No Priors Ep. 114 | With Duolingo CEO Luis von Ahn

Today we're joined by Luis von Ahn. Luis earned his PhD in computer science from Carnegie Mellon and went on to found reCAPTCHA, which was acquired by Google in 2009. He's now the co-founder and CEO of Duolingo, the world's most popular education app, with over 116 million monthly users, a market cap of $17 billion, and an owl mascot that faked its own death.

We're going to talk about AI for education, why motivation is the hardest problem in learning, taking risks with your company brand, why vibe cartooning is important, and the 16,000 A/B tests that got us here. Luis, thank you so much for doing this.

Thank you for having me.

Lots of people know what Duolingo is, but I would love to hear you describe it and, beyond the language-learning app it is today, what you want it to become.

Well, it's a language-learning app. It's the most popular way to learn languages in the world. As of the last couple of years, we also teach math and music, and very soon we will also teach chess.

The idea is that we're trying to be an app where you can go and learn the things that a lot of people want to learn but that also take a long time to learn.

You were a professor when you started Duolingo in 2011. I hope it is not offensive to say that lots of professors start companies, but few of them start gamified consumer companies. How did this happen?

It's not like I expected to start a gamified company. The way we got started is that I was a professor. I had a PhD student named Severin, who is now the CTO and was a co-founder.

We were looking for a PhD thesis topic for him, and what we agreed on is that we were going to work on something related to education where computers would teach you something. After a while, we agreed that a good topic to teach was languages, in particular English, because in most countries in the world, knowledge of English increases your income potential.

There are about 2 billion people in the world learning English. So we thought, okay, let's teach languages and let's teach them with a computer.

I made the first Spanish course because I'm a native Spanish speaker, and Severin is a native German speaker, so he made the first German course. We agreed that we were going to learn each other's language.

The problem that we ran into was that we couldn't get ourselves to do it because it was so boring. I was like, “Oh my God, I did not want to learn German. He did not want to learn Spanish.”

We were very worried because we thought, “Okay, if we can't get ourselves to do it, then we can't expect anybody else to do it.” The solution was to turn it into a game as much as possible, and by the time we launched, it was pretty fun.

It's mainly because we were trying to get ourselves to do it. Part of the reason that was the case is that neither of us likes learning languages. We actually are not language lovers; we don't like learning languages.

I think because of that, we made a product that works for the average person, as opposed to people who are obsessed with learning languages.

There's so much there that I would like to unpack, but I want to go back to my initial confusion. What makes this a good PhD topic? What was the computer science problem you were interested in?

The computer science problem was basically trying to teach things to people.

Is that a computer science problem?

Well, how to get computers to do it? There's a lot there: how to do adaptive learning, how to use the data to adapt to the student. There's also motivation problems and human-computer interaction. There was a lot there that we could have used.

By the way, at the time, AI was just starting to get to the point where we could have thought about training models—and this is not large language models, just training classifiers—to teach better or something like that. So there was a lot there that could have been a PhD thesis topic.

You've decided that you need to make it easier for normal humans who do not have infinite motivation to learn languages from computers. What was the first thing that you did that worked, that helped people get over the motivation hurdle of it being so boring?

The first thing that we did that worked was make lessons not 30 minutes long but 2 minutes long. It makes a big difference.

It's not because people were spending less time; it's that 2 minutes is amazing because at any point in time, you just think to yourself, “Oh, it's just 2 minutes.” But the time investment for 30 minutes—if I ask you right now, “Do you want to do something that takes 30 minutes?”—you're like, “Oh man, I don't know about that.”

Even though the time investment is only going to be 2 minutes, you may spend the 30 minutes, but it's just like, “Yeah, I can start right now because I can end any time.” So that's the first thing that worked.

We did many other things that worked, but that was the first thing that really was a big game changer.

Is there anything that really surprised you from a behavior perspective that worked?

A few things. I did not expect streaks to be this powerful. A streak is extremely powerful.

