Chris Pedregal——打造 Granola——[Invest Like the Best，EP.412]

Chris, I thought a fun place to begin our conversation today is with some of your ideas around the value of tools for thought that technology has given humans over the centuries. Obviously, you're building one of those tools now, and we'll get into that in great detail. But the first time we chatted, I was so intrigued by the way that you approached this and thought about this unlock of value for people. I think you used the XY plot as a good example of one of these tools for thought. Maybe you could riff for a while on this line of thinking and why you're so interested in it.

1. The Tools That Extend Thought

I love this topic. I think fundamentally, humans are toolmakers. It's one of the things that sets us apart from other animals. If you look back at history, there have been these inventions—tools that were invented that just enabled humans to do so much more.

The interesting thing about that is that some of those are explicitly tools for thinking. Examples could be writing, which is a great example, or different mathematical notation. 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, and that's fine.

My favorite example is this idea of being able to visualize data. What you brought up is this guy called Playfair. I think his name was William Playfair. It was something like 200 years ago that he was the first person to graph data visually, so you could use your eyes. Humans have evolved to bring in images and make sense of images really quickly.

The idea of mapping numbers to the visual plane and being able to intuitively feel, "That graph's going up or down, or it's going up much faster than it was before," is just crazy. Two hundred years before I was born, no one had done that.

All of this is to say that I'm sure we'll get into this in more detail. You have mathematical notation or writing, data visualization, then there's the computer, and I think with AI, we're just entering a new realm where the tools for thought will be exponentially more powerful and more useful. God knows what that's going to look like in 10 or 20 years. I guarantee it'll look nothing like it does today.

Maybe just talk about that transition. Mention what you're building at a high level first, and then we'll go into it in much more detail later. As we transition into understanding what new tools are possible, built on top of this new technology, how are you personally approaching that? What were the original things that you thought of when you saw some of these LLMs? Walk us through this phase change.

One observation I think is interesting about these tools for thought is that oftentimes what these tools do is let you externalize things that you have to hold in your head. One of the most ubiquitous tools for thought today is a notepad and a pencil. When you use a notepad and write things down, it means you don't have to hold everything in your head, and you can look at these ideas or look at these notes.

To use an analogy, it's a bit like extending your RAM. The amount of RAM that we have in our heads is hard-coded by physical limitations, and these tools basically give you more RAM, more memory.

I think what's incredible about LLMs—the real unlock here—is that you can use LLMs to bring extremely relevant context to a person in the moment they need it, and that context can be dynamically generated to match the needs of that moment.

I think it's as if being able to write your ideas on a piece of paper in a notepad makes you much more capable in a meeting or when you're talking with someone. Imagine if a computer can bring in all the relevant context to make you brilliant in that moment, so you have it at your fingertips. Because LLMs can rewrite content on the fly and pull that stuff in for you, I think that'll be an incredible unlock for people.

How does that manifest? Is it that everything in my life—everything I've read and every conversation that I have—is ultimately stored, and then there's some mechanism for me to feed my current context back to some system, and it serves up ideas or brainstorming concepts? Make this a little more real in terms of your vision.

Before you ask me to talk about what we're building at Granola, I can talk about that. I think in the realm of AI, it's easy to talk about the next two steps, and it's really, really hard to talk about what the world is going to look like 10 steps down the line.

2. Granola Is A Digital Notepad

AI, really simply, is like a digital notepad. Think of it like Apple Notes on your computer. It's an app on your computer. You can write notes. The main difference is that it's also listening to what's being talked about.

If you use it in a meeting, you can jot down whatever notes or thoughts you have. Granola is listening to the conversation and transcribing it in real time. Then, when the meeting ends, it'll take whatever notes you've written and flush them out to make them great.

You no longer have to write down everything that's important. You can really focus on the key insights or thoughts that you had in that meeting—the key judgment that you bring to that situation—and outsource all the busywork, the rote work of writing down information or facts to the AI.

What's so powerful about this—and we still don't fully understand how it's going to change the way people work—is that I know it's going to change the way people work because I work differently, and Granola users work differently. We're maybe 5% down the path to our vision.

When you look back at your notes, you have the full context of the meeting. You can then go and chat with Granola and ask it questions about what happened or pull out themes. Right now, we have a feature internally that we haven't launched publicly where you can look at all your meetings with a certain person or all your meetings on a specific topic and pull out themes across those meetings.

It makes this context that otherwise is lost or forgotten immediately accessible and useful. You wrote it down somewhere, but you don't know where that notebook is. You don't look it up when you're making a relevant decision, and it just makes it immediately accessible and useful.

Maybe talk about how you work differently than others, having been the person most exposed to Granola. What are the actual behavior changes so far? Then I want to ask about the 5% to 100%. But starting with just the 5% penetration, how has it most tangibly caused you to behave or work differently?

3. Work Gains A Context Window

This is something that I think will be widespread. Knowledge workers, folks like you and me, are constantly going to be thinking about, "What's the context I need right now to be the smartest I can be?"

For folks listening, when you're using something like ChatGPT or any LLM, there's this idea of a context window. You can put X amount of information into that context window, and it's basically like giving it, "Here's the situation. Here's the stuff you need to know to be able to think about it."

That way of thinking is also going to apply to us, to people. We're going to be thinking about that all the time. A concrete example: I need to write a blog post. Before, I would have just sat down with a notebook, scribbled down a bunch of ideas, and then tried to type it up.

What I did now was first talk to a few different people who had good advice on this blog post, and I used Granola. Now I have notes and the full transcript from those conversations. I then used the Granola app and walked around, speaking out loud about different ideas. I did a brainstorm where I was just recording it, and then I put all of that in a folder inside Granola and started chatting with the AI, asking it to pull out themes or suggested formats.

