 percent of Mac users at the time, but the choice made the product feel like a notebook always at hand and proved a better decision than expected.
“it was only possible to do what Granola does for users who are on, like, macOS 13, which was, I think, 15% of Mac users at the time.”
Granola spent a year before launching, which allowed the team to completely change the core interaction, a change that would have been impossible after public launch because users would have learned the old behavior.
“If we had launched that publicly, we never would have been able to switch it. There's no way. Because users would have learned a new behavior.”
The competitor Granola worries about most is the startup that has not launched yet, because it can learn from Granola's discoveries and execute faster from that point, rather than established big tech firms.
“the competitor that we have chosen to worry about at Granola is the one that hasn't launched yet.”
Educational tutoring should ideally be free and open-source because the incentives for profit in that space do not align with society's interest in widely available education, which is why Chris chose not to build an education business.
“I don't want to build a business there. The incentives around making money in that space versus I think what we kind of want for society aren't super aligned.”
In work settings a phone placed visibly on the table creates an easy social contract that makes clear recording is happening, and this transparency benefits everyone, unlike a hidden pendant listening constantly.
“I think in a work context the phone is great because you basically put it down on the table and it is an easy social contract”
AI at the app layer is an explore-mode problem, so the best investors bring deep product thinking and a shared long-term worldview rather than focusing on specific execution, which will change.
“I think AI as a whole is an explore mode problem. No one knows what the right thing is.”
AI-native users learn to supply the right context to AI tools, and when unsure they should give more context rather than assume; Chris's shrimp example showed a photo of the barbecue and the food yielded a critical insight that a text prompt would have missed.
“And actually when in doubt, you should probably give it more context and see what it's going to say as opposed to assume”
Defensibility in AI apps comes from building better products faster, and switching costs accumulate because the more context a tool like Granola holds, the more useful it becomes, so competitors must be much better to win users away.
“there's an inherent switching cost because the more context Granola has, the more useful it's going to be for you.”
The LLM idea of a context window, where a model is given the situation and what it needs to know, will also apply to people, who will constantly think about what context they need to be their smartest.
“That way of thinking is also going to apply to us, to people. We're going to be thinking about that all the time.”
Foundation models are moving toward exploit mode, but AI as a whole and especially the app layer remain explore-mode problems where no one knows the right answer, requiring product-centric sensibility.
“I think AI as a whole is an explore mode problem. No one knows what the right thing is.”
Once people use Granola for all important work calls they effectively outsource part of their long-term memory to it and come to expect to look up important information from any conversation.
“you basically are outsourcing some of your long-term memory to Granola.”
Granola's product philosophy centers on user control, so the notes are generated in an editable editor rather than a fixed PDF or email, because the user drives the tool and every design decision ties back to that principle.
“The Granola specific one is all about giving the user control. Granola is a tool to make you better, which means you drive the tool”
The standard advice to build the minimal thing and ship quickly can backfire when the team has not identified the real problem, because shipping a weak solution counts as progress while the actual problem remains unsolved.
“you end up shipping something crappy to people and you checked it off.”
Chris asks himself daily whether Granola is being built correctly, believing that all work tools will be rebuilt on LLMs and that a new class of workspace software for knowledge workers will emerge.
“We think that there's going to be a new class of software, it doesn't have a name yet, where people like you and I will spend all day in.”
Granola's guiding view is that AI should make people better by augmenting their abilities and intelligence, not by replacing a person or removing a task from them.
“we see the role of AI as being a tool to make you better.”
A well-designed product should deliver its core value moment within the first 20 seconds, but Granola's value only appears after a full meeting, which is a significant product disadvantage.
“Ideally, when you're building a product, you want that magic moment to happen in the first 20 seconds.”
Tools for thought work by letting people externalize what they would otherwise hold in their heads, so a notepad acts like extra RAM because human working memory is limited by physical constraints.
“it's a bit like extending the RAM. The amount of, like, RAM that we have in our heads are hard coded by physical limitations.”
