The Exchanges

Q Okay. How do you think about keeping a consistent user experience in a world precisely where those models, one, evolve all the time, two, behave differently, and three, as we all know, are stochastic, not deterministic. Do you expect the users, as you just said, to like be Be smart about it, and that's sort of theirs to figure out, or is that something that you abstract away for them?

A We abstracted away from, from people like the, the general design. At first we didn't, for the longest time, we didn't let users choose their model. Right. And we only let users choose their model on chat. On the note generation side, we completely abstract it away. And, and the reason we do that is every time a new model comes out, we have to completely change or tweak the prompts that we use for, for note generation to provide consistency of experience and an improvement of experience. And there's significant work that goes into that. Um, it's, I think one of, it's one of the value adds that Granola brings as opposed to just working with base models is that we take care of that and we make sure you get, Granola feeling or sounding notes consistently, um, and that they keep getting better over time.

AI assessment note: “On the note generation side, we completely abstract it away.”

Q on that topic that, uh, caught my attention, not having the bot first experience was actually a trade off in terms of like virality because you don't have the built in, um, expo product exposure because there's no bot. Showing up. Uh, so what have you done to, um, uh, sort of overcome that? And what are the, you know, viral sort of growth mechanics built into the product today?

A We haven't focused on growth. Basically what we really focus on is making the product really good for people. And turns out that that's actually, uh, led to a lot of viral growth, but that viral growth is from people telling each other. Uh, like, so like an interesting story that this, uh, This is, I never imagined that this could have happened, but I hear a lot now is, um, if you, if you have a one-on like, you know, you're meeting with someone on a zoom call and your AI bot shows up and you basically are told like, Hey, what are you, what are you doing with an AI bot? Like, why aren't you on granola yet? You know? And it's like, Oh, wow. The AI bot is now like a conversation start. Yeah. It's the weird thing. It's a conversation starter for a human to bring up granola. And to vouch for it, which is incredible. I never would have sat down and imagined that, that world. Um, so we always start from like a, like a value standpoint, like what is valuable to the users? Like, oh, we, we could email your notes to everybody in the meeting, like all the other companies do. But again, is that, is, is that a tool that you want to use? You know, is that acting like a tool for you or is that acting like, You know, like a growth engine. I, what, what we do have is, um, we do let people share granola notes. Basically, you can share notes on a link, and you can send that link to people, and w…

AI assessment note: “we do let people share granola notes. Basically, you can share notes on a link”

Q So a little bit on that note, uh, to, to close and, and, and, and zoom out, anything that you can talk about in terms of the roadmap or the futures? You mentioned a couple of times the idea of, um, searching through the history of, of meetings. So that's one thing, maybe double click on that, uh, and, or anything else that, uh, you can talk about.

A Yeah, absolutely. So I, I think that the world we're moving towards is, um, You have a bucket of context, and then you, you generate documents, um, or artifacts on, on a per need basis on the fly. And things, the types of things that we're working on are, uh, given my entire history of meetings, can you pull out really? So for example, a good question, um, that you could ask could be, um, that we could ask could be like, okay, out of everyone we've met in the last two years, like who are the firms who are most likely Uh, candidates to lead our series C, right? That's a question that I can't ask anywhere else in the world right now, but I have a version of granola that will go through my 2500 meetings and spit out a remarkably intelligent answer to that in 20 seconds. It's like a deep research mode, right? Um, the other thing that we've played around with is like, if you're, if you're dynamically generating Artifacts or UIs on the fly. Can those be shared, right? So this idea of we have a folder where all our sales calls, um, get put into and they're shared within the company. And then we have this artifact that you go to the URL. It doesn't have to be in the granola app. And it'll tell you, here are the most important things our, um, enterprise customers are telling us like today. And every time you reload it, it's up to date, but it's like a, it's, it's like a memo. Um, so the…

AI assessment note: “Granola that will go through my 2500 meetings and spit out a remarkably intelligent answer”

Q in terms of, again, for people, builders listening to this, what that means in terms of, um, What kind of team one needs to build an applied AI company these days, a company running on top of an LLM. So what was your level of sort of technical comfort working with LLMs, and at what point did you feel the need to start bringing people to do more technical stuff?