At the end of the day, I’m going to write the blog post, but that process was such an incredible way of synthesizing all this advice that I guarantee I would have dropped different parts along the way. That’s one example. I think another example—and this is something we’ve observed with Granola users—is just the way they approach notes is completely different from how they used to approach notes.

If you look at the notes of Granola users who use Granola a lot, they only write a couple of notes per meeting, and those notes are usually the internal thoughts that they had. It’s not the stuff that’s in the transcript. It’s like, “Oof, this person was a bit aggressive,” or, “They seem kind of down,” or, “I’m concerned about this area because they didn’t really answer my question.” These are really critical thoughts, and everything else is deferred to the AI transcription.

When they come back to use the Granola notes, they’ll often be chatting. Instead of reading lots of notes for your meeting, they usually have a specific question in mind or a piece of information that they’re looking for, and they find it more efficient to just ask that question and have a really high-quality answer written for them.

Maybe now talk a little bit about that 5-to-100 vision of what this could become. I know you can only think a couple of steps ahead with LLMs, but thinking 2 or 3 steps ahead, where do you think this goes next?

I think at the end of the day, the central question—what’s the information I need right now to be able to make the best decision possible?—is a central one. There’s this image that, if you’re a diplomat, you get this dossier before you go into a high-stakes negotiation that gives you all the background information that was crafted for that moment. I think we’re going to live in a world where everyone’s getting those in real time whenever they go into any meeting.

I think the interesting questions are: What context is useful for that? Is it just the previous meeting? Is it all your emails? Is it all the information in the world that goes in there? And then what does the actual interface look like?

I think my view for Granola right now is that Granola helps you generate the best meeting notes out there, but tomorrow, Granola should help you do all the work you want to do. You walk out of a meeting, and you need to write a follow-up email. You need to write an investment memo. You need to schedule an event with a whole bunch of different people. Granola, or a tool like Granola with all the necessary context, should be able to take you 80%, 90%, or 95% of the way there.

Something that’s very important to me and the folks at Granola is that we see the role of AI as being a tool to make you better. We think you can use AI to replace a person or take away a task from a person, or you can use AI to augment a person’s abilities, augment their intelligence, and augment their abilities. We’re really big believers in this idea of tools that make humans do more, achieve more, and think more.

Everything we’re building is based on this idea: Can you get Granola to do all the busy work of writing up that follow-up email? But then you add your judgment to it, which is actually what really matters here: “This is what’s going to convince this person, so I’m going to twist it slightly,” as opposed to worrying about all the specifics I need to get in there.

I’m curious about some of the nitty-gritty issues that you’ve encountered so far. One is just the recording aspect. How do you think the world will evolve, and how do you handle it today, where it seems to be becoming more and more normal that someone will ask to record a meeting? At first, it really turned me off, and now it’s just become normal.

Do you think we reach a point where the assumption is just that everything is being recorded? I know Granola handles this very thoughtfully. Maybe you should explain how you do it, but I also want to know where you think it’s going.

4. The Recording Privacy Tradeoff

I think as a society, we just need to be really thoughtful about the trade-offs. That’s the answer. I believe that in a couple of years—maybe 18 months, because the speed of AI is so fast—doing meetings and doing work without something like Granola will feel like such an impediment that no one’s going to want to do it. Everyone’s going to be using tools like this because they will be so useful.

There’s a real trade-off, like you said, on invasiveness and privacy, and I think as a society, we need to thread the needle where you get maximum usefulness from these tools with the minimum amount of invasiveness. Where that line is going to be and how we navigate that, I don’t know. I don’t know where we’re going to end up.

When we first created Granola, we made a very conscious decision not to record and store any audio. Even though Granola is listening to the audio, it basically transcribes in real time, but it doesn’t store any of the audio. Everyone kind of laughed at us for that. Why wouldn’t you? Wouldn’t the audio be useful?

Of course it would. You want to be able to go back and listen to exactly what someone said and what their tone of voice was. There’s definitely a loss of value for the user because we’re not recording the audio. But what it means, though, is that Granola is way less invasive than any of those other AI meeting bots that join your meetings.

Those bots record the audio, record the video, and store it. Who knows how long that stuff is around for? That feels completely different, in my opinion, from something like Granola, which is generating really nice notes, and you have a transcript, and it’s super useful, but it’s much less invasive and much less intrusive.

I think there’s a real question, which is: What’s it like when we’re walking around the real world? On a Zoom call, it’s one thing. Usually, there’s a specific reason for meeting, and people understand the context of that meeting and what the expectations are.

I think the norms in our social lives will be very, very different from those in the workplace, is my guess. I don’t know exactly where that will end up, but I see a pretty stark distinction. In the workplace setting, most people want all these things captured because the AI can provide so much value to the user, whereas in our social surroundings, I think that’ll be a very divisive issue.

We’ll see how that goes. I remember when Google Glass came out, there was a huge backlash. I could see a similar backlash happening when AI pendants start becoming popular. You have that one guy showing up at a party, recording everything, and pissing off everyone else.

How soon do you think it’s the case that in-person work meetings have the same expectations as a Zoom meeting? I’m actually already there. I’m already frustrated that, with no nefarious intent, I just wish I had a memory assistant with me so I didn’t have to think about notes. I could just be engaged in a conversation.

I wish that was the norm today, and maybe there’s just a social thing that happens where you decide, at the beginning of a meeting, whether it’s being recorded or not. I wish that was easy. I would wear the pendant now, just because I meet so many interesting people. I can’t keep all this stuff straight in my head. I furiously try to take notes afterward. It doesn’t feel that different. When do you think we get there?

