Jon Noronha: How Gamma’s big bet on AI paid off

0:04Hey everyone and welcome to Generative Now. I am Michael Magnano. I'm a partner at Lightspeed And this week, I'm talking with John Narona. He's the co-founder of Gamma, an AI-powered platform that helps users create interactive and engaging presentations, websites, and social media assets in minutes. Gamma started in 2020, but made a massive pivot to AI in 2023. We talk about what that shift looked like and how it led the company to quickly skyrocket to many millions of users around the world. So let's get into it. Hey, John. Mike, great to be here. Thanks for having me. Yeah, thanks for coming on.

0:40I've really, really been looking forward to this. I think Gamma is one of these really fascinating AI companies. Obviously, over the past however many years now, two, three years, there's been so much fascination with the model layer of AI. So much innovation has been happening. But now everyone is talking about the application layer and trying to make sense of where there's value, what companies and great products will emerge. And one of the products that I think of immediately is Gamma. I know you didn't start as an AI company. So tell us a little bit about founding journey in 2020 to almost like the ChatGPT moment and how you got to where you are today.

1:20Yeah, totally. It really feels like the application layer is having this moment. But as you said, we didn't start as an AI application. And I still don't know if we could truly call ourselves AI native, although we're always trying to be more AI native. So, yeah, we founded the company in 2020 out of sort of the wreckage of another company, Optimizely. So that's where my co-founders and I and where a lot of our founding team came from. Optimizely got acquired in 2020, and it was a great opportunity to try something new. But it was also a very weird time because it was the deaths of the COVID pandemic.

1:49The whole world was changing and feeling very shaken up. our original thesis for the company, we've always been going after presentations as our initial market. We thought that PowerPoint is just so ripe for disruption. PowerPoint's coming up on 40 years old. 40? In 1987. Yeah, 1987 is when the first PowerPoint came out. If you think about how the product has changed over that time, there've been some big changes, like color was one they added in the early 90s, and a few other things like that. They launched clip art at some point. But based on the paradigm has been unchanged, and other great slide tools have come along, but they're mostly just PowerPoint on different platforms.

2:22So Google Slides is PowerPoint on the internet, Keynote is PowerPoint for Macs, so on and so forth. We saw this opportunity to rethink presentations from the fundamentals. So don't just try to make a better editor for PowerPoints, try to make a new format for how people present ideas. That was the whole original thesis. It was our original pre-seed bitch tech. It was, we're building the anti-PowerPoint. That was kind of the whole idea. And we thought that COVID provided this amazing why now. It was this period, if you kind of cast your mind back to 2020 when the future of work was a really hot VC thesis.

2:53And in fact, there were companies like Zoom, Loom, Slack, just going crazy with usage, incredible growth coming out of nowhere. It felt like the door was finally open to revisit all of these sort of like standard workplace tools. And so we thought it was an amazing time to actually dive in and go after presentations. And so we did. Those first couple of years, we were all about rethinking presentations for an era of remote work where people are not all in the same room, maybe not even all experience and the content at the same time. So what does it mean to design presentations for recording? Our vision was actually create something that was a hybrid presentation and document, something that you could send around before the meeting and we'll have all the information.

3:30Send around after the meeting, it'll even capture all the notes in the discussion. So a living document that had kind of the rich visuals of presentations, but also the like density and real information that a presentation would have. So like really leaning into like new format, like you said. Yeah, really leaning into new format, leaning into different, trying to be as different as we possibly could versus sort of like save as possibly could. And we'll come back to that because that has been both good for us and bad for us. Okay. That was sort of the original thesis. I think we did our first beta launch in 2021, launched on product hunt in a bigger way in 2022.

3:59And we ended up right around 2022 in this sort of dead zone of like partial product market fit. I'm sure you see a lot of startups end up in this zone where we'd launched something that had, you know, a number of diehard fans, but our number of active users was in the thousands. Maybe if we're being really strict, maybe it was in the hundreds. So we had a couple of passion people, but it just wasn't there. Certainly wasn't there for like a venture sized outcome, not there to get us to be able to raise a series A. And so we were starting to look at our burn rate and see that we were probably going to go out of business in the next year or so just running out of runway, if we didn't change something drastically.

