# How Duolingo Build Product 10x Faster with AI

Duolingo's Biggest Lessons on Paywalls, Push Notifications and In-App Purchases · Why Small Teams are the Future of Product · Why PMs Will Become Extinct with Cem Kansu, CPO @ Duolingo

20Product · Jun 20, 2025 · 79 min · 15,801 words
Speakers: Harry Stebbings, Cem Kansu
Source: https://www.996.fm/episodes/20vc--ep-fcdc9a93/

## Cold open

**Harry Stebbings** [0:00]:

This is 20 product

## Intro

**Harry Stebbings** [0:01]:

with me, Harry Stebbings. Now 20 product is the monthly show where we sit down with the best CPOs in the world to discuss how they build their product teams. Today, we've got one of the best joining us in Cem Kansu, CPO at Duolingo, one of my favorites. At Duolingo, Cem leads product strategy for over 90,000,000 monthly active learners. Since joining, Cem has played a pivotal role in driving user engagement, revenue growth, and product innovation. He's launched Duolingo Math and the wildly successful Duolingo Music. Under his leadership, the company's consistently ranked as the number one education app globally. This is a masterclass in how Duolingo has created a product of true consumer delight. But before we dive into the show's

## Sponsor read

**Harry Stebbings** [0:44]:

are you struggling to beat model benchmarks or implement Gen AI in your product? If so, you need Turing. Turing is an AGI infrastructure company backed by incredible investors like Foundation Capital and Westbridge Capital, and they do two things. Number one, they help leading companies in AI labs like Salesforce, Anthropic, and Meta enhance their LLMs with advanced reasoning, coding, multilinguality, multimodality, and more. Number two, they combine human and artificial intelligence expertise to deploy cutting edge AI systems for awesome companies like Rivian and Reddit. Right now, Turing offers a free self assessment to help you pinpoint your place in the Gen AI journey, get tailored next steps to optimize your model strategy, and then finally learn how Turing can refine and implement your models for better performance. Take the guesswork out of Gen AI, visit turing.com/20vc to start your free assessment today. And once your Gen AI strategy is on point with Turing, now let's switch gears to the future of mobile and how one company is turning your screen into real returns. We spend nearly half our waking hours glued to our phones, upwards of fifty hours every week. Recently, one company transforming this reality stood out so much I personally became a shareholder, Mode Mobile. Mode Mobile created the Earn Phone, the smart phone that pays you for daily activities. Instead of big tech profiting billions from our attention, Mode returns over $325,000,000 directly to users through earnings and savings. Mode's revenues surged an incredible 3200481% in three years, recognized by Deloitte as twenty twenty three's fastest growing software company in North America. And here's why I'm excited. Mode's equity offerings have raised over $30,000,000 from 20,000 plus retail investors, one of twenty twenty five standout public raises. And you can now join me as a shareholder with as little as a thousand dollars at invest.modemobile.com/20vc. For a limited time, unlock up to a 100% bonus shares and a free earned phone. Email us for the investor brief at 20vc@modemobile.com or check out invest.modemobile.com/20vc. We've talked about making money from your phone. Now let's talk about making your product do exactly what you want it to. Whether it's the software you build or the software you buy, your tech stack should be creating results, not creating roadblocks. Well, Pendo's no code software experience management platform makes your software better with tools to see where users get stuck, guide them with in app messaging, and constantly improving your UI. It's so easy that over 14,000 businesses use Pendo to increase revenue, lower costs, and reduce risk. Businesses love the control. Engineers love the freedom. Everyone wins. Start for free today at pando.io/20product. You have now arrived at your destination.

## Conversation

**Harry Stebbings** [3:48]:

Cem, dude, I am so excited for this. I just mentioned it, but I freaking love the Duolingo product. I I told you I I love

**Cem Kansu** [3:56]:

that you love it.

**Harry Stebbings** [3:56]:

It's so nice for me because I interviewed lots of CPOs Yeah. And I don't love their product, and you kind of have to pretend. You know, it's like you're on a date where you're like, oh, really? That's what you do? Oh, fascinating. And here, it's like, I love it. And you know what I love? It's the attention to detail.

**Cem Kansu** [4:11]:

Yes. Yes. So

**Harry Stebbings** [4:12]:

special.

**Cem Kansu** [4:13]:

Yes. And you're learning Spanish. I'm learning

**Harry Stebbings** [4:15]:

Spanish.

**Cem Kansu** [4:15]:

Yes.

**Harry Stebbings** [4:16]:

Bruno s Americano. Mhmm.

**Cem Kansu** [4:18]:

Thank you. Attention to detail Thank for you so much. No. We take pride a lot in the pixel perfect details. Like, I think you'll see that in the product. Right? It's like every screen will have a very specific design language, characters will be very specific, etcetera. So, like, I'm glad you noticed it.

**Harry Stebbings** [4:35]:

Totally. Let's have some chronology because otherwise I'll just go off on a tangent. Can you take me to the moment you actually joined Duolingo?

**Cem Kansu** [4:42]:

Yes.

**Harry Stebbings** [4:42]:

Like, how did that come to be?

**Cem Kansu** [4:43]:

Yes. So I joined in 2016. It was kind of a dark, rainy day in September. I'll skip the funny part of the story. Anyway, like the joined the stories.

**Harry Stebbings** [4:53]:

Oh, I'm feeling out

**Cem Kansu** [4:54]:

September 2016 in Pittsburgh, Pennsylvania. That's where our headquarters is. And I was the first PM that was handled the baton for we were pre revenue, we would like to have revenue, and can you help us figure out how to monetize? Obviously, was an early career PM, and I knew ads. I had worked at Google before. My first experiments my claim to fame at Duolingo was putting ads on Duolingo. You can imagine, this is good yeah, fame exactly. Good fame and bad fame at the same time. And then I worked on monetization for about six years. So I built all of our revenue features, ads, subscriptions, in app purchases, and everything that's under our subscription umbrella. Subscriptions obviously became the lion's share of Duolingo's revenue, and then I kept on building our teams that work on revenue and building our features that work on revenue. And in the past three years now, I've moved on to being our head of products, so I lead the product management team, and now I dabble with all of our product efforts.

**Harry Stebbings** [5:55]:

I mean, my word.

**Cem Kansu** [5:56]:

How exciting. It's

**Harry Stebbings** [5:58]:

been a it's been a great journey. So you're the guy who puts ads in my Duolingo? I did that. Yeah. Yeah. Brilliant. Well, listen, I'm sorry. You're welcome. You know what? I think it's well done. Can I I am just jumping around to be fair, but my favorite dude on product is Gustave Sodashroom from Spotify? And he always told me that the details are not the details, they are the product. Yes. And I'm just intrigued when you think about product building. How do you think about the details being the product versus, for Christ's sake, we actually need to build product and move on?

**Cem Kansu** [6:27]:

I think I have a very similar saying I repeat a lot. Consumer products live and die in the pixels. Meaning, if you are working on a consumer product, you have to figure out how to make your change pixel perfect for the user. So, for example, product managers are sometimes viewed as these like, I'm the mini CEO. I decide strategy and then people execute. That is like the worst definition of a product manager that I've ever heard.

**Harry Stebbings** [6:51]:

Why? Because everyone says that on the show.

**Cem Kansu** [6:52]:

They love saying that. First of all, the CEO is the CEO of the product, not the product manager. The product manager has to understand what the user will respond to when you kind of give them what they're going to see in the app. So they have to understand user design really well. They also have to be able to stretch into business strategy, etcetera. That said, this is why consumer products live and die in the pixels, and attention to detail matters. If a button that's like the dark shade of green versus the light shade of green will make a difference in user behavior, a product manager or a product designer, for that record, like people working on the product, should understand what that means. And I think this is also how we build product teams at Duolingo, where everyone should understand pixel perfect design and why certain specifically what the user is going to see, how that's going to change user behavior.

**Harry Stebbings** [7:39]:

Can we start at the start of the process, which is the design process? Yeah. What does the design process look like for you today at Duolingo? And has that changed over time?

**Cem Kansu** [7:48]:

It's evolved. It's evolved quite a bit. I joined Duolingo when it was 50 employees. Now we're about eight fifty. So things have changed and we've gotten a little bit more processed. But at its core, the way it starts is you start with all of us use the app heavily. So most of the ideas that we work on come from our own usage of the app. Not all of them, but a lot of them. So for example, if we feel like and maybe this is the example we can kind of highlight, we felt like Duolingo didn't teach speaking enough. So when we were testing GPT-three, we were like, Okay, this is a good way. We've tried many different things to teach speaking and then GPT-three was this light bulb moment where we can actually teach speaking better if we use LLMs. So the design process got started with like, you have the technology and you have Duolingo, how do you bring them together? Through various iterations back and forth, we prototyped a thing where Lilly talks to you. We really liked it. We made a feature out of it. Now it's called Video Call with Lilly. Now a lot of our users have access to it and are using it.

**Harry Stebbings** [8:47]:

So I was walking through Bell which is a very smart part of London, the other day. And I was on Duolingo. And I was going, And this guy comes up to me and he's like, I'm sorry to interrupt your Spanish call, but I I love your show. And I'm like, you've literally been standing there listening to me go Bruno S. Americano and Voya la Playa thinking this guy is a fucking idiot. So yes, I agree. And it's fantastic integrating it. But that design process then, is it fast? And who needs the approvals? Don't laugh. Do you do Figma? Do you just go straight to prototyping now? How does that work?