Our iOS app is launching soon. Sam, my co-founder, and I built this because we wanted it ourselves. We thought it would be useful. We thought it was interesting. Quite frankly, we were really surprised by how it took off and by how, once someone starts using Granola for all their important work calls, they’re basically outsourcing some of their long-term memory to Granola.

You start to have this expectation that you can go back and look up these important things from any conversation. Some of the most upset emails we get from users say exactly what you’re saying: “Hey, a third of my meetings are in person, and I’m flying blind. I’m naked in those meetings. I desperately need Granola in person.”

I’m speaking about Granola right now because that’s what we’re building. Maybe it’ll be us, maybe it’ll be someone else, but I guarantee you a tool will be used by everyone, basically, in this context.

As to what the norms are going to be, I personally hate the idea of a hidden pendant that is listening to everything. I know in Silicon Valley that’s one of the visions for the future, and I personally don’t like that vision.

I think in a work context, the phone is great because you basically put it down on the table, and it’s an easy social contract with the people in that meeting about what’s happening. That’s how we do work at Granola. Basically, every meeting at Granola, it’s very clear if there’s a phone out and whose phone is taking notes.

I think the social contract really matters. It’s up to the individual to manage this, just as it’s up to the individual to manage everything in the work environment. If you put the phone out and you’re upfront about it, everyone benefits, and I think that change will happen much faster than you expect. Whereas in social circles, it will be very different.

One of the things that I’m so curious about right now in the world of AI application companies is this small-team meme, where some of the most incredible tools are built by teams smaller than 25 people, and as they scale their user base or their revenue, the teams really aren’t getting bigger.

They don't need bigger teams. Can you describe what it's been like in all its aspects, abstracting away a little bit from the product itself, but just building a company in this space relative to prior companies that you built or were a part of in the pre-AI era?

5. Small Teams Ride The AI Wave

The 2 defining characteristics that are different about this space in this moment are, 1, the speed at which the technology is getting better is nuts, and 2, Granola is built on top of LLMs, so it's an app-layer product. We get so much benefit from riding these incredible technological advancements that are happening at the LLM layer. We spend a lot of our time really thinking about what makes a great user experience end to end.

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. We really do benefit from that. That said, a lot of what makes Granola great is sweating the details of all these technical edge cases—stuff you'd never think of. It's like you're in the middle of a meeting and you take off your AirPods, and it's on a Zoom call that has multiple channels, and all of a sudden Granola needs to do something very specific to make that feel seamless that you never would have thought of until you built it and realized it felt crappy if you didn't do that.

We use as many AI tools as possible for as many things as possible inside of Granola, but some of the tools, at least on the development side, aren't quite there yet. We're so close to taking that end to end, so we still have to do a lot of work there. Again, I hate doing time-horizon guesses here because it's basically impossible to know. If you fast-forward 3 years, I think the way we would work and what we would be able to outsource to AI would be completely different.

Is that mostly engineering challenges, where you would expect that using Cognition and Cursor and whatever else, your team would be able to effectively be a manager versus an engineer and just tell it what to do, and you wouldn't have to actually engineer the endpoints?

That's right. Our CTO, Vas, has a goal. Basically, minimizing the number of lines of code every engineer writes at Granola every day is a goal of his. It's an active goal.

We just did this off-site, and the theme was basically, use AI everywhere for things you wouldn't expect to and push ourselves outside of our comfort zone. There's this great example. I was trying to barbecue some shrimp for the team. We bought some shrimp. This was in Spain, and I've never barbecued shrimp before. I'm typing into ChatGPT, "Okay, how do you barbecue shrimp?" And Vas was like, "No, give it the right context." So he's like, "Take a photo of the barbecue and take a photo of the shrimp."

He was totally right. I was like, "Yeah, yeah, yeah. Give it the context." So I did this. It turns out the shrimp was already cooked. We didn't realize it because it was in Spanish, so we didn't have to cook it at all. We just needed to heat it up, which I never, ever would have figured out if I had just typed it in.

An interesting point there is there's just a completely different intuition you need to have around how you use these tools and how you build with AI. Perhaps it's similar to when the web came along and people pre-web wouldn't automatically default to using Google; they'd go elsewhere. People who had grown up and were young enough when that happened would always default to using Google.

I think there's going to be a very, very, very similar divide here, which is basically that AI natives will just understand what context they need to give AI and how to work with AI. When in doubt, you should probably give it more context and see what it's going to say, as opposed to assuming you know what's right.

I'm 38. I'm very happy the team is constantly pulling me. I'm literally at the forefront in thinking about this all the time, and I don't use AI as much as I should be using it. If that's the case for me, think about the general population.

Is one of the key lessons there that a lot of what needs to get built, both technically and as an expectation for people, is context-gathering tools? You're doing one, obviously, for conversation, and that's one mode of input that's really, really, really important, especially for work. How do you think we'll capture the rest? Riff on context gathering as a function.

6. The AI Steering Wheel

Gathering the context, just getting all the data, isn't that hard. It's only a matter of time before you can plug all your email into Anthropic or ChatGPT, along with all your notes, all your company documents, and all your tweets, and it'll have all that. I think there's a different question, which is: Which of that context is really relevant for the thing I'm about to do right now? That may be a technical problem. That may be a UI problem. I don't know.

So that's on the context side. I do think a huge blocker for unlocking the power of collaborating with AI is: What's the UI? What's the interface for collaborating with UI? I really think we're in the terminal era with old-school computers, where you type in a command and then the computer would literally spit back a command. The way we work with ChatGPT, I don't think chat's going away, but I think it will feel archaic in how little control you really have as a user.