Can you explain what that is?

It's just a counter that measures the number of days you've done something in a row. For Duolingo, the Duolingo streak is the number of days that you've used Duolingo in a row to learn a language.

It's incredibly powerful. People really talk about their streaks. We have 10 million active users who have a streak longer than 365 days, so that means they haven't missed a day in the last year or longer.

I did not expect streaks to be this powerful. The other one that I did not expect was these notifications to get you to come back to the product.

We were sending you notifications that just said something like, “Hey, come back to learn Spanish.” If you did not come back the first day, we would send you a notification. If you did not come back the second day, we would send you a notification. But after 5 days of inactivity, we stopped.

At some point, it occurred to me that if we're stopping sending you a notification, we should probably let you know. So we started sending a notification on the fifth day that said, “These reminders don't seem to be working. We're going to stop sending them for now.”

It turns out this is very powerful at getting people to come back because they feel like we've given up on them. I did not expect that. We didn't even do that with the goal of getting people to come back, but it turns out that's actually very powerful.

I have not studied education, so this doesn't come from a place of knowledge, but this idea of making it easy and making the barrier to entry lower—I feel like there's some controversy around it.

It raises questions like, “Okay, learning is hard, and there is no way around it.” A smart friend of mine compared learning to going to the gym, but for your mind. It's very simple: you just have to do it, and there's no other way besides being willing to put in the effort and think really hard about it.

How do you react to that point of view?

I would be very, very happy if everybody who is working on an education app thought that way, because then we would have no competitors. It just turns out that the hardest thing about learning is motivation.

By the way, I believe that to be true of the gym, too. A lot of people talk about how you can spend all your time debating whether an elliptical machine is better than a treadmill, but in the end, what matters is that you do it. That's 90% of what matters.

It's the same with learning something. You can spend all your time debating whether you should read that book or this book, or whether you should do it through an app or with a tutor. There are, of course, differences in how effective each of these are, but what matters most is that you actually do it.

If you want to get people to actually do it, you have to make it as easy as possible to get started, to get in there, and to be motivated. I believe that is the reason why we have grown so much.

In fact, it's funny. We see a lot of language-learning apps that pop up and say, “Oh, we're like Duolingo, but without the gamification.” Every time I see that, I'm like, “Great, you do that. Carry on.”

Ninety-nine percent of the world's population just is not that motivated for any activity. There are the few people who are extremely motivated. Good for them. The vast majority of people are just not.

If you are an engineer or, in many other types of knowledge work, the idea of flow is almost holy, right? You need to be in the context and understand the codebase and really be thinking about it for a while, or you won't make any progress, much less learn something new.

Do you think about that at all? Because if people are like, “I'm on the subway; I'm going to do Duolingo for 2 minutes,” how does that factor into the experience?

We don't really think about that too much. I understand the concept and it makes sense, but for us, the single most important thing is to get people time on the app.

It turns out that for an English speaker, getting to a pretty good spot in Spanish takes about 500 hours.

Oh wow. Okay.

We just got to clock those 500 hours. We just got to clock those. That's it.

Is it 600 hours because you're not as in flow, or 400 hours because you're really in flow? Sure, there's probably variance there, but in the end, we have to clock something like 500 hours for you to get to a good spot in Spanish.

And by the way, that same number for Chinese is 2,000 hours.

Yes. As somebody with a master's degree in Chinese literature, it's not a language meant to be learned. It really doesn't seem to be.

I find your perspective on human behavior around education both a little bit dark—not particularly idealistic—and yet incredibly inspiring to me as an end user. I am working, I'm engaged in the work, I have 3 kids, and I theoretically do some other things. Yet the thing I would be most interested in doing with my free time is skill acquisition: becoming a better chess player, improving my Chinese, as fruitless as that seems, or learning more math—many different things.

And yet, if you ask me, “Are you going to carve out an hour and a half at the end of a day where you've been attempting to be productive for 16 hours and do more of that in a really painful way?” the answer is clearly no, right?