**Cem Kansu** [9:24]:

Yeah, great question. So we use Figma for a lot of our designs. That said, for something that is an experience, you gotta experience. Figma is good for us to say, This isn't headed in the right direction. Let's go to prototype. We have a product review process where our CEO, Luis, is involved, I'm involved, our head of design is involved, and our product leaders are involved, where teams come, they show their change, and we give feedback, they refine it, and they either go to prototype or they go straight to an AB experiment. Wow. So depending on the size of the change, we might do extensive prototyping. In the case of video call, we prototyped for, like, maybe two, three months. We did one change, we tested it, we didn't like it, we iterated, and then it was ready to go out to our users. What does extensive prototyping entail? Extensive prototyping is, I think you Let's say we have an idea. We think calling Lily is going to be a good way to practice our Spanish. The way that the first prototype is going to look is like, it's not going to be well animated. The illustrations are going to be janky. Maybe there's going to be, like, error codes in, like, some part of the UI, but we'll test the core interaction of can she give us smart answers. We'll try that. And then we'll be like, Okay, if she can, let's move to problem number two. It needs to look engaging. We solved that. So I think extensive prototyping is you prototype, test, and refine what the spec is. So it's like a step one utility, step two design, step three perfection. Pretty much, exactly. And I think for smaller stuff, like if it's not an entire new feature, you can do it all in one go. But if the extensive iterative prototyping is good because you don't have to invest, like, months of time before you see if it works or not. So you validate if it works, and then you kinda do the other steps.

**Harry Stebbings** [11:04]:

Is there subjectivity to if it works? And, like, a user behavior like Lilly, you you could be like, well, I I thought that was really engaging. And another person would be like, it's not.

**Cem Kansu** [11:13]:

Yes. Oh, a 100%. What do you do then? Well, I think we bring we bring a lot of smart people in the company together. We have them all tested and see if we can align for the most part. We don't all have to align. But generally, if enough high judgment people are like, yeah, this kind of works. Like, I can actually practice speaking better than everything else I could do on Duolingo, which was the case. We're like, we have something good on our hands. Let's make it better.

**Harry Stebbings** [11:38]:

You mentioned Lily there. I obviously spoke to Severin the other day. Wonderful, elegant man. And he told me about, like, two of the I think it was PMs who came up with the idea for chess and brought that forward, and I wasn't really sure. How do you think about the ideas you choose to pursue versus not choose to pursue?

**Cem Kansu** [11:56]:

This is a great question. You know, I would love to say we have this, like, super ironed out process. You know, you follow ABC and you have the perfect answer to the question of what to invest in. The reality is it's a lot messier than that. But Chess is a really good example of bottoms up innovation in the company where it's not an executive that had a great idea and said, Yeah, we should go in this direction. So it was a product manager and a product designer. They didn't have an engineer on the team. They went and prototyped basically. They're like, we think chess will be a good subject for Duolingo to do. We're like, kinda weird, doesn't exactly fit, but we're like, yeah, cool idea. They went and prototyped and almost like a full Chess lesson experience with no engineers. So, like, our PM vibe coded basically the whole thing. Our designer created excellent looking designs. We tested it and we're like, this is a pretty good Duolingo lesson. So they kept iterating actually on this vibe coded prototype, and then until a point where we're like, this is very much headed in the right direction, we really enjoy using this prototype. And then we staffed a team with engineers and now it launched this week. So in a span of eight months, it went from like, I don't know, this is kind of an interesting idea to like a full launch of a new Duolingo course.

**Harry Stebbings** [13:07]:

How does that

**Cem Kansu** [13:07]:

differ in terms of

**Harry Stebbings** [13:08]:

timeline from

**Cem Kansu** [13:10]:

traditional idea to launch? It's way shorter and way less people. Our experience with doing this was with math and music before. It was triple the amount of people and double the amount of time it takes to get a course out. The big unlock there is, I would say, maybe two things. One of them, AI coding certainly speeds up because you can prototype quickly, get feedback, iterate, rather than if you couldn't do that, you have to have an engineering team that's like building prototypes and that's a much slower process. The second one is, honestly, think having great people. I mean, we had a great PM and a great designer and they know how Duolingo works, what we think is good, so they can independently make really quick decisions. So I think staffing a great team with great tools was the big unlock.

**Harry Stebbings** [13:55]:

When you look at that process and you look at the team that went into it, what does the structure of the product team look like today? And how will that change over time?

**Cem Kansu** [14:04]:

This is a great question. I think today, the traditional product team is you're gonna have somewhere between five to 15 people. That's the traditional product team that has engineers, designers, and product managers at its core, and it might have other functions like data science, marketing, depending on what they're doing. What I expect is going to happen, which is maybe the Chess example is a good example, is these teams are going to be smaller because you might not need as many people to do the task at hand, which might be just like, we just need to create a prototype of this thing. We just need to figure out if this is the right design. We're probably not going to need as many people. So we'll have smaller number of people on each team. But honestly, I'm hoping by doing that, we'll be working on much more innovative bets quicker because, again, like, this chess example is a good one. If we can do that, like, 10 x more, well, Duolingo will hopefully be 10 x better over time.

**Harry Stebbings** [14:55]:

Will it be 10x better or will it just be 10x broader? Because there's a question of, like, you can go for breadth of subject or depth of engagement, and that is a challenge. I love a Matthew McConaughey talk, and this is why this show is so successful, because I bring in very wise leaders. But he says, you know, options make an enemy of us all, and actually constraints are sometimes good. Yeah. How do you think about ensuring feature simplicity without creep?

**Cem Kansu** [15:23]:

I think focus matters a lot in product. I think you can really distract yourself by going too broad. So our goal is not to 10x the number of subjects Duolingo teaches today. Making Duolingo 10x better in everything we teach is the main goal. That said, I think if we find a great subject that would be additive to Duolingo and we believe we can do it really well, I think you'll see us doing that. Right now, we teach four subjects like language, math, music, chess. The goal is not to get to 40 subjects and do them all equally the same. I think make what we have 10x better is the primary goal.

**Harry Stebbings** [15:57]:

Can I ask, when you think about design and when you think about product in eng, does what it takes to be 10x as a product person change now? Like you said that they've vibed coded.

**Cem Kansu** [16:10]:

Yeah. Do we have just hybrid people? I think so. I think what is likely to happen is there is clear work definitions of what an engineer does today, what a product manager does, and what a designer does. Think these are going to converge way more than what's happening today in technology organizations. So I don't know if they'll converge fully. I think that's maybe a little far, but they will converge in the sense that they'll be a lot more all of them will be capable of taking over each other's jobs in multiple ways. Because AI tools will enable them to design faster or PMs to start designing more or PMs to vibe code more or designers to do the same or engineers to stretch into design or PMs. So I think these roles will converge more. That's already happening, honestly. If you were advising, like,

**Harry Stebbings** [16:55]:

a young product person, say, 20 I'm your little bro. Yeah. I'm 23. I know we both have the good looks to carry off. But what would you advise me as a product person entering this very different world?

**Cem Kansu** [17:07]:

Learn AI tools in and out today, and they will look different almost in, like, four weeks. So I think you should be experimenting with how AI tools build product, like, very heavily. This is what I admire my own team. Like, we have basically told and aligned with our team, we will be AI first as a product team. This is very similar to when companies went mobile first in, like, 2012. It felt really artificial and weird because, like, tech companies looked at their user base and said, oh, 1% of our users are on mobile, 99% is on web. Going mobile first is kinda stupid. Like, we're investing in 1%, But, like, quickly in, like, three years, they realized that was the right bet to take, but it was quite artificial. Anyone that's going into products should bet on the same thing, which is you should build AI first, even though when you use these AI tools and say, Alright, create me a great app, you end up with a dinky web app. In my opinion, that will get solved in, I don't know, maybe twelve months. But I think if you know how to use these tools to create product, build product, and you can kind of harness their power to build great things, you'll come out ahead.

**Harry Stebbings** [18:15]:

There's a difference between idea and action. And I would like to say that with my team here. Obviously, we have a big media company, and I would like them to be bluntly trying tools. And I'm like, should I implement an hour a day testing time? Like, jokes aside. I think you 100% should. What do you do, like, deliberate to make sure people actually are? Because people are lazy, bored, and tired.

**Cem Kansu** [18:33]:

There's gonna be inertia. Right? I think it's like just saying, yeah, we should all use AI, I think that doesn't actually get groups to be like, yeah, like, I'll I'll go to this tomorrow. I mean, some people will because they're they're excited. I think what tends to work is, I think in one incentives, you should incentivize people to automate repetitive tasks. I think you should award and highlight people doing that. That's what we do. I think that helps. How do you reward people for automating? One is recognition. I think that's what we do right now a lot. So if one of our product managers automates the task, we encourage them to share it with others and we clap. I think that is one way to recognize someone that has done something great. I'd rather have 25 ks. I think you can

**Unknown** [19:12]:

Can give me a clap?

**Cem Kansu** [19:14]:

Money is another incentive, so I think you can reward people with compensation. That's another obvious one. And then I think the third is, in an organization, I think it's also in a tech company or like Duolingo, we bake it into how we think about performance. So if someone is actually creating more output with AI, well, they'll probably get promoted faster, so. What do we do in product today that we definitely won't do in five years? Oh, there's a lot. I think specs, like, you know, companies write product specs that, like, go from someone's idea that turns into a product spec, and then that goes into design, that goes into engineering. This process is, like, dead. Like, why would you write specs if you can go from your idea to an exact prototype? So this whole middle assembly line of creating product is a dead process.

