Duolingo联合创始人 Severin Hacker：AI如何影响未来的工作与教育

You should always try to raise money from the best Tier 1 VCs, even if the terms are slightly worse, because there's so much signaling. It's harder to raise $3 million than it is to raise $100 million. I can guarantee you Duolingo would not have been able to raise any money in Europe. We had 0 revenue for the first 5 years.

It was $3 million at $15 million. I think we only had 1 offer. It was either raise money or go back to university. We were ready to go.

Severin, I'm so excited for this, dude. I had Luis on 6 or 7 years ago, and I've heard so many great things before this show from Mihai Bing [?], from KP, Brad Feld, and from USV, among many others. Thank you so much for doing this.

Thank you for having me, dude.

1. Duolingo’s AI Pivot: What “AI-First” Actually Means

I would love to start with the news of the day. I don't like the whole context of “Tell me your life story” and normal podcast intros. Duolingo came out as being AI-first the other day, and I really wanted to start with what that means and what it does not mean, so people have a clear understanding.

Maybe taking a step back, why did we start Duolingo? What is our mission? Duolingo's mission is to provide the best education and make it universally available. That's why Luis and I started this company.

From day 1, if you think about it, if you want to build the best education, the best education in the past was only available to the richest people—the kings. They had private tutors for their kids, and that was the most efficient way to learn. By the way, this is still true today. If you want to become the best tennis player, you hire a 1-on-1 tennis coach.

You actually see this. I remember there was a US Open final, I think it was, and the 2 female players in the final were both daughters of billionaires.

Yeah, exactly. They got the 1-on-1 tutor. When we started Duolingo, we saw the potential of technology to do this for everyone. We can't all afford these 1-on-1 tutors; it's too expensive. But with technology, we believe there's a way that anyone can have access to the equivalent of a 1-on-1 tutor, with the same efficacy. That's why we started Duolingo.

Technology-first is what Duolingo was from day 1. Back then, it was software, and we called it software. Now we call it AI. In a way, not much has changed, except AI, as we now call it, has become a lot better.

2. The 12-Year Bottleneck Duolingo Crushed with AI

I totally get you. Take me to that discussion, then, when you're inside the room with Luis and the board and you decide, “We need to fundamentally change how we do everything and reshape the organization of this company toward being an AI-driven company.”

This has been going on for the last 2 and a half years. We were one of the launch partners of OpenAI when they first launched what was likely GPT-4, and we immediately saw the potential of this technology to help with our mission.

What was it that you saw that gave you such optimism?

We had early access, and we saw that you could really use it. I don't know when you first had access to it, but it was impressive. It was like an iPhone moment. You just realized, “This is the future. We're going to use this everywhere.” That was my first personal impression.

The second question was, “How do you use this at Duolingo?” The first thing we realized was that this could really help us accelerate content production. In the past, it would always take us a long time to produce new courses. It required a lot of effort to create these courses. With AI, there's the potential to just pump out a lot of these courses all at once.

In fact, we just did that. It took us, I think, 12 years to build the first 100 courses or so. Within 1 year, we built another 148 courses. That was just not possible before; it would have taken us decades to build this.

Can I be really weirdly granular? How are you using AI to 12x content creation within Duolingo?

There's still a lot of humans in the loop. For example, the curriculum design is all human-made. How do you structure the course? That's all human-made.

Inside Duolingo, there are a lot of short sentences—the ones you see in the lessons—and those are now produced with AI. There's also this concept of vocabulary: you want to introduce new words 1 at a time. You can give these constraints to the AI, saying, “Create a sentence that uses these words, doesn't use any of the other words, and only uses these grammar concepts,” and it can do that. That's the magic, and that's how we've been able to generate the sentence content within these courses.

With respect to curriculum design, when you look at students' completion rates and the data that you have in 6 months, along with success rates, satisfaction rates, and NPS scores, would it not be better to let AI do the curriculum design, knowing all of that, to actually build the curriculum itself?

That's a great question. I think the industry is definitely moving that way. One of the most amazing things about Duolingo is that we have this massive user base. We're by far the largest learning platform out there, by far—orders of magnitude—and we see how people learn. It's like the largest school.

Usually, in education, studies are done with 20 to 30 students. We have hundreds of millions, and that allows us to look at whether this curriculum design works versus that one. We can use that data to improve the curriculum design itself.

But not only that. Right now, we still have these courses, and your curriculum is the same as mine. Really, I think the future is going to be quite different. You might be interested in rugby, and you might have a trip planned to France. I think France is also big into rugby. You could then custom-design a course just for you that has exactly the vocabulary you want to use when you go to France, completely different from mine.

Why do we even create this batch on the back end? Why don't we create it on the fly when you need it? When you open the app, every single exercise could be custom-designed for you.

Personalization is the future of education, in your mind?

I believe so, and I think that's part of the reason why these 1-on-1 tutors are so good. They know where you are, they're really good at assessment, they know where you need to go, and they're very good at giving you feedback. They also know your interests and how you differ from other students. Think of the tennis coach again: they're really good at assessment.

I totally get you, and I agree. I was actually watching Zuck last night in an interview, and he was saying that the future of content is not passive consumption of videos like we have today. It's interactivity, where you will talk to the video and suddenly the fitness trainer will say, “No, you should have 12 grams of protein, not 8 grams of protein.”

It's that interactivity that will drive it. How do you think about personalization balanced with interactivity in the future of content and the future of education? To what extent is it multimodal?

One thing I love about Duolingo, which makes me sound weird, is that I'm walking around London and then I'm going, “Vorrei un Americano.” It's multimodal. It's my voice playing back, it's me typing, and there are many different modes.

Yeah, I think it's going to be multimodal, 100%. We have this feature inside the app called Video Call with Lily. You can call Lily. Lily is one of our characters; she's the one with the purple hair.

She's not a 1-on-1 tutor. She's more like a friend, but she remembers stuff. She remembers, for example, that I like cooking. She knows where I live, et cetera. She personalizes in that way, but it's a full conversation. You learn by conversing with an AI friend, and I think that's going to be super important.

There are always these modalities. Sometimes you're not able to talk—in every moment, you can't talk. Sometimes you're on the subway and you don't want to use audio.

How has the way that you build internally at Duolingo changed with AI? What tools do you use? How has it changed the cadence? How has that changed?

We use AI in 3 ways at Duolingo. Number 1 is content generation, and that's where we've seen the biggest success. It's completely changed how we produce content. This is the learning content.

Number 2 is AI features that you previously couldn't build. Again, this is Video Call with Lily—an interactive conversation with an AI bot that you couldn't build 2 or 3 years ago. Now we can do it, and it's actually the thing that was lacking in the product.

If you ask our users what's missing in Duolingo, it's this conversational piece: speaking. “I want to get better at speaking.” Now we can do this, and it's a really, really, really powerful feature. We have great adoption with the Video Call with Lily feature. That's number 2, and I think there's more to be done there, both making this feature better and building new AI features that we can now build.

The third is overall productivity improvements across the company.

So that's software engineers using Cursor. Then there's using AI for customer support. Do you mandate Cursor? Do you say, “Hey, we prefer you to use this or this”? How do you do that?

We don't mandate. Duolingo is surprisingly egalitarian and bottom-up, and we just say, “Use an AI tool or use whatever makes you more productive,” and we pay for it. You can use Cursor, you can use ChatGPT, you can use Copilot, and then we see what people actually use.

Where is AI? You said, again, LLMs are amazing at language.

Absolutely.

Where is AI much better than people think, and where is it not as good as people think and actually has quite a lot of room to improve?

Early on, there were a lot of these hallucinations—the AI making stuff up. Now, I think they have gone down, number 1, but then also, number 2, for us, it's not a problem if they hallucinate because we teach languages. As long as the language is grammatically correct, it is okay. If it makes up a historical fact, we don't teach history. Hallucinations are not a problem for our use cases.

I bet you have users, though, who are like, “No, it was 1914, not 1916.”

Yeah. We don't go that far into it.

But when you have so many users, you're like, “There's always one.”

Yeah. No, so hallucinations are not a problem. I would say there's a lot of Silicon Valley hyperbole around software engineering getting replaced very quickly. I'm less sure about that. I try all of these tools, and I think there's still a lot they can't do well.