I was looking this up. I was trying to find an analogy for this. The first cars that came out didn't have steering wheels. They had basically a stick that you could turn left to right, and it was fine if you were trying to go really slow. The moment you went fast, the stick was unusable. You'd move it too much, and you'd crash off the road, and it was a big security problem. Finally, someone came up with a steering wheel, and a steering wheel is a UI that gives you so much fine-grained control when you're trying to turn.

I think we still have to invent what the steering wheel is for when you're working with AI and collaborating with AI. Right now, we have some very coarse controls, and it's turn-taking. It's like I write something, then the AI does something, then I react back to it. I think it's going to be a lot more fluid and a lot more collaborative once we figure that out.

Bring that to life a little bit more for me, the fluidity aspect. How could you imagine that being, versus the back and forth?

It depends on the tool, but right now it doesn't feel like you and the AI are working on the same canvas. It's like we're working on 2 separate canvases next to each other. This is a very basic thing, but when you're using ChatGPT or Claude, you can't go and edit the response that the AI gave you. You don't go in there and be like, "Oh, actually, no, this point was dumb, and let's change the language here." You tell it, "Please make it shorter," as a command, and you hope that it rewrites it in the right way. That's just going to feel like madness not too long from now.

I guess there's a historical parallel here. In the early computing days, when you were in a text editor, the first text editors had this idea of modes. There was a mode where you were in text-insertion mode, and you'd go in and write some words. Then you'd exit that mode, and you'd go into deletion mode or copy mode. You'd have to enter that mode and make that change.

Larry Tesler basically went on a vendetta to change this. Now you should be able to type and delete and cut and copy and do all that fluidly without entering different modes. That was unthinkable before we made that jump. It's hard to imagine what that's going to be for AI, but I guarantee it'll feel completely different from what we have now. I think granularity of control and speed of collaboration are the 2 things that are going to go way up. It should be way more fluid.

Have you been surprised by any of the ways that users use Granola?

7. Granola Evolves Through User Feedback

There are a few things that have jumped out. One is the variety of use cases people use it for. We built it for work meetings. Very quickly, people started telling us, "My partner has cancer. We have all these meetings with doctors. Granola has become absolutely invaluable in that process. I actually don't know how I would've managed it before." There's the use-case thing that was unexpected.

Then the other thing is people are finding creative ways to get more context into Granola, even though it's just not designed for it. This is the, "I am brainstorming an idea, and I'm just going to create a meeting in Granola and a note in Granola and just talk to myself," or, "I need to plan out my day, so I'm just going to talk about the different things going on and then use Granola to help prioritize what I'm doing." Or, "I'm watching a YouTube video on a subject I'm trying to learn. I have Granola open, and I'm taking notes in there because of that."

That's probably the biggest surprise. The other behavior change, I think I mentioned this before, is that when people go back into Granola, less and less they read the notes that are there, and more and more they ask the Granola chat what they're looking for.

As an app builder, what is your perspective on the battle between model providers for your attention and business?

It's the best thing ever. It's fantastic. I fully support it. For us, we build on top of foundation models, and the speed at which models have gotten better over the last 3 years is incredible. I believe that companies like Granola benefit tremendously from the competition between the providers, and as a result, I think users are benefiting tremendously.

How is it built? Are you sort of hot-swapping the best model in, and that's just something that you could do in a morning every time? Anthropic apparently is coming out with this new model soon.

Will it just be a function of a quick eval and then hot-swap that thing in as the primary driver? And then switch again in the future if a new one comes out? Is it that simple?

That's exactly right. I think evals aren't simple, but what you described is exactly what we do. We don't just use one model in one place. We use lots of models in lots of different ways inside of Granola, but we will switch to whatever the best model is on any given day.

How do you think about the competitive dynamics of what you're building versus what might be achievable by using a model directly alone? Everyone always used to ask, “Won't Amazon just build this?” or “Won't Google just build this?” Now it's, “Won't Anthropic just build this?” How do you think about building in such a way that's protected from the future in which the model companies come to eat your lunch directly?

I don't have a crystal ball here, but here's the way I view this. There may be 2 axes that matter. One is, how common is this as a use case for me? Is this something I do once a month or twice a month, or is this something I do 500 times a day? And 2, how great do I need to be at this task?

I think everything that is low-frequency, where you don't need to be great at it, will be eaten up by the general assistant. I'd say most consumer use cases actually fall in that quadrant, because it's basically impossible to build a habit of using a new tool for a low-frequency use case. And if it's something where you just need it to be pretty good, then a universal assistant like Claude is perfect. Actually, the more you use that, the better that assistant will get for you.

I think the other end of that quadrant is basically a high-frequency use case where your output needs to be really, really good, and that's basically the power-tool quadrant. There'll always be that pro tooling for the people who really want to do a fantastic job at something. I think that's where Granola sits.

You might be like, “Oh, but why can't the general assistant do that as well if the model just gets smart enough?” My answer there is that it's not a question of intelligence. It's actually about how great the UI is optimized for this use case. If you have a product that is solely dedicated to being phenomenal at that use case, it will be a better experience than a general tool will be. I think the limitation there, what separates them, is really the product design and optimization of the user experience, not the underlying technology.

Do you have a crystallized product philosophy that guides your decisions?

My personal approach is that you can boil down most great product thinking and design to a very simple question: when you use a product, look at it and really ask yourself, “How does this make me feel?” Just keep asking yourself that question and really, really, really listen to the answer. Then, once you've done that 100 times, put that same product, UI, or button in front of another person. Just ask them that question over and over.

I think when you do that, you realize that within the first 500 milliseconds of looking at a product, you feel about 10 things. Oftentimes, those things tell you exactly: “Oh, it's too complicated. It's too cluttered. I don't know what to do. It makes me feel insecure.” There are so many emotions, and they go by in a flash of an instant.