4:34And so we had the choice of do we pivot the company or do we make any one more big bet on our current path, which is this sort of like presentations path. And frankly, it was lucky timing that right around that time of mid-2022 is when generative AI was finally on the cusp of getting good. These models had all been around in some form for a while, but they weren't really good yet. We even dabbled with them earlier on. But 2022, the key moment that happened actually was not ChatGPT for us. It was stable diffusion. So Stable Diffusion was the first really popular AI image model that came out. And it blew up on Twitter, I remember.

5:11Might even have been called Twitter at the time. It was people sharing these crazy pictures they created. And to be honest, they still weren't that good yet. They had all the weird hands and fingers and garbled text. But still, there was something magical about people creating illustrations for anything. And so we kind of had this realization that, like, gosh, what is painful about making presentations? So much of it is formatting and decorating. You're just putting all this time into making up clip art and pretty pictures to sort of like fill the space. And so we thought maybe this is the moment to really bet on AI.

5:42And so we started doing exactly that in summer 2022, really investing in it, really kicking the tires of what large language models can do. And I don't think we even fully realized this would be the direction for our company. I think we had this narrow idea, which was AI can help us with our cold start problem in our product. So we have this kind of new format. it's a bit hard for people to grasp our activation rate of like the percent of new signups that actually see the value is quite low. What if we could speed run you through that whole first, you know, hour of using the product and make it a first five minutes instead by using AI to generate that first presentation.

6:15That was kind of our initial narrow idea. And so we really bet on that. We threw sort of the whole team at it, a bunch of different ideas, not all of which worked, generating presentations, generating images, editing presentations. This idea of generating a presentation from a prompt turned out to be magical. And we basically put everything into launching that. And then we kind of had this another incredible stroke of luck, which was that year in November, Chachipiti came out. And we all know what happened next in terms of the whole world went crazy for AI, created this sort of insatiable demand for what can I do with AI, people talking about AI and social media, sharing what they made.

6:53It also meant that just in time for us to basically run out of money, the models got good. And so OpenAI released a new version of their model, GPT 3.5. So we ended up launching, basically, first few months in 2023, our AI generator. It was a total inflection point for us. We went from this sort of like middling zone of product market fit, like, I don't know if this is really a thing, to, oh my gosh, people are begging us to pay. They're begging us to monetize. And it's just been a totally different trajectory ever since. That's incredible. Yeah, it's funny. A couple of times, I think you mentioned the word luck.

7:27One thing I often think about with startups and definitely when I was building my company prior to being an investor, I feel like we too got lucky. And I want to take a step back and think about it. It's not really just luck. It's timing, right? You give yourself the space to exist at a certain moment in time when lightning strikes. And it sounds like you did a really good job of keeping the company around and keeping it going long enough for these moments to happen. Thank you. Yeah, I think you're totally right. It's the metaphor I've always heard that I really like is that luck in startups is like sailing.

8:00So you have to catch the right wind, but you also have to have rigged up your sail, put your boat in the right place to be able to ride the wind. And so for us, yeah, it was timing. In fact, it was, it was, you know, we had to be there when lightning struck, but we also had to have been building for basically three years ahead of that, building a product, building a team, building a technology base so that when it came, we were in the best position to seize that moment. You can build a product really quickly that takes advantage of tailwinds, but if it's not, you know, robust and the infrastructure is not stable and all these, like you can't scale.

8:30And so, yeah, those three years probably ended up being immensely valuable for Gamma before ChatGPT really took hold. So, you know, one of the things that I find really interesting about Gamma, not really even just Gamma, just sort of the space of creative tools is what you find often. if you look back at a lot of the successful companies, is if you just stick to one job or one format, it ends up not being a frequent enough use case to really support the company. You've expanded beyond presentations. And I wonder if that has anything to do with why you've expanded beyond presentations. You're totally right.