A I think the main thing that's changed here is that, um, It used to be that you need really strong technical chops just to build an MVP to understand if, if you, this is something people wanted or not. And I think the reality now is that, um, that isn't the case. You can, you can usually figure out MVP or like, is there, there, there may be even early product market fit potentially or signs there, uh, without a whole bunch of technical acumen. Um, if, as long as you're building on top of the models, like it's a completely different story, obviously, if you're building at the, at the model, uh, layer, but if you're building a, a rapper, a rapper company, like, like we are, then it's like, you can, you can learn a lot. And in, in those early phases, like when I was looking for a co-founder, I met Sam, but I was also, I met all the, all the LLM experts from Imperial and Oxford and Cambridge. Because I thought that was DNA we would need on day one. And as Sam and I started prototyping, we realized actually there wouldn't be much for that person to do until we figure out product market fit, until we maxed out on what like the, the base model, like the off the shelf models could do. And then we would need that expertise. And, um, and then we stopped looking for that person.

AI assessment note: “realized actually there wouldn't be much for that person to do until we figure out product market fit”

Q mode for about a year. So again, for builders out there, uh, how did you think about, When to launch, when not to launch, you know, there's this constant tension between, like, building public, you should be embarrassed by your first version, otherwise it means you launch too late, but on the other hand, you only get one chance to make a first impression. How do you think about this?

A Two thoughts about this. The first is, um, a simple way to answer this is, what is the fastest way for me to learn? So, like, Presumably, you start building something, you have some prototype, right? You have some early version of the product, and you say, will I learn faster if I launch publicly, or will I learn faster if I don't launch publicly? And the answer for us for about a year was, we'd learn faster if we didn't launch publicly, because we were onboarding users every day onto Granola, and it was painfully obvious what was broken about So launching publicly and getting, you know, 10,000 people telling us the exact same thing was actually going to slow us down rather than just fixing it based on what users were telling us. Um, there's a, there, there are many costs that come from launching publicly. Whereas like now you have users, you can't ship things with bugs. You can't, you know, you can't, if you pivot, it comes at a cost. So, um, we basically spent a year onboarding people, learning what was wrong about it, making fixes to that onboarding a new set of people. Fixing it and, and iterating. Um, and then basically the moment when we said, ah, like we now, we now have something that works and we're going to learn a lot more by having lots of people use it and realize who, who does it take off with, right? It's like maybe real estate people will love it in a way or use…

AI assessment note: “what is the fastest way for me to learn?”

Q And, uh, how does the prompting work behind the scenes?

A Granola takes in a bunch of signals about you. So who you are, what kind of work you do, where you work, who you're meeting with, uh, you know, what, what, where do they work? What are they trying to do? And, um, we've, we've put a ton of work into, uh, what are common meeting types for different With folks with different jobs and in those meetings, what are the things that really matter? So for example, I think the thing that kind of blew people's mind when granola came out, um, it was if let's say a VC, like an investor, uh, and a founder were both using granola in the same pitch meeting, let's say, The notes that Granola would generate for each of them look completely different, right? And as based, something as basic as I mostly care about what you said in the meeting, not what I said. Sometimes I care a little bit about what I said, but it's usually what the other person says. But also the kinds of things that I care about coming out of a pitch meeting is very different as a founder, as it would be as an investor. Um, and a lot of that is kind of, uh, hard-coded instructions that we, we build into the system.

AI assessment note: “a lot of that is kind of, uh, hard-coded instructions that we, we build”

Q Mm-hmm. As you get further into the enterprise, uh, do you get any kind of pushback or questions about, um, what it means for every conversation to be, uh, recorded and for it to be a, uh, I mean, quite literally, uh, like a track record or of, of, of everything that I was ever said from a, I don't know, legal perspective or, Any of those?