If there's an emotional recorder and you could play it back in slow motion, that would tell you all you'd need to do to make your product great. There are lots of other things that matter, but I feel like that one question is an incredible guiding force.

You gave the personal one. Is there anything that's different about the Granola-specific product philosophy?

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, and every decision we make ties back to that in one way or another.

Even the most basic one: it is an editor. Most AI apps that generate notes don't generate them in an editor where you can edit them. They give you a PDF kind of thing or an email: “Here are the notes.” There are tons of micro-decisions that all map to that idea.

Are you at all surprised by who your users are, what types of jobs they do, or do they tend to cluster in a couple of sectors? What have you learned just based on the raw data of who they are?

This is actually pretty interesting, and it might have implications. The people who use us are the people who are AI-forward, so it's folks who are leaning into these new tools and these new ways of doing work. That, interestingly, maps to a ton of founders, a ton of investors, and a ton of people across all disciplines who are working in the AI space. The number of AI startups where the marketing person is using Granola is extremely high.

It's interesting how there's a very stark line between the people who are leaning into these tools and those who aren't.

Yeah, it makes sense, right? It's very much in Geoffrey Moore's Crossing the Chasm framework: early adopters, natural early adopters.

One of the weird things—I remember when that happened. We launched Granola in May, so it was 8 or 9 months ago, and I've been building product for a long time. This was surreal, though. We launched it, and we were happy that some people tweeted about it. It wasn't a crazy big launch or anything, and we just expected to keep building.

A few weeks later, these really famous CEOs whom we did not know just started tweeting and then DMing me on Twitter with a whole bunch of product feedback that they wanted to give. It clearly resonated with a very specific type of persona, and then that persona was really loud on social media. My Twitter direct messages basically became a customer-support channel for CEOs of big tech companies, which is a really weird experience.

That's what I did to you. It's the same exact thing. This is such an interesting way to meet people quickly. I'm curious, in this whole building process, is there any plot twist that you look back on that turns out in hindsight to have been a blessing or a gift in your whole product-building experience?

We made the decision early on, at least with Granola, to make Granola a Mac app—an app that sits on your computer—rather than a bot that joins meetings or something on a website. There are lots of different ways you can build it, and that was a huge pain in the butt for a whole bunch of reasons. When we started off, it was only possible to do what Granola does for users who were on macOS 13.4, which was, I think, 15% of Mac users at the time.

The reason we did it, again, was this idea that we wanted it to be like a notebook and a pencil. We wanted you to be able to grab Granola and use it no matter where you are, whether you're on a Zoom call, in an in-person meeting, or in a huddle on Slack. We don't want you to have to think about it. An important thing about a tool is that it is reliable and works in a consistent way, so you know how to use it.

There have been so many downstream great things about being a Mac app, about being an app on your computer. It's so much more immediate and in your control, and it's so easy to get to. Basically, the way people use Granola, which I'd say is quite intimate, is largely a function of the fact that it's an app on your computer rather than a tab lost within 50 other tabs on a website that you have to find.

I think we can take a little bit of credit for that, but I think that was a way better decision than we realized at the time.

Has the process made you change your mind in a major way about anything?

Yeah. When we started off building Granola, we had a completely different interaction pattern in the app. The thing we pitched and the first version we built were very different. You would type in a keyword or 2 in Granola in real time, and you'd hit Tab, and then Granola would write the full note for you in real time.

It's a really cool demo. It felt kind of magical when you used it. I'd say something like, “Mac app.” You'd type in “Mac app” and hit Tab, and then it would write this: “Chris is really glad that he made the decision to build a Mac app.” We then basically spent 6 months trying to make this work, and we just couldn't.

What we found out was that no matter how great the notes we wrote were, if a computer is writing notes for you in real time during a meeting, you can't help but read it. What ends up happening is that you're incredibly distracted. The whole point of Granola is that you can be more present in the meeting, and what was happening was the exact opposite. People were just looking at the notes, and if they were not exactly how they wanted, they were editing the notes. Then they realized they had not been paying attention to the person speaking, and it was just really bad.

We ended up completely changing the interaction pattern to being something way more mundane, which is that during the meeting, it works just like a regular text editor, like a notepad. You type stuff, and then all the magic happens at the end. The magic moment—the value of Granola—you only realize after you've used it for a whole meeting, which is not great. Ideally, when you're building a product, you want that magic moment to happen in the first 20 seconds.

It just made it a way better product. Like I said, we spent 6 months trying to make this wrong thing work until finally we kind of accepted that there was a better way to do it.

If you think about the model providers as one vector of competition for the job to be done, how do you think about the other vector, which is other app builders, and the ways in which architecting the product might defend you because it's becoming more and more sticky and valuable to the user, so that even if another Granola 2.0 comes out that's a little bit better, they're not going to adopt it?

Do you think a lot about that sort of thing, even though you’re super, super young, and I’m sure mostly just focused on building something great for users? Does that line of thinking enter your mind?

8. Winning Means Building Faster

I think the only answer here really is you need to build something better than other people faster. In this space, there are switching costs and small moats, but I think the only way you win is you need to consistently build better stuff than other people faster than they’re building it. And doing that in a space that’s moving this quickly is not a small feat.

Something we talk about as a team all the time is that, with something like Granola, there’s an inherent switching cost because the more context Granola has, the more useful it’s going to be for you. Something would have to be much better, I think, for someone to switch off of Granola. But I think you get complacent for 3 months, you’re in trouble in this space.

Tell me how you do that with your team. I’ve heard a few different fascinating methods for engineering product velocity in a company building an app on top of AI. How do you think about it and do it? What’s worked? What experiments have failed? How do you engineer product velocity?