9:11And it's funny you say that, because when we were in particularly those dark times of wondering if we had a path as a company, one of the biggest things standing in our way was that so few investors wanted to invest in a presentations company a lot of bodies buried in prison yeah yeah yeah yeah it's a graveyard of startups exactly and there's this feeling of like despite the tam being extremely large i mean powerpoint google slides each have hundreds of millions of active users so there's no shortage of people but there's this feeling of like it's not a sticky enough use case it's episodic and you're going up against these massive suites like the office and google workspace suites where people are not even paying for the presentations product.

9:49Like you'll never sort of overcome that wall. Yeah, they just bundle it. They bundle it. And going up against bundling is a hard thing, no doubt. And so actually from very early on, even before we got into AI, we were already thinking about this idea of expanding the offering and going broad. I think maybe the best inspiration for us has been Canva, which has really shown the ability to build a blockbuster business by sort of diversifying across different assets and diversifying both in terms of making different stuff, but also winning over the same customer at different points in their week or their month.

10:19So taking someone from a weekly active user to a daily active user because they have five or six different things they use your product for. So we very much taken that approach. Presentations was sort of our starting point. It's still our biggest hook. But we've now added all these other formats that we support. So we have a document builder. And when we say document, think of something more visual than a Google Docs. So we make really shiny PDFs and brochures and e-books and that sort of thing. uh websites uh is one of our biggest ones and it sounds very different than presentations like oh my gosh websites it's a whole market in its own right but a lot of websites um are often quite simple and they are kind of just an online brochure if you think about like a small business owner or like a coach or something who's trying to represent themselves online actually their pitch deck and the website are almost the same thing just in slightly different shapes and so yeah we kind of lean into this idea of malleability which ai really enhances of start with a presentation turn it into a website, turn it into a PDF that you can leave behind by printing out at a client's site.

11:16And we just had another one in the last few months, which was social media. So basically, when people create social media graphics, like a LinkedIn carousel. And I think our goal is to go from two to four to maybe 10 of these different formats to turn people from those monthly to weekly to daily active users. Have you found so far that the user for these different formats is the same? Like somebody who needs a presentation is the same person that needs a website, who needs social media assets? Or are these actually bringing in like different customers completely? We have been amazed by the surprising similarity across different parts of our product offering and the ability to go horizontal.

11:52This is actually the other thing that came up in fundraising, which was so interesting, is so many investors, especially at that early sort of like pre-seed, seed stage, don't want to hear about your horizontal product. They want to hear about your vertical focus. So, you know, they say, tell me about the one persona that you're winning over. And I understand where this advice comes from, because if you try to serve everyone, you serve no one. And I think most founders who go after the horizontal thing crash and burn. And in fact, we almost did crash and burn. So I get where it's all coming from.

12:19Nonetheless, from the very beginning, when we've kind of pursued this vision, we have been shocked by the sort of like horizontalness of the space. We got the very start of Gamma. Actually, the very first thing we did before building anything was user research. We just reached out to our own networks. And we basically talked to 100 people. And we just asked them, tell me about the last presentation you made. And this was people in very different walks of life. It could be teachers or students or consultants or, you know, doctors, whatever it is. Turns out everybody had made a presentation because it's just common across many jobs.

12:49And they actually all had the same gripes. So it wasn't like the doctor needed very different features than the student. They all had trouble formatting things to look like. They all struggled with being judged for how visual their content was. They all struggled with keeping attention. It was like kind of all the same stuff. And now that we've expanded to more formats, it's even more true. The people who are making websites, now there is this distinction of like the designer developer versus everybody else. Obviously, designers and developers care a lot about getting all the details right. But 98 % of people are in the everybody else and they have the same needs.

13:22I just want to get something professional looking live without stressing about code and needing to pay somebody to maintain my site or change my phone number on it. And so we've been surprised by the regularity and surprised by, I would say, the level of crossover. When ChatGPT launched, and obviously you talked about stable diffusion early on, powering a lot of the images. Obviously, that was so magical to so many people. And I think it is still magical. But now, you know, across all of these formats, whether it's presentations or website, how much are you finding that people are relying on the actual AI versus just, hey, Gamma is a beautiful new format that makes it easy.

