Duolingo CEO Luis von Ahn thinks AI has a lot to teach us

Today, I’m talking with Luis von Ahn, the cofounder and CEO of Duolingo, the popular app that teaches languages. It’s an interesting time to be in the language business: if there’s anything the current state of AI tech can do, it’s babble away in different languages with people who aren’t quite fluent in what they’re hearing.

That means there are lots of opportunities to enhance a product like Duolingo with AI, and Luis and I talked about the new features in something called Duolingo Max, which offers chat conversations with some characters and even video calls with an AI avatar named Lily.

I wanted to talk about all of that, but I also wanted to talk to Luis about learning generally. If you’re like me, you’ve stopped and started using Duolingo several times; if you’re an overachiever, you’ve got a streak going and might even have a streak to maintain today. That streak is the key, and you’ll hear Luis come back to that as a big idea several times.

Engagement is the key, he says, because simply showing up is the cornerstone of actually making progress with language learning. You can’t teach someone who isn’t there, so over time, Duolingo has become more and more of a game, because people like to play games.

But there are real conflicts between gamification and actual learning. Luis is happy to admit that that conflict exists, and he’s given it a lot of thought. For him, the gamification is the important part because not only does it bring you back to Duolingo, keeping the business humming along nicely, but he says it also produces the results in language proficiency that Duolingo is aiming for. 

Luis got pretty deep into explaining where the money comes from. As you might guess, it’s from iPhone users in wealthier countries like the United States. And some technical decisions Duolingo made very early on mean the iOS version takes priority — it can take a year or more for features to roll out on the Android version of the app.

But Duolingo is a global product, where the biggest chunk of learners are actually trying to learn English — and those users are way more likely to use an Android phone and to want or need a free version of the product. There are a lot of tensions here, and you’ll hear Luis talk about his own childhood in a poorer country and how that informs his decisions.

This is a good one — Luis is the founder, and he’s helped the company to go public and now is helping it to embrace a pretty big technology shift. AI has a direct impact on the product he makes, and we talked about it all in a pretty direct way, with only a handful of jokes about founder mode. And of course, I asked him whether he approves of all the unhinged things the Duolingo owl says on social media. 

Okay, Duolingo CEO Luis von Ahn. Here we go.

This transcript has been lightly edited for length and clarity.

Luis von Ahn, you are the CEO and cofounder of Duolingo. Welcome to Decoder.

I often start by asking CEOs what their company is, but I feel like everybody knows what Duolingo is. How do you define Duolingo?

It’s an app that teaches languages. That’s what we’re mostly known for. As of the last couple of years, we have also taught math and music. It’s the most popular way to learn languages in the world. A fun fact: there are more people learning languages on Duolingo in the US than in all US high schools combined. This is true in most countries in the world. We teach languages to more people than the public school systems.

You have some big announcements coming up at Duocon that will be public by the time this episode airs. One of them is the ability to chat with characters like Lily and others.

Yes, the ability to video call with Lily.

How does that work? How are you making that happen?

We have this cast of characters that our users love. One of them is an emo teen with purple hair who is very unimpressed by you. You can talk to her now and practice your conversation skills, and you can have really good conversations with her. There are a lot of things that are amazing about her. First of all, she adapts to your level — we know your level because you’ve been learning on Duolingo, so we have a pretty good idea of what your level is. 

The other thing is that she has memory, so she remembers the last time you talked about something. For example, I just had a conversation today where she remembered that, last time, we talked about the fact that I like Nirvana. She was telling me that her favorite song is “Smells Like Teen Spirit.”

Easy choice, I have to say.

We’re dating ourselves on that one, but yes. These are pretty enjoyable conversations, and you get to practice your language. It’s entirely spoken, and it just works really well. We’re very happy because this is the first time that you think, “We really are not going to need humans for this.”

The animations and stuff, are those stock animations? Are they loops? How does that work? Is it in real time?

Yes. It’s animated in real time. We have a rig for her. We bought two animation studios in Detroit. This is why we have an office in Detroit. They’ve done a really good job. Her mouth moves when she’s speaking, and it’s tied to what she’s saying. She rolls her eyes at you. 

When you think about that investment — “We’re going to start building rigs and animations for characters. We’re going to do it all in real time” — I’m just coming back to cost. That’s a big investment. Do you think that’s going to make your existing users pay more money? Or is it going to get you new users?

I think it’s both. We see it as continuing to work on the app. There are a lot of places where we use a lot of animation, and we see it as continuing to work on the app. And generally, as we continue working on the app, we get more users and get more of them to pay.

The reason I ask that in this context specifically is that the economics of AI is just a series of question marks right now. I ask this of everybody who’s making the investments. How do you see it coming out on the other end?

For this particular feature, I think it’s an excellent use of large language models, and on our end, it’s working pretty well.

The other big announcement you have is called Adventures. It sounds like a video game. What’s going on there?

The way Duolingo works is that the homescreen is basically a path, and you’re just doing lessons. Some of the lessons are now going to be this thing we call an “adventure,” which is really just one of those video games where you move characters around. What’s cool about it is that you’re learning how to solve real-world situations on Duolingo.

For example, it’s like a little video game where you are one of the characters and you’re told, “Okay, go buy a pizza.” You move and have to ask around, and then you ask some people, and they tell you, “Oh, the pizza place is over there.” It’s super fun and it helps you learn to navigate the real world. So we’ve been working on that. What’s cool about that feature is that all the scenarios were mostly generated by AI. In the past, that feature would’ve taken a long time to scale, but we were able to scale it pretty quickly because of AI.

I played with Duolingo this morning. I have a long and complicated history with trying to learn Hindi. It’s free. I was using it for free today. How does the app make money?

It is free. You can use it entirely for free without ever having to pay. If you don’t pay, you may have to see some ads, and we make money from the ads. But also, if you want to turn off the ads, you can pay to subscribe, and it turns off the ads and gives you some extra features. We also make money from the subscription, and actually, the majority of the revenue comes from the subscription.

Is Duolingo profitable as a company?

Yes. As of relatively recently. 

I’m curious about this. I hear about this split from almost everyone we talk to: we start out, we want to grow our base of users, ads help us do that. It helps us keep the product free. And then the real money is going to come when we add value and we add paid subscriptions. Particularly with advertising lately with app tracking transparency on Apple platforms, with the massive influx of inventory from all the other platforms in the world, it seems like ads are even harder to make money on than ever. Has that been the case for you?

It’s probably true. Ads have never been a priority for us. I don’t know the exact number, but it’s something like 6 or 7 percent of our revenue comes from ads. For us, as long as they’re there, they’re a good reason for people to subscribe. But generally, we make about 80 percent of our revenue from subscriptions, even though, by the way, only a little under 10 percent of our monthly active users pay to subscribe. So 10 percent of our monthly active users give us more than 80 percent of our revenue.