What can they not do well?

On the software engineering side, I think what they're really good at is going from 0 to 80% for a simple app. It's like simple app builders. Then, the larger the codebase—even the one that generates itself—the larger the codebase gets, the worse they get. Adding the next 10% takes as long as the first 80%. They kind of create their own tech debt, which they also cannot really solve.

That might get better. The other thing they're really good at is isolated transformations. In a single file, you can add all the comments or change a function to use different input parameters or something like that. They're really good at that.

What they're not good at yet—and again, maybe this will change, maybe I'm wrong, maybe this will not age well—is large codebases and adding a feature, even a simple one, to a large codebase.

3. Will Duolingo Have More or Fewer Engineers in 5 Years?

Do you think you will have more or fewer software engineers in 5 years' time at Duolingo?

That's a great question. If you had asked me a year ago, I would have said 100% more. Now I'm a little bit less certain, but I think 2 things will change.

In a way, what these AI tools do is lower the entry level, the entry bar, to get into software engineering. Your mom can maybe build an app with Lovable or something like that.

She has to.

Right. So there you go. The other thing it does is lower the cost of lines of code, of producing software. Then there's Jevons paradox, which says if the cost of a good goes down, you will actually have more demand for it. I also believe some of this will be true.

I think there will be a lot more people writing apps. Whether they will actually touch the code or not, I'm not sure. The second thing related to this is, are we going to call them software engineers, or what are we going to call them?

Inside Duolingo, the 3 biggest functions are product, engineering, and design. Those compose probably about 70% to 80% of our workforce. But with AI, you can see that there's maybe a new role, kind of a product-engineer-designer—1 person who can do all of these roles to a certain degree. They can build a prototype, and then maybe at some point they need to hand it off to an actual software engineer or an actual designer, but I think that's going to be part of the future.

4. Will AI Kill the CS Degree?

So for everyone coming out of university today or contemplating university courses today, and who are told that CS in 5 years' time isn't going to be worth it—it's not? Do you still see immense value in CS as a principle and a mindset?

If I was your little brother and I'm 18, and I'm like, “Help me, Sev. Should I do CS at university?” what would you say?

I did my CS degree quite a while ago. They don't actually teach you coding. People think, “CS means coding,” but CS is actually—what they teach is the fundamentals of not even software engineering, but how computers work. Computer science is kind of applied mathematics.

I do believe there's still a lot of value in thinking logically, which is something that the best CS courses teach you. If you study this at university, they don't teach you Java or Python, and they shouldn't. I think that part is going to go away—all the details of actual coding—but the fundamentals of thinking about problems and problem-solving will still be valuable.

For the next 5 years, it's probably still a good investment.

If you extrapolate, once we have full AGI, we no longer have jobs, nobody has to work anymore, we have UBI, et cetera. It's kind of paradise. Do you actually buy that?

I believe that's going to happen. Again, the big question is when. I don't think it's going to be in the next 5 years. Maybe it's going to be in the next 10 years if you really see broad adoption of these technologies. More likely, maybe 20 years or so.

In that world, what is going to be valuable then? Today in the US, if you run into a stranger at a conference, the first question is usually, “What do you do?” Always. But in this future world, the first question might be, “What are your hobbies?” Because we don't have jobs anymore. Everything is going to be about not work.

I do question it, though. I think we can quite often get a little bit overexcited. When we had the agricultural revolution, it was like, “Oh my God, we had 100 people in a field, and now we have this combine harvester. We have 1 person driving it. We're all going to have no jobs.” Then we had computers and calculators. What are accountants going to do, and what are people going to do when we have them? We always find the next thing to do.

The idea that society will live in this UBI, lack-of-utility state—I hope that's not true.

I don't know how it's going to play out, obviously. But the other thing that's interesting about this, and this is more on a philosophical note, is if you just ask people, “Do you like your job?” Everybody here—I can ask anyone here—they all love their job.

You love your job.

I'm sure you love your job, right? You would do it for less money.

Most do. Dude, I did it for free for years. I literally get paid to hang out with amazing people. Are you kidding me?

Yeah. For us, it's not just income, right? It's purpose. We derive purpose from our job. But for a lot of people, the job is just something they do for the income. They don't derive purpose from it.

I think we are in this group—the people who derive purpose from their job, who work the most, et cetera: the tech people, the investors, and all of these people close to the AI revolution. I think we are most afraid of it because we feel we might lose our purpose. I think that's why we are sometimes not optimistic about this future world.

It's funny. I had someone likely named Jason Lemkin, who's a very famous SaaS investor, on the show recently. He said, “The hard thing is, with young people, they just don't want to work.” Fundamentally, they don't want to work.

AI is as good as they are at sales and at marketing, whether you're an SDR, whether you're a social media content creator for memes or GIFs, or whatever it is. In 12 months, we're going to see huge, huge unemployment among lower-level tech employees, and it's going to happen. Do you think that's true?

I'm not sure. Here's another thing. The general consensus is that these AI tools are really good for senior software engineers or staff-level software engineers because they have been trained, and now they can use AI and become much more productive. They're now a 10x engineer.

The general consensus is that AI can do the job of an entry-level engineer or an intern. Therefore, let's stop hiring entry-level engineers and only hire senior engineers. But I think that's a mistake, and we're actually not doing that at Duolingo. We still hire from universities.

One of the reasons is that we believe people who grow up with these tools and start using them early will become much better at using them.

So, in a way, yes. Why is that? Because young people have more plastic minds for new tools and new processes. What’s the thinking behind that?

I think it’s just like using these tools. I bet the person who spent 1,000 hours using ChatGPT is a lot better at using it than I am, and I think that’s the case for new grads.

Another thing that’s interesting is that Duolingo wouldn’t be where it is today in terms of our social media presence if we’d never hired from Gen Z. You need to know your target demographic. Zaria, who was on your show before, is Gen Z, and she just fundamentally understands it in a way that someone who’s not Gen Z could never understand it.

When I open TikTok and look at our own Duolingo stuff, I’m like, “I don’t get it. Why is this funny? Why do people watch this?” But it’s super popular.

Yeah, you’re not Gen Z anymore, Severin. I’m sorry, dude. You don’t fit in that demographic.

We mentioned how it changes engineering and product design. Another big area is customer support, and we chatted about it a little bit before. How does AI change customer-support processes and workloads within Duolingo?

We started using an AI tool for customer support. I think it’s another one of these early applications of AI where AI can really do a great job and transform an industry. We’ve found that AI can do 70% to 80% of the tickets that we get.

People are always like, “Oh, well, okay, you guys don’t need human customer-support agents anymore.” That’s not true. They’re still there because you still have the other 30% or so.

Not only that, now that you have this AI—kind of like an unlimited AI customer-support agent—you can actually give customer support to a much broader base. So far, we only give it to our subscribers, the ones that pay for Duolingo. But if we can reduce the cost of customer support by 10x or 100x, then we can actually give it to everyone, so everyone benefits from this.

Totally get it. Are you able to say which company? This is a classic example of a cost-down/demand-up paradox, right? As cost goes down, demand goes up. Who do you use for customer support?

We use Decagon.

Yeah, got it. I would love to have their CEO on the show. I think it’d be a fascinating one on the future of customer support. I’d love to do a panel with a couple of different providers.

5. Is AI Integration Causing Margin Pressure at Duolingo?

Seventy percent to 80% is really significant. Can I ask you, when you think about offering it to more people, and then also the implementation—you mentioned the features that you have, like implementing AI across the Duolingo product suite—it costs money at the end of the day. You have to pay OpenAI for a lot of these features. Do you see margin degradation by offering AI across the full product suite because you’re not charging more, but you are adding an additional cost from a model provider in terms of margins?

As I explained before, there are 3 use cases at Duolingo. Number 1 is content generation. Content generation is good for margins, right? It’s phenomenal for margins, and it’s also a one-off thing. You generate the content, then it’s done.

Number 2 is these AI features, like Video Call with Lily, where you actually have a unit cost. You directly interact with an LLM on the back end, and you have to pay for that API call, so there’s an actual unit cost there. You have to make that viable.

