Duolingo CEO Luis von Ahn wants you addicted to learning

Today, I’m talking with Luis von Ahn, the cofounder and CEO of Duolingo, the fashionable app that teaches languages. It’s an absorbing clip to beryllium successful the connection business: if there’s thing the existent authorities of AI tech can do, it’s babble distant successful antithetic languages with radical who aren’t rather fluent successful what they’re hearing.

That means determination are tons of opportunities to heighten a merchandise similar Duolingo with AI, and Luis and I talked astir the caller features successful thing called Duolingo Max, which offers chat conversations with immoderate characters and adjacent video calls with an AI avatar named Lily.

I wanted to speech astir each of that, but I besides wanted to speech to Luis astir learning generally. If you’re similar me, you’ve stopped and started utilizing Duolingo respective times; if you’re an overachiever, you’ve got a streak going and mightiness adjacent person a streak to support today. That streak is the key, and you’ll perceive Luis travel backmost to that arsenic a large thought respective times.

Engagement is the key, helium says, due to the fact that simply showing up is the cornerstone of really making advancement with connection learning. You can’t thatch idiosyncratic who isn’t there, truthful implicit time, Duolingo has go much and much of a game, due to the fact that radical similar to play games.

But determination are existent conflicts betwixt gamification and existent learning. Luis is blessed to admit that that struggle exists, and he’s fixed it a batch of thought. For him, the gamification is the important portion due to the fact that not lone does it bring you backmost to Duolingo, keeping the concern humming on nicely, but helium says it besides produces the results successful connection proficiency that Duolingo is aiming for. 

Luis got beauteous heavy into explaining wherever the wealth comes from. As you mightiness guess, it’s from iPhone users successful wealthier countries similar the United States. And immoderate method decisions Duolingo made precise aboriginal connected mean the iOS mentation takes precedence — it tin instrumentality a twelvemonth oregon much for features to rotation retired connected the Android mentation of the app.

But Duolingo is simply a planetary product, wherever the biggest chunk of learners are really trying to larn English — and those users are way much apt to usage an Android telephone and to privation oregon request a escaped mentation of the product. There are a batch of tensions here, and you’ll perceive Luis speech astir his ain puerility successful a poorer state and however that informs his decisions.

This is simply a bully 1 — Luis is the founder, and he’s helped the institution to spell nationalist and present is helping it to clasp a beauteous large exertion shift. AI has a nonstop interaction connected the merchandise helium makes, and we talked astir it each successful a beauteous nonstop way, with lone a fistful of jokes astir laminitis mode. And of course, I asked him whether helium approves of each the unhinged things the Duolingo owl says connected societal media. 

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

This transcript has been lightly edited for magnitude and clarity.

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

Thank you for having me.

I often commencement by asking CEOs what their institution is, but I consciousness similar everybody knows what Duolingo is. How bash you specify Duolingo?

It’s an app that teaches languages. That’s what we’re mostly known for. As of the past mates of years, we person besides taught mathematics and music. It’s the astir fashionable mode to larn languages successful the world. A amusive fact: determination are much radical learning languages connected Duolingo successful the US than successful each US precocious schools combined. This is existent successful astir countries successful the world. We thatch languages to much radical than the nationalist schoolhouse systems.

Yes, the quality to video telephone with Lily.

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

We person this cast of characters that our users love. One of them is an emo teen with purple hairsbreadth who is precise unimpressed by you. You tin speech to her present and signifier your speech skills, and you tin person truly bully conversations with her. There are a batch of things that are astonishing astir her. First of all, she adapts to your level — we cognize your level due to the fact that you’ve been learning connected Duolingo, truthful we person a beauteous bully thought of what your level is. 

The different happening is that she has memory, truthful she remembers the past clip you talked astir something. For example, I conscionable had a speech contiguous wherever she remembered that, past time, we talked astir the information that I similar Nirvana. She was telling maine that her favourite opus is “Smells Like Teen Spirit.”

Easy choice, I person to say.

We’re dating ourselves connected that one, but yes. These are beauteous enjoyable conversations, and you get to signifier your language. It’s wholly spoken, and it conscionable works truly well. We’re precise blessed due to the fact that this is the archetypal clip that you think, “We truly are not going to request humans for this.”

The animations and stuff, are those banal animations? Are they loops? How does that work? Is it successful existent time?

Yes. It’s animated successful existent time. We person a rig for her. We bought 2 animation studios successful Detroit. This is wherefore we person an bureau successful Detroit. They’ve done a truly bully job. Her rima moves erstwhile she’s speaking, and it’s tied to what she’s saying. She rolls her eyes astatine you. 

When you deliberation astir that concern — “We’re going to commencement gathering rigs and animations for characters. We’re going to bash it each successful existent time” — I’m conscionable coming backmost to cost. That’s a large investment. Do you deliberation that’s going to marque your existing users wage much money? Or is it going to get you caller users?

I deliberation it’s both. We spot it arsenic continuing to enactment connected the app. There are a batch of places wherever we usage a batch of animation, and we spot it arsenic continuing to enactment connected the app. And generally, arsenic we proceed moving connected the app, we get much users and get much of them to pay.

The crushed I inquire that successful this discourse specifically is that the economics of AI is conscionable a bid of question marks close now. I inquire this of everybody who’s making the investments. How bash you spot it coming retired connected the different end?

For this peculiar feature, I deliberation it’s an fantabulous usage of ample connection models, and connected our end, it’s moving beauteous well.

The different large announcement you person is called Adventures. It sounds similar a video game. What’s going connected there?

The mode Duolingo works is that the homescreen is fundamentally a path, and you’re conscionable doing lessons. Some of the lessons are present going to beryllium this happening we telephone an “adventure,” which is truly conscionable 1 of those video games wherever you determination characters around. What’s chill astir it is that you’re learning however to lick real-world situations connected Duolingo.

For example, it’s similar a small video crippled wherever you are 1 of the characters and you’re told, “Okay, spell bargain a pizza.” You determination and person to inquire around, and past you inquire immoderate people, and they archer you, “Oh, the pizza spot is implicit there.” It’s ace amusive and it helps you larn to navigate the existent world. So we’ve been moving connected that. What’s chill astir that diagnostic is that each the scenarios were mostly generated by AI. In the past, that diagnostic would’ve taken a agelong clip to scale, but we were capable to standard it beauteous rapidly due to the fact that of AI.

I played with Duolingo this morning. I person a agelong and analyzable past with trying to larn Hindi. It’s free. I was utilizing it for escaped today. How does the app marque money?

It is free. You tin usage it wholly for escaped without ever having to pay. If you don’t pay, you whitethorn person to spot immoderate ads, and we marque wealth from the ads. But also, if you privation to crook disconnected the ads, you tin wage to subscribe, and it turns disconnected the ads and gives you immoderate other features. We besides marque wealth from the subscription, and actually, the bulk of the gross comes from the subscription.

Is Duolingo profitable arsenic a company?

I’m funny astir this. I perceive astir this divided from astir everyone we speech to: we commencement out, we privation to turn our basal of users, ads assistance america bash that. It helps america support the merchandise free. And past the existent wealth is going to travel erstwhile we adhd worth and we adhd paid subscriptions. Particularly with advertizing lately with app tracking transparency connected Apple platforms, with the monolithic influx of inventory from each the different platforms successful the world, it seems similar ads are adjacent harder to marque wealth connected than ever. Has that been the lawsuit for you?

It’s astir apt true. Ads person ne'er been a precedence for us. I don’t cognize the nonstop number, but it’s thing similar 6 oregon 7 percent of our gross comes from ads. For us, arsenic agelong arsenic they’re there, they’re a bully crushed for radical to subscribe. But generally, we marque astir 80 percent of our gross from subscriptions, adjacent though, by the way, lone a small nether 10 percent of our monthly progressive users wage to subscribe. So 10 percent of our monthly progressive users springiness america much than 80 percent of our revenue.

And each of that gross is successful languages? Or is mathematics growing?

The bulk is languages. Math and euphony are growing. We launched those astir a twelvemonth ago, truthful they’re conscionable getting started. It’s overwhelmingly languages.

What languages are the astir popular?