And all of that revenue is in languages? Or is math growing?

The majority is languages. Math and music are growing. We launched those about a year ago, so they’re just getting started. It’s overwhelmingly languages.

What languages are the most popular?

English is the most popular by far. Forty-five percent of our active users are learning English. The second is Spanish, the third is French, and then there’s a big drop-off after that.

Are the majority of your users outside of the United States? Or are they inside the US?

US is about 20 percent of our users, and 80 percent are international.

So are 80 percent of your users trying to learn English?

About 45 percent are trying to learn English. Within the international segment, they also want to learn other languages.

There are a lot of languages offered in the app, and it seems like one way you could allocate resources would be by saying, “English is the most popular, we’re going to put the most resources there.” But that doesn’t feel like how the app works. How do you think about it?

We definitely do some of that. I was going to say we don’t do it as much as we should, but I don’t know if that’s the case. We don’t do it commensurately with the number of users because we would probably spend all of our resources on English, Spanish, and French. We spend the majority of our resources in the top eight languages to learn, and then we spend very little resources outside of that. The top eight are English, Spanish, and French. Then there is German, Italian, Japanese, Korean, Portuguese, I think, and Chinese. 

Mandarin. And after that, there really is a huge drop-off. For example, Arabic is a large language, but there are not that many people learning Arabic. So we do put some resources there, but it’s much less than for the larger languages.

How do you think about that kind of demand? I open Duolingo, I look at it, I’m like, “I should probably learn some Cantonese.”  I think, “Man, I should be much better at Hindi than I am.” Those are real things that I think all the time. I imagine there are a lot of people in my particular diaspora who feel the same way. But that’s latent demand. Do you ever go out and say to people, “You should learn some Spanish”? 

Do you ever say, “We should market Spanish in the American South”?

We don’t. We have remained neutral about that, but it is an interesting thing that demand for learning languages is not as correlated as you would like to see with the number of speakers or maybe even usefulness in a geopolitical world. For example, even though Chinese is one of our top eight languages to learn, only about 2 percent of our users are learning it. It’s relatively small, even as the most spoken language in the world. 

One of the things that goes into people’s calculus is how hard a language is to learn. Chinese, at least for English speakers, is just a lot harder. We have data. To get to a pretty good point in Spanish for English speakers takes, call it, 300 to 400 hours. That same level of knowledge for Chinese takes about 2,000 hours. The reality is, in the United States, if you’re just going for pragmatism, return on investment, Spanish is probably much better. In the US, you probably should learn Spanish. It’s quite an easy language to learn.

In the US, you should probably learn Spanish. That is a marketing message.

Well, we don’t say that. We’ve tried to remain neutral. We probably would get in trouble, or I would get in trouble inside the company, if we started pushing people on certain languages. 

Maybe not inside the app, but as a way to grow, as a way to capture new users. It seems like a lot of what Duolingo is right now is people know they should be multilingual or bilingual at least, and so Duolingo is there. But there’s also a huge portion of the population, at least in this country, who are like, “Screw it. I speak English.” The idea that there’s value in learning a second language is foreign to them.

There is, although I’m very happy with our results in the US. Historically, there hasn’t been a big desire to learn languages in the US and the UK. The thinking has been, “You can learn English. We are very happy.” In the US, 80 percent of our users were not learning a language before Duolingo. We’re growing the market in the US. It’s the same number in the UK. I’m very happy with that.

I think back to learning French when I was in high school in Wisconsin. There’s learning the language and then there’s all of the culture that comes with a language, particularly some of the regional languages. High school French is a lot of looking at a picture of a baguette. It’s totally foreign to whatever you’re doing. Do you think about that, inside of Duolingo, that there’s a huge cultural component here?

That’s more investment, right?

We do, and we try to add the culture. We don’t do it as much as maybe we should. We try to stick mostly with languages. It also depends on the language. Some languages are quite tight to the culture, and some are less so. Spanish is a good example. There are 20-some countries that speak Spanish, and some of them are pretty different than others. We do a little bit of culture, but we try not to be like, “Oh, you’re learning Spanish. You’re a Mexican person with a sombrero.” We try not to do that. I mean, we also have to not be offensive. But we try to add it a little bit. I would say that it’s not the primary goal.

The reason I ask that is that Duolingo is instantiated for most people as its mascot. We should talk about the mascot’s personality and its social media presence, but it’s fairly abstracted from a person teaching you the language. There’s not someone on the other side that’s like, “I’m teaching you this. Here’s the culture that comes with it.” You might have other teachers who might teach you in another way.

There’s an abstraction there that just feels interesting, especially as we’re obviously going to talk about AI and how you’re using that and how you’re expanding the platform. I want to push on that a little bit, that abstraction. Do you think it’s resulting in people who’ve learned a language or people who’ve learned how to communicate?

It’s been very much on purpose for us to not put humans in the app, as in human teachers. There’s nothing wrong with human teachers. It’s just the case that, from the beginning, we’ve been a technology company, and we’ve wanted to make it so that technology teaches you. There are a couple of reasons for that. One is that it’s a lot cheaper to teach you with technology than with a human teacher.

The other thing is, somewhere between 80 and 90 percent of language learners don’t want to talk to another human. They may tell you they do, but they don’t. It’s because when you’re learning a language, you’re pretty shy about it, and only the extreme extroverts are okay talking to a stranger on video in a language that they’re not very good at. The majority of people won’t do it.

We’ve done research studies over the years because, over time, we thought maybe we should add humans. But these research studies are some of the most amazing things that I’ve seen. When you talk to a user, you ask them, “What do you think could make Duolingo better?” Historically, in the past, they have said, “Well, more practice conversation with a real person.” They have said that. And then you ask the user, “Okay, so you’re telling me if I put a human on Duolingo, you would do that?” And then they say, “Yes, I would.” And you can even ask them, “Would you pay for it?” And then they’ll say, “Yes, I’ll pay for it.” And then you tell them, “Okay, do you want to do it right now?” And the answer invariably is, “No, no, not right now.” People just don’t want to do that.

That’s why we haven’t put humans in, and I think it’s been a good decision, especially now that we can do a pretty good job of getting you to practice conversation without a human with large language models.

I want to ask about this because I’ve been asking a lot of people on this show: what good are these large language models? What are the products you’re going to make? I understand you’re making the models, and it feels like Duolingo has a very natural solution, which is that you can talk to it and it’ll talk back. It doesn’t matter if everyone is hallucinating because all you’re doing is practicing talking.