Something we’re pretty explicit about is knowing, when we’re working on a feature, whether we’re in exploit mode or explore mode, because you need 2 completely different approaches to that. What that means is: Do we know what needs to be built here? Is there a clear idea, and is it just about executing it as quickly as possible? Or do we not know what the answer is here? Is this an unsolved, open problem where you need to do some exploration first and then figure out what the right solution is?

For the one where you know what you need to build, at least from our experience, it’s the basic advice that everyone hears, which is to build the minimal thing as quickly as possible, give yourself deadlines where you will ship it to real humans—maybe not to everybody, but to real people—and then try to increase the shipping iteration speed as quickly as possible.

I think we’ve gotten in trouble before, and it’s easy to not know what mode you’re in and use that philosophy for the open-ended problem. What ends up happening is you ship something crappy to people, and you tick it off. You’re like, “Oh, we shipped it in 2 weeks. This is great.” But actually, you didn’t solve the problem to be solved. The thing you did was ship, as opposed to figuring out what a great solution for people is and doing that.

Interestingly, I’d say that is extra important in this space because there’s so much pressure to move quickly that, every now and then, taking the extra time to think about how to do this is really important. A good example is that we were working on Granola for a year before we launched, and we were already so late to the AI note-taking game. We were 7 years late when we founded Granola, and we didn’t launch for a year.

You know how I talked about that interaction? We completely changed the core interaction of the product. 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. Users would have said, “Oh, this is cool.” The ones we would have retained would have liked it, but we wouldn’t have retained that many users. That would have been it.

I think it’s very important to protect your ability to change direction with the product until you have a lot of confidence that you’re in the right direction. How do you manage that while also moving really quickly in a fast-moving space? That’s the whole challenge.

How do you think about dialing your own degree of ambition? If it’s 1 through 10, where do you think it is, and has it moved a couple of points up since you started? What is the process of sussing out and dialing one’s own ambition? How have you experienced that?

I ask myself if we’re doing this correctly every day. Sam and I, when we started playing with LLMs, became convinced that all the tools for work that we use are going to be rebuilt or reinvented on top of LLMs. We became convinced that there’s going to be a new class of software.

In the same way that, if you’re a developer, you probably spend all day in Cursor or Visual Studio—some IDE—we think there’s going to be a new class of software. It doesn’t have a name yet, but people like you and I will spend all day in it and do our work there. Folks whose jobs revolve around people, communication, projects, meetings, and all that are going to have a new workspace, and that’s what we set out to build from day 1. That’s exactly what we’re setting out to build now.

I think the interesting question for us is that it’s really important, if you’re not OpenAI or Anthropic, that you are really, really good at a use case today. You can’t just be building a fantastic product for the future. You need to be damn useful at a very specific thing today, and every step along the way, you need to be super useful to people.

I think there’s a real tension there, which is how much time do you spend building the next obvious 5 things that are going to be really useful to people, versus taking the big swing? For us, we want to move from a world where you use Granola for notes to one where you use Granola to do most of your work. If you’re writing a document or a memo, it should be way easier to do that in Granola because of all the context we have about the work you’re doing that’s related to that. But that’s a really big swing. Getting that right is going to take a lot of work and a lot of iteration.

If you think about existing companies that do aspects of what Granola does better now, or may do in the future, which are the ones that you think about the most? If you were a VP at one of these companies, which ones should you be worried about because major disruption is coming?

My view on this is you can worry about a million things. You should choose selectively what to worry about, because there are very few things out of your control. The competitor that we’ve chosen to worry about at Granola is the one that hasn’t launched yet. It’s the startup that can look at what we’ve figured out, what other people have figured out, start at that point, and execute on it more quickly than us. That’s what we’re thinking about.

I was surprised at how quickly the big tech companies reacted to AI. There was this moment when ChatGPT went mainstream, and then you saw every big tech company pivot and try to adapt to that strategy. I was impressed by the leadership there. Just because you choose to do something doesn’t mean it’s easy for you to execute on it.

One of our investors has this saying: “If you list out all the AI features that you use on a daily basis, how many of them were built by big tech versus how many of them were built by startups?” I think a surprising number of those were built by startups, even though every big tech company is out there investing a tremendous amount of money to build AI features.

Does that get figured out over time? Maybe. Startups are oftentimes the R&D wing of all the big tech companies, and then, once something’s figured out, they can incorporate that into their large user bases. But generational companies figured something out earlier, and they were able to leverage that into becoming something massive.

If I was forcing you to put your mega-dreamer hat on and set aside feasibility as part of your consideration, what do you dream most about tools being available 5 or 10 years from now as tools for thought, which we opened our conversation with?

9. Tools That Make Us More Human

I want tools that make us more human and better humans. By that I mean tools that unlock our creativity and our ability to do all the things that humans are incredible at, that no one else can do.

I think the people who are building tools with AI need to be very intentional about that because there’s a fine line. You want to outsource all the rote work, all the boring stuff, all the mindless stuff, but you really don’t want to outsource the judgment.

When you were talking about generating ideas and asking AI to generate 100 different ideas so you can choose the right ones, that’s great. 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. That’s just one example, but it trickles down to everything.

It’s like, “Oh, okay, well, this idea of writing is thinking, and if AI is doing the writing for you, a lot of that writing is just rote work. There’s no value in it in any way.” But some of it is where you do your thinking, and if you’re not careful about what you outsource, I think there’s a real danger there.

The tool that I would want would be one that addresses the fact that, right now, we have so many silos of information and so many silos of where knowledge, inspiration, or information comes from. Oftentimes, I’m only really looking at data and information from one of those silos when I’m thinking about a topic.