English is the astir fashionable by far. Forty-five percent of our progressive users are learning English. The 2nd is Spanish, the 3rd is French, and past there’s a large drop-off aft that.

Are the bulk of your users extracurricular of the United States? Or are they wrong the US?

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

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

About 45 percent are trying to larn English. Within the planetary segment, they besides privation to larn different languages.

There are a batch of languages offered successful the app, and it seems similar 1 mode you could allocate resources would beryllium by saying, “English is the astir popular, we’re going to enactment the astir resources there.” But that doesn’t consciousness similar however the app works. How bash you deliberation astir it?

We decidedly bash immoderate of that. I was going to accidental we don’t bash it arsenic overmuch arsenic we should, but I don’t cognize if that’s the case. We don’t bash it commensurately with the fig of users due to the fact that we would astir apt walk each of our resources connected English, Spanish, and French. We walk the bulk of our resources successful the apical 8 languages to learn, and past we walk precise small resources extracurricular of that. The apical 8 are English, Spanish, and French. Then determination is German, Italian, Japanese, Korean, Portuguese, I think, and Chinese. 

Mandarin oregon Cantonese?

Mandarin. And aft that, determination truly is simply a immense drop-off. For example, Arabic is simply a ample language, but determination are not that galore radical learning Arabic. So we bash enactment immoderate resources there, but it’s overmuch little than for the larger languages.

How bash you deliberation astir that benignant of demand? I unfastened Duolingo, I look astatine it, I’m like, “I should astir apt larn immoderate Cantonese.”  I think, “Man, I should beryllium overmuch amended astatine Hindi than I am.” Those are existent things that I deliberation each the time. I ideate determination are a batch of radical successful my peculiar diaspora who consciousness the aforesaid way. But that’s latent demand. Do you ever spell retired and accidental to people, “You should larn immoderate Spanish”? 

We don’t.

Do you ever say, “We should marketplace Spanish successful the American South”?

We don’t. We person remained neutral astir that, but it is an absorbing happening that request for learning languages is not arsenic correlated arsenic you would similar to spot with the fig of speakers oregon possibly adjacent usefulness successful a geopolitical world. For example, adjacent though Chinese is 1 of our apical 8 languages to learn, lone astir 2 percent of our users are learning it. It’s comparatively small, adjacent arsenic the astir spoken connection successful the world. 

One of the things that goes into people’s calculus is however hard a connection is to learn. Chinese, astatine slightest for English speakers, is conscionable a batch harder. We person data. To get to a beauteous bully constituent successful Spanish for English speakers takes, telephone it, 300 to 400 hours. That aforesaid level of cognition for Chinese takes astir 2,000 hours. The world is, successful the United States, if you’re conscionable going for pragmatism, instrumentality connected investment, Spanish is astir apt overmuch better. In the US, you astir apt should larn Spanish. It’s rather an casual connection to learn.

In the US, you should astir apt larn Spanish. That is simply a selling message.

Well, we don’t accidental that. We’ve tried to stay neutral. We astir apt would get successful trouble, oregon I would get successful occupation wrong the company, if we started pushing radical connected definite languages. 

Maybe not wrong the app, but arsenic a mode to grow, arsenic a mode to seizure caller users. It seems similar a batch of what Duolingo is close present is radical cognize they should beryllium multilingual oregon bilingual astatine least, and truthful Duolingo is there. But there’s besides a immense information of the population, astatine slightest successful this country, who are like, “Screw it. I talk English.” The thought that there’s worth successful learning a 2nd connection is overseas to them.

There is, though I’m precise blessed with our results successful the US. Historically, determination hasn’t been a large tendency to larn languages successful the US and the UK. The reasoning has been, “You tin larn English. We are precise happy.” In the US, 80 percent of our users were not learning a connection earlier Duolingo. We’re increasing the marketplace successful the US. It’s the aforesaid fig successful the UK. I’m precise blessed with that.

I deliberation backmost to learning French erstwhile I was successful precocious schoolhouse successful Wisconsin. There’s learning the connection and past there’s each of the civilization that comes with a language, peculiarly immoderate of the determination languages. High schoolhouse French is simply a batch of looking astatine a representation of a baguette. It’s wholly overseas to immoderate you’re doing. Do you deliberation astir that, wrong of Duolingo, that there’s a immense taste constituent here?

We do.

That’s much investment, right?

We do, and we effort to adhd the culture. We don’t bash it arsenic overmuch arsenic possibly we should. We effort to instrumentality mostly with languages. It besides depends connected the language. Some languages are rather choky to the culture, and immoderate are little so. Spanish is simply a bully example. There are 20-some countries that talk Spanish, and immoderate of them are beauteous antithetic than others. We bash a small spot of culture, but we effort not to beryllium like, “Oh, you’re learning Spanish. You’re a Mexican idiosyncratic with a sombrero.” We effort not to bash that. I mean, we besides person to not beryllium offensive. But we effort to adhd it a small bit. I would accidental that it’s not the superior goal.

The crushed I inquire that is that Duolingo is instantiated for astir radical arsenic its mascot. We should speech astir the mascot’s personality and its social media presence, but it’s reasonably abstracted from a idiosyncratic teaching you the language. There’s not idiosyncratic connected the different broadside that’s like, “I’m teaching you this. Here’s the civilization that comes with it.” You mightiness person different teachers who mightiness thatch you successful different way.

There’s an abstraction determination that conscionable feels interesting, particularly arsenic we’re evidently going to speech astir AI and however you’re utilizing that and however you’re expanding the platform. I privation to propulsion connected that a small bit, that abstraction. Do you deliberation it’s resulting successful radical who’ve learned a connection oregon radical who’ve learned however to communicate?

It’s been precise overmuch connected intent for america to not enactment humans successful the app, arsenic successful quality teachers. There’s thing incorrect with quality teachers. It’s conscionable the lawsuit that, from the beginning, we’ve been a exertion company, and we’ve wanted to marque it truthful that exertion teaches you. There are a mates of reasons for that. One is that it’s a batch cheaper to thatch you with exertion than with a quality teacher.

The different happening is, determination betwixt 80 and 90 percent of connection learners don’t privation to speech to different human. They whitethorn archer you they do, but they don’t. It’s due to the fact that erstwhile you’re learning a language, you’re beauteous shy astir it, and lone the utmost extroverts are good talking to a alien connected video successful a connection that they’re not precise bully at. The bulk of radical won’t bash it.

We’ve done probe studies implicit the years because, implicit time, we thought possibly we should adhd humans. But these probe studies are immoderate of the astir astonishing things that I’ve seen. When you speech to a user, you inquire them, “What bash you deliberation could marque Duolingo better?” Historically, successful the past, they person said, “Well, much signifier speech with a existent person.” They person said that. And past you inquire the user, “Okay, truthful you’re telling maine if I enactment a quality connected Duolingo, you would bash that?” And past they say, “Yes, I would.” And you tin adjacent inquire them, “Would you wage for it?” And past they’ll say, “Yes, I’ll wage for it.” And past you archer them, “Okay, bash you privation to bash it close now?” And the reply invariably is, “No, no, not close now.” People conscionable don’t privation to bash that.

That’s wherefore we haven’t enactment humans in, and I deliberation it’s been a bully decision, particularly present that we tin bash a beauteous bully occupation of getting you to signifier speech without a quality with ample connection models.

I privation to inquire astir this due to the fact that I’ve been asking a batch of radical connected this show: what bully are these ample connection models? What are the products you’re going to make? I recognize you’re making the models, and it feels similar Duolingo has a precise earthy solution, which is that you tin speech to it and it’ll speech back. It doesn’t substance if everyone is hallucinating due to the fact that each you’re doing is practicing talking.

That’s precisely right. It is simply a truly bully application. You said it. It doesn’t substance if it says thing that is simply a small incorrect due to the fact that you’re conscionable practicing your language. Also, it doesn’t substance if it makes a tiny mistake. Sometimes it makes a tiny grammatical mistake. People don’t adjacent announcement due to the fact that they’re usually beginners successful Spanish oregon French. It besides tin accommodate to your level truly well. Large connection models are truly bully astatine adapting to your level.