**Harry Stebbings** [19:59]:

Do you lose the design process? Does Figma become obsolete in the world of rep, let's say?

**Cem Kansu** [20:05]:

This is a really good question. I don't think it becomes obsolete, but the way we use these tools, and maybe it'll be Figma, maybe it'll be a different tool, is going to change significantly. I think we'll go from basically text of an idea to a prototype of an idea rather than everything in between. Maybe Figma is the tool that we used to do that. Maybe it's something else. But then now there's, like, three, four steps to get there, or actually even more. Like, there's basically text to design and then design to code is kind of how things go. Actually, there's text to spec, spec to design, design to code. This is gonna all merge, is my guess. So this process is is changing already. Does this favor you as a big company? Because traditionally I like that you call us a big company.

**Harry Stebbings** [20:46]:

Well, you're a public company worth like $20,000,000,000. Know what I mean? And Cem, I think it's just like We don't

**Cem Kansu** [20:52]:

see ourselves as such, as you can see, but yeah. Sure.

**Harry Stebbings** [20:56]:

But like, it's $20,000,000,000.20000000000 dollars. Like, well done. And so my my point being there, like, traditionally, that would have been a bad disadvantage that you have. That process, respectfully, is cumbersome. And my team's seed don't need to do that.

**Cem Kansu** [21:11]:

Mhmm.

**Harry Stebbings** [21:11]:

Now you don't need to do that.

**Cem Kansu** [21:13]:

It's to our advantage. I mean, we can cut process. We're not doing process because we love process. I think if we can speed up all of our product teams, to us, that's an advantage. Like, we would love to disrupt our own process as soon as we can so that, like, every individual can ship product faster.

**Harry Stebbings** [21:28]:

Will you have more or less engineers in five years' time? Oh, this is a great question. Honestly, I don't

**Cem Kansu** [21:33]:

know.

**Harry Stebbings** [21:34]:

But that's a good answer.

**Cem Kansu** [21:36]:

Give me a yes or a no. Yeah, I'll give you my take. We will have more engineers. However, the definition of an engineer might look different in five years. Meaning, this hybrid, I'm gonna call it the prototype role, where it's like a hybrid of engineer, designer, PM will be way more common. And I think that will be a lot of the headcount growth that organizations experience rather than, I think right now when you say engineers, software engineering is like a very defined discipline. I think we'll end up with these undefined builders that can use AI tools really well. And I think companies will employ a large number of those.

**Harry Stebbings** [22:13]:

Duolingo, when we look at your Cursor or your Windsurfs or any of these tools that our engineering teams are using today, do they make 10x engineers 100x? Or do they make average engineers 10x?

**Cem Kansu** [22:23]:

I think my read right now is more of the latter. And also the magnitude of multiplication is way lower than 10x ing things. I think it's maybe in the realm of like 10 to 20% more productivity of an average engineering capacity. That's not that exciting. No. I don't so my read is, like, it's it's not that exciting. And I think there's a lot of hype generated around, like, you know, we were now replacing like, I I think some companies are putting out statements, okay, like, we're not gonna hire engineers anymore. Like, I've seen really hard

**Harry Stebbings** [22:53]:

HubSpot said the other day that they now produce too much code to ship. That sounds

**Cem Kansu** [23:00]:

insane.

**Harry Stebbings** [23:00]:

I hadn't heard that. Yeah. That was amazing. And then Salesforce, 20% of their code is AI generated. Microsoft is 30%.

**Cem Kansu** [23:08]:

I think these stats are very much of the autocomplete variety. Like, what Cursor is a lot of times good at is you start writing code and then you can autocomplete with AI. That's great. That's awesome. I mean, it's not a 10x unlock if you can do that. Like, meaning, if you're writing an article, if you can complete 20% of it with auto complete, it's the same it's like, yeah, it's good, but it's not a massive life changing accomplishment. What is a massive life changing accomplishment in that product paradigm then or productivity? The two things that really become 10x unlocks is when we figure out text to excellent looking, pixel perfect design with just prompting. I think that is not there, but I think it will get there. So, for example, we think that Bruno S. Americano is kind of the experience we should build, so we build a character like Bruno that you can talk to, and he's American, he teaches you American English. Let's say that's what we want to build. Without touching a single piece of code, can we create the best animated Bruno character that talks to you in American English? If we can do that, that is a massive unlock. And I think this is not super far. So that's like text to design. And then design to code, we have the design that we can build a prototype of that thing that is good enough to go into Duolingo and go out to real users, that is a 10x unlock.

**Harry Stebbings** [24:28]:

I feel this is going well. The joys of all things like end of day is you're just like, ah, fuck it. Why not? Why not? It's not 10AM. I can't I can't be asked to be formal. When you have that unlock, the amazing thing is it means that bluntly, everyone can create product to a new site.

**Cem Kansu** [24:49]:

I think we're about to enter a wild time with this. Yes.

**Harry Stebbings** [24:53]:

Does that excite you?

**Cem Kansu** [24:55]:

It I'm so excited.

**Harry Stebbings** [24:57]:

Or worry you? Because it now means that you have infinite supply of product. Yes. And that means that discovery becomes even more challenging.

**Cem Kansu** [25:05]:

Yes. On a personal level, I'm super excited. And, you know, for the discipline of product management, I'm also super excited. Maybe you've you've had multiple folks on this show, and I've I've heard this come up actually. It's like PMs do have impostor syndrome because when they're interacting with engineers, they're not technical enough. When they're working with designers, they're not creative enough. These tools level up everybody to a much They're maybe not at the same level, but to a similar leveling field where we're all operating with, like, strong tools, so it helps PMs be a lot more productive and independent for themselves. So I think you uplevel product managers, which is good, like, from my function, I love that, but you uplevel everybody to be way more productive. This unlocks much easier product building, and I think this is very similar to just like a 10x productivity gain where the best ideas will eventually win. So creating 10x more prototypes doesn't I mean, it creates maybe chaos at first, but then the users eventually get the best of 100 ideas rather than best of five ideas we can ship today. So I think just like creation goes up, and when creation goes up, the best things that go out to users are just going be better.

**Harry Stebbings** [26:13]:

Agree completely. When you look at the tools again, who wins, Cursor or Windsurf?

**Cem Kansu** [26:17]:

I'm gonna say Cursor. Why? Because internally at Duolingo, I'm seeing way more usage and way more creation come out of Cursor. That said, maybe we haven't experimented enough with Windsurf. Can I ask you, everyone says to me, I

**Harry Stebbings** [26:29]:

look at these tools and you know that neither of them have lock in or defensibility? And as we discovered from my apt due diligence on some companies we mentioned before, I naturally go, sounds very interesting. Tell me more. But I don't actually understand why. Why did they not have defensibility, and is that wrong?

**Cem Kansu** [26:46]:

I feel like these statements come out of, is there a data lock in? So like if you're feeding it more data, is the tool getting better? Or is there like a social lock in where, I don't know, you're collaborating with folks and switching is painful? For example, a lot of enterprise tools have heavy data lock in. Like, we use Jira for, for example, task management in the company. It's very hard to switch out of because all of our history has been now fed in there, and switching out means we kind of have to leave that data and go elsewhere. We use Jira for A la poo

**Harry Stebbings** [27:17]:

bell. I'm gonna introduce you to Carrie at Linear after this. Okay. That sounds good. Yeah. And he is amazing. And this is called Venture Value Add. Here all week.

**Unknown** [27:33]:

Honestly, they're brilliant. They're much better than Jira.

**Cem Kansu** [27:35]:

I mean, I've seen their designs. It looks awesome. I'll say the hardest thing is to switch out of the thing we have. So even if we love the product, switching out is gonna be a big debate because there's a huge cost. And I think people bring this up because you can switch from Cursor to Windsurf, you don't lose anything. If the Windsurf is better tomorrow morning, well, we'll start using that. Maybe these will change, meaning the more code you build with one will just make it so much better. For example, now ChatGPT has memory. I think that is an amazing lock in for ChatGPT because now since it remembers our chats, it can now give me smarter answers. That is a big lock in for me to not use a different product.

**Harry Stebbings** [28:13]:

I do have to ask, because you mentioned at the beginning, like, being central to the monetization. And I spoke to Severin, and he said that bluntly, it was your fault that they waited so long. Jokes aside, why did it take so long, you think? And how do you reflect on that decision?

**Cem Kansu** [28:32]:

The hindsight bias is easy to apply to our past and say, oh, we waited too long. But I think at the time, a few things were, I would say, maybe headwinds to starting to monetize. One of them is Duolingo obviously has a very strong mission today, and that certainly was the case then. When it first came out, Luis you can find Luis's talk. This was a lot easier to find on YouTube, and I think it was on our corporate pages. I think we removed it now because we moved away from that. But the monetization plan for Duolingo was going to be everything was going to be free. So as users went through, did language exercises, they would do translation, and then we would crowdsource those translations as a service to anyone who wanted to buy it. So if CNN wanted to translate an article from English to Spanish, they would feed into Duolingo system. That article would be broken up into chunks, fed into users and exercise, and as crowds translated that as an exercise, it would populate back a translation. That was the original plan. And as Luis announced Duolingo, he also said, we will never have ads, never in app purchases, and never subscriptions. And finally, that's the those are the exact three things I did as part of my job at Duolingo. And your lesson from that is never say never. That's lesson one, that things change and your view on these change. But I think, two, the reason there was a lot of resistance in the company for monetization is we launched with the view that if you did these things, you would screw up your amazing, great consumer product. I think what, honestly, my job eventually turned into is obviously building these things, but also changing the organization's mind that there is a way to do these without going against our mission of providing access to free education and doing them well. The big unlock there was, at the time, and honestly, mostly today, all education apps would lock all of their content behind the paywall because you know, that's so straightforward. Like, why did you come to this app to learn? Well, what are we gonna charge for? Let's charge for the learning content. We did the exact opposite. We're like, learning content is completely free, but we are gonna monetize the bells and whistles. We're gonna do freemium, so most of the stuff will be free, but the premium features could be part of the subscription. Even our investors at the time was like, oh, that's not a good idea. I've never seen that work.