That’s exactly right. It is a really good application. You said it. It doesn’t matter if it says something that is a little wrong because you’re just practicing your language. Also, it doesn’t matter if it makes a small mistake. Sometimes it makes a small grammatical mistake. People don’t even notice because they’re usually beginners in Spanish or French. It also can adapt to your level really well. Large language models are really good at adapting to your level.

So we tell it, “Okay, adapt to a beginner in Spanish.” We even tell it, “Hey, because we’ve seen this person learn on Duolingo, we actually know all the words they know.” So we tell the language model, “This person only knows these 200 words, so please mostly use only these 200 words.” It works really well for that.

How much investment into AI are you making? This is a new product. It’s very costly. Everyone is telling me about how much the Nvidia GPUs cost. You said you’ve only just become profitable. This feels like the thing that will immediately make you not profitable again as you invest in AI.

We’re investing a lot. Fortunately, it’s good for us in terms of profitability for two reasons. There are two places where we invest in AI. The first is generating data that is going to be used in our lessons. That data used to be generated partly with humans, and now it’s mostly generated with AI, and it’s a lot cheaper to generate with AI than with humans. It’s also a lot faster to generate it with AI, so we’re very happy with that.

And then the other big use is real-time conversation. That one is expensive. It’s expensive to provide a real-time conversation with a user, but what we had to do is add a higher-priced plan. We now have two subscription plans. We have Super Duolingo, which is our standard subscription, and we added a new one called Duolingo Max, which is about twice the price of Super Duolingo and gives you the conversation practice. It’s expensive, but people pay twice as much, so it really doesn’t cut into it. It’s worked out well for us.

Let me dive into the economics of that because, in general, I’m fascinated by whether any of this will result in profitable, sustainable companies. There’s a lot of money flowing into this. So, you charged twice the price to run inference. Is that someone else’s large language model? 

So you’re buying some capacity from OpenAI, you’re buying some tokens from them, and you’re reselling them to users for twice the price of your standard plan. What’s your margin on that resale?

I don’t know the percentages off the top of my head, but I do know that it’s good for us in terms of margins.

That’s the thing I’m curious about. I don’t know if it’s good for OpenAI all the way at the bottom of that chain. I don’t know if that’s profitable for them. But as you build products on this stuff, it seems like your economics depend on their economics, in some way, because you need to add a margin to that. That all seems very complicated and tenuous, especially if the AI features are what bring you new users.

The good news is, at the moment, the AI features are not bringing us new users. 

Yes, it is bringing us new revenue. There’s a good margin there. So, for a number of reasons, the price of the same exact call is going down over time, whether you do it through an OpenAI or whether you do it through a Microsoft. Everything is getting more efficient, and chips are also getting cheaper over time. At the moment, there’s a good amount of cushion, but we expect that there’ll be an even larger amount of cushion over time. At least for our application, I’m not particularly concerned in terms of margins. For our application, the margins work out pretty well.

Do you think the models are somewhat interchangeable? This is a thing that I’ve been hearing more and more is that the model business isn’t the thing; the product business is the thing.

I think the answer is yes, but the operative term is “somewhat.” They are “somewhat” interchangeable. We’ve tried to build our technology stacks so that you can interchange them, but the reality is that you start getting wonky stuff because you probably spent a lot of time testing your way into the right queries. You may have done some fine-tuning. You can interchange them, but if you do, you probably need to spend a few months making sure that the wonkiness goes away.

When you think about this investment over time, does it feel like you need to put the money in upfront and you’ll get more efficient on the back end? Or does it feel like, “Oh, this is going to be the future of the company, so we need to rebuild around the capabilities of a large language model”? 

It’s somewhere in between. I do think that large language models are going to be very positive for Duolingo — they already are, and I think they’re going to continue being very positive. What is not true is that large language models solve all our problems. One of the biggest issues that people aren’t talking about, particularly with education, is that large language models are good at teaching you stuff. They’re not good at engagement. And that’s the hardest thing with education.

The hardest thing about me trying to teach you something is just keeping you engaged. Somehow, people forget. I see some people saying, “You can learn quantum physics with ChatGPT.” And yeah, sure, but that’s just not that impressive. You can learn quantum physics with a book. The technology to learn that has been around for a long time. It’s called a book and it works. It’s just that people don’t really want to read a quantum physics book. And similarly, most people probably don’t want to go to ChatGPT and start asking questions about quantum physics. It’s the same thing for language learning. Large language models are very good at getting you to practice, but keeping you engaged is pretty hard.

I don’t know if large language models are going to help all that much with that part. In the end, this is a sad thing, but the reality is that Duolingo is very gamified. I wholeheartedly believe most people would rather spend more time playing Candy Crush than talking to others. That’s maybe a sad truth. And there are some exceptions. I mean, people love talking to someone they’re in love with, and sure, that’s nice, but the reality is, most of the time, most people would rather spend time playing Candy Crush or scrolling on Instagram than talking to others. I just don’t think large language models are going to help much with all of that. 

You have a long history in gamification. Your first project, which you sold to Google, was a gamified thing. You did reCAPTCHA, which is essentially gamifying training data in a particular way.

Do you think there’s an evolution in Duolingo — that the first thing that you worked on was the engagement and bringing users back to the app and having the character, and then the underlying content was language lessons? When I first started using Duolingo several years ago, I was like, “Oh, this is very familiar. It’s just that this bird won’t leave me alone and that’s why I’m back again.” And now you’re talking about this whole other spectrum of things: we’re going to use AI; we’re going to have these natural language conversations; we’re going to expand to mathematics.

When did you feel like you were making the transition from “we’re gamifying this very familiar thing” and “we’re using this new engagement mechanism” to “this is now a wholly new thing”?

From the beginning, this is a central thesis that we believe here at Duolingo: the hardest thing about learning something by yourself is staying motivated. In fact, that is probably the reason for the vast majority of our success is that we realized that early on. From the beginning, we have tried to have a thing that is enjoyable to use and that keeps you coming back. We have probably spent more effort on that than anything else.

Internally, our feeling is that learning a language is a lot like working out. It doesn’t matter all that much whether you’re doing the elliptical or a Peloton or a treadmill. By far, what matters the most is that you’re doing it every day, whatever the hell you’re doing. It’s kind of the same with Duolingo. Maybe some methods are more efficient than others, but what matters is that you’re doing it every day. We got very good at that. Now, once we got very good at that, we started trying to add more sophistication in what we teach, and we’ve been doing that for the last few years. But always, primarily, we are a motivation engine.

Is that the core of it still? 

I’m going to end up asking you about founder mode, but you’re the founder. How do you keep the focus on that part instead of everything else?