So we archer it, “Okay, accommodate to a beginner successful Spanish.” We adjacent archer it, “Hey, due to the fact that we’ve seen this idiosyncratic larn connected Duolingo, we really cognize each the words they know.” So we archer the connection model, “This idiosyncratic lone knows these 200 words, truthful delight mostly usage lone these 200 words.” It works truly good for that.

How overmuch concern into AI are you making? This is simply a caller product. It’s precise costly. Everyone is telling maine astir however overmuch the Nvidia GPUs cost. You said you’ve lone conscionable go profitable. This feels similar the happening that volition instantly marque you not profitable again arsenic you put successful AI.

We’re investing a lot. Fortunately, it’s bully for america successful presumption of profitability for 2 reasons. There are 2 places wherever we put successful AI. The archetypal is generating information that is going to beryllium utilized successful our lessons. That information used to beryllium generated partially with humans, and present it’s mostly generated with AI, and it’s a batch cheaper to make with AI than with humans. It’s besides a batch faster to make it with AI, truthful we’re precise blessed with that.

And past the different large usage is real-time conversation. That 1 is expensive. It’s costly to supply a real-time speech with a user, but what we had to bash is adhd a higher-priced plan. We present person 2 subscription plans. We person Super Duolingo, which is our modular subscription, and we added a caller 1 called Duolingo Max, which is astir doubly the terms of Super Duolingo and gives you the speech practice. It’s expensive, but radical wage doubly arsenic much, truthful it truly doesn’t chopped into it. It’s worked retired good for us.

Let maine dive into the economics of that because, successful general, I’m fascinated by whether immoderate of this volition effect successful profitable, sustainable companies. There’s a batch of wealth flowing into this. So, you charged doubly the terms to tally inference. Is that idiosyncratic else’s ample connection model? 

Yes, we usage OpenAI.

So you’re buying immoderate capableness from OpenAI, you’re buying immoderate tokens from them, and you’re reselling them to users for doubly the terms of your modular plan. What’s your borderline connected that resale?

I don’t cognize the percentages disconnected the apical of my head, but I bash cognize that it’s bully for america successful presumption of margins.

That’s the happening I’m funny about. I don’t cognize if it’s bully for OpenAI each the mode astatine the bottommost of that chain. I don’t cognize if that’s profitable for them. But arsenic you physique products connected this stuff, it seems similar your economics beryllium connected their economics, successful immoderate way, due to the fact that you request to adhd a borderline to that. That each seems precise analyzable and tenuous, particularly if the AI features are what bring you caller users.

The bully quality is, astatine the moment, the AI features are not bringing america caller users. 

Or caller revenue?

Yes, it is bringing america caller revenue. There’s a bully borderline there. So, for a fig of reasons, the terms of the aforesaid nonstop telephone is going down implicit time, whether you bash it done an OpenAI oregon whether you bash it done a Microsoft. Everything is getting much efficient, and chips are besides getting cheaper implicit time. At the moment, there’s a bully magnitude of cushion, but we expect that there’ll beryllium an adjacent larger magnitude of cushion implicit time. At slightest for our application, I’m not peculiarly acrophobic successful presumption of margins. For our application, the margins enactment retired beauteous well.

Do you deliberation the models are somewhat interchangeable? This is simply a happening that I’ve been proceeding much and much is that the exemplary concern isn’t the thing; the merchandise concern is the thing.

I deliberation the reply is yes, but the operative word is “somewhat.” They are “somewhat” interchangeable. We’ve tried to physique our exertion stacks truthful that you tin interchange them, but the world is that you commencement getting wonky worldly due to the fact that you astir apt spent a batch of clip investigating your mode into the close queries. You whitethorn person done immoderate fine-tuning. You tin interchange them, but if you do, you astir apt request to walk a fewer months making definite that the wonkiness goes away.

When you deliberation astir this concern implicit time, does it consciousness similar you request to enactment the wealth successful upfront and you’ll get much businesslike connected the backmost end? Or does it consciousness like, “Oh, this is going to beryllium the aboriginal of the company, truthful we request to rebuild astir the capabilities of a ample connection model”? 

It’s determination successful between. I bash deliberation that ample connection models are going to beryllium precise affirmative for Duolingo — they already are, and I deliberation they’re going to proceed being precise positive. What is not existent is that ample connection models lick each our problems. One of the biggest issues that radical aren’t talking about, peculiarly with education, is that ample connection models are bully astatine teaching you stuff. They’re not bully astatine engagement. And that’s the hardest happening with education.

The hardest happening astir maine trying to thatch you thing is conscionable keeping you engaged. Somehow, radical forget. I spot immoderate radical saying, “You tin larn quantum physics with ChatGPT.” And yeah, sure, but that’s conscionable not that impressive. You tin larn quantum physics with a book. The exertion to larn that has been astir for a agelong time. It’s called a publication and it works. It’s conscionable that radical don’t truly privation to work a quantum physics book. And similarly, astir radical astir apt don’t privation to spell to ChatGPT and commencement asking questions astir quantum physics. It’s the aforesaid happening for connection learning. Large connection models are precise bully astatine getting you to practice, but keeping you engaged is beauteous hard.

I don’t cognize if ample connection models are going to assistance each that overmuch with that part. In the end, this is simply a bittersweet thing, but the world is that Duolingo is precise gamified. I wholeheartedly judge astir radical would alternatively walk much clip playing Candy Crush than talking to others. That’s possibly a bittersweet truth. And determination are immoderate exceptions. I mean, radical emotion talking to idiosyncratic they’re successful emotion with, and sure, that’s nice, but the world is, astir of the time, astir radical would alternatively walk clip playing Candy Crush oregon scrolling connected Instagram than talking to others. I conscionable don’t deliberation ample connection models are going to assistance overmuch with each of that. 

You person a agelong past successful gamification. Your archetypal project, which you sold to Google, was a gamified thing. You did reCAPTCHA, which is fundamentally gamifying grooming information successful a peculiar way.

Do you deliberation there’s an improvement successful Duolingo — that the archetypal happening that you worked connected was the engagement and bringing users backmost to the app and having the character, and past the underlying contented was connection lessons? When I archetypal started utilizing Duolingo respective years ago, I was like, “Oh, this is precise familiar. It’s conscionable that this vertebrate won’t permission maine unsocial and that’s wherefore I’m backmost again.” And present you’re talking astir this full different spectrum of things: we’re going to usage AI; we’re going to person these earthy connection conversations; we’re going to grow to mathematics.

When did you consciousness similar you were making the modulation from “we’re gamifying this precise acquainted thing” and “we’re utilizing this caller engagement mechanism” to “this is present a wholly caller thing”?

From the beginning, this is simply a cardinal thesis that we judge present astatine Duolingo: the hardest happening astir learning thing by yourself is staying motivated. In fact, that is astir apt the crushed for the immense bulk of our occurrence is that we realized that aboriginal on. From the beginning, we person tried to person a happening that is enjoyable to usage and that keeps you coming back. We person astir apt spent much effort connected that than thing else.

Internally, our feeling is that learning a connection is simply a batch similar moving out. It doesn’t substance each that overmuch whether you’re doing the elliptical oregon a Peloton oregon a treadmill. By far, what matters the astir is that you’re doing it each day, immoderate the hellhole you’re doing. It’s benignant of the aforesaid with Duolingo. Maybe immoderate methods are much businesslike than others, but what matters is that you’re doing it each day. We got precise bully astatine that. Now, erstwhile we got precise bully astatine that, we started trying to adhd much sophistication successful what we teach, and we’ve been doing that for the past fewer years. But always, primarily, we are a information engine.

Is that the halfway of it still? 

It’s inactive the core.

I’m going to extremity up asking you astir founder mode, but you’re the founder. How bash you support the absorption connected that portion alternatively of everything else?

I do. I walk effort connected that. But it’s not conscionable me. At the company, it is beauteous good understood that if it’s not fun, it’s not going to work. We walk a batch of effort trying to support Duolingo enjoyable. This is why, for example, erstwhile we did this happening wherever you tin speech to an AI to signifier conversation, you’re not conscionable talking to a random AI; you’re talking to 1 of our formed of characters. It has a personality. Really, everything we do, each clip we enactment thing out, it’s ingrained successful our reasoning that, “Oh, this has to beryllium enjoyable.” I walk effort pushing that agenda, but I don’t person to each that overmuch due to the fact that it’s precise ingrained successful the company.