13:56I'm going to choose that over PowerPoint. Not because the AI features, just because I love Gamma. What is pulling people in now? Is the AI or the product? The AI is pulling them in, but the core product is making them stay. So it's kind of this dynamic of come for the AI. The AI is what drives the virality. It's something we can impress you with in the first five minutes of sign up. And it truly does answer that question of how is this 10x better than what came before? You can literally do something in two hours that took you 20 hours before. So that is still the promise. But I think this gets to your point around how focus has shifted from the model layer to the application layer.

14:27And I think in the application layer, a lot of this debate has come over this concept of a GPT wrapper. So are all these people just wrappers around GPT-3 or whatever it is? I think maybe what that perspective misses, this idea that often the wrapper is much thicker than the thing that is being wrapped. So it's almost like you have this magical jewel or something that you're putting into Thanos' arm. But you have to have the whole rest of the stuff, too. And so for us, yeah, that's core product. we spend much more of our time on features like having really good PDF export than we do on making the AI smarter, quite frankly.

15:01And that is maybe a misunderstanding about the AI application layer. I think in particular, when people think AI application layer, they think of these coding tools. They think of stuff like Cursor that's like totally blowing up, where actually the UI is pretty thin. It really is mostly just a way of getting clawed into your code base as fast as possible. That's right. Yeah. We have a bit of a different approach where presentations are a highly user experience intensive product. So is website building. Most of our work is actually UX design in front of engineering. It's how do we actually make it all work together?

15:31You see that actually in the composition of our team, which our team is like one third UX designers. Oh, wow. Zero ML engineers. So very different than maybe your typical AI company, but probably more similar to what you would see out of like a consumer software company. That's super interesting. So how do you think about like, are you guys doing any sort of research or developing your own models or even fine tuning your own models? Or is it really just all off the shelf and then make the product great with UX and design talent? It's off the shelf, but really heavy prompt engineering. So we've kind of dabbled in things like fine tuning, but we've actually found that prompt engineering is where a lot of the real power comes up.

16:07The challenge is that most users don't know how to prompt and don't want to learn to prompt. And so any interface that is just a prompt box with an output is not going to be very successful. And so there's an enormous amount of stuff in between what the user types into our product and what the AI model sees, which is where I think a lot of our value comes in. And that is prompt engineering, but prompt engineering is almost inseparable from user experience. It's like, what knobs are we choosing to expose? That also shows up in how we worked, which is our UX designers are also prompt engineers, are also front-end developers.

16:35We have these very messy overlapping roles where people have multiple skill sets that they combine to provide the right experience at the end of the day. So we're doing quite a lot of that. We do a lot of pre-processing and post-processing of what AI does to make it all feel sort of magical and consistent. A simple example of this is we now do a lot of AI image generation. Actually, we've generated almost a billion AI images. So we do an insane amount of image generation. But when we do that, we do a lot of work on top to make sure that those images feel coherent and consistent. So when you make a presentation that has 15 images in it, we do a lot of work to make sure they all have the same visual style and feel like they all came from the same place versus just a bunch of random prompts being sent off to Dolly or whatever it is.

17:17That's really, really cool. You're doing a bunch of great prompt engineering. I feel like one thing I keep hearing, I mean, I've been hearing it for two years now, but I feel like the voice just gets louder and louder is that if your defensibility is a prompt or a system prompt or prompting, you're toast. The models are just going to figure this out. Personally, I don't agree with this, to be clear. I invested the application layer. And so how do you think about that? Like, how do you think about the defensibility against, you know, something that can be recreated and just a prompt? I guess part of the answer, not to answer for you, is the UX.

17:48But like, maybe talk about building a traditional software company, like pre-AI, because it feels like that is what is defensible here. It's like just a great product. I think if all your product did was, for example, a blog post generator where you type in a line of text and you get out like a, I don't know, a 10 paragraph blog post, then I would tend to agree that you're toast. Like all you are doing is, you know, standing in between a really good text generator and the output. But I think it's different when you talk about like a real application that somebody lives in to do sophisticated work.