Right.

And yet, if you ask me, “Could you imagine putting 600 hours into it, 2 minutes at a time over the next few years?” Actually, that seems doable.

After working enough in consumer products, you realize a number of things. People don't read; they just don't. People are lazy. If given the choice, they may tell you all kinds of things, but the reality is that they'd rather scroll on Instagram or TikTok. That's just reality.

The thing that is positive in all of this is that you can take some of that behavior and do something useful with it. Anyway, perhaps that's what we're trying to do with Duolingo. That's what we're trying to do: use a lot of the same tricks that are used to keep people engaged with mobile games or whatever it is, but to get them to learn something. That's really at the crux of what we do at Duolingo.

Duolingo has a pretty unique position in that you were a company with some scale when the capabilities in machine learning started accelerating in their progress and expanding. How did you handle that, and how did you react internally as a product organization?

For us, it's been really positive. The fact that they're language models and we teach languages makes it a really perfect application. They've been really positive in 2 ways.

One of the things you have to do to teach something is create a lot of content, because it's the content that you're trying to teach. We used to make that content half by hand and half automatically. It was kind of by hand, but there was all kinds of automation around it to make the people creating the content go faster. Still, this was going relatively slowly.

Over the last couple of years, what we did is retool this whole content-creation pipeline to be entirely based on large language models. It's all based on AI, so humans are no longer involved, or very little, if at all. What that has allowed us to do is create massive amounts of content that were just not possible before. We're creating many more courses.

For example, that's one of the things that we're doing right now—we're launching it right now. We used to teach 40 languages, but the way we teach languages is that we teach Spanish for English speakers, which is a different course from Spanish for German speakers, which is a different course from Spanish for Chinese speakers, and so on.

We used to teach 40 languages for English speakers. So, if you were a native German speaker, you could only learn 4 of them. What we're able to do with AI now is go so much faster. At this point, we basically teach the 40 languages to every base language. That's a major increase in the number of courses.

The other example of something we can do now that we couldn't do before is practice conversation. Historically, we could teach you vocabulary and how to read, but for actual conversation, the only way we knew how to practice it was with another human.

We experimented with doing that, but it turns out most people don't want to talk to another human in a language that they're not very comfortable with because it feels bad. It feels bad. It gives you shame. It's not good.

But now, with large language models, you can actually practice conversation with an AI, and you do so without feeling judged. We're seeing a lot of uptake on that. It's been really good for us.

What else are you excited about in terms of AI changing the product itself—things like Roleplay and Explain My Answer? What do you think matters or works so far?

I'm excited about generally practicing real-world language, like conversation. I'm excited about that.

Outside of language, I'm really excited about teaching math. We can do a much better job at teaching math with large language models. We have a math course, but we're completely retooling it to be a lot more like a tutor.

Again, I'll go back to tutors. Tutors can be really good for learning outcomes. They have this one big problem: they're really boring. So, we're trying to find ways to make it a tutor, so to speak, but also gamified.

It turns out most people would rather play Candy Crush than sit there in front of a tutor. We're trying to come up with something that is as effective as a tutor, but as fun as Candy Crush. The reality is we'll probably come up with something that is 90% as effective as a tutor and 90% as fun as Candy Crush, but at least the combination of the two will be a lot more effective.

So, math, music, chess?

Chess now.

Chess now. Yes. How do you decide what Duolingo can teach beyond languages?

The first thing we're looking at for subjects to teach is a very large audience. We need something that hundreds of millions of people want to learn. There's a reason for that: we're an app at the end of the day, and we cannot charge $30,000 to each user. We charge you 10 bucks, so in order to make a significant amount of money from a new subject, it has to be learned by a lot of people.

Otherwise, these very niche subjects just aren't worth our time. If we charge you 10 bucks and there are only 100 people learning it, it's not worth our time. They have to have a large potential audience.