I do. I spend effort on that. But it’s not just me. At the company, it is pretty well understood that if it’s not fun, it’s not going to work. We spend a lot of effort trying to keep Duolingo enjoyable. This is why, for example, when we did this thing where you can talk to an AI to practice conversation, you’re not just talking to a random AI; you’re talking to one of our cast of characters. It has a personality. Really, everything we do, every time we put something out, it’s ingrained in our thinking that, “Oh, this has to be enjoyable.” I spend effort pushing that agenda, but I don’t have to all that much because it’s very ingrained in the company.

Where are you all located?

The largest office is in Pittsburgh, Pennsylvania. We have about 400 people there. We have about 250 in New York, and then we have offices in a few cities. We have one in Detroit, we have one in Seattle, one in Berlin, and one in Beijing. All of those offices have 30 people in them. But one key thing is we are not remote. We’ve got to come to one of those offices.

I just wholeheartedly believe that you can work better that way. Most of what we do, not 100 percent, but most of what we do, is creative stuff. It’s a lot harder to do so over Slack and Zoom. That worked out for about nine months during the pandemic, but it is actually impressive how when the pandemic started and we all had to go remote, we executed pretty well. But toward the end of it, our ideas had run out. We were executing the ideas, but we had run out of new ideas. It’s pretty amazing, as soon as we came back to the office, within three months, you would see all these ideas popping up, and it’s because, when you’re remote, you can’t sit in front of a whiteboard and talk about stuff. Also, we have lunch together here every day. In the lunch line, you hear people being like, “Hey, I haven’t seen you in a while. I thought of saying this to you.” It’s just something you would never send a Slack message for.

I think the combination of all of that makes it a better company. I don’t have much proof, but I am extremely convinced.

Sundar Pichai at Google told me at the very beginning of the pandemic that he was worried that the company would run out of ideas if they stayed remote too long. He said, “We know what we need to do for the next turn. I’m worried about what happens at the next turn.” Did you have controversy when you reimplemented return to office?

Do your employees think that?

We’ve done a lot of dumb things at Duolingo, but this was not one of the dumb things we have done, in retrospect. In my first message saying, “Everybody’s going to work remote,” I said, “But we’re going to come back to the office. I do not want Duolingo to turn into a remote company. We are not a remote company.” We kept saying that the whole time. A lot of companies did this thing where during the pandemic, they would hire people all over the world because, whatever, you’re remote. We never did that. When we hired people, we would say, “I get that you’re not coming to the office right now, but your job is in New York, and we expect you to be in New York because at some point hopefully soon, we will be back in the office.” We never stopped repeating that. By the time we said, “Okay, time to go back to the office,” this was not a surprise for anybody. I don’t think we lost a single employee from that.

Do you think that the markets you’re in help you with that — being in Pittsburgh and Detroit? If you were in San Francisco, I think a lot of people would say, “Screw you, I can go get another job.”

That’s probably true. We are not in San Francisco, and that’s probably true. Although the New York office is now the second largest office, and we also didn’t lose people in New York.

Do you find that people are demanding more flexibility even with a full return to office?

Sure. I mean, compared to before. For example, we’re not here five days a week in the office. The way we work is Tuesday, Wednesday, and Thursday, you have to be in the office. Monday and Friday, it’s optional. What happens, in practice, is that about half the people come in on Mondays, and around 20 percent of the people come in on Fridays.

We are talking on a Friday. 

I am here. I am here in the office, though.

Oh, you’re there. Oh, very good.

I am here in the office. I come. But I don’t feel like I have the political power inside this company to say, “All right, people, you’ve got to come in five days a week.” I feel like that would not go over well.

One of the other pieces of the pandemic puzzle and return to office is that there was a suppression of demand to travel and explore. I have friends who, at least from their Instagrams, haven’t set foot back in the United States in two or three years. Have you connected to that group of people who want to learn languages on the go? Has the reexplosion of travel had an impact on your business?

Travel is interesting. Now that we’re a publicly traded company, people have hypothesized all kinds of stuff about travel with us.

They’re like, “Travel’s opening up. That may be good for Duolingo.” Or they say, “Travel’s drying up. That may be bad for Duolingo.” The reality is that travel does not affect us all that much. I can have hypotheses for why that is, but we have not seen traveling closing down or opening up affect us all that much.

We have a lot of learners that have different motivations, but the two big buckets are not travel. One of them is a hobby, and that’s the biggest bucket in the US. If you ask people in the US why they’re using Duolingo, the most common answer is, “Well, I used to play a lot of Candy Crush or I used to do a lot of Instagram and now I’m doing Duolingo and at least I’m learning some Spanish. It’s just a hobby.”

And then the other huge group is people learning English. For them, it’s not about travel. They just need to learn English either for educational opportunities or for job opportunities. Those two big buckets account for 90–95 percent of our users. Travel just doesn’t affect it very much.

We started talking about latent demand. What are people coming to you for? And then there’s growth, which is, “how do we go create some demand?” When you think about the structure of the company, would you ever say, “We’ve got to go do marketing to make travel happen”? Or is that just not how you think about it?

No. In our marketing, we haven’t thought about that. Ninety percent of our users come in from word of mouth, and that will keep happening, I think. We also spend very little in marketing comparatively. Our entire marketing budget for the whole world, and we really do operate in every single country in the world, is 50 million a year, which is quite small for a company with our revenue. But whatever we’re doing with marketing seems to be working pretty well, and we don’t spend a lot of money on it.

I feel like I have yet to ask you the Decoder question. So, as long as we’re talking about your marketing spend, how is the company structured?

We have functions. There’s the marketing function, there’s the engineering function, there is the product management function, design, etc. We have functions and each function has a function head. To give you a relative idea of the sizes of the functions: engineering, product management, and design combined account for about 70 percent of our employees. Design is weirdly large for our company. We have 850 total employees but about 130 people in the design department. So design is large, but we have engineering, product, and design account for about 70 percent.

Those are the people working on the product. If you look at that group, it’s structured into areas, and each area is related to one of the things we’re trying to optimize. For example, with language learning, the three main things we’re trying to optimize are: engagement, so how fun Duolingo is; teaching [the material] better; and how much money we make. We have an area for each one of those things. Then, in each area, there are teams, and in each team, there are people. It’s a little bit of a matrix structure.

One important thing that I think has worked really well for us — but it’s not that easy to do — is that our areas and our teams are not feature-based. What I mean by that is that most software or app companies usually have a team for each feature. So this is the login team, which owns login. Or if you have a leaderboard, this is the leaderboard team and they own the leaderboard. We do not have that. Our teams don’t own features. Our teams own metrics. So we have a team for subscription revenue. We have a team for daily active users. And they can touch whatever they want in the app. All they have to do is continually increase the metrics. There are positives to this, which are very aligned to metrics. There are negatives in that no team owns certain features. When something breaks, there are a lot of people being like, “It’s not my feature. I don’t know.” There are positives and negatives, but all in all, this has actually worked out really well for us.