What I want is a tool that will pull out the most relevant and best stuff from my personal life and my context, but also from out there—from what humans have figured out—and present that to me dynamically, on the fly, in a way that I can interpret and make use of in real time. What that looks like, I don’t think anyone knows.

I saw this amazing demo a friend of mine made. There was a microphone hooked up to something like Midjourney, but it was running at something like 5 or 8 frames a second. What it was doing was, in real time, as you were talking—for this conversation, for example—it would project imagery on the wall that was related to what we were talking about but slightly divergent.

He was using this for a Burning Man creative experience. You could imagine something like that in a work context, where it's helping you think out loud, but it's also extending and bringing in ideas or useful information that you wouldn't have had otherwise. I think doing that in a way that's helpful and not distracting is really, really hard, and there are a lot of these ideas in science fiction that sound fantastic and then, in practice, don't work for really silly tactical reasons, like notes being written for you in real time and being distracting.

I think there's a lot about the human experience that defines what works and what doesn't. I can talk about this for hours. I gush on it. I just think it's such an incredible moment to be alive and to be building things.

Micky Malka, the great investor, has an art installation that does what you just described, where, as you talk in the conference room, it visualizes—

Oh, really?

...what you're talking about. It is quite distracting, I will say, in a good way. You can't look away from it. It's just so mesmerizing. But to extrapolate that, I saw that 6 months ago or something, and these things get better at an alarming pace.

One question is always: What are these models bad at? Everyone's very bullish. Everyone's very excited. They're great at a million things, and they're going to get better and better. I think everyone is coming around to that. Is there anything across the model generations that you've been surprised isn't getting better? Things that they just don't do well and consistently haven't done well that are real limitations?

I think it's good to separate the reality today from what's a reality that will persist, and what limitations will persist in the future. It is surprising to me how unpersonalized any of these models feel today. If you ask it a question versus me asking it a question, the answers are going to be identical or almost identical. Given we're X number of years into this cycle, I think that's really surprising.

This is a small thing we do at Granola that people like, but if you were using Granola in, let's say, a meeting and I were using Granola in that meeting, your notes and my notes would look completely different, and that's just because we built it that way. We're like, “Okay, the things that matter to Patrick in this meeting, we think, are this. The things that are going to matter to Grace are this.” But the low level of personalization is surprising to me.

What advice would you have for investors? You've raised money from great investors, and I'm sure you've talked to a ton of them. Most investors in the technology world and in private markets are mostly or entirely focused on investing around this wave of AI technologies, and so I think they're all trying to answer the question: What is the best, most productive way to interact with company founders and new applications and all that?

I'm curious what advice you would give to those people who are trying to do their best job of allocating capital to the highest and best use. What would you tell them? Maybe the way to answer is: What have the best investors you've encountered done with you, and what have the worst ones done that we could avoid?

I'm not an investor, so it's hard for me to give advice to investors. I can tell you what speaks to me. The same way I talked about how, when you're building a feature, you need to know: Is this exploit mode or explore mode? I think AI as a whole is an explore-mode problem. No one knows what the right thing is. I think maybe foundation models are now more in exploit mode, but everything else, especially at the app layer, is total explore.

When you're in explore mode, you need to have a certain sensibility there, which is, in my opinion, very product-centric, and a certain exploration and depth of thought around what's actually going to be a good product or good for people. Not many investors talk about that or think in a deep way. The stuff that stands out from the noise for me—there have been some really good ones—but if I get a cold email and they write a very specific insight about their usage or product behavior in the space that they've thought about, maybe something Granola gets right or we get wrong, that really makes me pay attention, because if something's hot, you just get inundated with messages.

My inbox is hard to manage right now, and that's just because AI is exciting right now. It may not be exciting tomorrow, and what I want, at least, when I partner with an investor is a partner I'm going to work with for a very long time. I want us to agree on an outlook on the world and how we think about a problem. All the specific execution, all of that's going to change. It's an adapting world, but do you have a similar worldview on how you should go out and solve problems? I know that's a very generic answer, but I have that with my investors. I think they're great product thinkers, and I think they can engage at a bunch of different levels, which is a huge unlock.

If I forced you to build something else in this space—Granola ceases to exist, and you're not allowed to build Granola 2.0—what's your instinct on where you would go to get into explore mode?

Before I started Granola, I was thinking about what I should start. My previous startup was an education AI app called Socratic, and everyone was like, “Oh, why don't you go into education?” I was like, “Oh, I think there are a whole bunch of reasons why I don't want to start another education company or an education AI company.”

But I've been playing with GPT-4 voice mode—you know, the Scarlett Johansson voice thing? With my kids, you can actually turn the camera on, and they were playing hide-and-seek with ChatGPT, which is kind of nuts. My kids are 5 and 7. They were hiding behind the table and then peeking out, and she would be like, “Oh, I can see.”

Tutoring is one thing, or what's going to help you get good grades, but that interaction was something that caught me off guard. I just haven't seen an interaction like that between a kid and technology. I don't know what the product would be, but there's definitely a there there, and I think the way you design that really matters.

What's hard about education? What did you learn building Socratic that you'd caution others building in that space about, or encourage them to do?

The holy grail in edtech is basically building one-to-one tutoring. There are all these studies that show if you have a one-to-one tutor, the median student actually performs at the level of a top 5 or 10% student, and that's kind of been true throughout history. A lot of the great people we read about in history books had a tutor. Was it Peter the Great who had Aristotle as a tutor? I mean, of course you're going to do well. That's an unfair advantage.

I think that's the holy grail. Everyone wants to have a one-on-one tutor. It should be free. It should just be an open-source model. It should be free. Everyone should build on top of it. It's just better for everybody. I don't want to build a business there. The incentives around making money in that space versus what we want for society aren't super aligned, and I think you're also going to get competition from the generic assistants.