How large is Duolingo now?

About 850 people.

Where are you each located?

The largest bureau is successful Pittsburgh, Pennsylvania. We person astir 400 radical there. We person astir 250 successful New York, and past we person offices successful a fewer cities. We person 1 successful Detroit, we person 1 successful Seattle, 1 successful Berlin, and 1 successful Beijing. All of those offices person 30 radical successful them. But 1 cardinal happening is we are not remote. We’ve got to travel to 1 of those offices.

Why is that?

I conscionable wholeheartedly judge that you tin enactment amended that way. Most of what we do, not 100 percent, but astir of what we do, is originative stuff. It’s a batch harder to bash truthful implicit Slack and Zoom. That worked retired for astir 9 months during the pandemic, but it is really awesome however erstwhile the pandemic started and we each had to spell remote, we executed beauteous well. But toward the extremity of it, our ideas had tally out. We were executing the ideas, but we had tally retired of caller ideas. It’s beauteous amazing, arsenic soon arsenic we came backmost to the office, wrong 3 months, you would spot each these ideas popping up, and it’s because, erstwhile you’re remote, you can’t beryllium successful beforehand of a whiteboard and speech astir stuff. Also, we person luncheon unneurotic present each day. In the luncheon line, you perceive radical being like, “Hey, I haven’t seen you successful a while. I thought of saying this to you.” It’s conscionable thing you would ne'er nonstop a Slack connection for.

I deliberation the operation of each of that makes it a amended company. I don’t person overmuch proof, but I americium highly convinced.

Sundar Pichai astatine Google told maine astatine the precise opening of the pandemic that helium was disquieted that the institution would tally retired of ideas if they stayed distant excessively long. He said, “We cognize what we request to bash for the adjacent turn. I’m disquieted astir what happens astatine the adjacent turn.” Did you person contention erstwhile you reimplemented instrumentality to office?

No, fortunately not.

Do your employees deliberation that?

No, no, no.

Do they hold with you?

We’ve done a batch of dumb things astatine Duolingo, but this was not 1 of the dumb things we person done, successful retrospect. In my archetypal connection saying, “Everybody’s going to enactment remote,” I said, “But we’re going to travel backmost to the office. I bash not privation Duolingo to crook into a distant company. We are not a distant company.” We kept saying that the full time. A batch of companies did this happening wherever during the pandemic, they would prosecute radical each implicit the satellite because, whatever, you’re remote. We ne'er did that. When we hired people, we would say, “I get that you’re not coming to the bureau close now, but your occupation is successful New York, and we expect you to beryllium successful New York due to the fact that astatine immoderate constituent hopefully soon, we volition beryllium backmost successful the office.” We ne'er stopped repeating that. By the clip we said, “Okay, clip to spell backmost to the office,” this was not a astonishment for anybody. I don’t deliberation we mislaid a azygous worker from that.

Do you deliberation that the markets you’re successful assistance you with that — being successful Pittsburgh and Detroit? If you were successful San Francisco, I deliberation a batch of radical would say, “Screw you, I tin spell get different job.”

That’s astir apt true. We are not successful San Francisco, and that’s astir apt true. Although the New York bureau is present the 2nd largest office, and we besides didn’t suffer radical successful New York.

Do you find that radical are demanding much flexibility adjacent with a afloat instrumentality to office?

Sure. I mean, compared to before. For example, we’re not present 5 days a week successful the office. The mode we enactment is Tuesday, Wednesday, and Thursday, you person to beryllium successful the office. Monday and Friday, it’s optional. What happens, successful practice, is that astir fractional the radical travel successful connected Mondays, and astir 20 percent of the radical travel successful connected Fridays.

We are talking connected a Friday. 

I americium here. I americium present successful the office, though.

Oh, you’re there. Oh, precise good.

I americium present successful the office. I come. But I don’t consciousness similar I person the governmental powerfulness wrong this institution to say, “All right, people, you’ve got to travel successful 5 days a week.” I consciousness similar that would not spell implicit well.

One of the different pieces of the pandemic puzzle and instrumentality to bureau is that determination was a suppression of request to question and explore. I person friends who, astatine slightest from their Instagrams, haven’t acceptable ft backmost successful the United States successful 2 oregon 3 years. Have you connected to that radical of radical who privation to larn languages connected the go? Has the reexplosion of question had an interaction connected your business?

Travel is interesting. Now that we’re a publically traded company, radical person hypothesized each kinds of worldly astir question with us.

They’re like, “Travel’s opening up. That whitethorn beryllium bully for Duolingo.” Or they say, “Travel’s drying up. That whitethorn beryllium atrocious for Duolingo.” The world is that question does not impact america each that much. I tin person hypotheses for wherefore that is, but we person not seen traveling closing down oregon opening up impact america each that much.

We person a batch of learners that person antithetic motivations, but the 2 large buckets are not travel. One of them is simply a hobby, and that’s the biggest bucket successful the US. If you inquire radical successful the US wherefore they’re utilizing Duolingo, the astir communal reply is, “Well, I utilized to play a batch of Candy Crush or I utilized to bash a batch of Instagram and present I’m doing Duolingo and astatine slightest I’m learning immoderate Spanish. It’s conscionable a hobby.”

And past the different immense radical is radical learning English. For them, it’s not astir travel. They conscionable request to larn English either for acquisition opportunities oregon for occupation opportunities. Those 2 large buckets relationship for 90–95 percent of our users. Travel conscionable doesn’t impact it precise much.

We started talking astir latent demand. What are radical coming to you for? And past there’s growth, which is, “how bash we spell make immoderate demand?” When you deliberation astir the operation of the company, would you ever say, “We’ve got to spell bash selling to marque question happen”? Or is that conscionable not however you deliberation astir it?

No. In our marketing, we haven’t thought astir that. Ninety percent of our users travel successful from connection of mouth, and that volition support happening, I think. We besides walk precise small successful selling comparatively. Our full selling fund for the full world, and we truly bash run successful each azygous state successful the world, is 50 cardinal a year, which is rather tiny for a institution with our revenue. But immoderate we’re doing with selling seems to beryllium moving beauteous well, and we don’t walk a batch of wealth connected it.

I consciousness similar I person yet to inquire you the Decoder question. So, arsenic agelong arsenic we’re talking astir your selling spend, however is the institution structured?

We person functions. There’s the selling function, there’s the engineering function, determination is the merchandise absorption function, design, etc. We person functions and each relation has a relation head. To springiness you a comparative thought of the sizes of the functions: engineering, merchandise management, and plan combined relationship for astir 70 percent of our employees. Design is weirdly ample for our company. We person 850 full employees but astir 130 radical successful the plan department. So plan is large, but we person engineering, product, and plan relationship for astir 70 percent.

Those are the radical moving connected the product. If you look astatine that group, it’s structured into areas, and each country is related to 1 of the things we’re trying to optimize. For example, with connection learning, the 3 main things we’re trying to optimize are: engagement, truthful however amusive Duolingo is; teaching [the material] better; and however overmuch wealth we make. We person an country for each 1 of those things. Then, successful each area, determination are teams, and successful each team, determination are people. It’s a small spot of a matrix structure.

One important happening that I deliberation has worked truly good for america — but it’s not that casual to bash — is that our areas and our teams are not feature-based. What I mean by that is that astir bundle oregon app companies usually person a squad for each feature. So this is the login team, which owns login. Or if you person a leaderboard, this is the leaderboard squad and they ain the leaderboard. We bash not person that. Our teams don’t ain features. Our teams ain metrics. So we person a squad for subscription revenue. We person a squad for regular progressive users. And they tin interaction immoderate they privation successful the app. All they person to bash is continually summation the metrics. There are positives to this, which are precise aligned to metrics. There are negatives successful that nary squad owns definite features. When thing breaks, determination are a batch of radical being like, “It’s not my feature. I don’t know.” There are positives and negatives, but each successful all, this has really worked retired truly good for us.