**Harry Stebbings** [30:42]:

Well, if I was one of your investors, I'd say, hey. Growth is everything. Engagement is everything. Retention is the fucking holy grail. Don't stick in ads when we don't have a massive company yet.

**Cem Kansu** [30:52]:

I think that was also the common wisdom at the time. There was the Facebook model. Just grow. It doesn't matter. Ignore monetization. And honestly, it was good advice, but I think what we ignored is you can still have high retention, you can keep growing your product, but also monetize at the same time. I think it felt like it would kill things, and it didn't. It was fine. Like, I mean, putting ads on Duolingo, if you do it tastefully, doesn't kill your retention. So you can still have very high retention and monetize the bells and whistles, but have high growth.

**Harry Stebbings** [31:22]:

Can you just take me to the moment where you were like, okay, we're gonna move from a zero revenue company to trying to make money? Yeah. Was there a meeting? Can you take me to that moment?

**Cem Kansu** [31:31]:

It probably wasn't one meeting, but there was a few key moments where I remember being like, how how am I gonna figure this out? One of them was sitting with Luis in, honestly, probably one of my first weeks, and I'm a very early product manager trying to figure stuff out. And I'm like, Alright, Luis, here's our ads roadmap. I calculated how much money this is gonna make. Here are the experiments I'm gonna run. And he's like, Alright, what is it gonna look like? And I show him the prototype, or we look at it, he's been looking at it, he's like, these ads look like shit. Shit is his first response. He's like, I'm not sure if we should do it. And I'm like, okay. Where did we I like, I hear you. I understand that this doesn't look like the pristine every other page of the app. It's different. It's another advertiser's property, but we can improve the look. We can change the design. Anyway, it was an iterative process from there to be like he's like, let's test. Let's see what happens. And the other worry was like, oh, this is gonna kill retention. We tested. It's fine. It's, like, neutral on retention on the first experiment. And we're like, okay. Our users are not running away. We are improving how the ads look and when they surfaced. And then I think over time, it started snowballing to be like, okay. We can do this. We can do ads. Once we did ads, the second phase was, well, no ads was like a request people started sending us being like, can I pay you guys to remove ads? We're like, well, let's try that. That's a good start.

**Harry Stebbings** [32:47]:

Before we move then into the paid then, do you accept that you cannot do ads without hindering user experience a little bit? There's certainly a cost. I had it the other day, and you have the timer in the top left. Yeah. And I'm like, listen, it's annoying. I stay, of course, and I love Duolingo just the same.

**Cem Kansu** [33:05]:

I can't say I love ads as a user either. So I think you sacrifice because it is not a property you control. Meaning, if I bring an advertiser, they have to show their asset and that asset comes with a five second timer on, I don't know, a video for selling you shorts, whatever it might be. It's a cost, I think, that you take. That said, there are many different ways to do ads. So I think treating ads as this black box carrier of like it's only bad and it can only hurt is a bit too simplistic. What's the lessons on what it takes then to do ads well if it's too simplistic to just black box? I think ad load is the key thing there. We stick to one you can you will only see one ad at the end of a lesson, and that's not mid lesson, that's not start of the lesson. You finish your lesson, and that's you take a breather. That's the only time we show you an ad is what we aligned on, versus you can put a banner ad on top of you know, there's many tasteless ways to do it. Which you

**Harry Stebbings** [33:58]:

optimize for cash, though, because you can charge more for in

**Cem Kansu** [34:02]:

play versus post, obviously. Correct, but I think that is short term thinking at the end, because you maximize a lot of maybe revenue now because you're putting a lot of ad clicks across the app, but the moment you ruin your free user experience, you kill your growth engine.

**Harry Stebbings** [34:16]:

Can I ask you, do you get paid on views or on clicks?

**Cem Kansu** [34:21]:

We get paid by impressions, so views, I guess. So the rate that these ad networks pay you depend on how clickable these impressions are. It kind of both touch eventually the ad revenue you make if you're running programmatic ads on your products.

**Harry Stebbings** [34:36]:

Okay. And so I then say, You know what? Fuck this. I want no ads. And you're like, What happens then?

**Cem Kansu** [34:44]:

A lot of users obviously were like, I love Duolingo, but I'll pay you guys whatever. Can I remove the ads? That no surprise because that's a common model. That was the idea for our subscription. So it's like we had ads. We had a no ad subscription. That was the start. And what we saw also at the time was there was a lot of user love for Duolingo. It sounds like you you also love Duolingo. And we saw users just buying the subscription because a lot of our users was I was using Duolingo for five years. I just wanted to contribute something back, so I bought the subscription even though I didn't know what was really in it. No.

**Harry Stebbings** [35:18]:

I'm not that Christian.

**Cem Kansu** [35:19]:

Definitely not. The $20,000,000,000 market

**Harry Stebbings** [35:22]:

cap made me feel.

**Unknown** [35:24]:

I was like, no.

**Cem Kansu** [35:26]:

But yeah. So but there basically, the no ads was a starting point, it and was feature light, honestly, the time. The subscription was kind of a nice to have more than anything else.

**Harry Stebbings** [35:35]:

What was if you were advising me as a founder on doing that, like, premium subscription, what would you say, having been through that?

**Cem Kansu** [35:43]:

It really depends on what business you're building. But if you are building a consumer app business, I would very, very highly advise you to build a freemium model, meaning you should have a very great free user experience without asking anybody to pay. The reason for that is if you have a premium product, let's say it's all paid, you can only use the app if you subscribe or it's very limited, what you have to build is also a paid marketing engine. So you have to spend a lot of performance marketing to acquire subscribers, generate revenue from them, and spend it back on performance marketing. That's how paid products kinda have to grow. Free products, if you build an amazing free product, it takes way longer. But what you end up building is an organic growth flywheel, Meaning, you tell your friends about Duolingo. I don't know if you do. I tell my friends about Duolingo. Let's go off of me. They can go experience the full thing for free. Whether they pay or not, doesn't matter. If they like it, they will also tell their friends. It just expands the entire pie of people who can use your product without having to pay. And we have a lot of users that use the app for free for five years. Eventually, one January, they're like, oh, I'm gonna go ham on my French this year. I'll subscribe. It doesn't matter. But they use Duolingo. So to build basically a large user base, and if you wanna build a large business, it's way more defensible to build a great free user experience and do it freemium rather than full premium. The other problem with premium is competition. So, look, the reason Duolingo has grown quite a bit in the past ten years or even gained initial success was because it had an amazing free user experience. If you created the exact equivalent of Duolingo today, what you would probably do is, okay, like, you wanna make some money, apps have tried this actually, you kinda copy Duolingo exactly as is, you make it fully paid. You're like, I'm gonna go fully paid and spend on paid performance marketing. Well, what happens is the users who you can acquire will land on your app and say, okay, you asked me to pay $80, $100, whatever it is, or $10, or I can just go to Duolingo and do it for free. So your paid costs go really high, so it makes it really hard to compete with a free, amazing product. So this is a very roundabout way to say, if you're building a consumer app business, I think you wanna make sure you have an amazing, free growth flywheel, makes you defensible, and it kind of spreads your product without you having to spend marketing money.

**Harry Stebbings** [38:01]:

How do you think about the right time to hit them with a paywall? You wanna build the trust, the love, the loyalty without being like, okay, level one paywall.

**Cem Kansu** [38:09]:

Exactly. So I think it's twofold. One of them is honestly, like, hitting with a paywall is kind of the wrong approach in the sense that I think you want people to discover the paywall by experiencing features. That is the best way to honestly, like, convince someone to buy something. Meaning, let's say I would like to convince our users to buy Duolingo Max and its key feature is video call. The best way to do it is have them try video call, actually. So, a lot of our ways to do that is maybe show them a video of how video call works and then say, if you're interested, here is Duolingo Max. Or, have them go into the experience, have a call themselves, and then at the end say, this is part of Duolingo Max. That is kind of the best way when it's contextual and users discover that there is something to be paid to experience. I think just having people showing kind of what I would call generic ads works way less better. And it looks spammy in the app. So we try to do everything contextual. For example, if we show you an ad, we will say right after, there's an ad free version if you'd like to upgrade. And that is in a way, like, you've discovered an ad. Okay. Maybe you didn't love it. Okay. But you can also find an ad free upgrade.

**Harry Stebbings** [39:20]:

I just wasn't expecting Pornhub Premium to be an advertiser Tell that. Honestly, for a second, was like, oh, shit. Did did did that happen? The ad network surprised us.