The other thing is that we look for things that take a long time to learn. If you can learn something in 2 hours, you probably should just go watch a YouTube video, and that does a perfectly good job. We look for things that really take hundreds of hours to learn.

Then, we look for things that we think are good for the world and that we can do a good job with in a mobile app. There's one extra thing: internally, we need to have somebody who's really excited about it. This has been true for all of these other subjects. There's somebody who's very excited about music, somebody who's very excited about chess, and that's what has made it happen.

Even beyond what makes sense for Duolingo as a business, from a technology perspective, do you think there are things that are hard to teach with computers for now?

I don't know if there are things that are hard to teach with computers for now, but there are things that are certainly different from what we do at Duolingo. For example, history—the way you teach history is not with these drills. Duolingo is really good at drills. The way you teach history is probably with really well-produced videos. That's probably the best way to do it.

Maybe AI can get better at that, but I think there are just things that are different from Duolingo. Ultimately, I'm not sure that there's anything that computers can't really teach you. I think overall, we'll be able to teach everything really well with computers.

Do you think there's any version of the world where AI is a threat? Are large language models a threat to Duolingo?

Sure. That's one of the things that's scary about the world that we live in with AI and large language models. We're undergoing a platform shift of some sort. I don't know what's going to happen on the other side. I'm not super worried, but you just never know.

It's not just Duolingo. It could be all kinds of things. It could be a threat to Netflix. It could be that a large language model just presses a button and makes you the perfect movie. I don't know if that will happen or not. I just don't know.

It's a similar thing: who knows what will happen? At the moment, it doesn't look like it's a major threat, but your guess is as good as mine.

I think you guys have a huge amount of distribution, a huge amount of data, and an engineering culture to go after it.

The way we see it is that we have large distribution. We have data on how people are learning languages that is unique. We have a lot of people. It's also the case that brand ends up mattering quite a bit, and we do have a good brand for language learning. We're hoping that the combination of all of this will be good for us.

Can we talk about brand for a minute? Duolingo has a very unique one. There is perhaps more risk and more distinctiveness in the brand voice than the vast majority of public companies might take. You are, of course, a consumer company, but there are plenty of companies that don't take those risks. Where does that come from for you guys? Why did the owl die?

He didn't die. He just faked his death.

Okay, that's even weirder.

It's not like on day one we decided, “We're going to have a brand that's unhinged, where the owl does weird stuff.” It just kind of evolved over time. Our brand voice evolved over time.

What happened was, in the product, the owl was a little pushy to get you to do your lessons. Then the internet just started coming up with memes about the owl doing crazy stuff. They started coming up with the owl being willing to kidnap your family to get you to do a lesson. That was all invented by the internet, not by us. But as we saw that was happening, we started leaning into it. We thought, “Why not?” And the more we leaned into it, the better it worked. So we just found something that resonated. That got us going.

A few years ago, we had a very junior marketing employee that we had just hired who said, “Hey, we have this old owl suit here that we had used for events inside the company. Can I make some TikToks out of it?” And I actually was against it. I was like, “I don't know if anybody's going to be interested in this.”

It turned out that when we put it on TikTok, the owl was really just doing dumb stuff, like twerking and falling down or whatever. A lot of these videos started going viral. It was interesting because none of the videos said, “Learn a language on Duolingo” or “Subscribe to Duolingo” or anything. It really was just the owl doing weird stuff. It just took on a life of its own.

We formalized a lot of it, and we have a pretty good idea of the types of risks we're willing to take. But we do take more risks than most companies, certainly than most public companies. One thing that I think helps us is that we have a mascot, but I think it also helps us that we're an education company because, ultimately, you can't really say that education is bad. It's hard to argue against education.

That allows our marketing to do things where we can't be criticized that much. If we were a company that was trying to just really extract your money, or if we were a vaping company or something, people would probably push back more on our marketing. But given that we're an education company, I think it gives us a little more leeway.

Given this unique scale of Duolingo, I think it's more than 100 million people monthly. Is that right?