Right now, Video Call with Lily is in our Max tier. We have 3 tiers: the free tier, Super, and Max. It’s mostly in our Max, the higher-tier subscription, because of that reason. We would like to give this to everyone, but the reality is, today, there’s a unit cost to AI, and therefore it’s in our higher-tier subscription.

We believe all of these costs will come down. The batch-processing cost has already come down. There’s a lot of competition there, and in the real-time, multimodal models, we hope there’s going to be a lot of competition. Costs will eventually come down. I think we’ll be able to give this to a lot more people, or almost everyone.

When you think about model relationships, you mentioned obviously being early access with OpenAI. How do you think about the flexibility of moving between models, with different generations of models and different specialties? Do you think you will have relationships with 5 model providers, or do you think it’s, “Oh, no, we’re tied to OpenAI here”?

I think there’s going to be a lot of competition. In more specialized use cases, like visual models—or, for us, audio is very interesting, multimodal—there’s less competition, and you can see there could be companies that keep a lead.

For our voices, we also work with ElevenLabs, and they’re just amazing—amazing models, amazing voices. I can see that that might be harder to just replace with one model or the other.

When I hear you say ElevenLabs there, or Decagon there, or even think about Duolingo, the thing that strikes me is, “Ah, but OpenAI are going to do that.”

I’d love to go to the moment where I think Duolingo stock actually took a hit—and I never talk about public-market prices, so don’t worry, we’re not going there—but I think it actually took a hit when there was concern that OpenAI and ChatGPT would move into language learning very significantly. Can you take me to the internal discussion around, “Oh, shit, are ChatGPT and OpenAI going to move into the application layer for language learning?”

When ChatGPT came out, one of the first killer use cases for it was homework cheating, for lack of a better term. If you were in the homework-cheating business, that was devastating. That was really bad.

Some of our competitors—I’m not going to name them—call themselves education, but they’re really more in the homework-cheating business.

I love this homework cheating. I’ve never heard of this sector of the economy before, but let’s roll with it: the homework-cheating business.

Yeah, they struggled with that. Then I think there was actually a similar thing where people said, “Oh, well, Duolingo is in education. Therefore, they’re going to be disrupted.” Obviously, the opposite is true because we’re not in the homework-cheating business. We’re in the language-learning business—or, you know, the general-education, actually-teaching-you-new-things business—and that’s completely different.

The other thing that I think you see happening in the public markets—and this is my only comment about this—is this bifurcation of the market into, “Is this an AI winner or is this an AI loser?” I think, by mistake, we were first put into the AI loser category, and now it changed to, “Oh, no, Duolingo is going to be an AI winner,” and there’s a lot of tailwind from AI. I outlined some of this.

So, there was never a concern for you internally? You never had a call with Luis where you were like, “Oh, shit, is Sam going to do this?”

Well, I think nobody is stopping them.

I assume here you mean OpenAI or ChatGPT building a really good language tutor. One hundred percent. I’d be like, “You know what?” Can they take away a lot of your very casual users by doing a very simplistic “chat with a friend, choose your language”?

Yeah, that’s definitely somewhat of a concern. I think there are a bunch of things to be said about that.

First of all, what is the hardest part about learning a language?

I would say the discipline.

Yes. It’s the actual motivation. By the way, this is one of the key insights of Duolingo. We’ve been sharing this with the entire world for the last 10 years, and still nobody seems to believe us: the hardest part is motivation.

Duolingo, because it’s so gamified, is really good at this. It’s a motivation engine. It really gets you to do this. It gets you to come back every day and do Duolingo. If you stick to it and do the hard material, you will actually learn a language. That is the power of Duolingo, and in general, people underestimate how important that is.

If you want to build something that is as good as Duolingo, you have to solve this somehow. Also, if you ask our users what they would do if Duolingo went away, they never mention any of our language-learning competitors. They say, “Oh, I would spend more time on social media.”

The hyper-casual users love Duolingo because it’s so motivating.

Well, if you're sitting there as Sam, you're thinking there's a couple of core areas which are winners: chat, which we already own—the chat interface; customer support, which I think they'll buy someone very soon; coding, which they've obviously just bought someone for; and I'd say language learning is the fourth. As you said, it's so perfectly aligned—I mean, LLM is kind of in the name. It is so perfectly aligned to the use case of what they're best at. It's not a stretch.

Yeah. Well, we'll see. We shall see.

The other thing is that Luis always says there's a difference between being a first-class citizen and a second-class citizen as a feature, right? Duolingo works because we have these gamification mechanics: the streak, the XP, the leagues, and all that. You can build all of this inside a different app, but if it's just a second-class citizen buried somewhere, it's hard to make that experience as sticky as if you built an app just for it.

Is there anything within Duolingo that's a second-class citizen that you wish you could make a first-class citizen?

There's one thing, maybe. We have these stories inside Duolingo, so every now and then you get a story. I think they're actually becoming really good again with AI, and there's kind of an entertainment aspect to them. Maybe you could make them more prominent and go into more reading, including your first language, L1, or maybe something for kids, like learning how to read. But we haven't done that yet.

Yeah, when it's not at the level of Harry Potter. One thing that strikes me when I use Duolingo every day is exactly what you said there: the little things that make the product experience delightful. It reminds me of someone likely named Gustav Söderström from Spotify, who always told me, “The details are not the details; they are the product.” There are so many little things that make me feel that when I use Duolingo.

To what extent do you believe that and advise that to founders who call you, versus, “Just don't spend too long on the way that something pops up; just get it done in the early days”? It's a fine balance.

I think ultimately product quality is details, and Luis is obsessed with details. He can spot something that's not properly aligned within half a second. The details really, really matter.

If you look at retention, which is the main metric we care about, the details really matter. It also matters in terms of how likely you are to recommend it to a friend, right? You're not going to say, “The details are really good,” but, “It's a great app,” or, “It's fun,” and all of that. That's all details. That's the difference between a good app and a great app: details.

I love that. I also love the first-class, second-class attention to detail.

Yeah, yeah.

Can I ask you where you spent time that, with the benefit of hindsight, you're like, “Ah, we shouldn't have spent time there”?

One thing that comes to mind is that when we started with languages, the mission was always all of education. At some point, we thought, “Let's build new subjects. Let's add new subjects.” Then someone said, “Let's add math.” Math is another one of these fundamental skills that we'll probably need to learn for the next 100 years or so. That's our belief, or my belief, so let's add math.

The first idea was to build a new app. In fact, we built a new math app completely standalone. There was Duolingo, the main app, and there was an entirely new math app.

Completely standalone?

Completely standalone.

Did it have a name?

Well, yeah, it was in the App Store. I'm actually not sure if it's still in the App Store. It was the Duolingo Math app.

Creative. Well done, Severin. Wonderful.

Then we realized this was looking more and more like the main Duolingo app. We were about to add a leaderboard and league system and streaks, and it was like, “Why are we doing this? Why are we replicating this whole functionality?”

That was insight number 1. Insight number 2 was that, for the user, it was also not a great experience because, first, you had to know that Duolingo had a math app. Then you had to find it in the App Store, download it, install it, and grant it permissions again to send you push notifications. Then it was like, “Wait, is this streak different from my Duolingo streak?” It looks and feels very much like Duolingo, but it's not quite Duolingo.

Then we said, “No, no, no more new apps. If it's education, it goes into the main app.” Internally, we called this a super-app strategy. We probably spent 2 hours working on this separate app. The good thing is that we could just transfer all the content and merge it with the main app.

6. Chess in Duolingo

Can you talk to me about chess? We mentioned it outside, and it was a great story. How did chess come about?

Chess is an interesting one because is it really education? It helps you think. We had this discussion before about whether you should still study computer science. I think chess is similar in a way: it helps you think and helps you think logically. We believe it fits within the mission of Duolingo.

Then it was 2 people who were just super passionate about chess, and it was like, “Hey, let's try to build a chess course.” There were already 2 people internally.

Yeah, yeah.

We were not convinced at first.

Why were you not convinced?

We were like, “This is just a game. We don't want to build just a game, right? There needs to be some educational aspect.” That's always a trade-off between, okay, are you just a game, are you an educational game, or are you just a textbook, boring experience?