This sounds, one, like a reaction to working at Google where teams do own things like the login screen and they endlessly communicate about how they’re going to change logins. But it also seems like it might work for a small company where one person can see the whole product or understand the whole product and how it all works together. And then you’re going to get inevitable collisions as two people try to change something to increase two different metrics in different directions. How do you resolve those collisions?

There are definitely collisions. There are a couple of things that help us here. One is, every change to the app passes through this review process called product review, which is not just one person. There’s a group of people that have a lot of knowledge about how the whole thing works. They serve as a little bit of a semaphore, a little bit of like, “No, you should not do that.”

And then the other thing that is really important is we have guardrail metrics. So here’s how that works: If you are on the team that’s trying to improve subscription revenue, your goal is to improve that metric. But we tell you, “You can’t mess up any of the other metrics.” For example, if you do an experiment that improves subscription revenue by a million bucks or whatever but it decreases daily active users, you can’t launch it. That has really helped teams police themselves. They at least won’t go do anything that really messes something else up. The combination of these two things has helped.

You are right that if we had 100,000 employees, I don’t think the structure would work. That said, I don’t think that a company like Duolingo, at least with the products that we have, needs 100,000 employees. I think we’ll grow and we’ll continue growing. Maybe we’ll get to, I don’t know, 5,000 employees, but I doubt we’ll ever get to something like 100,000 employees.

How often do these collisions come all the way up to you? How often do you have to make a tradeoff?

Not that often. Teams police themselves a lot. I do see every single change that passes through the app. I do see that, but usually, I’m not making tradeoff calls. The main thing that I’m looking for is making sure that everything’s high quality.

At the beginning of the show, you were joking about founder mode because I called you the cofounder. That’s Brian Chesky, who has been on the show. He’s talked a lot about how he refactored Airbnb. He was the conductor of the orchestra. That has gotten whatever amount of attention it’s gotten. 

Do you see yourself in that kind of role, that you’re the person who can see the whole app, that you are the person who understands how all these tradeoffs are getting made?

Yeah, it’s definitely true. The good news is our employees stick around for a very long time, and our leadership, particularly in the product areas, has been the same for the last eight years. I have a view of everything, but the reality is that our head of product, Cem [Kansu], has a view of everything, too. Our head of design, [Ryan] Sims, also has a view of pretty much everything. Yes, I am in that mode, but we have a number of people who could probably play that role as well.

When the three of you disagree, how do you resolve it?

The good news is there’s little disagreement, which happens for a few reasons. The first is that we’re a very metrics-based company, so usually, we just let the metrics speak. If we run an A/B test and the metrics say something, my opinion doesn’t matter all that much unless it’s something that we think is really like a dark pattern or something. But generally, my opinion or their opinion doesn’t matter all that much. That’s one reason. The other is that we’ve been working together for so long that we’re pretty aligned on everything. And then the last thing, I have the saying, “If we’re going to go by opinion, let’s go by mine” Generally, when we have disagreements, I see how deeply they believe in their thing, and sometimes I just disagree and commit. But if we believe with equal strength on something, I will just go with my thing.

That might be the most succinct definition of founder mode I’ve heard yet, actually.

What is — that “if we’re going to go with an opinion, let’s go with mine”?

“Just do what I say” is really the answer to what founder mode is.

But the majority of things, we don’t really go by opinion. The majority of things are just by data.

So here’s the other Decoder question. This is a good foundation for it. How do you make decisions? What’s your framework? 

For the company or for me? They’re similar, but they’re not identical.

Some people don’t think they’re different and some people think they are very different, so answer whichever way you want.

For the company, the decisions are very much tied to return on investment. With most things, there’s a return on investment calculation. Even if we don’t sit there and write the numbers down, there’s how much effort you’re going to put into something and how much you’re going to get back. That drives most of our decision-making. There’s another thing that is not unique to Duolingo but I think is not the norm at most companies, which is, usually, when you’re doing a project, there are three things that matter: how much does it cost?; how fast are you going to do it?; and what’s the quality? Usually there are tradeoffs between these things. At Duolingo, the most important thing is quality, the second most important thing is speed, and the third most important thing is whether we’re on budget. In many companies, it’s the other way around, where the most important thing is budget, then speed, then quality. Here, quality is the most important thing. So that’s another component of our decision-making.

That’s for the company. For me, it’s very gut feeling-driven, which I used to find myself trying to justify. I have stopped doing that because, at this point, I’m like, “Look, this is what I think we should do. I can tell you reasons that I can probably come up with after the fact, but the reality is that my gut says we should do that.” Because I’ve been working on Duolingo for 13 years, my gut’s pretty good. It’s not 100 percent correct. I make mistakes, but it’s pretty good. I mainly do things based on gut feelings, and then I tell people the justification afterward. But everybody around me knows that these justifications are after the fact. They’re not rational thoughts.

That obviously works for a startup founder for a private company. You’ve been a public company CEO for over three years. Is that working for you as a public company CEO?

Yes, because again, the majority of decisions that we have to make, there’s a clear answer. It’s just like, “Well, look, this is going to lose us money. Let’s not do that.”

But do you have to change the way you communicate? I’ve heard this from a handful of CEOs who’ve taken their companies public and now they’re on the quarterly reports cadence and they have investors. Elliott Investment Management might show up on your doorstep and be unhappy that you’re not marketing to more Chinese speakers or whatever. Have you had to change that attitude now that you run a public company?

No, it’s been very fortunate. First of all, we hired an amazing CFO before we went public, Matt Skaruppa, and fortunately, he deals with most public company stuff. I don’t do a lot of that, and I’m very thankful for that because I don’t have a finance background. I have a PhD in computer science. That’s what I’m good at, not finance. So there’s that.

The other thing we’ve been fortunate about so far as a public company is that we’ve executed well. I think that has given us a little bit of latitude in that basically we don’t get asked very tough questions because we’ve executed very well. I am sure that won’t be the case forever. I’m sure at some point we’re going to miss a quarter. We haven’t so far, but I’m sure we will. And when that happens, I’ll probably have to answer some questions and I’ll probably have to tell people, “Sorry, we’ll be more buttoned up from now on.” But so far, I show up to earnings calls in a T-shirt. The day you see me show up to an earnings call in a suit, you’ll know that we fucked up.

Yeah, it’s time to get out.

[Laughs] Like, “Oh, so sorry.”

The other thing I hear from public company CEOs is something that relates to your emphasis on quality first. You have a lot of metrics, which means your investors can see a lot of your metrics or demand a lot of your metrics in different ways. Quality is not measurable in that way, right?