Yeah.

What do you know about how people learn that other humans don't that you think is interesting?

A lot of it is just codified in what the algorithms have adapted to. We have adaptive algorithms that try to figure out how to teach, and I think it's hard to verbalize a lot of it. We know a few things that you can verbalize that may not be that surprising: the farther your language is from your native language, the harder it is to learn.

We have a model that can predict whether you're going to get an exercise right or wrong, and we're actually extremely accurate. When we give you an exercise, we know if you're going to get it right or wrong. We're very accurate. The way we do that is we just watch everything you're doing, and we see what you're good at and what you're bad at.

For example, we know, for each user, “This particular user is bad at the past tense.” One of the things you may think at first is that the right thing to do is, because this user is bad at the past tense, we should give them more past tense. That's roughly true, but it's not exactly true, because if all we did was give you lessons of things that you're bad at, these would be very, very horrible lessons for you.

So, there's a whole—I was going to say art, but it's actually more of a science. There's more of a science about when to give you the things that you're bad at. For example, whenever we give you an exercise, the right thing to do is to give you an exercise that you have about an 83% chance of getting correct. It turns out that maximizes enjoyment, and maximum enjoyment means I'm going to get to my 500 or 600 hours, and that's really the outcome.

Exactly. And it just keeps you motivated.

It's funny: it's not 100%, because 100% is too easy. But it's also not 50%. It really is closer to 100%. You have to mostly win, and that seems to be what works.

What do you think are some of the implications of what you guys are doing or learning for traditional education and schooling?

First of all, I do think that education is going to change over the next several years. I can't say one year, but it's probably less than 20 years. Something's going to change, and the reason for that is it's just a lot more scalable to teach with AI than with teachers.

By the way, that doesn't mean the teachers are going to go away. You still need people to take care of the students, and I also don't think schools are going to go away because you still need child care.

In your view, schools could be child care, but everybody's Duolingoing?

I think it's going to be something like that. It may not be Duolingo, but I think it's going to be something where there's 1 teacher and 30 students. Each teacher cannot give individualized attention to each student, but the computer can. The computer can actually know with very precise knowledge what this 1 student is good at and bad at—knowledge that the teacher just has no chance of having because there are 30 students, and they just cannot give you that.

I do think that it would be more effective if some of that time were spent with essentially an AI teaching you. I do think it's going to get to that. It's also the case that there are extremely good teachers, for sure, but there aren't very many of them, and certainly most everybody in the world doesn't have access to a good one. So, I think there's going to be some change like that.

I do think it's going to take a while because changes in education are very slow. But I do think it's going to be like that. I think a lot of what we're doing will apply in terms of keeping people motivated and so on.

I do think that in a formal education setting, some things do need to be different. You have a little more control over the students actually doing this stuff when they're in school. Instead of expecting that they're only going to do 2 minutes, you can probably expect that they're going to do 20. So, there are some differences in the school setting, but I think a lot of what we're learning here will apply.

Is there anything you feel like people misunderstand about Duolingo?

When we became a publicly traded company, investors thought we were a COVID phenomenon. Turns out we were not. The biggest misunderstanding is how important motivation is. I really don't think people understand—I mean, even our competitors, not just competitors in language learning, but people who run education companies. It is amazing to me that most everybody we talk to does not seem to understand how important motivation is. I think that's one thing.

I think another one is just how much sophistication there is. A consumer may not know because the app is so cutesy, with little animations and everything, but there's a lot of sophistication about when to give you even the animation. There's a lot of sophistication about that.

I don't think people understand that Duolingo is the result—I checked the other day—of having run 16,000 A/B tests over the history of the company. I don't think people understand that it has taken 16,000 A/B tests to get to this point. I think that's probably the biggest misunderstanding.

Any predictions about what the large-scale changes in learning mean for society, or even how you might encourage people to think about learning for themselves or their children?