Users cannot edit AI responses in tools like ChatGPT or Claude, so the human and AI work on separate canvases rather than a shared one, which will feel like madness once fluid collaboration exists.
“it doesn't feel like you and the AI are working on the same canvas. It's like we're working on two separate canvases next to each other.”
William Playfair, about 200 years ago, was the first to graph data visually, and mapping numbers to the visual plane lets people intuitively grasp trends such as rising or accelerating growth in a way raw numbers do not.
“He was the first person to graph data visually so you could use your eyes.”
Within roughly 18 months, doing meetings and work without an AI note tool like Granola will feel such an impediment that almost no one will want to do it.
“in a couple years, maybe 18 months speed of AI so fast, doing meetings, doing work without something like Granola will feel like such an impediment”
Granola should evolve from meeting notes into a tool that does most post-meeting work, such as follow-up emails, investment memos, and scheduling, taking the user 80 to 95 percent of the way with the user adding final judgment.
“Granola or a tool like Granola with all the necessary context should be able to take you 80, 90, 95% of the way there.”
Granola's CTO Vas has a standing goal of minimizing the number of lines of code each engineer writes, pushing the team to use AI for work they would not expect.
“our CTO Vas, he has a goal basically minimizing the number of lines of code every engineer writes at Granola.”
Humans are fundamentally toolmakers, and a central category of tools are tools for thinking, such as writing, mathematical notation, and data visualization, which let people do cognitive work they could not do unaided.
“fundamentally humans are toolmakers. It's one of the things that like just sets us apart from other animals.”
Roman numerals limit how much arithmetic a person can do in their head, whereas modern positional notation allows long division of very large numbers, showing how a notation system itself is a thinking tool.
“With Roman numerals, you can only do math up to a certain number in your head without an abacus, whereas with the notation we use now, you can do long division of massive numbers”
Much of what makes Granola good lies in handling technical edge cases, such as switching audio inputs mid-call, that users would never anticipate but that feel broken if handled poorly, and some needed development tools are not yet good enough to automate that work.
“a lot of what makes Granola great is sweating the details of all these technical edge cases, stuff you'd never think of.”
People are limited by information silos and only see data from one source at a time, so an ideal tool would surface the most relevant ideas from personal context and humanity's accumulated knowledge dynamically during thinking.
“we have so many silos of information and so many silos of where knowledge or inspiration, our information comes from.”
Real-time AI that augments a live conversation, such as generating visuals or notes, sounds compelling in sci-fi but often fails in practice because distraction undermines the benefit, which makes the human experience the key constraint.
“a lot of these ideas in sci-fi that sound fantastic and then in practice don't work for really silly tactical reasons”
Chris is happiest when building something he believes in and important, and he finds meaning through active realization of human potential, both in work and in parenting.
“I am happiest when I am trying to build something that I believe in and that I think is important”
The best investor relationships are long-term partnerships built on a shared worldview about how to approach problems, and investors who show deep product thinking in specific cold outreach stand out.
“what at least I want when I partner with an investor is I want a partner I'm going to work with for a very long time.”
Instead of worrying about general data access, builders should consider what kind of data is ungettable, since that is what provides durable advantage in a world where most data can be obtained at moderate cost.
“I'd think about what kind of data is ungettable.”
Users unexpectedly adopted Granola for work-adjacent personal needs, such as a partner with cancer who uses it to manage a series of doctor meetings, which they found invaluable.
“My partner has cancer, we have all these meetings with doctors. Granola has become absolutely invaluable in that process.”
Users creatively employ Granola as a personal brainstorming, planning, and learning notebook, such as talking to themselves to prioritize a day or taking notes on YouTube videos, which was not part of the original design.
“people are finding creative ways to get more context into Granola that it's just not designed for.”
Patrick suggests a world where a 10 billion dollar company has around 20 employees is plausible, because AI tooling makes large headcounts unnecessary for many businesses, even though Granola itself will still need many people.
“maybe we are entering a zone where there could be a 10 billion company that has 20 employees or something like that?”