**Unknown** [39:31]:

It was the first useful ad I've had on Duolingo. It's the only ad you click turns out. I'm like, sign up. Sign up. Sign up. The joys of doing the show is also, like, you hear the show and you're like, god. It's so serious. And then behind the scenes, you're like, what a ride. Behind the scenes, born up preview.

**Harry Stebbings** [39:52]:

Yeah. It's much more fun that way. Can I ask you, when you look at the different paywall pages that you've experimented with, are there lessons on how to bucket different offerings? Oh, yeah. What what are those lessons?

**Cem Kansu** [40:05]:

There's a lot. The biggest one I will say is what you say in your copy in strings makes a massive difference. And, obviously, you don't wanna kind of sell something that you don't have, But how you talk about, entering that experience makes a massive difference. I'll give you an example. We have tested many ways to on the button that is that is going to start your subscription. We tested probably endless number of things. Turns out, if you say, try for zero dot zero zero, which is not a normal human, like, I wouldn't talk to you. I'm like, would you like to try something for zero dot zero zero pounds? Like, I wouldn't say that. But in the string, if you write multiple zeros and you hint at the fact that there's a free trial with multiple zeros, conversion goes up a significant amount. Over

**Harry Stebbings** [40:49]:

start your free trial?

**Cem Kansu** [40:50]:

Over anything else. Start my free week. Start my free trial. Start your free trial or come up with any other combination. We've tested probably many different things. Turns out if you put multiple zeros, I think it, like, lights something up in the brain that you're like, oh, yeah. I get it. It's very free. It's because it's like zero dot zero zero. So copy matters. That's one. Because there's with subscription, there's natural aversion to entering a thing that might be hard to get out. We also have seen talking about how easy it is to cancel is like an unlock in in conversion. Do you guys send reminders? We also send reminders. Because I signed up to Max,

**Harry Stebbings** [41:25]:

and I signed up because of the reminder.

**Cem Kansu** [41:27]:

There you go. So that that actually that actually makes a big difference. And, this is one of those changes I love because it's user friendly. If you don't like the experience, I wanna remind you to cancel. We don't want people who forget and pay because of it. But that relieves the anxiety of entering a subscription free trial.

**Harry Stebbings** [41:44]:

And the thing about the different packages, like give people just one option, paid versus not paid, max versus premium versus free, anything on like choice. I'm always fascinated by the paradox of choice. I

**Cem Kansu** [41:56]:

think comparison is super important. What we've also seen is if you give one, for example, we have multiple packages now. There's super, there's family plan, there's max. We think you are a great fit for super individual. If we only show you one package and you have no anchor for different packages, it converts way less because it feels like if that's not for you, you're just gonna decline. If we give you more than three packages on one page, it's too much. Like, too much choice, you also convert way less. I think two to three is we've seen as a sweet spot. So you want some anchors to compare to, and most of the time, if you give a choice of three, you want the middle option to be the option that you really recommend because that makes choice very easy. Humans love, I mean, on a menu, too, if you think about it, if you have the lowest price, middle price, and highest price, middle price feels like the safe choice if you don't know what to choose. Very similar to if you're kind of offering someone a subscription choice, I think you want to make the choice easy by recommending one, and recommended one being the middle one feels safe to choose. What percentage of people take the recommended? Very high. I don't remember the exact percentage, but certainly more than it's the majority, more than 50%.

**Harry Stebbings** [43:06]:

Wow.

**Cem Kansu** [43:07]:

Yeah.

**Harry Stebbings** [43:07]:

Any other big lessons on copy matters? I'm sure you've been through many different iterations on copy.

**Cem Kansu** [43:12]:

Oh my god. We've tested so many things. Yeah. Copy matters. You know how every company eventually puts, like, a checklist in front of you where, like, oh, if you do this package, you get, like, two check marks, but if you do this other package, you get, like, 10 check marks. This has become, like, a very common way of subscription purchase flows. Weirdly, that also works because it's the psychological difference you can see visually. I'm like, oh, what I have now has one check mark. What I don't have has, like, 10 check marks. Well, 10 check It marks feels

**Harry Stebbings** [43:40]:

better.

**Cem Kansu** [43:42]:

Feels like you're getting a lot, and I think that that feeling of getting a lot feels better. I think the other thing that we've learned is there's a lot of emotion on purchasing anything, certainly something that is an upgrade to a consumer experience, like Duolingo going from Free Duolingo to Super Duolingo or Duolingo Max. And I think you want to convey that emotion in your design, Meaning Duolingo takes pride in, like, a fun being fun and animated. While our flow is being animated on Duo doing funny stuff through the flow, it works well because it, like, conveys that this is not just a boring subscription as a service purchase, it's like part of your experience and it's an upgrade. So anyway, if you use illustrations, animation, and things like that that make it a fun flow, that moves metrics at the end and increases conversion. What percent of revenue is subscription today?

**Harry Stebbings** [44:31]:

Today, it's more than 80%. What revenue line does Duolingo not have today that will be meaningful in five years?

**Cem Kansu** [44:38]:

We do have this revenue line, but it's incredibly small today, but I think it'll be way bigger, is merchandising. We have a store. You can buy Duolingo merch. Our most selling item is the Duo plushie. It's like a very funny plushie of Duo. I think we are one of those brands where our characters are quite Duo is quite recognizable at this point, and we're trying to build more IP around our characters. There's a lot we can do on creating better merch and expanding our merch offering. Surely content is it. You want to become the Disney of the next generation. Why not? Exactly. We would love to become the Disney of the next generation. I think that is hard for sure, but that would be awesome.

**Harry Stebbings** [45:18]:

But I'm actually the founder of this idea and now company, because that was my idea. So thank you, Luis. I will take over. Would you say that you're an education company, an entertainment company, a content company?

**Cem Kansu** [45:31]:

I know it's all semantics. Honestly, neither is the reality. We have maybe somewhat of an identity crisis, is is also the reality. We don't identify as an education company, but we also don't identify as a game entertainment. So we're kind of in our weird own league and own market that we've created. The reality is what we compete for, if you're like, okay, but what are you? Who do you compete with? We eventually compete for screen time. We would like to make your screen time very useful for you to learn something and get better. The reason we don't fit into education very much is if you look at the education space, honestly this is a bit of a hot take, but it's all boring ass products. So I think we would like to not be a boring ass product. As you can see, we try to do everything super fun. And I think if you say, well, you do guys do a lot of fun, then are you a game? Well, we're not a game either. We also want to teach you something. So we're kind of in this zone league of our own is how we see it. But what we eventually compete for is your screen time. So everything that takes up your screen time, that might be social media, that might be gaming, that might be streaming, whatever it is, at the end is what Duolingo competes for. It's also a hard one on multiples. Gaming multiples, are shocking. Education, tough. And also, that's not the product we're building. We're not building just for entertainment, but we're also not building just to educate and ignoring the fun element of it. So that's why it's like right in the middle.

**Harry Stebbings** [46:54]:

Okay. The final element of this tool, we've got the ads, we've got the premium, and then we have in app purchases. How do we think about what worked there and the biggest lessons to doing that so successfully?

**Cem Kansu** [47:06]:

So in app purchases for us was very similar to how games have designed in app purchase mechanics. So we have a virtual currency, it's called GEMS. It's not named after me, it's the actual Diamond GEMS. Sure, sure, yeah. But you were the one who came up with monetization. So, yes, GEMS was one of the first features I built, actually. So there's some tie in there.

**Harry Stebbings** [47:27]:

In '20 VC, we hand out Harry Stars, actually. It's like I said, but I didn't come up with it, actually. No connection whatsoever. No connection at all, but it just happens to be that.

**Cem Kansu** [47:38]:

And

**Harry Stebbings** [47:38]:

you can You're gonna listen back to this and go, this guy was a fucking joker.

**Cem Kansu** [47:43]:

Yeah. I am curious what you what you cut and what you keep, so I was gonna end up really being interesting. Oh, it's gonna be fun. Surprise for you and me. Scramble. But so you give gems? We give gems, and you can either buy gems or earn gems, and you can spend those gems in virtual features. So you can for example, one of the most popular features is a streak freeze. So streak, as you might know, is a very popular feature. People take pride in their streak. They don't wanna lose it, but if for one reason or another you can't do Duolingo that day, you can use a streak freeze. This is a very valuable item. People love it. And if you earn GEMS, you can spend your earned GEMS on it. If you're out of GEMS, you can buy GEMS to purchase a streak freeze. Items like that exist, and that's part of our gem economy. Have you

**Harry Stebbings** [48:25]:

ever had someone sorry to interrupt you. Have you ever had someone lose a streak and then contact you and be like, listen, I was tracking up the, like, Everest, and, you know, I've dropped my phone and and so I couldn't do my streak. All the time. What's your High favorite

**Cem Kansu** [48:39]:

volume request. I think that most most of the stories are time zone switches, and honestly, this has happened to me, so I actually have full empathy for this, where I actually did my lesson, but the reason is I flew to Australia, I was on the plane for fourteen hours, and when I landed, the day had already moved forward, so I lost my streak. So you had to cool up the head of product. Completely get it. And I we fix those. So, like, if you actually reach out to customer support and they were like, I have this legitimate reason, we're, like, lenient. We're like, of course. Like, we'll fix your streak. Don't laugh. Do you have to prove it? Honestly, we really don't ask that many questions, but and and we're lenient. But, yes, there's there's there's, like, time switches, and there's, like, I was on a mountain with no connection. What was I supposed to do? And, of course, like, we understand. Off the fucking mountain. We understand. But, also, yeah, like, use a streak freeze maybe and then get get off Buy the the freeze. That's why they're

**Harry Stebbings** [49:31]:

there. But okay. So lessons on how to do that effectively. Is it just create addictive apps that people don't want?

