Final Boss: A talk with Duolingo’s Luis von Ahn

Co-inventor of the Captcha. MacArthur genius grant awardee. His elegant crowdsource-assisted digitization of the New York Times 150-year archives became the basis for a technology he ultimately sold to Google in 2009. (The undisclosed sum he received for the sale, according to a 2023 New Yorker profile, “was sufficient to insure that neither he nor his future children would ever need to work.”)

In 2021, Luis von Ahn, 46, added a new title: Chief executive of a publicly-traded company. And shares of Duolingo, the language-learning app he co-founded in 2011, have been on tear lately. So, we asked for a few minutes of his time last week to talk over the growth outlook for the company. The following conversation has been edited for concision and clarity.

Sherwood: The stock has surged over the last month or so. I know CEOs don't like talking about short-term stock fluctuations in general, but, given your academic background in crowdsourcing do you think that gives you a different view on interpreting market moves?

Luis von Ahn: The day-to-day moves of the stock are extremely random to me. In the long term, I'm hoping that because our company is doing well, that the stock would do well. But on a given day, we cannot explain it.

Sherwood: It seems like a lot of the excitement is AI-related to some extent.

Von Ahn: If you ask me, I think it is partly AI. I think it's also the case that our growth rates, of both revenues and users, are good. I mean, most consumer companies at this moment don't have the kind of growth rates we have. I think the other thing is that, generally, investors have come to trust that will do what we say. I mean, so far we've been quite good at forecasting and giving guidance. So I think the combo of these things has just made it so that we're, you know, a good stock.

Sherwood: I looked at your revenues. It's like the platonic ideal of a sales chart. Even aside from the angle of incline, the steadiness is fairly remarkable. It’s been such a volatile time, you know, for the economy over the last four years since Duolingo came of age as a public company. How do you account for that steadiness?

Von Ahn: What's interesting, this was true even before we were a public company. If you look at whatever metric you want to look at, revenues, daily active users, whatever, we never had explosive growth. We never had that. But we've had this continuous growth that’s just monotonic for the last ten years. It’s just every year is bigger than the previous one.

Sherwood: Can you talk a little bit on the AI opportunity for Duolingo?

Von Ahn: We're very excited by it. There are two places where AI is quite useful. The first one is in generation of our content. So when when we do teaching, some of the content that we teach with, historically, that's been kind of half done by humans, half done by automation. But with large language models, that can be almost entirely done by computers. What that does is that saves us money. That's good. But the more interesting part is that it allows us to go a lot faster and therefore do things that were kind of not feasible before.

So let me explain what I mean by that. About five years ago somebody proposed this new lesson type where you would get to listen to a, kind of, animated little story. I asked how long it was going to take to make the data, because that data was going to all be made by hand. They said it was going to take five years. And at that point I told them, don't do this project.

Sherwood: By “make the data” you mean, essentially, make the stories that would be included in the lesson?

Von Ahn: It was basically making hours and hours and hours of cartoons in all our languages for all of our proficiency levels. So it’s like basically making cartoons. It was going to take five years to make all of this. And I'm like, “Okay, I don't do that.” That’s just crazy. About a year ago, same exact project, we figured out how to do it with large language models, and the estimate was that it was going to take three months. As we did it, we just did it. So it's this amazing thing that just kind of allows you to do things you couldn't do before.

The other place that I think is a really big thing for us is being able to practice live conversation. One of the aspects of learning a language is learning conversation. And historically, we just couldn't teach conversation very well with a computer. It required a human and we never really got into putting humans in there for many reasons.

One of them is that they're expensive, but another one is that people actually don't want to talk to humans when their language skills are not very good. But now with large language models, you can actually mimic conversation with somebody and you can use that to practice. It's really, really good. So we’re launching this this thing where you have basically a video call with one of our characters. It has a lot of really good things about it, the first one is that people are not embarrassed to talk to an animated character as opposed to a human.