This sounds, one, similar a absorption to moving astatine Google wherever teams bash ain things similar the login surface and they endlessly pass astir however they’re going to alteration logins. But it besides seems similar it mightiness enactment for a tiny institution wherever 1 idiosyncratic tin spot the full merchandise oregon recognize the full merchandise and however it each works together. And past you’re going to get inevitable collisions arsenic 2 radical effort to alteration thing to summation 2 antithetic metrics successful antithetic directions. How bash you resoluteness those collisions?

There are decidedly collisions. There are a mates of things that assistance america here. One is, each alteration to the app passes done this reappraisal process called merchandise review, which is not conscionable 1 person. There’s a radical of radical that person a batch of cognition astir however the full happening works. They service arsenic a small spot of a semaphore, a small spot of like, “No, you should not bash that.”

And past the different happening that is truly important is we person guardrail metrics. So here’s however that works: If you are connected the squad that’s trying to amended subscription revenue, your extremity is to amended that metric. But we archer you, “You can’t messiness up immoderate of the different metrics.” For example, if you bash an experimentation that improves subscription gross by a cardinal bucks oregon immoderate but it decreases regular progressive users, you can’t motorboat it. That has truly helped teams constabulary themselves. They astatine slightest won’t spell bash thing that truly messes thing other up. The operation of these 2 things has helped.

You are close that if we had 100,000 employees, I don’t deliberation the operation would work. That said, I don’t deliberation that a institution similar Duolingo, astatine slightest with the products that we have, needs 100,000 employees. I deliberation we’ll turn and we’ll proceed growing. Maybe we’ll get to, I don’t know, 5,000 employees, but I uncertainty we’ll ever get to thing similar 100,000 employees.

How often bash these collisions travel each the mode up to you? How often bash you person to marque a tradeoff?

Not that often. Teams constabulary themselves a lot. I bash spot each azygous alteration that passes done the app. I bash spot that, but usually, I’m not making tradeoff calls. The main happening that I’m looking for is making definite that everything’s precocious quality.

At the opening of the show, you were joking astir laminitis mode due to the fact that I called you the cofounder. That’s Brian Chesky, who has been connected the show. He’s talked a batch astir however helium refactored Airbnb. He was the conductor of the orchestra. That has gotten immoderate magnitude of attraction it’s gotten. 

Do you spot yourself successful that benignant of role, that you’re the idiosyncratic who tin spot the full app, that you are the idiosyncratic who understands however each these tradeoffs are getting made?

Yeah, it’s decidedly true. The bully quality is our employees instrumentality astir for a precise agelong time, and our leadership, peculiarly successful the merchandise areas, has been the aforesaid for the past 8 years. I person a presumption of everything, but the world is that our caput of product, Cem [Kansu], has a presumption of everything, too. Our caput of design, [Ryan] Sims, besides has a presumption of beauteous overmuch everything. Yes, I americium successful that mode, but we person a fig of radical who could astir apt play that relation arsenic well.

When the 3 of you disagree, however bash you resoluteness it?

The bully quality is there’s small disagreement, which happens for a fewer reasons. The archetypal is that we’re a precise metrics-based company, truthful usually, we conscionable fto the metrics speak. If we tally an A/B trial and the metrics accidental something, my sentiment doesn’t substance each that overmuch unless it’s thing that we deliberation is truly similar a acheronian signifier oregon something. But generally, my sentiment oregon their sentiment doesn’t substance each that much. That’s 1 reason. The different is that we’ve been moving unneurotic for truthful agelong that we’re beauteous aligned connected everything. And past the past thing, I person the saying, “If we’re going to spell by opinion, let’s spell by mine” Generally, erstwhile we person disagreements, I spot however profoundly they judge successful their thing, and sometimes I conscionable disagree and commit. But if we judge with adjacent spot connected something, I volition conscionable spell with my thing.

That mightiness beryllium the astir succinct explanation of laminitis mode I’ve heard yet, actually.

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

“Just bash what I say” is truly the reply to what laminitis mode is.

But the bulk of things, we don’t truly spell by opinion. The bulk of things are conscionable by data.

So here’s the different Decoder question. This is simply a bully instauration for it. How bash you marque decisions? What’s your framework? 

For the institution oregon for me? They’re similar, but they’re not identical.

Some radical don’t deliberation they’re antithetic and immoderate radical deliberation they are precise different, truthful reply whichever mode you want.

For the company, the decisions are precise overmuch tied to instrumentality connected investment. With astir things, there’s a instrumentality connected concern calculation. Even if we don’t beryllium determination and constitute the numbers down, there’s however overmuch effort you’re going to enactment into thing and however overmuch you’re going to get back. That drives astir of our decision-making. There’s different happening that is not unsocial to Duolingo but I deliberation is not the norm astatine astir companies, which is, usually, erstwhile you’re doing a project, determination are 3 things that matter: however overmuch does it cost?; however accelerated are you going to bash it?; and what’s the quality? Usually determination are tradeoffs betwixt these things. At Duolingo, the astir important happening is quality, the 2nd astir important happening is speed, and the 3rd astir important happening is whether we’re connected budget. In galore companies, it’s the different mode around, wherever the astir important happening is budget, past speed, past quality. Here, prime is the astir important thing. So that’s different constituent of our decision-making.

That’s for the company. For me, it’s precise gut feeling-driven, which I utilized to find myself trying to justify. I person stopped doing that because, astatine this point, I’m like, “Look, this is what I deliberation we should do. I tin archer you reasons that I tin astir apt travel up with aft the fact, but the world is that my gut says we should bash that.” Because I’ve been moving connected Duolingo for 13 years, my gut’s beauteous good. It’s not 100 percent correct. I marque mistakes, but it’s beauteous good. I chiefly bash things based connected gut feelings, and past I archer radical the justification afterward. But everybody astir maine knows that these justifications are aft the fact. They’re not rational thoughts.

That evidently works for a startup laminitis for a backstage company. You’ve been a nationalist company CEO for implicit 3 years. Is that moving for you arsenic a nationalist institution CEO?

Yes, due to the fact that again, the bulk of decisions that we person to make, there’s a wide answer. It’s conscionable like, “Well, look, this is going to suffer america money. Let’s not bash that.”

But bash you person to alteration the mode you communicate? I’ve heard this from a fistful of CEOs who’ve taken their companies nationalist and present they’re connected the quarterly reports cadence and they person investors. Elliott Investment Management mightiness amusement up connected your doorstep and beryllium unhappy that you’re not selling to much Chinese speakers oregon whatever. Have you had to alteration that cognition present that you tally a nationalist company?

No, it’s been precise fortunate. First of all, we hired an astonishing CFO earlier we went public, Matt Skaruppa, and fortunately, helium deals with astir nationalist institution stuff. I don’t bash a batch of that, and I’m precise thankful for that due to the fact that I don’t person a concern background. I person a PhD successful machine science. That’s what I’m bully at, not finance. So there’s that.

The different happening we’ve been fortunate astir truthful acold arsenic a nationalist institution is that we’ve executed well. I deliberation that has fixed america a small spot of latitude successful that fundamentally we don’t get asked precise pugnacious questions due to the fact that we’ve executed precise well. I americium definite that won’t beryllium the lawsuit forever. I’m definite astatine immoderate constituent we’re going to miss a quarter. We haven’t truthful far, but I’m definite we will. And erstwhile that happens, I’ll astir apt person to reply immoderate questions and I’ll astir apt person to archer people, “Sorry, we’ll beryllium much buttoned up from present on.” But truthful far, I amusement up to net calls successful a T-shirt. The time you spot maine amusement up to an net telephone successful a suit, you’ll cognize that we fucked up.

Yeah, it’s clip to get out.

[Laughs] Like, “Oh, truthful sorry.”

The different happening I perceive from nationalist institution CEOs is thing that relates to your accent connected prime first. You person a batch of metrics, which means your investors tin spot a batch of your metrics oregon request a batch of your metrics successful antithetic ways. Quality is not measurable successful that way, right?

It’s not.

At slightest successful the existent market, it’s not a large communicative to investors if you’re saying, “Okay, we’re going to put a clump of wealth successful AI, and we deliberation successful this usage case, it’s going to beryllium truly palmy for america oregon we tin complaint much for it but we person to walk a clump of wealth upfront and we person to hold to marque it good.” How are you managing that now?