**Cem Kansu** [49:39]:

Well, like, I think the core experience, of course, matters a lot. Addictive sounds like the bad version of this, but habit building, like you want to build good habits. And the reason we use that term is that if you want to learn a language, you have to practice, if you want to be fluent, thousands of hours. And to get to thousands of hours, you need a daily habit. You can't binge your way to fluency, so that's why everything at Duolingo is built around the daily habit. Once you've done an experience that people want to use daily, you have good daily retention, then you kind of layer on these benefits that are kind of extra, like streak and a streak freeze. You don't need it, but it is really core to your daily habit, and then freezes this extra benefit that you can monetize. That's how we've thought about it.

**Harry Stebbings** [50:21]:

Totally. Yeah. That that's fascinating. What do you think people do wrong with in app purchases? You see so many. You're a great product mind. What do people do wrong?

**Cem Kansu** [50:29]:

I think people do really aggressive things, and that ends up hurting the experience. So for example, streak freeze is so extra. Right? We're talking about it because it's like if you don't buy it, your Duolingo experience is gonna be exactly the same, but it's this value metric that you can choose to freeze. And if you don't, nothing really changes. It's all up to you. Like, your French is gonna be the same either way. I think what sometimes apps do is they make really core functionality an in app purchase, and I think then you start killing the fun, the retention producing elements of your app. So imagine if Duolingo was every new unit of a lesson was in that person, it was like a dollar per unit. You would get pretty sick of it pretty quickly. It's like buying a chapter of a book for a dollar every time, you could do it, that's a monetization model, but it's a pretty stupid and bad one. So I think you want to find things that people find joy in buying that doesn't hurt the free user experience.

**Harry Stebbings** [51:22]:

I think it actually goes to trust, which is when you do that, I actually lose trust in you, and it makes it a transactional relationship. Because I don't even feel those I'm a premium. Like, the one hit pain of, I'm a premium, fine, you pay the toll and now you enjoy the holiday. And you feel good. And you feel it. But if it's like consistent pain inflicted, fuck that.

**Cem Kansu** [51:40]:

No, it's very annoying. Like we talked about with ads the same way, tasteful monetization really matters. It builds user trust, and I think user trust goes a long way for the business because you tell your friends, they download the app. What are the metrics that matter most to you? There's two metrics that drive the business. So if you really squint at a business like Duolingo, you have to look at two metrics. One of them is daily active users, and the second one is revenue. But if you just wanna spreadsheet a business, you have to look at DAUs and revenue. But I think for me what matters is retention. Because that's like company metrics. Correct. So if you think about product metrics, the inputs that drive the outputs. Retention. Retention is everything when you think about building product because it is the best metric to distill down if people are finding value or liking a feature or a new app, whatever it is. What is your daily retention? I mean, if it's a daily app, daily retention. If it's a weekly app, you look at weekly retention. For us, everything we build around daily usage. So it is daily retention. So if we say, for example, Chess, where we just launched Chess, the thing we're looking at day in, day out is what is the daily retention. Because if it's really high, well, it's gonna compound to high DAUs. We know it's going to. For example, you could look at DAUs, do you have many inputs that are muddying the waters of is this a sticky habit building product? So daily retention is the king of all metrics for consumer products.

**Harry Stebbings** [53:07]:

And just so I understand, daily retention is do I come back tomorrow?

**Cem Kansu** [53:10]:

Correct. It's high bar. High bar. Exactly. With mobile products, that's why it's so important, because you have many different things that can steal your time on your you don't, maybe. I didn't have anything. You're You're like, I don't know what apps are. Don't know what apps are.

**Harry Stebbings** [53:28]:

Okay. So we have daily active users as well, like, Yeah. Core What is good?

**Cem Kansu** [53:34]:

I think it's gonna really depend on your product. There's no, like, easy benchmark you can put. And also, it really depends on what your user base looks like. If it's all new users, you have a different metric than if you have a bunch of users that are already settled users, I'm gonna call them, like, Duolingo, where they've already been using for, like, three years or above. When

**Harry Stebbings** [53:51]:

they're settled users, do those cohorts in terms of usage look quite like Candy Crush users? Like, the very, very settled gaming users?

**Cem Kansu** [54:00]:

I think so. I'm not knowing Candy Crush's metrics myself at all, but what one tend to see is if you've reached a thirty day streak, so you've built up a thirty day streak, your retention is very close to 100%. It's not 100%, but it's very high because you've built a habit. Like, it's not crazy because you've done it for thirty days without breaking your streak. You have built a habit, so you're very likely to continue with the habit. But if you look at retention of users who are streak less than seven days, it's way lower because they actually haven't built a habit. Or maybe they're unsure of Duolingo, they're just playing with it. One way or another, they haven't built a habit. Do you see wildly different retention numbers in paid versus free? We do, however, the difference is a little hard to parse on It's causation versus correlation. Do they retain better because they are a subscriber, or they were likely to retain because they were more engaged and that's why they decided to subscribe? It's hard to parse in the data, but if you look at just pure users who choose to subscribe, their retention, and users who choose not to, subscriber retention is much higher.

**Harry Stebbings** [55:03]:

What caused the biggest change in retention from a product perspective? Three things. Wow, that was very Yeah, this is something I've thought about a lot. You had that answer ready to go, didn't you? No, no. I was like, that's a good question. You're like, I got that one.

**Cem Kansu** [55:18]:

So one of them is the streak, but no surprise. Having the streak and making the streak exciting so when you finish your first lesson, there's this flame animation that lights up, there's a haptic, there's a sound. We make the streak like one of the most important things on Duolingo. The details being the details, it's exceptional, really. The details matter a lot. So every tiny thing that might not even make sense to talk out loud, but when you see it, that made the streak more exciting has helped improve retention, actually. So the streak is this incredibly important mechanic that instills daily habit, helps you learn a language or learn a subject. That is one of the biggest things for retention. I would say second, as a feature, was leaderboards. Turns out, no surprise in hindsight, but if you see your name ranked on a leaderboard, there's this human tendency to be not be at the bottom of that leaderboard, but instead move towards the top. So a lot of our learners, even though when they came to Duolingo, they didn't come to compete in French, we have a leaderboard feature where you see your name among a cohort of similar learners, you want to rise to the top, a lot of users want to at least, so that drives a lot of usage, lot of retention. Who's like, no, I'm happy languishing at the bottom? I think some people are like, I don't wanna compete. But a lot of users, when you see your name, you're like, I would rather move to the top, not to the bottom. I'm one of them, for sure.

**Harry Stebbings** [56:40]:

For sure. Do not understand that people aren't happy just being like, well, you're not getting hired, are you? It's like the worst hiring trait ever. Thoroughly uncompetitive, happy losing. Exactly, exactly. Okay. That's we love that.

**Cem Kansu** [56:54]:

So that's the second one. Third is push notifications, actually. For every app, it is an incredible mechanic that you have to figure out how to make amazing. Still. Still. Maybe ten years ago, was email was a heavy channel. Email is a dead channel for most products at this point. Who do do it? Dead is maybe an overstatement. Like, we obviously invest a good amount of time in making our emails

**Harry Stebbings** [57:13]:

team, we still love you, don't we? Yes. We do love you. We hope you enjoy the dark basement with the leak in the corner. It's fine. Luis will let you out next year.

**Cem Kansu** [57:23]:

Yeah. Where the push notification team is like popping champagne. Yeah. They're the Belmont in Portofino

**Harry Stebbings** [57:30]:

celebrating twenty b. So, yes, I

**Cem Kansu** [57:34]:

what does it take to do push well? You can get incredibly complicated, and this was kind of news to me over my career, incredibly complicated with limited real estate and and copy you write in a tiny push notification. For example, timing is a key factor. Content is a key factor. Personalization of that content is a key factor. Emoji usage is a key factor that moves metrics. So if you get into the full details of really optimizing push notifications, which we have, you can do a lot with it. And every year, we're like, okay, we invested a lot. Are you gonna keep improving? There's a lot to do on improving push notifications. I'll give you one insight that was kind of like tested in and out, then this turns out to be the winner. We schedule so let's say you did your Duolingo lesson today at one p. M. Schedule a notification almost exactly at twenty three point five hours later from that time. And we've tested many different timing And the notification will remind you to practice your Spanish because we you just did a Spanish lesson. We've tested many different times, but the twenty three and a half hour works way better for retention because turns out humans do lessons. If you're a morning person, you do in the morning. And if we schedule it almost thirty minutes before ready to do your morning lesson, you're very likely to do a Duolingo lesson at the same time. So based on your behavior, we time the next notification. That has done really well for retention. So when you get into, like, this type of experimentation of, like, how do we actually figure out how to make it better for each of our users, there's a lot you can do.

**Harry Stebbings** [59:02]:

Wow. That's fascinating. Okay. So that's like timing. Emoji usage, you said?