This feature, we have to put it behind a paywall. Most of our features are really accessible, but this one, we have to put it behind a paywall because it costs us money because you have to query OpenAI. But what's really nice is, I really do think that this will get a bunch more people to subscribe, and I like that.

Sherwood: That brings me to a question, I’ve looked at your R&D spend over the last few years and I think a few years back it was around $10 million a quarter, but the last quarter was somewhere around $50 million. So is that the way you've structured the business to try to offset some of that, by charging for it?

Von Ahn: What we're saying is over the long term, we're going to get efficiencies on pretty much everything except for R&D. It's not that R&D is going to become a larger percentage of our revenue, but it's going to probably stay around the same as a percentage of our revenue. All the other things, you know, marketing, G&A, we're going to get efficiencies. R&D, I mean, that's basically what we spend most of our effort on, just making the product a lot better.

Sherwood: Okay, you have this incredible niche in the language learning space, but will language learning itself become obsolete as a result of AI?

Von Ahn: People have brought this up, you know. I don't buy it. If you look at our users, the majority of them fall into one of two big buckets. The first bucket is people who are learning language as a hobby. And for them, it kind of doesn't matter whether a computer can do it or not. It's a hobby. It's like learning chess. Computers can play chess, but people still learn chess.

And then the other big group of people are people who are learning English. And that group of people just need to learn English to go to a university or to get a job or to move to an English speaking country or whatever. So neither of these groups is going to be affected by that, I don't think. And so, you know, we are just not particularly worried about this.

Sherwood: The way your business is structured, it reminds me almost of Netflix in the sense that you’ve got a cheaper tier with advertising and a premium tier without advertising. Is there an opportunity to pull the lever to boost some of the advertising a bit to generate higher value?

Von Ahn: So there's there's definitely opportunity in ads for us. I mean, our ads, we're not very sophisticated with ads. It's a small fraction of our business as well. You know, check the numbers. But it's somewhere like 6% or 7% of our business is ads.

We are a subscription business. We like subscriptions a lot more, not only because we have a direct relationship with the consumer. But it's also extremely predictable. Most people subscribe for a year, as opposed to a month-by-month basis. It's also not susceptible to weird stuff that ads are susceptible to, like, Apple changed some privacy things so now your all your ad prices go down. Also ads seem really susceptible to the economy, whereas subscriptions have just been a lot more robust. So we much prefer subscriptions.

Sherwood: Can you talk about one part of the business I don’t think is particularly well understood: the Duolingo English Test?

Von Ahn: Yes, it's not as well known. Particularly in the United States, because in the United States, most people don't need to take an English test. But it turns out that English testing is a pretty big business in the world. So we started this test called the Duolingo English Test about ten years ago. The big innovation — at the time it was considered to be crazy — is that you could take our test online. It turns out we can actually make it so that people can take the test online securely.

But then at the beginning of Covid, what happened was that the testing centers for offline standardized tests all closed because of Covid. And so those tests could not be delivered and then universities started to freak out because they were like, “We still need international students.” So they went around and they looked and it turned out there was this one test, which was our test.

It's only 10% of our business. But what I love about this test is that it's strategically important. And let me tell you why. So the test has a score from 0 to 160, and that score means something. So, for example, the score that is needed to get into Stanford is 120.

We want to get to the point where we own the score by which people say how much they know a language. So right now, when you ask somebody, “How much French do you know?” the the most common answer is, “I took four years of French in high school,” which doesn't really answer your question. What we want is for people to say, “I am a Duolingo 65.” And we want that to be understood.

The way we're going to get there is we’re starting to put in the app the score of how well you know a language. It’s not there for all languages, but it is there for some of them. So that will give us literally more than 100 million people who have this score. And then on the flip side, because this score will be the same as the test score, the test will give the score legitimacy. It means something. It is like, look, 120 on this is good enough to get into Stanford.

The combination of legitimacy and large scale, which is what the app gives us, I'm hoping that over the next few years we'll be able to own this proficiency score. And that's a very defensible position.

I mean, if you happen to own the score, it's very hard to dethrone you.