Chris Pedregal: Revolutionizing Meetings with AI

0:04Hey, everyone, and welcome to Generative Now. I am Michael Mignano. I am a partner at Lightspeed And this week on the podcast, I speak with Chris Pedregal, the CEO and co-founder of Granola. Granola is an AI-powered notepad for meetings that works with you by combining your notes with AI-generated notes in order to provide you with insights, takeaways, and next steps. I speak with Chris about what makes Granola different, the AI startup scene in London, and how he sees the future of AI-driven products as human and AI collaborations. Take a listen. so i saw on your your twitter your ex rather uh that there's a new banksy in london that that i guess what is right right near the granola office a block away yeah it's one of the it's one of the perks that we our office is in shoreditch which is east london it's the uh yeah the most new yorky part of london it's like the creative side of things and uh yeah a new banksy popped up two days ago about a block from our office.

1:06So is that a thing? Like, does that happen? Is that a thing that happens in London? Like we're like famous street artists. They just, their work just happens to pop up all over. I guess it kind of happens in New York sometimes. It happens a lot in Shoreditch. Banksy kind of goes on, I think every couple of years, he goes on a rampage and drops a bunch of artwork. So it's definitely a thing. I actually didn't look closely at it. What was the, what was the piece? Swinging monkeys on a, on a overground bridge. Yeah. Some monkeys. It's a bunch of animals. He's putting animals all over the place. Why?

1:37Why animals? Is this a theme? Don't know. We don't know. Yeah. Shoreditch is really cool. I haven't spent a ton of time there. I hung out with you there once. And yeah, I know what you mean by it feels New York-esque. Yeah, I love it. It's a lot of the tech startups are over here. West London is, you know, beautiful. East London's a bit edgier. It's where all the chefs who want to experiment new stuff, They'll do all that new stuff that you're testing out and around with creative energy is usually happens on the east side. So it's got a nice vibe to it. What is it like building a startup in London versus building a startup in New York City?

2:17Your other startup, Socratic, you built that in New York. The whole journey was a New York team, not distributed, not remote. Now you're doing it in London. Same thing. What are the differences or what are the similarities? I would love to hear you just talk about it. There are fewer, let's say, consumer or kind of consumer-esque success stories out of the UK than there are in New York and obviously the US, right? And I think that kind of permeates. So a lot of the... There aren't a ton in New York either, right? That's true. The fintech, like the financial apps, the Monzos, the Wise, the Revoluts, they really kind of like set the bar in terms of consumer tech in London.

2:57So I think a lot of the culture is kind of defined by that. Um, there's amazing technical talent, incredible technical talent in London, specifically in AI. Um, there's less of a startup culture here. And I guess there's the startup culture difference. And then there's also just British culture is a little different, right? Our team's very cosmopolitan. Yeah. Um, but I'm conscious that I'm, I'm the American, you know, in, in the room. And I might talk about things a little bit more differently, a bit more loudly, a bit more directly than would be the norm in British culture. But it's also really, it has a little bit of that, in my opinion, the energy that New York had when I was building Socratic.

3:43Whereas there's a lot happening, but it's a little bit the underdog, if that makes sense. So there's a lot of camaraderie between the startups in London. So you mentioned there's a ton of technical talent. Where does that come from? I mean, in Silicon Valley, it obviously comes from decades of these generational companies being built and then people leaving them and starting new companies. In New York, it comes, I think, similarly from some of these incumbents which have had their roots in the city for many, many years. Where does it even come from in London, especially on the AI side? A lot of the big American tech companies, the first office they open is in London.

4:21right so you have the the googles the metas the um i mean stripe has an office here snap has an office here so it's like if you're going to open one office in europe it's usually london and you get talent from across all of the eu to come here um specifically ai there's like four places it's basically there are really good ai programs at oxford oxford cambridge ucl um and imperial and then DeepMind was started here, right? So DeepMind kind of became a center of gravity. Meta opened up an AI research division here. So basically it's kind of been a leader in AI before the recent waves of AI. I have so much I want to talk to you about with Granola.