With chess, it's different. Chess is not like a racing game or something. Chess helps you think, so that's why we're good with it. But then it's like, how does this fit into the path? How does this look inside Duolingo? Can we make it fit into the Duolingo design language and all that? That was another concern.

This is, again, super interesting, and I want to share this story. The 2 of them built this prototype with AI.

Okay.

There was a PM and a designer, and the PM had some engineering background, so he knew how to think logically. He hadn't written much code before Duolingo or at Duolingo, but he was able to build this first version just using Cursor and a bunch of AI tools. The prototype is what convinced us that we should do this.

If it had just been a Figma design, we probably would have said, “Ah, maybe, maybe not.” But the prototype was so good that it convinced everyone. It was like, “Yeah, let's build this.”

What was it about the prototype that was so good?

It was fun. It was progressive in terms of the mindset and the way you could play it, and it just felt like this belonged inside the app. When you just see a screen, it's like, “Yeah, maybe you have to imagine it,” versus if you have the prototype, it's like, “Well, this feels and looks like Duolingo.” Because of AI, again, we were able to build this within 9 months. This is the fastest we've ever created a new course.

What do you mean?

9 months from start to finish.

Yeah, wow.

Normally, it would be, well, years.

You said that you've added languages, you've added math, you've added music, and now we have chess. What's the next one that you'd love to have?

For me right now, I think what I'm most passionate about is the social aspect of the app. We have this leaderboard mechanic, but we now have this massive user base and all this brand love. People love Duolingo, and I feel like there's untapped potential in connecting the people who are on Duolingo. I don't know how to do it, but I think there's a lot of potential there.

What would that look like? Duo Social? Duo Date?

Duo Date's not a terrible idea.

It's not.

7. Would Duolingo Open a Physical School?

Well, yeah, I don't know. It could be something like that. It could be learning together. But it could also be something else. I don't know. I think there's a lot of untapped potential that not that many apps have.

Would you ever open a physical school? I know that sounds completely counter to what Duolingo is all about, but having, in New York, a branded, amazing experience where people could go for a week, meet the owl, sit with people from their leaderboard, and build that community and play—that is really special and different.

It's a good question. I'm not sure about that directly, but one thing that's interesting about education is that there's K–12, and then there is higher ed: universities, college, et cetera. One thing that's interesting when people talk about education is that we always assume it's the instruction. It's like, here's a skill, I learned this skill, and now I go from A to B, right? That's kind of how we think of education.

But higher ed, for example, is actually more than that. Going to college is 3 things. There is the actual instruction, the actual teaching, and the learning of new skills. That's number 1.

But then there is also the degree—the credentials, right? The university says, “Harry, you got a degree. You successfully finished this course, and you can put this on your LinkedIn or résumé.” Then an employer can say, “I guess Harry did take this computer science class, and therefore we can hire him as an engineer.” That's credentials. Number 2, very important, because otherwise you would always have to prove again that you know something, right? That's the purpose of the degree.

And number 3 is the social aspect. A lot of people meet their partner at the university, and it's their friendship network. Most companies in edtech just focus on number 1. But I think if you really want to disrupt the Harvards of the world, you have to solve all 3. That's 1 thing, by the way.

Do you want Harvard?

No. We want to achieve our mission. That's why we started this company.

But by achieving your mission, are you not disrupting Harvard?

Yeah, we would be. But on the second thing, we have not solved the social thing, right? That's what higher ed does. We have not solved the social thing, but on number 2, the credentials, 1 thing we're also very passionate about is the Duolingo Score.

I'm not sure if you noticed it within Duolingo, but we now give you a score for how good you are at a language. We want this to become the way people talk about language proficiency. It's like, “Hey, my Duolingo Score is 61 in French,” and then people know, “Okay, I guess you can have this kind of conversation,” et cetera. We're getting into the credentials or the degree aspect of it with this score.

It's actually pretty helpful. The number of times you're asked, “Do you speak French?” and you're like, “A little bit.” It would be helpful to say, “I'm actually a 40,” and they're like, “Okay, not really. He's stupid,” or whatever it is. Do you know what I mean? It quantifies it a lot more easily.

Can I ask you—and this is jumping around—but you mentioned the push-notification element. Does that still make a difference? We have so many push notifications, with the oversupply of push notifications. You're like, “Ah, that was ruined. Was it good or not?”

It still makes a huge difference. We know because whenever the service that sends the push notifications goes down, you can see the dip.

Really?

Yeah.

Yeah. 100%.

It's huge. It's massive.

Yeah. Wow.

Going back to the other thing, there were the 3 things for higher ed. For K–12, people also say, “Okay, my kids go to school to learn things,” but again, it's not just the learning. In fact, a lot of K–12 is child care. Even if you solve 1 aspect of it, like if you have an AI tutor that is really good, you still need to solve the child-care problem.

It's unlikely that people will just go to a school without any humans or teachers and just use apps and then go home. That just doesn't work. I don't think we would have some sort of Optimus in a school. A lot of it still needs to be done by humans.

By the way, I do believe that AI tutors will become really good, and that it's going to be the way most people learn most things in the future. I do believe that.

However, what does the future of schools look like then? Is it 10 children sitting in a class with AI tutors?

I think there's still going to be this human element of guidance and coaching and motivation: What do I need to learn? Also, why should I learn this? An AI will never tell you why you should learn something. It will help you learn it, but it's not going to tell you why you should learn it or why it's important.

Do you know what I wish you had? Sorry. You have leaderboards, which are social-status or social mechanics that drive motivation and incentives. I'm sure you've studied this. I have, because I'm a loser.

Financial incentives are interesting. When you offer someone money, they are much less likely to do something. When you take money away from them, they are much more likely to do it. Basically, the fear of losing is much greater than the gain of winning.

I wish Duolingo had this so that, for every day you don't do your streak, it would take a committed amount off you. Say I commit to £100 a day.

Dude, I would do that. 100%.

You can just give me the money whenever you lose your streak. That would be the best side gig. Another £100. Harry's just giving me money. Turn off his push notifications, guys. No more.

Yeah.

The streak is exactly this loss-aversion mechanic, right? It's motivating to build up your streak, but you absolutely do not want to lose it. You'll do a lot of things to keep it.

You had crazy stories of people who do crazy things.

Oh, yeah. It's crazy. We also sometimes get emails like, “I was on this trip to climb Mount Everest, and I couldn't extend my streak. Can you please restore it? I had no internet access.”

It depends, by the way. When there was the power outage in Portugal and Spain, we proactively extended people's streaks so they wouldn't lose them.

It is amazing, isn't it? I mean, that's so cool.

Listen, I do want to talk a little bit about how you work, because you're fascinating in how you work. I saw in some of your posts that you said you're not good at finishing the final 20% of a project or a bit of work. That seems problematic, doesn't it? How do you reflect on that?

It would be problematic if it were just me and everybody was like me, right? Ultimately, you need to have complementary skills within the founders, but then also within the whole executive team and the whole company.

If everybody were really bad at finishing the last 20%—the last 20% are basically details, by the way—then you need to have someone who's obsessed with details. In Duolingo, that is Luis to this day. From day 1, he was obsessed with details and really cared about them.

It's not that I don't care about details, but he is an order of magnitude better at that, I think.

Has he ever been obsessed with details to the detriment of the company? Dude, we don't need to obsess over it that much. Come on.

That's a yes, but I'm not sure whether I should say it. He mostly works with the PMs or the product team, and they would probably say something like that.

We still have this product-review process where every change goes through review, and Luis sits in most of those. He still does. That's founder mode for him: he's still in most of the product reviews.

I think that's rare for the CEO of a company of our size, where the CEO sits in most of the product reviews.

So you decided to delegate that to the product team on whether he obsesses or not. I love that.

You said “founder mode” there, and it drove so much engagement and discussion. How do you reflect on it? When you look at your own style of management and leadership, how do you think about that?

My role has changed practically every year since the beginning, since we started.

Tanzeem told me this. I spoke to him at General Atlantic. I stalked you. I told you. He said your role has changed so significantly.

Yeah. So early on, I was implementing the thing and building Duolingo. Then it morphed more into code review, architecture, how we build the backend system, and all that system design. Then it morphed more into management and hiring.

I think I'm actually quite good at hiring, particularly hiring engineers. But then there's also other stuff that people wouldn't expect from a CTO. I was the one who pushed for hiring our first product manager.