At least in the current market, it’s not a great story to investors if you’re saying, “Okay, we’re going to invest a bunch of money in AI, and we think in this use case, it’s going to be really successful for us or we can charge more for it but we have to spend a bunch of money upfront and we have to wait to make it good.” How are you managing that now?

You’re right, quality is not measurable. The way we make decisions about that is that, particularly in our design department, we have people who are very much sticklers about quality. We’re like, “Nope, that’s just not good enough.” We do that a lot.

In terms of investment, I mean, honestly, with the public markets, we don’t talk much about that. We talk about the metrics. We don’t talk too much about how it turns out we spent an extra month working on this feature just because we didn’t really like the way the owl was animated. We don’t talk too much about that. I think that’s fine. But my guess is that if we went on earnings calls and spent all of our time talking about how much effort we put into animating our owl, I don’t think people would like that.

I honestly think more companies should spend more time talking about how much time they spend making things good. That would be, I think, a significant upgrade to American capitalism.

I would like that. But yeah, the reality is we do spend an inordinate amount of time on things. If you look at our app, it, over time, has become very animated. We spend an inordinate amount of time looking at the precise frames of the animation. We’re like, “See, this is not smooth enough.” I’m not claiming that our app is perfect, but we at least try really hard for it to get as close to perfect as possible.

One thing that also seems hard to measure, or a metric that might lead you in different directions, is how successful Duolingo is. Maybe the most important metric of all is: Are people getting good at Spanish? Are they leaving this experience with the ability to communicate in Spanish? Do they not just know the language but can actually communicate? Can you measure that?

Yes, we can, but not as effectively as you would like to measure it. So the answer is, yes, Duolingo works. We have measurements. I’ll tell you how we measure it. Unfortunately, this is the only way we know how to measure it reliably. It’s not that great of a way, but it’s this: You take somebody who’s just starting Duolingo, you ask them a bunch of questions about their previous knowledge, and you also give them a test to measure how much they know. Then you have them use Duolingo for a long time because it takes a while for you to actually learn stuff in the language. You have them use it for a year or two. And then, at the end of that period, you ask them questions about whether they used other resources and you also give them a test to figure out how much better they got. It turns out that people who knew nothing before and used Duolingo and did not use other resources learn about as much as or more than in a classroom. It varies by the study, but we’re pretty happy with that. The results actually work.

The problem with this way of measuring is that it’s very slow. It takes one or two years for us to get a new measurement, and I really don’t like that. But we have not been able to come up with a better way despite the fact that we have tried. We’ve done things in the app where we’re like, “Okay, we can do micro-measurements of whether you’ve learned this word.” It’s been super complicated to do that and never given great results. So we just rely on these old-school, “pre-test, post-test” methods. That’s it.

This is where you veer right into lengthy society-level debates about education and how we measure the performance of schools and teachers. Do you feel like you’re participating in that system? You’re using their measurements, right? This is how schools do it. They test you.

We are using their measurements, and efficacy is really important. We spend a lot of effort trying to make sure that we’re efficacious. The other good news, even though the timescale here is in years, is that you can plot how effective Duolingo is. If you look at it over the last 10 years, every year it is more effective than the previous year, for sure.

That’s on this test-based measure. More people are passing the test?

Yes, more people are getting higher scores on the test. Basically, people are learning more on Duolingo every year. And there are a number of reasons for that because we work to try to teach [the content] better. But it is definitely true.

At this point, when we compare ourselves, we know we are as good as or better than a classroom environment. We know we’re not yet as good as a good one-on-one human tutor. Our goal is that we can do that over the next three to five years: be as good as a one-on-one human tutor in terms of efficacy. But we’re way better in terms of getting you to stick around. But in terms of efficacy, if you have the money and the strength to continue going, a one-on-one human tutor does better.

Do you think that there’s a conflict between gamification and engagement — the things that you’re historically successful at — and education?

How do you manage that conflict?

Very easily. Always go with engagement.

I mean, presumably, you’ve heard both sides of the argument. Why have you made this decision?

I’ll give you many arguments, but the one that works the most is this: It doesn’t matter how effective you are. You can’t teach somebody who’s not there.

That’s it. People leave. The reality is it’s not always true that engagement and learning outcomes are at odds. But when they are, we usually prefer going for engagement. I’ll give you an example. There are some things that are frustrating, and frustration makes you leave. We actually just smooth them. By that I mean, if I could force you to sit there, I may be able to teach the material to you in five minutes, but it’d be a very frustrating five minutes. Instead, what we do is teach it to you in two hours — just way slower, but the whole time, things are animating on the screen and you’re getting dopamine hits or whatever. Even though a really good teacher could have taught it to you in five minutes, watching you make mistakes, it would have been frustrating. We much prefer to keep you around.

Part of the reason is because we’re in an app setting as opposed to a school setting. In a school setting, the truth is the kids are held hostage there. They can’t leave. With an app setting, the tiniest frustration, people are like, “You know what? I’m going to go to TikTok now.” We just can’t lose those users. So we always opt for engagement, but that doesn’t mean we won’t teach the material to you. We’ll just take it a little slower.

It’s clear that you have thought about this a lot.

We’ve spent years thinking about this.

I want to round this out a little bit because you have a very clear answer and a very clear point of view. What do you think the specific tension between gamification and education is? What are you losing when you always pick gamification?

Probably the thing you’re losing is efficiency — by that I mean, the amount of content learned per unit time. The truth of the matter is, I grew up in the third world a while ago. Some of the stuff that I grew up being taught, my teachers were hitting me while teaching it. I’m not kidding. They would hit me. The reality is that I probably learned really fast because when you were learning penmanship, if you did the wrong thing, they hit you with a ruler. You have a real incentive to get that done very fast. You just learn really fast because you’re like, “Whoa.” So I think it’s true: you can learn faster if you’re in an environment where you’re forced to do so and nobody cares whether you feel good about it. But in our case, I’m okay with slightly slower learning as long as you’re still engaged.

Efficiency, I understand that one. I had some pretty strict teachers in my time, but I was really good at taking the tests. My wife and I went to college together. She’s much smarter than me. 

But you’re a good test taker.

She was always mad at me because I could just show up at the end and take the test. This is truly probably why she didn’t date me for years, because of that core frustration. This is what I mean by education. That human teacher can evaluate whether you’re good at taking a test or whether you’ve actually learned something. That’s the tradeoff that I was getting at, is that if it’s all a gamification engagement, people might just learn to play the game. They might not have learned anything.