5:05Obviously you and I talk often. I am, you know, full disclosure for the listeners. I'm very lucky to be an investor in Granola on behalf of Lightspeed. But I feel like something when we talk, we we talk very tactically, right? That's that's sort of the relationship between kind of like founder, board member, investor, like we're talking about the next project, the strategy, etc. I feel like we don't often have a chance to get sort of into the nitty gritty of like product design or how you made certain decisions about the product. And, you know, I think one of the things that I certainly feel about Granola, which I know many other people do as well, definitely many of your users, is that it is this perfect example of an AI application.

5:49Right. There's been so much talk and love and investment and capital poured into the foundation model layer. but there are a few examples of really, really high quality products. So things at the application layer of AI. And again, I'm biased, I'm an investor, but it feels like people, many, many people who are not investors in granola have commented on it being one of these great examples. And so that's where I think it would be great to spend some time today talking because I'm sure that you have lots of insights that you can share from the first couple of years of the company. I know what Granola is.

6:31I'm sure many of the listeners do as well, but for the benefit of those who don't, maybe just start with explaining the story of Granola and what it is. Yeah, absolutely. I'll start talking about what the product is and thanks for all the the kind of things you said, Mike. So Granola is an AI notepad for meetings. So it's specifically designed for people who have lots of back-to-back meetings. And it looks like Apple Notes. It's a simple text editor, and you can take whatever notes you want during a meeting. And the difference with Granola, what makes it so powerful, is that when you're in a meeting, it's also listening to what's being talked about.

7:14So it has the whole context of the discussion. And the moment the meeting ends, Granola will take whatever chicken scratch notes you've written and rewrite them to be really nice, high quality and understandable to other people. They still feel like your notes though. And it's in a text editor. So you can go in and you can edit and change. The whole idea is that it's augmenting your note taking ability as opposed to taking notes for you or on your behalf. that's the basic idea of granola well how did how did you get there i mean you know your your experience your history before granola wasn't rooted in well i guess you it was in many ways you've been working on sort of ai before it was cool with socratic but how do you go from building a product for students um and homework and and education to building this like you know llm powered super notes.

8:09In March 2022, I, is that right? Yeah. March 2022, I played with, it was the Instruct version of GPT-3, right? And within a week, I was just, it blew my mind. I was like, this is, this changes things. I think we have all had that moment at different times, at different moments. Maybe it was when ChatGPT came out, but for me, it was Instruct Update to GPT-3. I was like, this is going to change the tools we use for work. This is going to change how we interface with computers. That was really clear to me. I didn't know how, but I wanted to work on it. And in short order, I quit my job and started building prototypes and playing around with it.

8:55LLMs are extremely good at manipulating information, right? That's like one of their superpowers and taking information that's in one format and translating it into a different format. And that different format might be useful for different needs. Like in our example, it can take 30 pages of transcript, which is impossible to read. I don't know if you've ever, well, you probably have because of granola. But if you try to read a transcript from a meeting, it's like gobbledygook, right? And LLMs are extremely good at being able to take all of that and turn it into different artifacts that might be useful from that meeting.

9:29Anyway, so it became clear to me that LMS can have the superpower for manipulating information. I really wanted to work on tools that help us think better, like tools for thinking, right? Because that is like, in essence, what language is, what notes are. And I teamed up with my co-founder, Sam, because he was in this world also thinking about how can we build tools that help us think better. and when we started interviewing users real people we realized like the main place where people take notes and they generally take terrible notes are in meetings but like that is the killer application of this so if you could build a tool that would elevate the information or your thinking during meeting but also free up you know your cognitive cycles to be a better version of yourself in that meeting, it's a double win, right?

10:19Because you can perform better in the meeting. You can be smarter, pay more attention, but also have the higher quality notes and outcomes. So once that became, when it became clear that that was a really big opportunity, because there's so many AI meeting bots that had already launched at this point, right? So we kind of entered the space begrudgingly, but we kind of got at it from our first principles of like, how could LLM technology be most useful to people? Like what is a great application for it? And we entered it. And then we just spent a tremendous amount of time building prototypes, putting it in front of people, figuring out what works and what doesn't work.