A Yeah, absolutely. So I think, I think there's two lines of questions here. One is, um, like it's important in the enterprise context that everyone knows that you're using granola, right? And we, we are like, we have functionality that posts this in the chat, uh, right now you can turn this on and be like, whenever you join a call, like post in the chat, let everyone know I'm using granola and we're going to launch a whole bunch of stuff that makes that better. But that, that's like one line of questioning and like, You should always tell people use canola. It's like the right thing to do. Um, regardless of what the laws of where you are, you know, I, I think it's like, we're also moving towards a world where like these tools, if they're well designed and not too invasive, like it'll be normalized in certain types of contexts. Um, then the other question is like, And this is not just a question for granola, but this is for AI in general, which is like, there's your liability footprint, right? Like the more you want to at Google, our emails were deleted after two years or three years, right? So you literally couldn't go back and search and see why a certain decision was made on a product. You know, like, why did we do this on Gmail three years ago? You, you, you couldn't find that in the email. Um, And, and that's the limit, the liability footprint in a world of AI, where all tha…

AI assessment note: “So I think, I think there's two lines of questions here.”

Q on hide all works behind the scenes, the, the tech stack, the mechanics of it all. So starting with the model, so I'm seeing on the, on the website and as a user, um, you use lots of different models. So as a first question, do you, uh, exclusively at this point use third party models? So have you built some of your stuff, uh, from a pure AI perspective?

A No, we were, we, our philosophy is to use the best model that is on the market as quickly as possible. There's so much, uh, value in focusing on the product, like low hanging fruit when you focus on the user experience today and the base models are getting so much better and smarter so quickly. Like our strategy has been to use the latest and greatest. And when we feel like we hit a wall and the only way to make the experience better is then to, to fine tune or train models. And we will do that. Um, what we found is that there's so much alpha and like the improving models and making sure you get the most out of that, that that's kept us busy thus far.

AI assessment note: “our philosophy is to use the best model that is on the market”

Q Yeah, but call it, you know, a board meeting, which is four hours, right? And there's, like, rate limits and all the things. Or maybe it's no longer a problem at all.

A No, no, no, that's not a problem. The problem becomes now, well, okay, there are two things. One, notes aren't very, notes are short, transcripts are long, single meeting is fine. It's when you have lots of meetings and a large corpus information. That's where this problem comes up a lot. And the, like, The interesting The trade-off or the thing that's tricky here is that if you care about information lookup, right, then, then you can do rag, you can do, well, either like some form of keyword search or cosine similarity. What we found is that a lot of the most interesting queries that people have, um, like would completely fail with that type of method. So for example, query might be like, what are all the things, um, I didn't do a good job explaining. Um, or, or tell me what are all the bugs that this user encountered, uh, that, you know, in this user call? Um, and the only way, the only way you can, you can get a good answer to that is if the model has the full context. Um, and so we, and this is very costly, but we, we generally tend to put lots of context into the context windows. Um, and we are on that side. Um, and we have some, some really cool stuff in the works where we look at full context across thousands of meetings, which I've, I've told you about, but again, it's, uh, it is costly with today's technology. Like there are trade-offs and the trade-offs are basically …

AI assessment note: “No, no, no, that's not a problem.”

Q And to the tangible versus untangible point, how do you decide effectively even to play back some of what you just said, the 50% that you need to cut? Are you looking for qualitative feedback? Or do you look at quantitatively, uh, what people actually do with the product? Uh, which part is science, art, uh, taste versus Data measurement.

A Yeah, it, it, it's all of the above. So our, our general our philosophy that's gotten us here, and it may not get us there as we scale, is, um, uh, we make most product and design intuition, sorry, decisions based on Intuition. And so like, what, what do we think makes sense is kind of this vision of the product and that we're headed towards. And, um, that like, we just kind of make the decision based on like, does this feel right? Does it feel like it's in line with the vision and what we do to make sure we're not divorced from reality. Is it, think of us as like an LLM. It's like, we try to fill our context with as much real, um, user feedback, user opinion as possible. And some of that is quantitative. Of course, everything we launch, we measure and we look at the graphs and it was like, Ooh, people, a lot of people were asking for this and not that many people use it. It's like, Oh, it must've been the loud minority. Um, and you know, that's a very important tool in the tool chest, but what I think is even more important is constantly talking to people. And that's something where it's like Sam and I, and I'm talking about Sam and I because we work very closely together, but, uh, like most of the team actually, they, they do regular user calls. And we, we, we aim to do, I think Sam and I aim to do four to six calls a week, uh, with, with users, but constantly not like, oh, w…

AI assessment note: “Yeah, it, it, it's all of the above.”