The good news—I don't know if it's good or bad news, maybe probably bad news for Duolingo but good news—is that this change is going to be slow. I don't think you're going to see a change where, next year, everybody's learning completely differently. So, I think we'll have some time to adapt.

What's the drag force? I feel like that's actually a controversial point of view, that it's going to be slow.

Go to a real school. They're doing stuff from 30 years ago. The drag is just—it's like government. It's just slow.

If you live in New York or Silicon Valley, you'll see that there are schools that are doing really progressive stuff. But the vast majority of the education system is very slow to make changes like this. School systems are regulated. In Texas, they're trying not to teach evolution. There's just weird stuff that goes on in school systems. So, I think in general, it's going to take some time to do this.

You probably will see some private schools move there faster. There's also an interesting thing about private schools—the very fancy private schools. You pay $50,000 a year to go to a private school, and it's hard for them to say, “Well, what our kids do is use Duolingo,” because it's like, “Well, why am I paying you $50,000?” So, it's an interesting dynamic that's going to happen.

But I do think that some of the private schools are probably going to be the first ones to start really moving toward this. I think you'll just see a lot better learning outcomes in general. The other place where you may start seeing this is that there may be some countries that leapfrog—some countries that are, at the moment, probably a little behind.

You know, for them, this is the only way in which they can scale their education, so you may see that.

Yeah, I tend to think very tactically also about the impact on the entrepreneurial ecosystem. Maybe it was all about finding the idea, but you studied math and then CS, got a PhD, were a professor, and had done some body of work before you became an entrepreneur, including starting another company—actually, a really interesting one, reCAPTCHA. But one of the things that is interesting to me, even, let's say, 15 years into the technology ecosystem in my own career, is that people are becoming experts at much younger ages.

Oh, yeah. I think, obviously, if you can set yourself on a learning journey—first it was just because of the internet, right? You can learn from content and forums and finding community and whatever else—but when I think about some of the dropouts that we work with who are like, “Well, I wrote a textbook on wireless technology as a sophomore,” and I'm like, “Man, I waitressed at Outback Steakhouse when I was 16.”

Right. One of the things that I think would be really exciting as we look forward 5 or 10 years is that, if people can get to skill acquisition and learning outcomes that are much cheaper, broadly accessible, and easier to motivate, you will get experts who are doing really interesting things by the time they're 15 or 20.

Yes. And not only will that be true, but I think they'll be experts and they'll have the tool of AI. So I think it'll be the case, at least in the foreseeable future, that that'll just magnify their expertise. So they're experts, and in addition to that, they can go a lot faster.

Compared to waitressing at Outback Steakhouse or whatever dumb thing it is that I was doing, I did not have access to the internet. The first time I had access to the internet, I must have been 15 years old. I don't know what I was doing before that, but I think I was playing with Legos. I mean, that's basically what I was doing. So, yeah, I think we're going to see a lot of that.

So it sounds like course authoring is a place that Duolingo's gotten a lot of leverage. When you think about just using AI broadly as a tool within your company, be it building product or everything else you do, where else do you think it can have the most impact?

The visual style of Duolingo. We're very—I mean, unlike most—well, I mean, a lot of apps are kind of well-designed, that's for sure, but we're very animated. We're like a cartoon. Making that has required a lot of human effort. It turns out we can do a lot of that with computers now.

That doesn't mean we're not going to employ the artists. The artists are still here, but they're working so much faster. What one artist could do before in a month, now they can do in a day. It's really unleashing their creativity because they're not spending their time on the mechanics of, “Is this shadow just right?” No, they're just unleashing their creativity. And I think that's pretty awesome.

And so we're seeing a lot of that. A lot of our animations and a lot of illustrations are now computer-generated.

That's not what I expected you to say, but I am really thrilled for the future of math and chess and 140-some new languages and vibe cartooning for Duolingo. It'll be great.

Yes.

Thank you so much for doing this, Luis.

Yeah, thank you. Thank you, Sarah.