**Cem Kansu** [59:06]:

Yeah. Well, emoji breaks up pattern is why I think it works. So a lot of notifications suffer from they fall into a pattern and now a sea of everyone sends so many notifications that it blends in. So you wanna do something that is more interesting than the 85 other apps sending you notifications a day. Emoji is one way to break pattern. The other thing we do is kind of go full unhinged on notifications. So we write them from Duolingo's voice, and we make it sound like Duo is basically screaming at you, or to do your not always screaming, sometimes sending you a subtle message, but it's like a personalized message rather than Duolingo saying, Do your Spanish lesson or Practice now. That kind of doesn't go very far. But if Duo says, Yo, what's up? Are you doing Spanish today? And it looks like almost a message from Duo, that triggers you to be like, Okay, yeah, I'll respond to Duo rather than this app that is sending me sanitized notifications.

**Harry Stebbings** [59:57]:

It's like a replacement girlfriend, isn't it? You don't need someone to shout at you. You've just got Duolingo being like, yo, bitch, get in I mean,

**Cem Kansu** [60:04]:

you would respond to that more, so, you know. You

**Harry Stebbings** [60:07]:

respond

**Cem Kansu** [60:07]:

because

**Harry Stebbings** [60:07]:

it's different.

**Cem Kansu** [60:08]:

It's different. Yeah. Exactly. It's

**Harry Stebbings** [60:09]:

a purple cow.

**Cem Kansu** [60:10]:

Yeah? Yeah. It's different, and it doesn't feel like a company is sending you a thing. It's a human would send you a thing that way.

**Harry Stebbings** [60:19]:

We we can do a growth session, me and you. Okay. So we've got a couple of fun programs we can do. We can do a Duolingo School in New York. Physical school, great for community, great for vibe and

**Cem Kansu** [60:30]:

Did this come from Severin? This is his passion project of like, yeah, Duolingo should have a physical school. I agree with him. That show Don't.

**Harry Stebbings** [60:38]:

I'm I'm one of the founders, so you you wouldn't know. But we were talking about this about Disney. This founder talk. Okay. Got it. Yeah. You can do a duo movie. Go in the Disney direction.

**Cem Kansu** [60:48]:

Okay.

**Harry Stebbings** [60:49]:

Really? I

**Cem Kansu** [60:49]:

do like that idea for the record.

**Harry Stebbings** [60:51]:

Really, really cool. I

**Cem Kansu** [60:52]:

think if we wanna build characters, we have to expand more beyond the app.

**Harry Stebbings** [60:56]:

You can do DuoDate. DuoDate? Yeah. Which is like dating, connecting people, learning the same language. You've mentioned before people getting married, before who maybe met on leaderboards. Which one do you love? Which one do you hate?

**Cem Kansu** [61:09]:

I hate the school one, if I'm being honest, because building a physical business seems like that's not our strength. I like the movie idea. I love the date idea. That's how I would pick. Wow. You love the date idea. I do love the date idea. I think it is the riskiest one because, you know, what could go wrong by just building a dating platform? However, there's already signs that people date on Duolingo. So the best way to build product is capture latent demand that users can't do on your product today. We have couples that got married because they met competing on a leaderboard, and then they found each other on Facebook and got married. This behavior is already happening, so it's not out of the question for us to figure out, well, how can we do matching better?

**Harry Stebbings** [61:55]:

Wow. That's pretty cool. I

**Cem Kansu** [61:56]:

mean, it is out of the question. For the record, we're not doing that, but I love the idea.

**Harry Stebbings** [62:00]:

There's so many interesting lessons to be had from my hindsight. When you look at the product decisions that you didn't do, what didn't you do that with the benefit of hindsight

**Cem Kansu** [62:10]:

you wish you'd done? We talked about the obvious one, so I'll start with that, which is we didn't monetize in 2014. We started monetizing kind of two years later than we could. So I think that's one. We talked about it, we don't need to unpack it. I also think we this is, in hindsight, easier to say we waited too long to apply the Duolingo magic to new subjects. Meaning, we first launched a subject that is not language, that was math, and then music followed fairly quickly, but math, we first started working on math three years ago. We didn't have to wait that long. Like we kind of had the Duolingo mechanics in place. We could have honestly started our journey of figuring out other subjects earlier. In hindsight, this is very easy to say. This is a long journey to figure out product market fit for new subjects. However, we have the competency, we kind of knew eventually we would do something like this, but instead we were investing in various projects that now in hindsight we shouldn't have. Like for example, we were trying to build a school's dashboard for teachers to use and that was a complete different product problem. That is a very hard problem to solve. Instead of doing that, we should have probably done math and music and chess way earlier.

**Harry Stebbings** [63:25]:

I was gonna ask what did you do that you wish you hadn't done? A

**Cem Kansu** [63:28]:

school's dashboard. Yeah, there's another one. I think at the VC level this makes sense, but when you get into the product level it doesn't, which is why is Duolingo not catering to language teaching in the classroom? Like, why does Duolingo not have a product that fits into the classroom? So we wanted to solve that, so we tried to come up with a version of our product that teachers could assign teach certain parts of a curriculum and assign homework through the Duolingo interface, and then they could track if assignments are being done, etcetera. The problem with this is every language teacher wants to teach a different way, and the Duolingo curriculum does not fit any of it, so you end up having to customise how people should go through Duolingo, and that's not the product we have. So even if teachers could assign assignments, that doesn't match the textbooks they're working on in the French classroom. Anyway, it didn't work.

**Harry Stebbings** [64:20]:

What what how long did you assign to it?

**Cem Kansu** [64:22]:

Oh, we worked on schools. By the way, this this product still exists. It's just, like, not a product mark it doesn't have product market fit. So you can still use it? You can still use it. Like, the school's dashboard exists. Like, if you'd like to try it out, it it exists. It's there. It just hasn't found the adoption we wanted to because it doesn't fit the classroom very well. But we worked on it for, I don't know, three to five years, maybe.

**Harry Stebbings** [64:45]:

Why still have it live? Why not pull it?

**Cem Kansu** [64:47]:

Because there's still usage. It's not zero usage. Some people find use out of it, but it's thinking about the market of classrooms teaching languages, it's a very sensitive no one on it, like, maintenance. Right now, there's no investment. That doesn't mean we won't invest again in the future, but today, we've chosen not to. It's a complicated problem.

**Harry Stebbings** [65:04]:

Anything else that we did that we

**Cem Kansu** [65:05]:

There's

**Harry Stebbings** [65:05]:

should

**Cem Kansu** [65:06]:

a big one now in hindsight. It feels like, oh my god, why did we do that? So right now, we're talking about how Video Call and Lily are speaking feature. There was a short lived moment where we had a high tier subscription, was basically an Uber for a Spanish tutor. So you could press a button to speak Spanish to a Spanish speaker immediately. So rather than scheduling, you could press, so we figured out how to connect you immediately with someone online, and it would be like a two, three minute condensed conversation. Honestly, at the ideal, I was like, oh, this is great. Sounds pretty great. Yeah, we solved

**Harry Stebbings** [65:38]:

it. Natural cool.

**Cem Kansu** [65:39]:

Exactly, we have solved the tutoring problem, which is the biggest problem is scheduling, and it's expensive, etcetera. We made it part of a subscription so you could do unlimited calls on demand. Very cool idea. And, also, users have been asking us to do this for years. Like, we're like, okay. Finally, we figured out how to do it. There was two massive problems in hindsight now we know. One is it is super scary to actually press a button and talk to a human immediately in the language you're not proficient in. Even though you think you want it, what we've seen when we observe users doing it, press a button and they're like, okay, I'm ready, I'm gonna do it. And then the person would dial in, start speaking Spanish, and then they'd be like, hang up. Go away. It's what would happen to most users that ever tried it. So it was a scary experience to be thrown in, and that's why AI is way better for this because you can stumble, you can talk shitty Spanish, you're not embarrassed. A lot of users found it embarrassing to stumble when there's a human looking at them on video. That was problem number one. Problem number two is you have no quality control. So you can't run an AB test and say, hey, let's do the opener in English. I mean, you do, but then you can't control if the human delivered experience is what you thought it would be. So we found it kinda impossible to operate a human based product experience. It's not a problem, it's very hard. And around that time, GPT-three came out and we're like, oh, hold on. We can solve this. So we pivoted

**Harry Stebbings** [67:09]:

away from that. What would you most like to change in the product? But for some external reason, shareholders being public, execs not wanting to do it, you can't do it.

**Cem Kansu** [67:19]:

I would like to make Duolingo way more social. And I don't think there's any blocker or anyone against it. I think we need to figure out a way forward and the best way to do that and then staff teams accordingly to solve that problem. So right now, Duolingo has social features. So I can add you as a friend. If you reach a certain streak, I can congratulate you. You can do it back. So we have this social layer, which is very congratulatory and nice. We can do a friend quest together. We can do friend streaks. So it's a lot about collaboration and congratulations to achievements. That said, I mean, I think this is kind of why I loved your dating idea. It builds on the social potential of Duolingo. Duolingo today taps into this reality of you're doing something good for yourself and there are others doing the same, but if you go through the experience, you feel like it's a single player experience, you are doing it solo. And language learning is inherently a social thing. Most people are learning language to do something social with others, maybe order their coffee in Italian when they go to Italy or date someone or whatever it might be. There's a social need. And I think social just fits into language quite well. So we have to figure out how to make Duolingo more social. We can go in many different ways, but I think that's one thing that we have to invest more. And which way would you most like to go? Dating is actually not the direction I would go first, but I would love to go in a way where we build better features that you can do with others. So for example, gifting is one area people know that I'm passionate about. So for example, if I can send you more virtual items when you do something well, you can send something back, we build this reciprocity. Imagine I send you a streak freeze, you send me a thank you back. When you're in need, I send you another one. We can kind of create more of a positive gifting environment. That would be awesome. But social is one of those hard things where right now, not a lot of our users have friends. So when you build these features that are not easy to get right, they don't kind of hit massive adoption because not everyone has friends, but it's the chicken and egg. We don't have a lot of features, that's why people don't add friends as much. So we have to invest in both.