As you were asking before, what kind of use cases are going to get eaten up by the ChatGPTs of the world? I think most of education will fall under that category.

Can you imagine a successful tool that doesn't have a data advantage, either unique data that it has access to or first-party data like the data you've built, where, as a person uses it, they're building a dataset that's custom to them? Is it possible to imagine a dataless AI application that is nonetheless still very successful, or do you think data is just an absolutely critical component of sustainability and edge?

With a lot of this data, you don't need that much of it anymore, and getting a little bit of data isn't that expensive or that hard. The way the world's going, you get these foundation models that can understand the world and do a whole bunch of different things, and then, with a little bit of data on top of that, you can really hone it into a use case.

Whereas before, with the old machine-learning paradigms, you'd need millions and millions and millions of examples of something. Now, it's crazy that we can get away with 50,000 examples, and even if it's a very expensive data type to get, 50,000 isn't that hard. I think about what kind of data is ungettable. I don't know. I guess I'm split.

This idea that soon anyone is going to be able to build apps, I think that's going to happen, and I think that's going to happen relatively soon. It's less clear to me what the effects on the world are going to be. I've been thinking about historical examples. Does that make people who are really good at building apps less valuable or more valuable? I don't actually know.

In the beginning of photography, it was almost impossible to take a photo. If you just had a camera, that's it, you're winning. Then cameras became more accessible, but they were still expensive, and you had to spend a lot of time to get good at it and learn about different lenses. Then everyone had a phone in their pocket.

In a lot of ways, everyone's a photographer, and it's amazing what people can do. At the same time, I feel like there's a premium on taste now. If you're actually really great and you can stand out in that, it's almost like you're more valuable. I'm curious what you think. What's going to happen with software? What's going to happen with apps? Is that it?

I think about this a little bit like music. I would be surprised if, in the future, everyone just has all their own music.

I think there's some shared consciousness, shared experience thing that matters for how good something is. In the same way, there's a social-proof thing, or something like the wine studies where the label and knowing how much it costs makes it taste better. Knowing how popular a song is might make you like it more.

Maybe something similar applies to software. Of course, I don't know, but it seems hard to imagine that everyone's going to have the will and interest to build their own version of an app versus just being lazy and clicking the app that everyone else uses, even if it's not entirely perfect for them. I don't think everyone's going to be an app builder in the future, because not everyone's an entrepreneur now.

With Stripe Atlas and cloud providers and all these things, it's massively easier to be an entrepreneur, and not everyone's an entrepreneur. That's what I think. I think the future will often look a lot like the past, and it's really exciting because I can't wait to build some stuff with it. That's my tendency, and other people have different tendencies. I don't know. We'll see.

Thankfully, people like you are building this stuff that's going to make it possible. Another question it brings to mind is that we talked earlier about the small-team meme and how many people are going to be required to build very big businesses. Can you imagine a world where Granola has 1,000 employees? Is that still going to be— I mean, it is a thing objectively. There are plenty of AI companies that have big employee bases, but for you specifically, maybe we're entering a zone where there could be a $10 billion company that has 20 employees or something like that?

I think so. Here's a very real example for us: We just made our first customer-experience hire. We have lots of people writing in, and we interviewed a ton of candidates. 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?” Maybe this is the year.

The ones created after 2025 are going to look completely different. They're probably going to be a lot smaller in terms of people. The way they use tools and what those people do will be very different. I think the departments that are created before then will have trouble. It's much harder to change something that already exists than to build something from scratch on a new paradigm.

We're very ambitious at Granola, so I think we're going to need a lot of people. But when you read about these companies that have thousands and tens of thousands of employees, the world in which that's necessary at Granola seems very small.

This has been so much fun. I'm so interested in what you're building and how you're building it. I think it's such a great example of the new things that are possible and how they're being built in this new world. Thank you for doing this with me.

When I do interviews, I ask everyone the same traditional closing question: What is the kindest thing that anyone's ever done for you?

My dad spent a lot of time giving me a lot of feedback on things, oftentimes critical. I always felt very loved and supported, but oftentimes quite criticized. Now that I'm in his shoes with my kids, I realize just how hard and tiring that is, and there's not a lot of upside for you as an individual to do that.

Sometimes something just needs to be said to someone, and there's a lot of upside for the individual who gets the feedback and only downside for the person giving it. I appreciate just how hard that must have been and how kind that was, because it was really all for my benefit.

How do you think you're most different, in terms of how you think and behave, from how you would be had he not done that?

I think I have a much more honest assessment of myself. People talk about first principles, and I think that phrase gets overused. It's easy to hide behind justifications or philosophies to feel good about something, but I think oftentimes the reality is pretty straightforward.

I can hold his voice in my head quite often, which is interesting because he was never an entrepreneur. He never worked in tech, none of that stuff. But the number of times I hear his voice saying, “That sounds like bullshit”—maybe it's bullshit I'm telling myself, or something someone else is saying—it's in there a lot.

Maybe in closing, how does all that translate into how you would articulate the why behind building Granola?

The most honest answer to that is that it's a very personal thing. I'm happiest when I'm trying to build something that I believe in and that I think is important, and I'm pretty unhappy when I'm not. I'm just wired that way.

I think a boss I had early in my career put this philosophy well. He was like, “Aristotle believed in the active realization of human potential.” That phrase stuck in my mind. When do I feel like my time is well spent? Do I feel like I'm actively trying to realize my potential, but also humanity's potential?

I think that comes for me primarily through my work, but also as a parent, which is something I didn't expect but kind of makes sense now that I'm on the other side.

A beautiful place to close. Chris, thanks so much for your time.

Thank you, Patrick.