10:56I think the fundamental question in Granola was, so you have a text editor, right? And you have all the context that Granola is aware of from the transcript or from the calendar events or from bios of the people who are in the meeting. So there's all this context that the AI knows about. How do you let the user make use of that or interface with all that context without being overwhelming and distracting or pulling you in directions you want to go? So that was a lot of trial and error and testing and experimenting and iterating. It does feel, and you mentioned, it does feel like there are lots of other products that are doing this.

11:37There are these meeting bots. There are these things that sit in the Zoom meeting. I think Zoom themselves is doing this. I think Google meet. There are a lot of these things. You don't really hear people talk about loving these products. And it does seem like there's been kind of this outpouring of love for granola. So what is it about granola that you're doing differently that is sort of leading to that user love? Is it just the fact now that you also have the scratch pad that pulls in context? Like that doesn't seem like that big of a difference. So where is the difference? Or maybe that is it.

12:10I don't know. On the surface, Granola and an AI meeting bot, they look very similar, right? They help you generate notes from a meeting. I think the moment you zoom in a little bit, you realize there are different categories of products, right? So I think what an AI meeting bot is trying to do, I think the premise is actually pretty flawed in my opinion, is you're going to say like, hey, I don't want to take notes. I don't want to have any control over the notes, right? This bot, this other entity is going to join this meeting and they're going to take perfect notes and the same notes will work for everybody.

12:44And now I don't take notes anymore. I've outsourced that to this AI entity and that's going to be great. And I think the position we took with Granola is one that was completely different. We don't want to take away note taking from you. right we want to augment your abilities so i think like a good metaphor here would be uh let's say you're used to riding a bike right and you know how to ride a bike and you know how to you're really good at bike riding you've been doing it your whole you know whole career uh and then all of a sudden granola comes along and it's like a it's like an e-bike right so now you you pedal once and all of a sudden boom like the bike goes right so you still have total control but you can get you can you can do whatever you want you can just do it with way less effort way faster.

13:35And I think if you come at it from that place where it's like, how do we augment what you can do so you can be the best version of yourself in a meeting or in the notes that you're trying to capture, you end up with a very different product. Something that's like really important to call out is that writing is a big part of the act of thinking. Like you talk to a lot of people and they're like, no, no, I write notes in large parts so that I'm processing information and so I'm smarter. And that's like a really important part of the process. You take that away from people. Like they might not be thinking in the same level that they'd want to be.

14:08What we want to do with Granola is we want, there's a very big separation between work that's just busy work and that work that requires judgment. Right? And I think this is like a, not just a question for Granola, but for all AI products where you want to automate as much of the busy work as possible. But you want to keep the judgment, like you want the human in the driver's seat, everything that's like nuance and judgment, because that's why you're in the meeting. right? You're there, right? So it's very easy to be like, oh, okay, now an LLM can write this thing for me. So we're just going to have it write this thing for me.

14:43There's the real danger of throwing the baby with the bathwater out, right? Which is like, you want to automate the busy work. You outsource your thinking to a bot. I don't think that's a future any of us want to live in, right? You want the thinking to be done by humans and you want AI to elevate or augment our abilities. Yeah. So it sounds like what you're saying is it's actually very, very important for the human being, the user in this case, to feel like they're in control. And even though that the bot is maybe now doing, I don't know, 50 % of the work, maybe even 75 % of the work, the human being does feel like they have agency to either drive the quality of the notes or change them or augment them in some way.

15:21And in fact, if we were to go all the way to 100%, well, now we're not even really thinking, right? We're just kind of, we're pointing the bot in a direction and just having it do all of our work, which kind of defeats the whole purpose of why we even have our job in the first place. I think that's exactly right. And I think an example is sometimes the most interesting notes from a meeting are not things that were ever said out loud. They're thoughts that you have. If you think about a hiring interview, oftentimes the most interesting notes are like, oh, I noticed this behavior. I'm going to write that down.

15:57There's no way a bot can read your thoughts. Maybe one day with Neuralink or something like that, they'll be able to. But for now, that high quality stuff needs to somehow still get into your notes. Sorry, did you mention that the bot also has context of other elements? Like it looks up bios of the participants and talk a little bit about that. What other signals do you grab to increase the quality of the notes? We found that the best way to think about LLMs is to think of them as a human, right? And you have to think like how, what context or what instructions would I have to give a human for a human to do a really good job at this?