Q like the Zooms of the world doing that. So for the builders listening to this who may look at a category and see a few companies and try to decide whether they should build in this category or ignore it and find a category which is less crowded. How did you guys think about, oh, we can come up with something that's going to be better than all of them?

A It's a great question. So, um, I think the answer really just comes down to Like what were we trying to build when we, we set off to start Granola? There have been like meeting transcription or recording products like Otter and Fireflies, I think are like nine years old, right? They've been around for, for a long time. Like that's not a new idea. Um, there are all these tools that are like, okay, we can now record meetings, right? And, and, and try to make something useful there. That's not at all where we started with Granola. Like that's not, we don't think like a meeting recorder, that's not what we're building. We, We want to build a tool for thought. Like the, the, the genesis of Granola was I, uh, I quit Google because Google bought my last startup. So I quit Google knowing I wanted to do a new startup and I came across LLMs for the first time and they blew my mind. And I was like, this is going to change everything. This is absolutely going to change the tools we use for work or productivity tooling. And I met my co-founder who had come from a tools for thought knowledge management Uh, space. And we basically said, ah, AI is going to let humans work differently, think differently. There needs to be a tool that supports that. And that's what we want to build. So this idea of a contextually aware, um, uh, workspace, like AI powered workspace, like that's what we wanted to …

AI assessment note: “That's not at all where we started with Granola. We want to build a tool for thought.”

Q And, and, uh, and then more importantly, what, what has it been like?

A It's funny. I, um, we're here. We were in London for personal reasons. My wife's, um, my wife's English. Uh, we moved here. I knew I wanted to do a startup. We chose London because there, there's a, There's amazing engineering talent here. There's enough to, there's nothing of ecosystem here to really have a go at as a startup. And then, um, when I decided I wanted to build an AI startup, I said, oh my God, yes. Like deep minds here. Like a lot of like modern AI was like invented here. Uh, some of the best programs, like I said, UCL, uh, Cambridge, Oxford, Imperial, they have amazing AI programs. And then, and then I had the realization that actually, you know, product and design and just general product taste and building is super important. Didn't need to hire those folks, uh, early on. I think there are trade-offs, right? I think there, there are very real trade-offs. There's like a center of gravity of talent in Silicon Valley. I think we're in a very lucky position to be, I'd say, one of the, the most, um, Uh, visible and desirable, like AI, like consumer facing AI startups in London. So for, there's a, for the continent of people over here, like they, they find us, which is, which is incredible. And I think, I think in an era of where taste matters and product sensibility matters, there's a, there's just amazing talent here for that. Um, as well as amazing engineering tal…

AI assessment note: “The upside is probably that it's a little quieter over here.”

Q How do you think about, uh, guardrails? Make sure that the... System doesn't spit out, um, you know, things it shouldn't, for example.

A I guess for every product, the idea of like, what should the system not, what, what, what would be harmful or negative for the system to spit out is a little bit different. I think if you go to, um, if you go to something like Google or ChatGPT and, and you ask it for something, it's like an open-ended place where you're, you're looking for guidance or help or health advice or what have you. There's some really, really, uh, bad scenarios there. Um, for our case, it's, it's, it's a little bit different. You're usually going back and asking questions over your, your, You're meeting data. So there, there's a question of, we get something wrong, or it hallucinates, um, but what we found there is that the best thing to do They're never going to get it a hundred percent, right? We can never be like, oh, you know, we make no mistakes, you know, you just trust us. Um, so obviously we do the best we can to avoid, uh, those mistakes, but really what's important is the way you, you design your product needs to let the user kind of like, like view source, kind of look behind the curtain and be like, wait, where, like, how did you construct this answer? Like where, what are all the citations? So we spend a lot of time thinking about citations, about letting you, uh, view original transcripts and quotes, and there's a lot more we want to do there, but that's really been the, The, the way you…