There’s probably a little bit of that. It’s very hard to measure, of course. But the reality is, ultimately, it works. Duolingo works. Just as an example, for me, I’ve been using only Duolingo to learn French for the past few years. At this point, I can watch Netflix shows in French — with no subtitles or anything. I just watch them, and it works. So you’re right, there’s probably a tradeoff. It’s probably pretty hard to measure. But what we’re looking for here is that people are actually using their time well.

I want to try to tie all of these themes and ideas together. You have a big vision for Duolingo. You’ve talked about it a lot. Being available to teach everybody languages around the world, being in a number of countries. And then there’s the fact that you’re a public company. You’ve got to make money. You’re still showing up in T-shirts. The first thing that comes immediately to my mind is, you’re launching new things like math and music, and they’re not available on Android, which is the single most popular operating system in the world. It’s used by the majority of lower-income people in the world. That feels like an immediate tension, that the best experience of Duolingo is on the platform that wealthier people use. How do you resolve that?

It’s a good point. First of all, math and music are about to be available on Android, or by the time this airs, they will be available on Android. We are about a year behind on Android. This has been true on Duolingo almost since the beginning. Android has been six months to a year behind iOS. There are a number of reasons for that, but probably the biggest one is that it has been harder to find really good Android developers when compared to iOS developers. There are just more really good iOS developers, so we have more of them at Duolingo.

The way we work is that we experiment with most new features on iOS. Because a new feature is usually not that great off the bat, you usually have to do trial and error to try to make it better. By the time it’s good, we port it to Android. That’s how we operate.

We understand the importance of Android. You are right. There are more people with Android phones than iPhones. Generally, all features are going to make it to Android, just about six months behind, and we feel okay about that. It’s also easier to develop on iOS for a number of reasons, not just that there are more developers. So that’s it, we’re just ahead.

In retrospect, given the technology that there is today, maybe we would be doing something that is cross-platform where we develop on all platforms at the same time, but we’re locked into being native on both ends. We have a native app for iPhones and we have a native app for Android phones. That was the best thing we could do 10 years ago, and we’re locked into that.

When you think about growing the company, supporting multiple platforms, that’s just double the effort all the time. Is it on your mind that, “Okay, we’re going to intentionally slow down development here so we can keep this team smaller”?

It is, and we don’t have one huge project where we’re going to stop all development and be like, “You know what? We’re going to now be in a single platform kind of thing.” But we are slowly getting there. I don’t know how long it’ll take. The hard part with this is that if we were to start from scratch right now, the decision would be clear, but you also have to keep the plane going. It’s such a big investment to do this that we will probably have to stop all development for a year and a half or something. I don’t even know the timeline because it’s just so big. So far, we have decided to do this piecemeal rather than all at once.

Your premium subscription tier, the Max tier, only just arrived on Android.

Literally in the last few days.

One thing I’ve heard over and over again since the dawn of the modern smartphone is that iOS users spend more money.

Yes, four times as much. At least for Duolingo, a given user spends four times as much per capita.

So the majority of your money is coming from iOS users, is what I’m getting from this.

Yes. More of our money comes from iOS than from Android. Even though we have more Android users than iOS users. It’s just hard to overcome that 4x.

Is that demographically that they have 4x the income? Or is it that demographically, iOS users are spending four times the money in your app?

No, it’s 4x the income. It’s that a user spends four times as much. We have more Android users, but it doesn’t balance out in the end. We make more money from iPhone. I’m going to give you a number here just to give you an idea. The split of users is 60 / 40, so 60 percent Android, 40 percent iOS. The split of revenue is the other way around. It’s basically 60 percent iOS, 40 percent Android. Those are very rough estimates.

So when you think about expansion, again, a public company, if most of your users are on Android and Android is the biggest operating system in the world but all of your money is on iOS, how do you resolve that tension?

Is it that we’re going to make our money on iOS and not Android? The big mission is to bring free language education to all of those other people, so is the iOS user subsidizing the free mission?

I mean, that’s one way of thinking about it. It’s not quite true that all our money is on iOS. It’s just that more money is on iOS, that’s for sure. But it is a little bit true, regardless of Android versus iOS, if you just look at who pays for Duolingo at the moment, they are usually people who are well off. They may not be millionaires, but they are people who live in countries like the US that are wealthy, countries that have salaries like $100,000 a year. A person with a good, stable job in a wealthy country, that is who pays for Duolingo. The people who use Duolingo for free are usually in poor countries. They may not have a stable job, so it is true that we’re getting the wealthy people to subsidize the education for everybody. That is the case, and that’ll probably always be the case.

Now, on our end, we also need to get better as a business to get more people in some of these developing countries to pay. As a good example, Netflix has done a really good job of getting people in Brazil and India to pay. We have not done as good of a job, and part of the issue is that we’re freemium. Again, I grew up in a poor country. Even if the price is scaled down to match the GDP per capita and it’s much cheaper, the problem that you have in a poor country is that the attitude is, “I won’t pay unless I have to.” That’s just the attitude. It doesn’t matter if it’s just a dollar, and I do happen to have a dollar. I just won’t pay unless I have to.

What you see is extreme tolerance for ads. For example, we can put 10 ads at the end of a lesson and they still won’t pay. This is why, for example, Netflix does so well in some of these countries because, with Netflix, there’s no free. They’re just like, “Look, whatever, you have to pay.” And people are like, “Fine, fine. I’ll pay.” So we have to figure out what to do as a freemium product in these countries, and we have some ideas, but the reality is, we have not really succeeded at strong monetization in countries like India. 

Do you think that that is the next logical place for you to grow as you think about English education?

For sure, we are spending a lot of effort on that. And it is growing, which is nice, but it’s a massive opportunity. Language learning is another funny thing where the largest market is a country like the US: rich countries. Language learning as a whole, not Duolingo, but language learning as a whole, the largest market is actually developing countries: the Indias, Vietnams, Brazils, and Mexicos of the world. They’re learning English, and that’s the largest language-learning market, but we have not cracked it. We have cracked the smaller one, which is the US and Western Europe and richer countries. We’ve cracked that in terms of monetization and in terms of users. We have a lot of users in India; they just don’t pay us.

I feel like I have to ask you about the owl. It’s very important to everyone that I ask you about the owl. At least as expressed in this country, the owl is a very online, very culturally defined character. If you took the owl outside of United States social networks and dropped it anywhere else, it wouldn’t make any sense. Is the owl expressed culturally in different markets, or is it just one owl?

I don’t know how to answer the question. It’s in between. We started using social media with the owl a while ago. It grew mostly in the US through TikTok because the owl does unhinged stuff on TikTok.

Wait, hold on. The owl doesn’t do anything. How big is the team that writes and performs the owl?

Okay. And they work at Duolingo?

I’m assuming they’re in New York City? 

Actually, no. Most of them are in Pittsburgh.