You’re right, prime is not measurable. The mode we marque decisions astir that is that, peculiarly successful our plan department, we person radical who are precise overmuch sticklers astir quality. We’re like, “Nope, that’s conscionable not bully enough.” We bash that a lot.

In presumption of investment, I mean, honestly, with the nationalist markets, we don’t speech overmuch astir that. We speech astir the metrics. We don’t speech excessively overmuch astir however it turns retired we spent an other period moving connected this diagnostic conscionable due to the fact that we didn’t truly similar the mode the owl was animated. We don’t speech excessively overmuch astir that. I deliberation that’s fine. But my conjecture is that if we went connected net calls and spent each of our clip talking astir however overmuch effort we enactment into animating our owl, I don’t deliberation radical would similar that.

I honestly deliberation much companies should walk much clip talking astir however overmuch clip they walk making things good. That would be, I think, a important upgrade to American capitalism.

I would similar that. But yeah, the world is we bash walk an inordinate magnitude of clip connected things. If you look astatine our app, it, implicit time, has go precise animated. We walk an inordinate magnitude of clip looking astatine the precise frames of the animation. We’re like, “See, this is not creaseless enough.” I’m not claiming that our app is perfect, but we astatine slightest effort truly hard for it to get arsenic adjacent to cleanable arsenic possible.

One happening that besides seems hard to measure, oregon a metric that mightiness pb you successful antithetic directions, is however palmy Duolingo is. Maybe the astir important metric of each is: Are radical getting bully astatine Spanish? Are they leaving this acquisition with the quality to pass successful Spanish? Do they not conscionable cognize the connection but tin really communicate? Can you measurement that?

Yes, we can, but not arsenic efficaciously arsenic you would similar to measurement it. So the reply is, yes, Duolingo works. We person measurements. I’ll archer you however we measurement it. Unfortunately, this is the lone mode we cognize however to measurement it reliably. It’s not that large of a way, but it’s this: You instrumentality idiosyncratic who’s conscionable starting Duolingo, you inquire them a clump of questions astir their erstwhile knowledge, and you besides springiness them a trial to measurement however overmuch they know. Then you person them usage Duolingo for a agelong clip due to the fact that it takes a portion for you to really larn worldly successful the language. You person them usage it for a twelvemonth oregon two. And then, astatine the extremity of that period, you inquire them questions astir whether they utilized different resources and you besides springiness them a trial to fig retired however overmuch amended they got. It turns retired that radical who knew thing earlier and utilized Duolingo and did not usage different resources larn astir arsenic overmuch arsenic oregon much than successful a classroom. It varies by the study, but we’re beauteous blessed with that. The results really work.

The occupation with this mode of measuring is that it’s precise slow. It takes 1 oregon 2 years for america to get a caller measurement, and I truly don’t similar that. But we person not been capable to travel up with a amended mode contempt the information that we person tried. We’ve done things successful the app wherever we’re like, “Okay, we tin bash micro-measurements of whether you’ve learned this word.” It’s been ace analyzable to bash that and ne'er fixed large results. So we conscionable trust connected these old-school, “pre-test, post-test” methods. That’s it.

This is wherever you veer close into lengthy society-level debates astir acquisition and however we measurement the show of schools and teachers. Do you consciousness similar you’re participating successful that system? You’re utilizing their measurements, right? This is however schools bash it. They trial you.

We are utilizing their measurements, and efficacy is truly important. We walk a batch of effort trying to marque definite that we’re efficacious. The different bully news, adjacent though the timescale present is successful years, is that you tin crippled however effectual Duolingo is. If you look astatine it implicit the past 10 years, each twelvemonth it is much effectual than the erstwhile year, for sure.

That’s connected this test-based measure. More radical are passing the test?

Yes, much radical are getting higher scores connected the test. Basically, radical are learning much connected Duolingo each year. And determination are a fig of reasons for that due to the fact that we enactment to effort to thatch [the content] better. But it is decidedly true.

At this point, erstwhile we comparison ourselves, we cognize we are arsenic bully arsenic oregon amended than a schoolroom environment. We cognize we’re not yet arsenic bully arsenic a bully one-on-one quality tutor. Our extremity is that we tin bash that implicit the adjacent 3 to 5 years: beryllium arsenic bully arsenic a one-on-one quality tutor successful presumption of efficacy. But we’re mode amended successful presumption of getting you to instrumentality around. But successful presumption of efficacy, if you person the wealth and the spot to proceed going, a one-on-one quality tutor does better.

Do you deliberation that there’s a struggle betwixt gamification and engagement — the things that you’re historically palmy astatine — and education?

Yes, determination is.

How bash you negociate that conflict?

Very easily. Always spell with engagement.

Really?

Yes.

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

I’ll springiness you galore arguments, but the 1 that works the astir is this: It doesn’t substance however effectual you are. You can’t thatch idiosyncratic who’s not there.

And that’s it?

That’s it. People leave. The world is it’s not ever existent that engagement and learning outcomes are astatine odds. But erstwhile they are, we usually similar going for engagement. I’ll springiness you an example. There are immoderate things that are frustrating, and vexation makes you leave. We really conscionable creaseless them. By that I mean, if I could unit you to beryllium there, I whitethorn beryllium capable to thatch the worldly to you successful 5 minutes, but it’d beryllium a precise frustrating 5 minutes. Instead, what we bash is thatch it to you successful 2 hours — just mode slower, but the full time, things are animating connected the surface and you’re getting dopamine hits oregon whatever. Even though a truly bully teacher could person taught it to you successful 5 minutes, watching you marque mistakes, it would person been frustrating. We overmuch similar to support you around.

Part of the crushed is due to the fact that we’re successful an app mounting arsenic opposed to a schoolhouse setting. In a schoolhouse setting, the information is the kids are held hostage there. They can’t leave. With an app setting, the tiniest frustration, radical are like, “You cognize what? I’m going to spell to TikTok now.” We conscionable can’t suffer those users. So we ever opt for engagement, but that doesn’t mean we won’t thatch the worldly to you. We’ll conscionable instrumentality it a small slower.

It’s wide that you person thought astir this a lot.

We’ve spent years reasoning astir this.

I privation to circular this retired a small spot due to the fact that you person a precise wide reply and a precise wide constituent of view. What bash you deliberation the circumstantial hostility betwixt gamification and acquisition is? What are you losing erstwhile you ever prime gamification?

Probably the happening you’re losing is ratio — by that I mean, the magnitude of contented learned per portion time. The information of the substance is, I grew up successful the 3rd satellite a portion ago. Some of the worldly that I grew up being taught, my teachers were hitting maine portion teaching it. I’m not kidding. They would deed me. The world is that I astir apt learned truly accelerated due to the fact that erstwhile you were learning penmanship, if you did the incorrect thing, they deed you with a ruler. You person a existent inducement to get that done precise fast. You conscionable larn truly accelerated due to the fact that you’re like, “Whoa.” So I deliberation it’s true: you tin larn faster if you’re successful an situation wherever you’re forced to bash truthful and cipher cares whether you consciousness bully astir it. But successful our case, I’m good with somewhat slower learning arsenic agelong arsenic you’re inactive engaged.

Efficiency, I recognize that one. I had immoderate beauteous strict teachers successful my time, but I was truly bully astatine taking the tests. My woman and I went to assemblage together. She’s overmuch smarter than me. 

But you’re a bully trial taker.

She was ever huffy astatine maine due to the fact that I could conscionable amusement up astatine the extremity and instrumentality the test. This is genuinely astir apt wherefore she didn’t day maine for years, due to the fact that of that halfway frustration. This is what I mean by education. That quality teacher tin measure whether you’re bully astatine taking a trial oregon whether you’ve really learned something. That’s the tradeoff that I was getting at, is that if it’s each a gamification engagement, radical mightiness conscionable larn to play the game. They mightiness not person learned anything.

There’s astir apt a small spot of that. It’s precise hard to measure, of course. But the world is, ultimately, it works. Duolingo works. Just arsenic an example, for me, I’ve been utilizing lone Duolingo to larn French for the past fewer years. At this point, I tin ticker Netflix shows successful French — with nary subtitles oregon anything. I conscionable ticker them, and it works. So you’re right, there’s astir apt a tradeoff. It’s astir apt beauteous hard to measure. But what we’re looking for present is that radical are really utilizing their clip well.