AI assessment note: “we spend a lot of time thinking about citations, about letting you, uh, view original transcripts”

Q why wouldn't open AI do that. Uh, but, um, Uh, you know, zoom as a, as a, as a product. Um, and it seems to be so fundamentally important what you, what you're doing and so horizontal and so, you know, transformative that it feels like all those great companies should focus on this at one point or another. How do you, how do you think about navigating that tension?

A Yeah. Well, I guess it was really helpful that, um, most of our competitors had Some kind of AI note taking feature, but when we launched, so it all, that was already the case and, and somehow Granola was able to stand out and like win people's hearts and, and, and grow. I think again, the failure mode here is to think about the world as it is today and the product capabilities as it is today. And I think my view is that, um, Notes are, are kind of useful. I think they are a stepping stone to the way we're going to work in the future. And the way we're going to work in the future is with AI that has really deep personal context about you. And the product experience that is, that Granola will be a year from now, two years from now, will look radically different from what it is today. Hopefully it'll still be very simple, but it'll, it'll help you do a lot of work. And I think, um, no one, no one's built that. Yeah. Like a lot of people are racing towards that, and I, and I, and AI is an incredibly competitive space, but when we're talking about really, uh, when you're talking about products that are like native to a new medium, right, oftentimes, uh, startups have, have an advantage.

AI assessment note: “oftentimes, uh, startups have, have an advantage.”

Q where, like, directionally you guys are, are, are at where you for now operate on, uh, lower gross margins, and I will not ask you for any specific numbers, uh, but again, you're building for the world of tomorrow where you have higher gross margins. Um, is that, is that, is that, is the world of like meetings different in terms of like, um, token needs, uh, versus AI coding?

A So the most expensive thing about our business is actually transcription. Uh, and historically it's been actually transcription and high quality transcription versus, um, LLM inference. And, um, the, we basically use the best transcription Real-time transcription on the market at any time. Um, and the cost of transcription has, has fallen dramatically, uh, over the last couple of years, and I suspect we'll continue to do so. Um, so yeah, we're, We're not at a negative gross margins right now, and, but what I do expect is I expect the cost of inference to stay the same or go up as we allow users to do much more complicated queries over much larger data sets. It'll be an interesting race to see, like, you know, does, does the cost of inference go down, um, faster than the user desire for more complicated, more intelligent features goes up.

AI assessment note: “So the most expensive thing about our business is actually transcription.”

Q And do we think long-term that, uh, what happens? Like we become, we have more time to actually Think and reason, but, uh, it sort of feels like models are doing that for us as well.

A I mean, I, I think this is the, this is why I'm so excited to be building Gridola right now, to be working in the space, because I, I think that future is kind of up to us, right? In terms, like there's this great quote, which is, we shape our tools and thereafter our tools shape us. And when you think about AI and, um, the future world and how it fits into our society, I, I think there's this big question of like, what, you know, what do we outsource to AI? Where does AI replace humans and, and where does it augment humans? And, um, where like me personally, Sam Granola, we're really big fans of the augmentation ideas. It goes back to Douglas Engelbart in the fifties, right? Augmenting human intelligence and his, his view, honestly, his stuff, I feel like people don't talk about him enough. It's just so inspiring. Like he's known for being the inventor of the mouse. And I think the mouse is like the, the least important thing he's come up with. And, and basically this was, Um, when computers barely existed, like the computers that existed were in the military, in the Navy, and they, they like filled up whole, um, uh, whole floors, right? And there are a, a, a bunch of folks actually who are true visionaries at that time who imagined this world where computers would be accessible to people, they'd be common, and they would be tools for work and tools for thought. And the way En…

AI assessment note: “we're really big fans of the augmentation ideas. It goes back to Douglas Engelbart”