Okay. I didn’t realize Pittsburgh had this many terminally online people. Godspeed.

Yes. But it started out with TikTok and it was mainly in the US. That was several years ago. What has changed in that time is, first of all, we are no longer just relying on TikTok. It is now on YouTube, YouTube Shorts, Instagram, etc. So it’s everywhere on social media. That’s one big thing. The other one is that we learned how to localize this to different markets. So we started Duolingo accounts for a bunch of countries: Mexico — well, Spanish speaking — Japan, Brazil, Germany, France, China, etc. We have figured out how to make all of them succeed. I was dubious at first when somebody told me we were going to open an account in Germany. No offense to Germans, but I thought, “These people don’t have a sense of humor.”

They do! In fact, it’s one of our more successful accounts. [The international accounts] are a little different. It’s not that different, but they are a little different.

And these are the same five people locally?

No, we have a global team, which is these five people, and in each country, we have a small number of people, probably one or two people, that localize this stuff. And “localize” doesn’t mean we take the exact same videos and in Mexico put a sombrero on. That’s not that. We have themes and we have figured out what themes work globally and also what themes work in certain countries. For example, the German one, we had a really big thing on Oktoberfest. Also, at some point, because there’s this dance club scene in Berlin, I guess they all went to one of those 24-hour dance clubs. Each country does different stuff, and it’s worked out pretty well.

What’s the hiring process like to be the writer for the Duolingo owl? Do you just read people’s Twitter accounts and say, “You’re unhinged enough to do this”?

It’s a lot of that by now because we are such a presence online. By the way, I didn’t know this until recently but there are weeks when our video on TikTok is the most watched video on all of TikTok that week. That happens. By now, our accounts are so well known that we have our pick in terms of [recruitment], and a lot of people want to work for that team. Typically, we just look at what they’ve done before. It’s a small group of really good creators, so we hire from that group. Usually, these are pretty funny people that are even funnier online, but when they’re offline, they’re not as funny. They’re still funny, but when they’re offline, you’re like, “It’s you? It’s you who came up with that?”

And you measure everything, it sounds like. Is this working? Are you getting lots of new users because of the owl?

Yes, this works. By the way, this is not paid. All that social marketing is not paid. It’s free. We make our videos and they go viral. About 15 percent of our users are coming in from social media. Now, if you look at social media views of our content, which we measure in the billions, there’s a roughly equal number of social media views of our content versus the content that is about us but not made by us. Also, there are a lot of other people just making content about Duolingo, but they’re not us, and they combined have about as many views as we do.

Have you ever told the team to pump the brakes? Have you ever looked at something they’ve made and said, “We just can’t do this”?

Yeah. There’s our review. There’s an approval process. We’re close to the line in some of the stuff that we have put out, and we have in fact gone across the line and published things that we shouldn’t have. Since we did that, we now have a pretty strict approval process. This is a whole layer, and the last step is basically me. But stuff doesn’t come to me often because there are people before me, like the CMO, so there are a lot of steps. 

What’s the last one the CMO was like, “I don’t know. Luis has to approve this one”?

I’m trying to remember what that one was.

What’s the last one they convinced you to do even though you were skeptical?

I don’t remember the exact video, what it was, but I know that the last one that I approved, I was wrong, as in, I shouldn’t have approved it. A lot of this you only know in retrospect. You don’t know until it happens because you put it out and then you see this reaction. I don’t remember what it was, but I know I approved it and I know I was wrong because I didn’t imagine that it was going to have that reaction. We haven’t had that many faux pas. It’s been like three or four videos that were like, “You probably shouldn’t have done that.”

The other thing happened about a year ago. We had made this crazy video. It was insane. We were a little hesitant about it, and we ended up cutting it. There are all these memes online about how the owl really wants you to learn a language, and it goes to great lengths, including kidnapping your family. This was a video about kidnapping, and we were a little hesitant about it, and then the October 7th attack happened, and we cut it. And then we cut it last year, and we thought, “Well, you know what? We may use it next year.” This year came along and again we cut it. And then we came up with an internal thing that a year when we can play that, it’s probably been a good year for humanity.

Yeah. We’re probably never going to play that.

The world context of that one needs to be substantially improved, I think. All right, I have to end with a feature request. You’ve given us a lot of time, and then I’ll let you get out of here. We talked a lot about India. We talked a lot about emerging languages. Can you put Gujarātī in this app?

This is the language that I can understand and speak like a baby, but I can’t read or write it, and I would love to close the loop.

You’re asking for languages. That’s a hard one.

It’s the native language of Gandhi, of the current prime minister of India.

There is this unfortunate thing about being a huge language versus the desire to learn it. It’s a pretty big difference. Hindi is probably the one that has the most desire to learn it in terms of Indian languages. It’s a tiny number of people who are learning it. It’s got to be, I don’t know exactly off the top of my head, but it’s certainly well below 1 percent of our learners are learning Hindi. I’m going to guess 0.1 percent of our learners are learning Hindi. That’s the hard part about adding languages, that we have to maintain them, we have to do a really good job with them, and then, in the end, we just don’t get a lot of usage. So, sorry.

All right. That’s a hard no. It’s one of the first times a CEO has given me a hard no. That’s again, founder mode.

Well, it’s just really hard to say yes to. In the past, I used to say yes to this stuff, and we made a lot of mistakes adding languages that, in retrospect, we probably shouldn’t have added.

Have you ever cut languages?

We have. We cut, what was it? I think it was Afrikaans. But the cut in part was because there was very little demand. The biggest reason was it was just a low-quality course, and at some point, we thought this was a bigger brand risk than anything else. We made the decision, we’re like, “Well, could we improve it or what?” And we made the decision it was not worth improving.

Do you think AI is going to help you add languages?

Maybe, but unfortunately, at the moment, AI is really good for big languages and really bad for smaller languages. There’s a pretty high correlation with the languages we offer. AI is very good at the languages we have: the Spanish and the French. It’s not super good at your Esperanto or Navajo or smaller languages.

AI is notoriously bad at math, or at least the current LLMs are pretty bad at math. Are they going to help you with that?

The good news is that, in the constrained environment that we have, it can help quite a bit. It’s been helping quite a bit. A lot of the data that we generate for our math course is with AI. The other thing is, some of it is without AI, but it turns out, just computers are good at math.

It’s funny how many times I’ve asked this question and someone fails to bring up the idea that there’s a computer. I’m happy you did that.

Computers are good at math! And I understand AI is not so good at following a pattern or whatever. It may not be so good at that. But the data that we generate for our math course is a lot of stuff like fractions and multiplication. Computers are pretty good at generating that data.

All right. Well, Luis, you’ve given us a lot of time. Thank you so much for being on Decoder.

Thank you for having me. And great questions.