**Harry Stebbings** [69:21]:

When you look at chess, music, maths, and languages, is 85% still languages?

**Cem Kansu** [69:28]:

Yes. Majority of our users are doing languages. Chess is very new, like counted as kind of zero still. It just rolled out like a few days ago. So but math and music are a small percent of our business. Which subject? You can do one more. You're

**Harry Stebbings** [69:41]:

capped at five forever. Which one would you add? This will be the next Business Insider headline.

**Cem Kansu** [69:46]:

I think the one I would do is physics. Because, you know, as surprising as that might sound, it's very similar to math. Meaning, we can kind of apply a lot of the stuff we're doing on math, and we're doing a lot of product development for for teaching math. We can apply it to physics. I'm not sure if it's the largest market, or it's certainly not the sexiest market. It's probably the nerdiest thing we can teach. But it's certainly we have a competency, like, we're we're building the competence for math that we can apply to physics.

**Harry Stebbings** [70:13]:

God, fascinating. That's an audience grabber, isn't it? Speed equals distance over time. Where do I sign up? Gem, that's You

**Unknown** [70:20]:

sound excited. I'm

**Harry Stebbings** [70:22]:

so excited. I'm so excited. Okay. Are you ready to do a quick fire? Let's do it. Okay. So which other product do you look at and you're like, yeah. That's smart. I like what you did

**Cem Kansu** [70:31]:

there. The products I love are the ones I use a lot. So I'll mention two, and they are everyone uses these products a lot. One of them is I'm an Apple fanboy. IOS itself is a beautiful piece of software that is so well designed, so sleek, amazing animations, attention to detail, next level. Another product I love is Spotify. Amazingly built product. Obviously, very successful business on top, but I think they also take care and attention to detail is high. I'll mention a third one, which is we actually were chatting about this is Bump from Ammo. It's just a high craft, incredibly creative app. So those are the the apps I'm when I use them, I get ideas for me to think about for Duolingo. What was the biggest

**Harry Stebbings** [71:15]:

change PMs will have to encounter in the next five years?

**Cem Kansu** [71:19]:

The biggest change will be how they do product is going to be significantly different than today, and the tools they use will be obviously a lot more AI based. They will have to learn, basically, how to be a lot more independent rather than reliant on other functions to build the thing they have a vision for, Meaning, they will be independently coming up with prototypes themselves rather than today, they rely on machines that involve a lot of other humans to even visualize an idea and then turn it into prototype. AI tools will be the main thing they use to do that. What have

**Harry Stebbings** [71:55]:

you changed your mind on in the last twelve months? An example for me, was like, you know what, chat GPT, whatever, fuck it, we'll go to Grok, Perplexity, you name it. I don't think there's any lock in there. I'm like unwaveringly chat GPT focused now. I have memory. I love the brand.

**Cem Kansu** [72:11]:

Similar to you, I've actually changed my mind about the same thing, but I'll pick a different one. I thought this change in how we build product was maybe five to ten years out. Twelve months ago, if you had asked me like, Oh, how do you think the PM job will change? I'm like, Well, maybe in ten years, we'll be using more AI tools. I think it will get there slowly, would be my answer. Honestly now, probably in twelve to twenty four months, the job of a PM will look entirely different. So I think we're talking about five years because that feels safe, but honestly, my guess now, or what I've changed my mind on is it's going to be in two years, is my guess.

**Harry Stebbings** [72:47]:

Where do you most need to improve as a product leader?

**Cem Kansu** [72:50]:

What I'm trying to improve on is quicker decision making. One of my fallbacks is I can debate something for a very long time because I would like to try to, in my head, reach super 100% clarity or perfection, but that is a massive mistake because you're slowing your decision making down. I think you want 80% clarity and faster decision making. And, you know, really everything comes back to AI, but it does. ChatGPT is actually really good at helping me do that because it can be my sounding partner to bounce back and get to clarity quickly or enough clarity, so I'm actually doing that.

**Harry Stebbings** [73:24]:

Totally agree with you. It actually makes me think of the stat, which is you can learn a language to, I think, 95% proficiency if really devoted, but it takes five years to get to 98%.

**Cem Kansu** [73:34]:

Oh, yeah. The last mile is high investment.

**Harry Stebbings** [73:36]:

Totally. Yeah. Okay. Final one. In twenty years, where's Duolingo?

**Cem Kansu** [73:41]:

Twenty years is so impossible to think about. But here's what I imagine it will be. I don't know what platform it will be on. Assuming we're all still using mobile phones, it'll be on mobile, but if it's if the different consumer platform emerges, it will be on that. If we're all wearing VR headsets and living in the metaverse, which I don't think it will happen, it'll be on that. But Duolingo itself is going to be incredibly personalized in twenty years, meaning rather than, I think, following standard patterns, I mean, it's already headed there heavily, but this will compound. That's one. It will teach many more subjects. I think that is just going to happen. It will teach you, hopefully, everything that's good for you. Maybe dating by then will be part of the thing, we teach you how to date better, or maybe you'll be dating AI agents, who knows? But it will teach you way more things, and since it's personalized, it will teach all of those things with much higher efficacy. So an hour of time spent will give you much more yield than it does today. That's what I assume it will do. I think the other thing that will kind of emerge is these AI first experiences. And what I mean by that is now when people think of AI first, they think of, oh, it's ChatGPT. It's a text box or a voice input. What hasn't emerged much is UI creation that is also AI first. So I think you'll see a lot of on the fly UI generation off of what you're doing instead of hard coded UI that comes into play. So I assume in twenty years, this will be a thing that we experience where let me paint a picture, you enter a Duolingo lesson. Duolingo has not built a world where you you wanna engage in hiking up a mountain. However, that's where your conversation leads. That's where your interest is. UI can generate you hiking up a mountain and then kinda put you in that position, maybe teach you the word for mountain while kind of you immersed over the experience that. I think that type of experience will happen more, which will make it makes the whole thing way more entertaining. I look forward to Duolingo date. At

**Harry Stebbings** [75:36]:

that point, I'll take one with that.

**Cem Kansu** [75:38]:

It's fair enough. You're like, I want one future. I just this is why I

**Harry Stebbings** [75:41]:

did the show. That's why I did it. Listen, dude. Honestly, I, you know, I do a lot of shows. I I don't have this much fun in shows. You hear them. There's not this much fun. So I really appreciate you making it so fun, and this has been such a joy. Love it. It's it's been a pleasure. That was one of the most fun shows to do I've done in a long time. If you wanna see that episode, you can find it on Spotify by searching for 20 v c on Spotify. It makes such a difference if you leave a review or a five star. I would so appreciate that. But before we leave you today,

## Sponsor read

**Harry Stebbings** [76:15]:

are you struggling to beat model benchmarks or implement Gen AI in your product? If so, you need Turing. Turing is an AGI infrastructure company backed by incredible investors like Foundation Capital and Westbridge Capital, and they do two things. Number one, they help leading companies in AI labs like Salesforce, Anthropic, and Meta enhance their LLMs with advanced reasoning, coding, multilinguality, multimodality, and more. Number two, they combine human and artificial intelligence expertise to deploy cutting edge AI systems for awesome companies like Rivian and Reddit. Right now, Turing offers a free five minute self assessment to help you pinpoint your place in the gen AI journey, get tailored next steps to optimize your model strategy, and then finally learn how Turing can refine and implement your models for better performance. Take the guesswork out of Gen AI. Visit turing.com/20vc to start your free assessment today. And once your Gen strategy is on point with Turing, now let's switch gears to the future of mobile and how one company is turning your screen into real returns. We spend nearly half our waking hours glued to our phones, upwards of fifty hours every week. Recently, one company transforming this reality stood out so much I personally became a shareholder, Mode Mobile. Mode Mobile created the Earn Phone, the smartphone that pays you for daily activities. Instead Instead of big tech profiting billions from our attention, Mode returns over $325,000,000 directly to users through earnings and savings. Mode's revenues surged an incredible 32481% in three years, recognized by Deloitte as twenty twenty three's fastest growing software company in North America. And here's why I'm excited. Moe's equity offerings have raised over $30,000,000 from 20,000 plus retail investors, one of twenty twenty five stand out public raises. And you can now join me as a shareholder with as little as a thousand dollars at invest.modemobile.com/20vc. For a limited time, unlock up to a 100% bonus shares and a free earned phone. Email us for the investor brief at 20vc@modemobile.com or check out invest.modemobile.com/20vc. We've talked about making money from your phone. Now let's talk about making your product do exactly what you want it to. Whether it's the software you build or the software you buy, your tech stack should be creating results, not creating roadblocks. Well, Pendo's no code software experience management platform makes your software better with tools to see where users get stuck, guide with in app messaging, and constantly improving your UI. It's so easy that over 14,000 businesses use Pendo to increase revenue, lower costs, and reduce risks. Businesses love the control. Engineers love the freedom. Everyone wins. Start for free today at pendo.io/20product. As always, I so appreciate all your support, and stay tuned for an incredible episode coming on Monday with Larry Asherbrooke at g squared.