It was actually interesting because Luis, at the time, was like, “Oh, we don't need product managers.” His previous company was acquired by Google, so he had seen and worked with some of them at Google. He said, “I don't think they provide that much value, and we don't need this. I can do this myself, and we just need a bunch of designers.”

I said, “No, no, we really should. If every single company has a product manager at our size, maybe we're wrong. Let's hire one, and if it doesn't work out, then we can still go without that.”

Obviously, that was the right decision. So, hiring: what do you think makes you good at hiring that is non-obvious or not typical?

We have a really strong consumer brand. Literally, everyone knows Duolingo at this point, worldwide.

Given the strength of Duolingo's brand, do you think Duolingo is underpriced today?

I think we dramatically underprice brands. When you think about global brands that billions of people know, they're underpriced in terms of monetary value.

Like, “Oh, the stock price?”

Yeah, I cannot talk about that. What I mean is, it's just amazing.

Do you see what I mean by that? People often think about Manchester United. It is a global brand, for example, and I think it's drastically underpriced, given the fact that you have children in India, Nepal, and Brazil wearing a Man U shirt, and it's a $5 billion company.

I don't know. I think Duolingo is a massive brand. At some point, we were bigger on TikTok than Nike, which is a brand that has existed for a very long time. We all obviously know Nike, so it is a very strong brand.

The other thing is that I always say we run the world's most efficient marketing organization. People think, "Oh, yeah, they're big because they spend a lot of money on advertising and social media influencers and stuff." Actually, by impact, we must be the most efficient marketing organization. It's a small team.

Give or take, I think 40 people or so. The reason for it is that, for a very long time, we had no marketing budget. We didn't advertise—nothing, like a $0 marketing budget—because we weren't profitable. We also didn't have any revenue. So it's like, no, we can't spend money that we don't have.

That was the constraint, but then the constraint was really good at forcing us to come up with good ideas. We always say Duolingo marketing is not powered by money. It's powered by ideas. Good ideas. It's what you see on our TikTok channel: having the green owl, our mascot, having the costume, and sometimes doing unhinged things, and also making fun of ourselves in a way. That is why it is so effective.

8. Which Conventional Rules Are Worth Following?

Yeah. I love that—the creativity that comes from the constraints. You mentioned not having PMs, and if all the biggest companies have them, maybe there's something to it. It makes me think of this: conventional rules are conventional for a reason, sometimes.

What conventional rules do you find are conventional for a reason? 1. And then, what 1 conventional rule do you find yourself thinking, "What the fuck? That's just the way it's been, but it's not right"?

First of all, I would say you should always—not always, but every now and then, especially when there's a big technology shift like now—question all of these assumptions, right? How you build companies. I think we're in the middle of this massive transformation. You can build companies with probably 10 times fewer employees. So that convention is completely overthrown: that you need to hire a lot of people.

That's 1 thing. Early on, there was this belief that you can't run a completely flat organization: no managers, no hierarchies, no levels. I think that's not really possible. At a certain scale, you need to have managers or layers, or you need to have some sort of hierarchy. You need to have some sort of career progression and stuff like that.

I think the oldest organization in the world that is still around—the Catholic Church—is very hierarchical, and I think there's probably some reason for it.

Does that not go against founder mode and Luis sitting in the product reviews? When you think about managers on managers and layers, and the diffusion of knowledge and power that comes with those layers, does that not go against founder mode?

Yeah, I think it's always this: how do you strike the balance between the 2, right?

What do you still do today that you don't think you should do?

For me, I listen to every single show we do.

Yeah. 10 years in. Why?

It's because it's my product. I'm not shipping product that is not perfect. I'm actually quite good at delegating things and having self-awareness of what I'm good at, what I'm bad at, what I like, what I don't like, what is beneficial to the company, and what is not beneficial. That's also why my role has always changed, and 1 of my principles is: reduce, automate, delegate.

Can you walk me through that?

The main idea is, maybe once a month, once a quarter, really think about: does this need to be done at all? What happens if you don't do it? If you just don't do it, is it the end of the world, or is there an actual negative effect? So that's number 1. In particular, what happens if I don't do this?

Number 2 is, what you cannot reduce, what you cannot remove—let's say you decide, okay, this actually needs to be done—then the second question is, can I automate it? Now, again, with AI, it's a lot easier. Can I automate this part of my job? Can I use ChatGPT to do this report or answer this question?

That's the second. That's software engineering; it's very natural. Can we build a tool that does this for me, or the company, or whatever? What you can't automate, you delegate. I try to delegate to other people in the organization. I've handed off most of the day-to-day engineering to our head of engineering. She's really good with this. I'm now a little bit out of the weeds.

I'm thinking more right now. Actually, I'm only focused on 2 things. 1 is this question with AI: what's going to happen? What does the future look like in our space? What are the implications for Duolingo? What actions do we need to take? That's number 1. I probably spend 80% of my day thinking and acting on this AI question.

Number 2 is more M&A. It relates to the first question: M&A—what companies should we buy? Where do we invest externally? That's basically all I do right now.

Where have you not invested that you would like to invest?

We did make a bunch of acquisitions.

How many have you made?

Not that many. 4 or 5.

Any lessons from doing that?

It's hard.

I remember having someone likely named Dax Dasilva on from a payments provider in restaurants. They've done about 18, and there were a lot of lessons there.

It's actually very interesting. 1 insight, in hindsight, it's kind of obvious, but they say the best predictor of 2 founders working out—that they're not going to get into a fight or founder conflict—is whether the 2 founders worked together before they started the company. The number 1 thing to look for is: have the 2 founders worked before they started this company? Have they worked together?

For example, Luis and I had been working together on research for 2 years prior to starting Duolingo. By the way, it's not the same as friendship. You can know someone for a very long time, but they might be very different at work. Prior work experience of working together—that is number 1.

With M&A, it's the same thing. I actually think it's: do you have prior experience working with this company, yes or no? I would say that's the number 1 advice: have you worked with them before? Do you incorporate that? It doesn't have to be a customer-vendor relationship; it can also be an investment relationship.

Do you do Duolingo Ventures? Do you invest early?

9. How Do You See Competition and Value Distribution in EdTech?

We just made some minority investments, but it's not a lot. It's always with a view to buy, essentially. We have to figure this out. It's a competitive space.

I've seen many competitors. I remember, I think you came pretty close to doing a round in Praktika. How do you view the competitive landscape today and the distribution of value across it? Is it like Uber and Lyft, where Uber takes 90% and everyone else takes 10%, and that's Duo and everyone else? Is it a much more distributed value dispersion? What does that look like in consumer?

It's much more winner-take-all kind of mechanics. We've now been running this for 12 years. I would say every single year there was 1 or 2 companies that grew really fast in language learning. 1 or 2. Inevitably, they go down again.

What happens is they use venture money or some other funding to grow through paid acquisition. So they spend it on Facebook ads, Google ads, and they grow really fast. But ultimately, the only thing that matters in our space is retention. You can spend as much money as you want. If you don't have retention, it's a leaky bucket. All these people are going to leave.

When we see 1 of those now, we're not scared, because we've seen this play out a million times. The 1 company I'm scared of is 1 that has higher retention than Duolingo.

What's the day retention number that matters to you? Of the millions and millions of people that use Duolingo today, how many come back tomorrow?

Got you. All of these correlate with each other. If your D1 retention is very high, you also have very high D7 retention, and you also have very high D30 retention, but you get a data point every single day.

That's absolutely fascinating. It reminds me of someone likely named Alex Schultz, who says exactly the same thing: retention, retention, retention. So I totally get you there.

10. Do You Think Europe Is Becoming Irrelevant?

I do want to ask, before we move into a quick fire: you're a European. You are also based in New York. When you look at Europe today, I sit here and I'm very aware of our fading into irrelevance, which is why we do things like Project Europe.

How do you think about Europe? Are we fading into irrelevance? And how do you see that, having the perspective of being a European in the US?

11. The Real Value of Tier 1 VCs (Even at Worse Terms)

I'm very conflicted internally. I want Europe not to fade into irrelevance, of course. However, if you asked me, “If a young European founder came to me and said, ‘Hey, I have this brilliant idea of how to use AI, or a new model, or something like that. What should I do?’” I would say, “100% go to Silicon Valley. 100%. You absolutely maximize your chances.”