A demo that projects imagery from live conversation at five to eight frames per second suggests tools could extend thinking by bringing in divergent ideas during discussion, though making this helpful rather than distracting is hard.
“it was running at something like, I think, five or eight frames a second. And what it was doing is like real time.”
Startups in fast-moving AI must decide how much effort to spend on the next obvious features versus taking a large swing, and Granola's large swing is moving from notes to doing most of a user's work, which requires substantial iteration.
“how much time do you spend building the next obvious five things that are going to be really useful to people versus you, take the big swing.”
Chris wrote a blog post by gathering advice from people, recording a spoken brainstorm in Granola, and using the AI to pull out themes and formats, a process that synthesized advice he would otherwise have dropped.
“that process was such an incredible way of synthesizing all this advice that I guarantee I would have dropped different parts along the way.”
The shrimp barbecue example shows that giving an AI the right context, such as photos of the grill and the shrimp, produced a far better answer than a typed text question, because the shrimp was already cooked.
“give it the right context. So he's like, take a photo of the barbecue and take a photo of the shrimp.”
An investor cold email that shows a specific, thoughtful insight about how people use products in the space stands out from the flood of inbound messages in a hot market.
“if I get a cold email and they write a very like, specific insight about their usage or like product behavior in this space”
Chris's young children playing hide and seek with GPT-4 voice mode, with the camera on, showed an interaction pattern between a child and technology that he had not seen before and that suggests an unexplored product category.
“I just haven't seen an interaction like that between a kid and technology.”
Chris's father's frequent critical feedback, which Chris now understands was costly to give, gave Chris a more honest self-assessment and an internal voice that questions self-justifications.
“I think I have a much more honest assessment of myself.”
Chris personally rejects a future in which a hidden pendant listens to everything in a person's life, preferring a model of visible, consent-based capture.
“I personally hate the idea of a hidden pendant that is listening to everything I know in Silicon Valley.”
The two dimensions that will grow most in AI interfaces are granularity of control and speed of collaboration, making future interaction much more fluid.
“I think granularity of control and speed of collaboration are the two things that are going to go way up”
Granola's user base is concentrated among AI-forward people such as founders, investors, and marketers at AI startups, with the marketing person at AI startups using Granola at an extremely high rate.
“The people who use us are the people who are AI forward.”
After a modest launch in May, well-known CEOs unknown to the founders began tweeting and direct-messaging product feedback, so the founders' DMs became a customer support channel for big tech CEOs.
“my Twitter direct messages basically became a customer support channel for CEOs of, like, big tech companies”
Chris's claim that Granola will need a lot of people despite small-company momentum is tempered by the view that organizations with thousands of employees will become less necessary.
“The world in which that's necessary, Granola seems very small.”
Most education use cases will be absorbed by generic AI assistants like ChatGPT, so standalone education AI products face strong competition from general tools.
“I think most of education will fall under that category.”
Competition among model providers is a strongly positive development for application companies like Granola, which benefit from rapid model improvement and so do users.
“I believe that companies like Granola benefit tremendously from the competition between the providers.”
In a fast-moving AI market, the only durable way to win is to consistently build better products faster than competitors, since even three months of complacency can put a company in trouble.
“the only way you win is you need to consistently build better stuff than other people faster than they're building it.”
Some of Granola's most upset user emails say that about a third of their meetings are in person, leaving them flying blind without the tool.
“a third of my meetings are in person and I'm flying blind.”
Big tech companies pivoted quickly toward AI after ChatGPT went mainstream, but choosing a strategy is not the same as executing it well.
“I think just because you choose to do something doesn't mean it's easy for you to execute on it.”
Granola is an AI notepad that transcribes meetings in real time while the user writes notes, then fleshes out the user's notes at the end of the meeting so the user can focus on key judgments rather than copying facts.
“AI really simply is like a digital notepad. So think of it like Apple Notes on your computer.”
Even though Granola is designed to be small, it will still need many people because the company is very ambitious, while the world where thousands of employees are needed will feel very large compared to Granola.
“We're very ambitious at Granola, so I think we're going to need a lot of people.”