18:22And so I'll just focus on our use case. Although I think this applies more broadly in the application layer. Actually generating the text of the presentation is not the hard part. And in fact, ChatGPT is already really good at it. Models just do that. Yeah. But we see all of our value in being what you do with that text. So for us, that includes some things that we use AI for, but it's very domain specific. So actually laying out content in a presentation. So that means structuring ideas into multiple slides, figuring out where content goes on each slide, coming up with different ways of visualizing stuff.

18:51So not just boilerplate AI images, but actually our own sort of structured visualizations, things like a timeline that are very sort of like PowerPointy that we do. abstracting away all of that so that the user can give very simple prompts. And also, I think crucial for us, it all still needs to be human editable at the end of the day. So, you know, there's currently this craze of all these like, sort of web app builders, like your bolts and your lovable. I think they're really cool products. But they generally help somebody go from idea to prototype. And then they have the prototype. And it's like, okay, but can I really use this as my full app?

19:24Then there's a disconnect, because it's not really a fully editable app. And in fact, to make it a fully editable app, you almost need an interface like Webflow or something. where you have this full point and click no code type interface. A lot of what we have really invested is that no code interface all around it, where you can manipulate the content, edit very easily and quickly, and then do a bunch of stuff with it. So like we have a present mode, we have PowerPoint export, we have publishing a website to a real custom domain. That stuff is actually where all of the pain goes into. I see all of the sort of like, you know, coding work that our team is doing every day, all the pull requests, and 99 % of it is that kind of thing.

19:58not to mention making the whole application scale and all the things that can go wrong when you actually have to make this work across you know 100 countries uh across very different user types and with very large numbers of requests every day am i hearing you correctly in that you can't just build a product that takes something that's in a model and gives it to the user like it that's a wrapper right like a wrapper is a thing that basically just takes chat gbt and makes it look a little different. What you're saying is just like in pre-AI times, you have to provide value by cheating together workflows.

20:32That's what you're doing. You're building a product. Yeah, sure. It's getting some content from AI, but at the end of the day, that's not why people are using it. They're using it because it's a great product. Absolutely. And that product is great because of the way that it fits into their life and their workflow. It's because you are establishing a trusted relationship with that user that they can come back again and again and get a predictable result by working with you that they'll be happy with. And that's hard because they are working with unproductive technology under the hood. And so, so much of the magic is smoothing over the bumps of what AI can do and exposing it in a way that has like nice clean handles for people to grab onto.

21:07They're not using it because it's AI. In fact, I would argue customers probably don't even care about that. Like they're not, nobody uses a product for the technology, right? Absolutely right. One of the things that I often think about is the fact that every company wants to leverage AI now And every company wants to be an AI company. So when you talk about using presentations as a way to then ladder into websites and as a way to ladder into social media assets, all with sort of AI as the wedge and the thing that you mentioned drives the virality, you could see a world in which this product does look a lot like Canva or, I don't know, name some other creative suite.

21:46But as we all know, these companies want to go in the other direction as well. They want to take the suites they have and they want to inject AI everywhere they exist so they can get the benefit of the wedge also. So, yeah, just how do you think about that? Are you are you headed sort of directly at these really, really large incumbents now? And how do you think about differentiating yourself? I think inevitably, yes. You know, you either die in obscure startup or live long enough to fight the big suites. And so our hope is to live long enough to fight the big suites. And so far, I would say we're doing well in that.

22:18And I think what that really comes down to is the advantage of being lean and nimble. Imagine if AI development were to totally pause. So all capabilities just freeze at the moment they are now. I agree with many folks who believe that there would still be a huge amount of AI innovation to be squeezed out just from what we've already gotten. But I think in that world, what you would see is that the PowerPoints and Google Slides of the world would then sort of like have the predictability to be able to invest and plan something. But that's not quite the world we live in. We live in a world where the capabilities are changing actually quite drastically and quite frequently.