By the way, this is why the environment is so much stronger. I can say this because we built Duolingo in Pittsburgh. We're not a Silicon Valley company. We don't have an office in Silicon Valley, and we succeeded despite not being in Silicon Valley. There are also pros to not being in Silicon Valley, and we can talk about this, but I can guarantee you Duolingo would not have been able to raise any money in Europe.

We had $0 revenue for the first 5 years. Zero. There's no European investor who would have invested—not in Switzerland, not anywhere in Europe.

In fact, in the US, do you think that's true today? I'm not being rude, dude. You're a really smart guy. You came out of great institutions. Luis, I know, is a super-smart guy too, out of great institutions. There's so much money in Europe now. Dude, I'll give you a term sheet today.

I think it has changed to a degree, but you still need to look at where the most ambitious people and the founders go. It's a little bit sad because, in Europe, I see a lot of European founders and I often think, “How successful would they be if they had stayed in Europe?” For most of them, I would say at best half as successful, or maybe some of them 0%.

I would like to change that. That's why I invest in European VCs. I'm part of Project Europe. We need to change that.

How do you think, then, about Synthesia, ElevenLabs, and something incredible like Lovable in Sweden being built from Europe?

That is great. They deserve all of our support. The sad part is also that, in Silicon Valley, being a tech founder is really—you’re mostly a good person. Of course, it's also inspiring, right? People want to be that.

In a lot of Europe, people are like, “Why are you doing this?” They look at it as, “Oh, yeah, they're just doing it for the money.” There's almost a suspicion that you can't be too successful, or that you shouldn't be too ambitious. I think that is very detrimental to the European startup ecosystem.

You shared an insight with me before that I'd never thought about, about having offices in San Francisco as a non-San Francisco-headquartered company. What was that insight, and how did it lead to your not wanting to have an office in San Francisco?

Duolingo does not have an office in Silicon Valley or anywhere in California. The reason is that, when we thought about opening one—and, of course, all of our investors said, “You've got to open a San Francisco office”—we did go to San Francisco and looked at office space. We were this close to opening an office there.

Before we pulled the trigger, we asked a bunch of founders who had companies headquartered outside of Silicon Valley and then opened an office in Silicon Valley. We asked them, “What do you think? Was it a good decision or a bad decision?” All of them said, “Worst single decision ever.”

I said, “Why?” They said, “What happens is it creates this internal funnel of your best employees moving from your headquarters to Silicon Valley.” By itself, that's not a problem. The problem is that they then get recruited away by whatever the hot Silicon Valley companies were at the time. It could be Airbnb or Uber at the time; maybe now it's OpenAI or Anthropic.

You basically create a hiring funnel for Silicon Valley, and you lose your talent. That was also new to me, and I thought, “No, let's not do that.”

I love that. I hadn't thought about that at all. So you will not open an office in San Francisco?

Nope.

Do you think I would be more successful if I was in San Francisco? I know that's an unfair question to ask, but it's something that I've always been told, Severin, for years and years: “You have to. Why are you in London?”

As a founder, I would say 100%.

Let me tell you my rationale, and you can see if you think I'm right and you can call bullshit on me. In San Francisco, I'm competing against Peter Fenton, Marc Andreessen, Vinod Khosla, Elad Gil, the Collisons, and Daniel and Nat Friedman. I rate myself. I buy my product. But that is a fucking hard market.

Marc Andreessen getting you around to his house for breakfast on a Saturday morning—and, rolling into Europe, it's probably 2 good VCs.

Yeah. By the way, I started angel investing with a very similar thesis in Pittsburgh, because there was no one other than me and Luis. But then a lot of those companies eventually moved to Silicon Valley. That was exactly my investment thesis in Pittsburgh.

But, so, I think as an investor, it's not clear. As an investor—and, you know, doing the show—I think it's not clear because, again—

Yeah, I think that this is much less crowded. You can really become the top investor that founders want to work with in Europe.

I think I agree, but I was explaining this as a founder. If you want to start an AI business and you have this crazy idea, would you stay or not? I would move.

But that's exactly the problem. Part of the problem—this is a little bit not obvious—is that it's so easy to move to Silicon Valley. If you have a good degree and you can raise money, the US is actually really good at getting these brains to come to Silicon Valley.

In a way, the best thing that could happen to Europe is if the US made it really hard to immigrate for founders. If they restricted immigration such that people like me couldn't move there anymore, then I would build it here.

What I would do first is say, “Let's first change the environment here to make it easier for startups to succeed,” because right now there's also no pressure from the industry to change the regulation to make it easier to start companies, because I can just go. I can vote with my feet.

What would you change about Europe? Is it the regulation? Is it the level of capital? Is it the mindset that we tell young people?

All of the above, obviously. All of the above. Regulation is definitely hurting. The regulation is definitely hurting.

No great entrepreneur was stopped by regulation, though. Travis wasn't stopped by regulation with Uber. Elon with SpaceX wasn't stopped by regulation. Regulation does not inhibit the greatest entrepreneurs.

I disagree. Again, I think it depends if there's an alternative or not, right? As a founder, I can say, “Okay, I can work with the EU regulation on AI,” which is, I think, some of the dumbest regulation ever written.

It's literally like the Ottoman Empire when they invented the printing press: “Oh, this is not good for our community. It's not good for us. Let's not do the printing press.” I feel like the EU AI Act is that. It's like, “Let's not do AI,” because there could be negative repercussions or negative side effects or something. It's that dumb.

If there was no alternative, then I would agree. By the way, in Uber's case and Airbnb, there's a lot of local regulation. There are actually lots of alternatives. Even if 1 city doesn't work—say, New York—this is the regulation; you have 200 other cities in the US.

12. Two Biggest Mistakes Duolingo Made

If you could go back to the early days of Duolingo and do something differently, what would you do?

We're pretty open about this. I think the 2 mistakes we've made at Duolingo—they were not fatal, but obviously, at 20 billion, I would say the 2 big ones were, 1, we waited too long with monetization. We just didn't take it seriously enough early on, and we were operating under this business model of, “Let's just raise more money.”

People would ask, “What's your business model?” We'd say, “Oh, it's like a venture capital.” Eventually, we did take it seriously, and we're very good at this now. We have this internal concept of the Green Machine: you run a lot of experiments, see what works, and then double down. That's what we did with monetization. It's super-good now, but we waited too long.

Was that because of naivety? Was that because of a push for user growth and engagement?

Yeah, a combination. It was also that the investors didn't push us to monetize early, which was counterintuitive to me. Why didn't they ask for monetization? But I never would, because I'd want engagement, user behavior, and habit-forming.

Don't put a paywall in front of them. Our investors also invested in Twitter and some of these big platforms, and for those platforms, it was always the right decision: just grow users first. That's what we did.

Number 1 is that we knew we could always grow users more. Number 2 is that we knew we could always raise more money. The other one was our mission: to provide the best education and make it universally available. For a long time, we didn't know how to build a business model that was compatible with this mission.

For a long time, it was, “Well, everything has to be free.” But then we realized, no, what really matters is that we don't exclude people. We want to be inclusive. You should be able to learn everything that's on Duolingo without a bank account, and that's still true today. In fact, most of our users do not pay Duolingo; they use the free tier.

That's actually another thing I'm very proud of. It's not just a great business; it's a great business with a great purpose. But that's also why we probably waited a little bit too long. We should have started 2 years earlier.

Number 2 is that we also waited too long to hire more senior managers. Up until around 30 employees, it was completely flat. It was a lot of people from CMU, and it was pretty chaotic.

That goes against everything everyone says to me on the show. Everyone always assumes, or says, “Don't hire the senior manager if they've been there and done it before. The playbook's not transferable. You want youngsters that you bring up through the ranks.” They want to be involved more in the weeds and the hands-on. Beware of the manager.

Again, you have to strike the right balance. You cannot run a 1,000-person company with all new grads. You need a little bit of both. By the way, today Duolingo is still a very young organization, so we still hire a lot of people directly from university.

You mentioned you always knew that you could raise more money. I love that. What was the business model?