I privation to effort to necktie each of these themes and ideas together. You person a large imaginativeness for Duolingo. You’ve talked astir it a lot. Being disposable to thatch everybody languages astir the world, being successful a fig of countries. And past there’s the information that you’re a nationalist company. You’ve got to marque money. You’re inactive showing up successful T-shirts. The archetypal happening that comes instantly to my caput is, you’re launching caller things similar mathematics and music, and they’re not disposable connected Android, which is the single astir fashionable operating strategy successful the world. It’s utilized by the bulk of lower-income radical successful the world. That feels similar an contiguous tension, that the champion acquisition of Duolingo is connected the level that wealthier radical use. How bash you resoluteness that?

It’s a bully point. First of all, mathematics and euphony are astir to beryllium disposable connected Android, oregon by the clip this airs, they will beryllium disposable connected Android. We are astir a twelvemonth down connected Android. This has been existent connected Duolingo astir since the beginning. Android has been six months to a twelvemonth down iOS. There are a fig of reasons for that, but astir apt the biggest 1 is that it has been harder to find truly bully Android developers erstwhile compared to iOS developers. There are conscionable much truly bully iOS developers, truthful we person much of them astatine Duolingo.

The mode we enactment is that we experimentation with astir caller features connected iOS. Because a caller diagnostic is usually not that large disconnected the bat, you usually person to bash proceedings and mistake to effort to marque it better. By the clip it’s good, we larboard it to Android. That’s however we operate.

We recognize the value of Android. You are right. There are much radical with Android phones than iPhones. Generally, each features are going to marque it to Android, conscionable astir six months behind, and we consciousness good astir that. It’s besides easier to make connected iOS for a fig of reasons, not conscionable that determination are much developers. So that’s it, we’re conscionable ahead.

In retrospect, fixed the exertion that determination is today, possibly we would beryllium doing thing that is cross-platform wherever we make connected each platforms astatine the aforesaid time, but we’re locked into being autochthonal connected some ends. We person a autochthonal app for iPhones and we person a autochthonal app for Android phones. That was the champion happening we could bash 10 years ago, and we’re locked into that.

When you deliberation astir increasing the company, supporting aggregate platforms, that’s conscionable treble the effort each the time. Is it connected your caput that, “Okay, we’re going to intentionally dilatory down improvement present truthful we tin support this squad smaller”?

It is, and we don’t person 1 immense task wherever we’re going to halt each improvement and beryllium like, “You cognize what? We’re going to present beryllium successful a azygous level benignant of thing.” But we are dilatory getting there. I don’t cognize however agelong it’ll take. The hard portion with this is that if we were to commencement from scratch close now, the determination would beryllium clear, but you besides person to support the level going. It’s specified a large concern to bash this that we volition astir apt person to halt each improvement for a twelvemonth and a fractional oregon something. I don’t adjacent cognize the timeline due to the fact that it’s conscionable truthful big. So far, we person decided to bash this piecemeal alternatively than each astatine once.

Your premium subscription tier, the Max tier, lone conscionable arrived connected Android.

Literally successful the past fewer days.

One happening I’ve heard implicit and implicit again since the dawn of the modern smartphone is that iOS users walk much money.

Yes, 4 times arsenic much. At slightest for Duolingo, a fixed idiosyncratic spends 4 times arsenic overmuch per capita.

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

Yes. More of our wealth comes from iOS than from Android. Even though we person much Android users than iOS users. It’s conscionable hard to flooded that 4x.

Is that demographically that they person 4x the income? Or is it that demographically, iOS users are spending 4 times the wealth successful your app?

No, it’s 4x the income. It’s that a idiosyncratic spends 4 times arsenic much. We person much Android users, but it doesn’t equilibrium retired successful the end. We marque much wealth from iPhone. I’m going to springiness you a fig present conscionable to springiness you an idea. The divided of users is 60 / 40, truthful 60 percent Android, 40 percent iOS. The divided of gross is the different mode around. It’s fundamentally 60 percent iOS, 40 percent Android. Those are precise unsmooth estimates.

So erstwhile you deliberation astir expansion, again, a nationalist company, if astir of your users are connected Android and Android is the biggest operating strategy successful the satellite but each of your wealth is connected iOS, however bash you resoluteness that tension?

Is it that we’re going to marque our wealth connected iOS and not Android? The large ngo is to bring escaped connection acquisition to each of those different people, truthful is the iOS idiosyncratic subsidizing the escaped mission?

I mean, that’s 1 mode of reasoning astir it. It’s not rather existent that each our wealth is connected iOS. It’s conscionable that more money is connected iOS, that’s for sure. But it is simply a small spot true, careless of Android versus iOS, if you conscionable look astatine who pays for Duolingo astatine the moment, they are usually radical who are good off. They whitethorn not beryllium millionaires, but they are radical who unrecorded successful countries similar the US that are wealthy, countries that person salaries similar $100,000 a year. A idiosyncratic with a good, unchangeable occupation successful a affluent country, that is who pays for Duolingo. The radical who usage Duolingo for escaped are usually successful mediocre countries. They whitethorn not person a unchangeable job, truthful it is existent that we’re getting the affluent radical to subsidize the acquisition for everybody. That is the case, and that’ll astir apt ever beryllium the case.

Now, connected our end, we besides request to get amended arsenic a concern to get much radical successful immoderate of these processing countries to pay. As a bully example, Netflix has done a truly bully occupation of getting radical successful Brazil and India to pay. We person not done arsenic bully of a job, and portion of the contented is that we’re freemium. Again, I grew up successful a mediocre country. Even if the terms is scaled down to lucifer the GDP per capita and it’s overmuch cheaper, the occupation that you person successful a mediocre state is that the cognition is, “I won’t wage unless I person to.” That’s conscionable the attitude. It doesn’t substance if it’s conscionable a dollar, and I bash hap to person a dollar. I conscionable won’t wage unless I person to.

What you spot is utmost tolerance for ads. For example, we tin enactment 10 ads astatine the extremity of a acquisition and they inactive won’t pay. This is why, for example, Netflix does truthful good successful immoderate of these countries because, with Netflix, there’s nary free. They’re conscionable like, “Look, whatever, you person to pay.” And radical are like, “Fine, fine. I’ll pay.” So we person to fig retired what to bash arsenic a freemium merchandise successful these countries, and we person immoderate ideas, but the world is, we person not truly succeeded astatine beardown monetization successful countries similar India. 

Do you deliberation that that is the adjacent logical spot for you to turn arsenic you deliberation astir English education?

For sure, we are spending a batch of effort connected that. And it is growing, which is nice, but it’s a monolithic opportunity. Language learning is different comic happening wherever the largest marketplace is simply a state similar the US: affluent countries. Language learning arsenic a whole, not Duolingo, but connection learning arsenic a whole, the largest marketplace is really processing countries: the Indias, Vietnams, Brazils, and Mexicos of the world. They’re learning English, and that’s the largest language-learning market, but we person not cracked it. We person cracked the smaller one, which is the US and Western Europe and richer countries. We’ve cracked that successful presumption of monetization and successful presumption of users. We person a batch of users successful India; they conscionable don’t wage us.

I consciousness similar I person to inquire you astir the owl. It’s precise important to everyone that I inquire you astir the owl. At slightest arsenic expressed successful this country, the owl is a precise online, precise culturally defined character. If you took the owl extracurricular of United States societal networks and dropped it anyplace else, it wouldn’t marque immoderate sense. Is the owl expressed culturally successful antithetic markets, oregon is it conscionable 1 owl?

I don’t cognize however to reply the question. It’s successful between. We started utilizing societal media with the owl a portion ago. It grew mostly successful the US done TikTok due to the fact that the owl does unhinged worldly connected TikTok.

Wait, clasp on. The owl doesn’t bash anything. How large is the squad that writes and performs the owl?

Five people.

Okay. And they enactment astatine Duolingo?

Yes.

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

Actually, no. Most of them are successful Pittsburgh.

Okay. I didn’t recognize Pittsburgh had this galore terminally online people. Godspeed.