16:33So in a meeting, if you just say, hey, there are five people in this meeting and they're talking about something versus saying, here are the people in the meeting. Like a good example might be like, here's a founder that is pitching an investor, right? On like trying to get an investment for their deal. That context really matters in terms of what are the quality of the notes and what you want to focus on. So Granola does, I'm not going to tell you all the different contexts we do, but we're tied into your calendar, right? So you can think of who's in the meeting, what are their bios, what type of jobs they have.

17:07Yeah, that's super interesting. I actually didn't know that about Granola. Another thing we do is, just so you're aware, is depending on your job, you have different types of meetings. and we're pretty opinionated about, okay, what would good notes for this type of meeting be? Or if the, for example, like if you're an investor and you're meeting with another investor versus a founder, whether that investor is a colleague, right. Or someone you're just gossiping about deals or like an LP, like those are very different meetings as well. Um, and Granola has an opinion of what, what's important in each of those.

17:42Yeah. That's super interesting. So it sounds like what you're saying is kind of taste and opinion is really, those things are really important. It's not about just like building the rails and handing it over to the user. It's about guiding the LLM in like a very opinionated way. And by opinionated, I mean the opinions of, I guess, you and your team. LLMs are really interesting because they're very good at a very wide range of things, right? Which means that you can build a decent or okay feature experience for an extremely wide range of things. I think what separates your ability to build something that's like okay or so-so to something that's really fantastic requires you to be very specific about, hey, what's the problem or the user flow that we're trying to solve for?

18:33And then be very opinionated on how to do the best job on that user flow. There's this trap, and I think there's this trap that we definitely fell into or flirted with before launch, and then we ended up cutting half of the app that we had built, where it's very easy to build features. It's very easy to build features that are like 60, 70 % of the way there, because with prompts, it's extremely, extremely easy to do that. And because it's so easy or so tempting, you're like, oh, it's silly that granola can't do X or can't do why, right? Like, it's silly. Like right now, we don't consolidate all your action items automatically for you in Grinnell, right?

19:11That seems like a no-breaker. But, and we had that, and we cut it before launch. And we cut it before launch because it worked okay most of the time, but sometimes it totally failed. And like action items are really important. And we can't, you have to get that right. And you have to design that in the right way. And you can't design and build everything from the get-go at a high standard. So you have to be very opinionated about what you what you prioritize. So, so it sounds like, you know, what I'm hearing from you is like constraints are really, really important. You know, you could build a million different things.

19:42LLMs, as you've noted, like they make it easy to build stuff. But you got to resist that temptation and you got to stay focused and you got to stay really, you know, laser, laser focused on, on the most important thing. And I guess in this case, it's, it's the notes, it's the notes, right? Is that, is that, is that a way you think about it? In a world where everything's possible, right you decide what's important by the constraints you place on yourself and and the users right so there's a constraint in terms of you know we're going to focus on this user flow or this experience and make that great there's also a bunch of opinions in granola but you talked about user love and i think a lot of what what resonates with people in granola is that we make a lot of decisions on your behalf based on what we think is important so we don't give you too many choices.

20:27We don't let you do a million different things. We do let you edit text, right? But we really try to reduce the number of decisions you have to make as a user because you're trying to rush to this meeting, right? The mental model we have is like, you are three minutes late to an important call, right? I don't really want to think about my software. I just want it to work and I want it to be there and support me when I need it, but don't make me think. So we, granola is not the best product for everything, right? But it's, it's really optimized for people who need to jump into a meeting and capture the important stuff and make sure that their thinking is all there and, and to leave with, with great notes.

21:07Maybe on the topic of great notes, I'd love to, love to pull on that a little bit. You know, one of the things that strikes me about meeting notes and LLMs is that the stakes are really high, right? If you're outsourcing a lot of your work to a bot, you want to make sure it's great. But LLMs, as we all know, are not yet imperfect. They hallucinate, they get things wrong, just like humans. How do you manage that? I kind of feel like if the notes were to get something wrong, I would have a hard time trusting Granola. So how do you make sure they don't get things wrong when we know they do all the time.