Venture capital. Harry's like, “Thank God I still have a job.”

When you think about the fundraisings that you did, what was the hardest? Was it a Series A? Talk to me about that.

We started as a research project at Carnegie Mellon University, and the belief was, “These are 2 academics. Luis was a professor and I was his PhD student. They're going to build something that's not going to make any money, and they're going to sell to Google.”

That belief lasted until we actually made money. For the first 5 years, it was, “These 2 researchers are going to build this thing, and eventually they're going to sell it. They're not going to become a public company or anything.” That was 1 reason.

The other reason was that this was 2011. We were raising on a prototype, and it was a website for a Series A. It worked because my co-founder had 2 prior exits, so he had a track record. The other belief was, “Even if this one doesn't work out, it allows us to invest in Luis's next company.” They were basically buying a ticket for the next company.

Including the first one, they didn't believe that this was going to become this big. How much was the Series A?

I think it was $3 million at a $15 million valuation. That was a Series A.

This is pre-seed today. That's like your friends-and-family round. Who did a PayPal round?

13. Why Union Square Was the Only VC to Say Yes

We never raised a seed or pre-seed. We had these NSF grants, so we built this project inside the university. Series A was our first round.

Who did that round? Can you take me to the moment when you signed the term sheet?

Union Square Ventures. It was hard. We may have to cut this afterwards, but I think we only had 1 offer. It was USV or back to university.

But isn't that amazing, though—that USV, one of the best firms in the world, recognized you? Statistically, as proven by the data, there was only 1 offer. This is what I see so often, though. I don't think there's anything controversial about that; it just shows quality in picking.

Yeah, absolutely. They were amazing investors.

What do you think they saw that others didn't?

One thing that was true back then is that we did the Silicon Valley road show. We actually got a lot of interest from Silicon Valley VCs at the time, but they all made it a requirement to move to Silicon Valley.

They were always asking, “You've got to move here, right? When are you going to move here?” They'd say, “We're super interested. We want to make a deal here, blah, blah, but when are you moving to SF or the Bay Area?” We said, “We're not moving,” and then they immediately lost interest.

By the way, some of them later invested in later rounds, but they didn't want to be the first one in a non-Silicon Valley company. USV didn't care. In fact, they said, “We're the only investor that doesn't care that you're in Pittsburgh.” That's true.

Do you think things have changed completely? Do you think the best founders need investor value?

No. I would say the best founders do not need investor value. However, investors need to have your back. That is important.

I would say you should always try to raise money from the best Tier 1 VCs, even if the terms are slightly worse. There's so much signaling. The Series A was the hardest, but because USV invested, every other investor was like, “Okay, what does Union Square do?” They had an amazing track record and amazing performance. If USV invests in your company, people will look at it.

That was really powerful, so every other round was a lot easier.

That's the weird thing: it's harder to raise $3 million than it is to raise $100 million. I totally see that. What was the easiest round to raise?

The latest ones. Everybody wanted to invest.

Was there a question with you and Luis where you were like, “Do we think $5 billion is a bit high?” This is obviously at the time. How do you reflect on price and overpricing when you have such demand?

I think the mistake you can make is not so much—I mean, we didn't want to have a down round. By the way, we also never had a layoff. We were always growing fast, but we never went on a hiring spree, and we never wanted to have a down round.

The problem with raising at a super-high valuation, or raising a lot of capital at a super-high valuation, is that the expectations become very high. You're obviously not worth that much at the time, but they want you to grow into that valuation, right? That's why they invest.

Today, if you take your ARR multiple or whatever—which, by the way, was infinite, the joys of raising with no revenue—you're obviously not worth that much, but we believe that you're going to be worth that much a couple of years down the road. I think that's 1 thing we wanted to avoid: a down round and having an investor in it that just has extreme expectations that we cannot live up to. That was our philosophy.

The other thing I learned on fundraising—and Luis did most of this—is that I thought every deal was custom. Your Series A is different from my Series A, and your Series B is completely different from my Series B.

What I realized is that there are only 3 possible deals. 1 is no deal: nobody wants to invest. That's the default. Number 2 is that you get a standard deal, which is what most companies get at your stage. I don't know what it is today, but it's about 15% to 20% dilution, whatever the price and amount are. That was basically our deal at the time.

Then there's the extreme outlier deal. That's the deal that 1 or 2 companies of a generation get. Facebook probably got that one. Uber and Airbnb probably got some of those.

You see it more and more today, which I think is brilliant for companies. They raise, say, $100 million at a $2 billion price, and you're like, “Wow, $100 million is a lot of money to bring into a company financing-wise, but only diluting 5% is a small amount for quite a large and significant amount to bring in.” I think they're good rounds to do for companies.

Yeah, but that changes things. Those are the 3 types.

Those are the 3 types. Did you do the 3rd type?

We never did the 3rd type. We always got pretty much a standard deal.

I think the last couple were more competitive. By the way, we could have raised a lot more at higher valuations, but we didn't do it for the reasons mentioned before.

Who was the most helpful investor?

USV was great. Bing is actually great. Bing is a rare type of investor because he is very product-minded. He uses the product and gives us a lot of feedback on it. He's obsessed with product.

Most investors are more business types and are like, “Okay, investment, return, EBITDA, and all that.” Bing is a very different type of investor.

14. If Duo Were Private Today, Would You Still Go Public?

Do you like—I'm going into a territory where you might go, “Harry, wave the white flag”—do you like being public? I have a new theory: Why would anyone go public today when, if you have a consumer brand, you never need to?

What I mean by that is, if you're the Collisons, if you're Ali at Databricks, or if you're Elon at SpaceX, you don't need to go public today. The expansion of private markets is so significant that there's just no need. If Duolingo were private today, would you go public?

Yes. I'm in the camp of, yes, you should go public. It's actually the better thing for the ecosystem. In fact, I believe there should be more European companies going public, because whether it's in the US or somewhere else doesn't matter, as long as it kind of flows back to the European ecosystem.

For Duolingo, it's been a great journey, actually. Going public and ringing the NASDAQ bell was a career highlight by far. It's incredible. I grew up in this small town in Switzerland, made it to the US, got a PhD, and started this company. It's my first job; I've never worked anywhere else.

Then this became this huge hit, with a massive audience and millions and millions of people loving the product. Great purpose, unicorn, decacorn, public company—it's just, you feel like it gives you this achievement. It feels like you won a big tournament.

You're not done, right? It's not done.

It's not done. It's an achievement. It's a milestone.

Yeah, it's not done, and I think that's also really important: that you don't think it's done, right?

I'm still with the company. Most founders who are not CEOs leave after 4 or 5 years, once they're fully vested, or at least when they get acquired or IPO, right? I'm still with the company.

Can I ask you, going back? The one thing I think is true is that these private markets are now much, much, much bigger, and there are lots of pros to staying private, right?

For us, before we went public, our finance team, for example, was 2 people. Now it's, I think, 25 or so. There's a lot of overhead, and the financial markets really care about—this was also new to me—predictability.

I feel like they care even more about that than they care about the actual profit margins or growth or whatever. They really care that you can predict. I think the way they see this is, if you can't predict the future of your business, you're not a good operator. You probably don't understand what drives your business, and they kind of lose trust in the management team.

So, someone like that, do you not think that's the biggest weakness now more than ever before? This is a time where there's no predictability, where models can change, where a new model can come out and completely change how we think about China's approach to AI or open source or whatever that is.

Actually, respectfully, I don't think we can predict very well. I think, for you now, doing M&A, investing, and thinking about the future of AI, that's a negative place to put a premium on predictability.

Yeah, there's so much uncertainty in the market right now. If you're a private company, there's a stronger argument for staying private. But for the much more mature businesses, like Stripe, I think you want to go public because you create this liquidity for your employees.

There are other ways to do it, and you can do secondaries and all that. But one thing I am also proud of is that we hired all these people and made lots of millionaires, and that money flows back into the ecosystem. Now it flows back into the Pittsburgh ecosystem, and maybe some of it into the New York ecosystem.

If you had more European companies, like Spotify, go public, making lots of employees some money, and then maybe having them start their own companies, that's what we need. That's what we need in Europe.