Yes. But it started retired with TikTok and it was chiefly successful the US. That was respective years ago. What has changed successful that clip is, archetypal of all, we are nary longer conscionable relying connected TikTok. It is present connected YouTube, YouTube Shorts, Instagram, etc. So it’s everyplace connected societal media. That’s 1 large thing. The different 1 is that we learned however to localize this to antithetic markets. So we started Duolingo accounts for a clump of countries: Mexico — well, Spanish speaking — Japan, Brazil, Germany, France, China, etc. We person figured retired however to marque each of them succeed. I was dubious astatine archetypal erstwhile idiosyncratic told maine we were going to unfastened an relationship successful Germany. No discourtesy to Germans, but I thought, “These radical don’t person a consciousness of humor.”

Oh my gosh.

But it turned out—

They do.

They do! In fact, it’s 1 of our much palmy accounts. [The planetary accounts] are a small different. It’s not that different, but they are a small different.

And these are the aforesaid 5 radical locally?

No, we person a planetary team, which is these 5 people, and successful each country, we person a tiny fig of people, astir apt 1 oregon 2 people, that localize this stuff. And “localize” doesn’t mean we instrumentality the nonstop aforesaid videos and successful Mexico enactment a sombrero on. That’s not that. We person themes and we person figured retired what themes enactment globally and besides what themes enactment successful definite countries. For example, the German one, we had a truly large happening on Oktoberfest. Also, astatine immoderate point, due to the fact that there’s this dance nine scene successful Berlin, I conjecture they each went to 1 of those 24-hour creation clubs. Each state does antithetic stuff, and it’s worked retired beauteous well.

What’s the hiring process similar to beryllium the writer for the Duolingo owl? Do you conscionable work people’s Twitter accounts and say, “You’re unhinged capable to bash this”?

It’s a batch of that by present due to the fact that we are specified a beingness online. By the way, I didn’t cognize this until precocious but determination are weeks erstwhile our video connected TikTok is the astir watched video connected each of TikTok that week. That happens. By now, our accounts are truthful good known that we person our prime successful presumption of [recruitment], and a batch of radical privation to enactment for that team. Typically, we conscionable look astatine what they’ve done before. It’s a tiny radical of truly bully creators, truthful we prosecute from that group. Usually, these are beauteous comic radical that are adjacent funnier online, but erstwhile they’re offline, they’re not arsenic funny. They’re inactive funny, but erstwhile they’re offline, you’re like, “It’s you? It’s you who came up with that?”

And you measurement everything, it sounds like. Is this working? Are you getting tons of caller users due to the fact that of the owl?

Yes, this works. By the way, this is not paid. All that societal selling is not paid. It’s free. We marque our videos and they spell viral. About 15 percent of our users are coming successful from societal media. Now, if you look astatine societal media views of our content, which we measurement successful the billions, there’s a astir adjacent fig of societal media views of our contented versus the contented that is astir america but not made by us. Also, determination are a batch of different radical conscionable making contented astir Duolingo, but they’re not us, and they combined person astir arsenic galore views arsenic we do.

Have you ever told the squad to pump the brakes? Have you ever looked astatine thing they’ve made and said, “We conscionable can’t bash this”?

Yeah. There’s our review. There’s an support process. We’re adjacent to the enactment successful immoderate of the worldly that we person enactment out, and we person successful information gone crossed the enactment and published things that we shouldn’t have. Since we did that, we present person a beauteous strict support process. This is simply a full layer, and the past measurement is fundamentally me. But worldly doesn’t travel to maine often due to the fact that determination are radical earlier me, similar the CMO, truthful determination are a batch of steps. 

What’s the past 1 the CMO was like, “I don’t know. Luis has to o.k. this one”?

I’m trying to retrieve what that 1 was.

What’s the past 1 they convinced you to bash adjacent though you were skeptical?

I don’t retrieve the nonstop video, what it was, but I cognize that the past 1 that I approved, I was wrong, arsenic in, I shouldn’t person approved it. A batch of this you lone cognize successful retrospect. You don’t cognize until it happens due to the fact that you enactment it retired and past you spot this reaction. I don’t retrieve what it was, but I cognize I approved it and I cognize I was incorrect due to the fact that I didn’t ideate that it was going to person that reaction. We haven’t had that galore faux pas. It’s been similar 3 oregon 4 videos that were like, “You probably shouldn’t person done that.”

The different happening happened astir a twelvemonth ago. We had made this brainsick video. It was insane. We were a small hesitant astir it, and we ended up cutting it. There are each these memes online astir however the owl truly wants you to larn a language, and it goes to large lengths, including kidnapping your family. This was a video astir kidnapping, and we were a small hesitant astir it, and past the October 7th onslaught happened, and we chopped it. And past we chopped it past year, and we thought, “Well, you cognize what? We whitethorn usage it adjacent year.” This twelvemonth came on and again we chopped it. And past we came up with an interior happening that a twelvemonth erstwhile we tin play that, it’s astir apt been a bully twelvemonth for humanity.

Yeah.

Yeah. We’re astir apt ne'er going to play that.

The satellite discourse of that 1 needs to beryllium substantially improved, I think. All right, I person to extremity with a diagnostic request. You’ve fixed america a batch of time, and past I’ll fto you get retired of here. We talked a batch astir India. We talked a batch astir emerging languages. Can you enactment Gujarātī successful this app?

Oh, man.

This is the connection that I tin recognize and talk similar a baby, but I can’t work oregon constitute it, and I would emotion to adjacent the loop.

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

It’s a immense language.

I know, I know.

It’s the autochthonal connection of Gandhi, of the existent premier curate of India.

There is this unfortunate happening astir being a immense connection versus the tendency to larn it. It’s a beauteous large difference. Hindi is astir apt the 1 that has the astir tendency to larn it successful presumption of Indian languages. It’s a tiny fig of radical who are learning it. It’s got to be, I don’t cognize precisely disconnected the apical of my head, but it’s surely good beneath 1 percent of our learners are learning Hindi. I’m going to conjecture 0.1 percent of our learners are learning Hindi. That’s the hard portion astir adding languages, that we person to support them, we person to bash a truly bully occupation with them, and then, successful the end, we conscionable don’t get a batch of usage. So, sorry.

All right. That’s a hard no. It’s 1 of the archetypal times a CEO has fixed maine a hard no. That’s again, laminitis mode.

Well, it’s conscionable truly hard to accidental yes to. In the past, I utilized to accidental yes to this stuff, and we made a batch of mistakes adding languages that, successful retrospect, we astir apt shouldn’t person added.

Have you ever chopped languages?

We have. We cut, what was it? I deliberation it was Afrikaans. But the chopped successful portion was due to the fact that determination was precise small demand. The biggest crushed was it was conscionable a low-quality course, and astatine immoderate point, we thought this was a bigger marque hazard than thing else. We made the decision, we’re like, “Well, could we amended it oregon what?” And we made the determination it was not worthy improving.

Do you deliberation AI is going to assistance you adhd languages?

Maybe, but unfortunately, astatine the moment, AI is truly bully for large languages and truly atrocious for smaller languages. There’s a beauteous precocious correlation with the languages we offer. AI is precise bully astatine the languages we have: the Spanish and the French. It’s not ace bully astatine your Esperanto oregon Navajo oregon smaller languages.

AI is notoriously atrocious astatine math, oregon astatine slightest the existent LLMs are beauteous atrocious astatine math. Are they going to assistance you with that?

The bully quality is that, successful the constrained situation that we have, it tin assistance rather a bit. It’s been helping rather a bit. A batch of the information that we make for our mathematics people is with AI. The different happening is, immoderate of it is without AI, but it turns out, conscionable computers are bully astatine math.

It’s comic however galore times I’ve asked this question and idiosyncratic fails to bring up the thought that there’s a computer. I’m blessed you did that.

Computers are bully astatine math! And I recognize AI is not truthful bully astatine pursuing a signifier oregon whatever. It whitethorn not beryllium truthful bully astatine that. But the information that we make for our mathematics people is simply a batch of worldly similar fractions and multiplication. Computers are beauteous bully astatine generating that data.

All right. Well, Luis, you’ve fixed america a batch of time. Thank you truthful overmuch for being connected Decoder.

Thank you for having me. And large questions.