21:45The most important thing here is that the relationship with the LLM is one of collaboration, right? Going back to the meeting bots, they just give you the notes and they're like, here are the notes, right? Take them or leave them. You had no say on what was important or not, which is like a very different world. I wouldn't trust a lot of people to take notes for me, let alone an AI bot. So the fact that it's an editor is really important. But there are a bunch of principles that we kind of came across, I guess, through experimentation that matter. So the first is, if you see a note that Granola writes, you should be able to see where it came from in the transcript, right?

22:28So that's like this idea of being able to view source or view provenance is super important. I guess the high level idea here is LLMs, like you said, will get things wrong. or will hallucinate, or will do things that are maybe not exactly what you want. So you have to design for that. And the ability to kind of view the source or view the provenance helps a lot. So in Granola, if there's a note, there's a little like a magnifying glass that you can hit, and then it'll show you where in the transcripts, sometimes it's multiple places in the transcript that it pulled together to make a point. So you can kind of see the original source and be like, if you agree with that note.

23:04The other one is we try to make it clear to you what notes came from you and what notes came from the AI, from Granola. So notes, because like I said, it's a text editor, you can write notes and then Granola will enhance them or add other notes. Your notes show up in black and the AI written notes show up in gray. So at a glance, you can kind of quickly see like where did this come from? And is this something that I, you know, So it's like, what I want to double check, just having an understanding of like what came from your brain versus what came from, from the system is I think super important.

23:39So speaking of that and like where the notes are coming from, was it the human or the LLM? I feel like one thing that I hear talk about often in product discussions about AI right now is sort of figuring out like what the right interface is for engaging with LLMs, right? Like obviously, you know, a chat makes a lot of sense. you know you're chatting with this agent but at the same time chatting is a pretty like primitive interface for things and in many ways it's very it's it's very efficient for some things like fast communication it may be very inefficient for getting tasks done how are how are you and the granola team thinking about designing uh the interface of interacting with an llm as an industry or like it's very early days, right?

24:28And I think we have like the first UIs, people often talk about the metaphor of like the chat UIs a little bit like the terminal was, you know, in the early days of computing. I think the interesting question is how, what are the UIs where you and the AI are collaborating on the same thing as opposed to working side by side on different things? And I think that's an unsolved problem. And I think that will be different depending on your medium, right? If you're working on an image or text or a video or what have you. I think if you like chat is great because it's universal. You can use it to ask all kinds of questions.

25:13God, does it feel weird sometimes? So let's say I'm asking ChatGPT to help me write some text. it feels so weird that I have to then like write another instruction to make a small change in it as opposed to I can't go and edit what chat GPT wrote right this it feels a little bit like I'm trying to make a sculpture or like a like I'm doing pottery or something like that and there's like the AI next to me building it's doing its own pot and like like we're working on two things next to each other as opposed to both of us working on the same pot or being like oh can you can you change this on here?

25:43And I think that's a lot that needs to be figured out. For Granola, we're just the early days there. The idea that there's one editor, you write text, then the AI takes that, enhances it, but you can still see what came from you and what came from the AI. And I think we're starting to discover some things that seem to click. I think this idea of black text and gray text, which I think was pioneered by IA Writer, and those guys are great. That seemed to be one of those paradigms that might stick around because it's a it's like a simple way of like quickly seeing provenance but i think we have i don't know five six seven more kind of discoveries like that that are going to come along and become become ubiquitous so in this conversation we've talked a lot about kind of focus and really focusing on the use case and and just nailing that um and i think that makes a lot of sense how do you think about kind of the future and where granola can go So like, do you just want to focus on meeting notes or, you know, are there tangential use cases?

26:43And how do you even think about tackling those when it is so important to stay laser focused on the task at hand? It's a balancing act. I think they're right. If you look at this in terms of features, it becomes a really hard decision to make because there are so many features that our users ask us for that we want to build. and we know they'd be useful and some of them would be really awesome. So it's like, how do you decide this or this or this or this? I think if you zoom out and look at it from a jobs to be done perspective or a user journey perspective, it becomes a little bit easier, right?