I agree. I just don't know if that mechanism is IPOs. I think you need a consumer brand to have that extension in private markets. If you're a boring B2B company that does whatever that no one cares about, you kind of have to go public because I don't think there's the appetite for late-stage demand in the private markets.

Duolingo could stay private longer, though, and there are so many different ways of doing share buybacks. There are so many people who want to have that late-stage secondary access. I think you can do it. That's a big question for the next few years.

I think the other big question is Deliveroo. Deliveroo sold for $2.9 billion; it was priced at $1.5 billion on the UK stock exchange. That's a $1.4 billion delta in how they value a company.

15. Quick-Fire Round

Mega. Listen, dude, I want to do a quick fire. I could talk to you all day. What's the biggest near-death experience with Duolingo? Was there a day where you thought, “Oh my God, that was the hardest day”?

It's not a single day, but it's the first 5 years we had.

When you said 5, I was like, “5 days?”

5 years. We just didn't know how to find a business model that aligned with our mission. It was hard.

What's the best angel investment you've made, and why do you think?

I'm not a good angel investor.

Why?

I think there's an overlap between being a good founder and a good investor, but ultimately they're different things. First of all, the investment cycle is 10 years, whereas as a founder, at least at the early stage, you get data points every day. You can learn; the feedback loops are so much faster.

I think I'm better with the fast feedback loops than the 10-year investment horizons. I've had a few successes, like a 10x or something.

But you've had a few 10x's?

Yeah, that's pretty rare. I've had several, but I would not consider myself an investor. Investors talk a big game.

What do you hear a lot from investors that makes you think, “Really?”

They ask this question—it's not just investors—“What is the secret sauce? Why is Duolingo so successful?” They look for an answer that's a 1-word explanation, and you can tell that when you give them the actual answer, they're disappointed.

Okay.

They wish the answer was, “It's the streak. That's it. That's the secret sauce. It's this 1 mechanic. That's why we're so successful. We invented it, or we maximized it,” or whatever. Or it's the leaderboards. Or it's, “We're so good at this particular market. It's Brazil. That's the secret sauce. We figured out how to crack Brazil.”

Luis keeps telling them the true answer, which is running thousands and thousands of A/B experiments. That's it. The streak mechanic, for example, is something we added—that was 1 experiment—but then we've probably run 300 experiments fine-tuning the streak mechanic. That's where you get the gains in retention, and that's what ultimately drives the growth of Duolingo, which is still super fast. We're still growing so fast.

It's kind of the growth machine. It's what I mentioned before: We run a lot of experiments. This is also true for marketing. It's not A/B-controlled experiments, but it's more like, “Let's try this on TikTok. Let's try this on Instagram,” and then double down on the stuff that works.

That's it. That's the secret sauce. It's the process. It's not a single feature.

The thing I find hard with that process, though, is the compressed timeline on outcomes. What I mean by that is content is a really good example. It takes time to know if content works, and so you have to do things with no obvious gain in the short term for compounding long-term advantage.

Does that make sense? It's almost a bit like the gym. You don't gain muscle on day 1, day 2, or week 1. You'd be like, “Harry, quit this. This A/B test is failing.” But 3 months in, it shows.

Yeah, I would agree with that. I think the trap you can run into is that you only run tiny experiments, right? Like, on your purchase page, or you run these tiny experiments on copy or something.

I think it's very important that you have a portfolio of changes. You have lots of low-risk, small changes, but you also have some big changes. The big change could be adding chess or adding math, and that's how you stay innovative and relevant.

You can't just do that; otherwise, you end up at a local maximum where you can't get out because you optimized yourself into it.

What have you changed your mind on most in the last 12 months?

AI and the impact of it. I’ve gone back and forth on basically any of the fundamental questions, like, is it going to change the world? Are we going to have more engineers or fewer engineers?

One thing I think gets lost in the conversation is that, okay, let’s assume AI really drives productivity gains. We all become more productive. You become more productive. I become more productive. Everybody we work with becomes more productive. Let’s say 20% or 30%. Is that good or bad?

Good.

That’s great. One of the big problems we have is demographics. We all need to become more productive because we don’t have enough kids. How are we going to finance our welfare systems? How are we going to finance our governments? We need every single productivity gain we can get.

In a way, I feel like in the US there’s this huge deficit. There’s all the debt. But if you can grow the economy 10% year on year instead of 1%—I guess right now we’re in a recession—but if you can grow it 10% or more, all these problems will go away. It will really be an age of abundance.

Will we have AI friends in the future, more than normal friends?

No. I don’t believe that. In fact, I think human-to-human interaction—the value of that—will go up significantly. Why? I just believe there’s going to be a pushback. Again, there’s going to be value in verified human work, even if it’s not as good.

For example, in the art world, I believe that art made before 2022 is going to increase in value because it’s verified human. Picasso—we know it was Picasso, but it wasn’t Sam Altman masquerading as Picasso.

Exactly. I totally get that. Who’s the most underrated founder in AI right now? We talked about Decagon. I think it’s an amazing company.

Whose life do you secretly admire, and why?

There’s this Naval Ravikant quote: “The true measure of intelligence is the ability to get what you want.” If you get what you want out of life, by that definition, I’m maybe 80% successful or intelligent.

What do you not have? You said 80%. What’s the 20%?

I have an investor friend. I’m not going to share who it is, but I feel that guy has kind of figured it out for himself.

What do you not have?

One thing I would say—and that’s one of the reasons why I said yes to this podcast, or why I wanted to be on this podcast—is I feel like I’m reduced to just Duolingo. My identity is Duolingo. It says “founder of Duolingo.” Every introduction starts with, “Oh, it’s the founder and CTO of Duolingo.”

I feel there’s much more to me—to my personal identity—than just Duolingo. That’s the other 20% that I have not explored or haven’t been able to share, really.

I think the hardest thing is that your identity is tied to your company. When your identity is tied to your company, the performance of your company determines your happiness. I’ve only ever had 1 job, which is 20VC, and when we’re not doing well—either the fund or the media company—I’m depressed, like, down.

Yep. That’s hard. So true. I’ve talked to many founders who left their company, and half of them, I would say, fall into a depression because their identity is the company.

Could you ever leave? Don’t laugh, but when it is so tied to you, like yours is to you and mine is to me—11 years; yours is even longer—the thing is, I still enjoy it so much. Sometimes I wish I hated my job; then it would make it easy to leave and do something else.

Another point was that, early on—and it’s still somewhat true—when an employee leaves or quits, it’s like a punch in the gut. It’s like, “Employee? Sure, sure.” But it’s like, “Why did they leave? What did I do wrong? Do they not believe in this company anymore?” I would think about this for weeks.

Someone likely named Neil Mehta asks a brilliant question from Green Oaks. He likes to go into a company and ask the employees, “Are your best days ahead or behind you?” When I ask you that for Duolingo, why would you say they’re ahead of you?

Well, thanks for assuming that they’re still ahead of us.

No, I think so. Going back to the very beginning of the conversation, the mission is to provide the best education and make it universally accessible. But now I think we can really do it. We can really build an AI tutor that is as good as the best human tutors, and we can do it for everyone, not just the rich. That will change the world.

I really think this is the moment. We had to wait 10 years to get to this stage, but I think this AI wave is going to really enable us to do this.

Final one on public market caps and stuff. I don’t know your ownership structure, which I probably should. I believe you’ll probably technically be a billionaire, so well done.

My question to you is, do you think about money? How do money and happiness correlate, or not correlate, for you?

Yeah, and that’s an entire half-hour conversation or so. I’ve also spent a lot of time thinking about this question, and the weird thing is, I wish I hadn’t. I feel like it’s not productive.

I thought a lot about this: am I happier now than I was when I was a grad student? Is it because of the money, or is it because of the success, or is it because I know myself better? It’s very hard to disentangle all the different factors that go into happiness.

I would say, though, that there are different levels of wealth. They actually did this study, I think, with post-exit founders. They asked them, “How much is enough? How much is enough? When do people switch their mindset and say, ‘Okay, now this is enough?’” It’s probably still subjective, but the answer they got was—do you know what the answer was?

$20 million.

No, $100 million.

$100 million. Jesus, inflation is real. Whoa.

That was the answer of those people. It’s kind of like when you stop, I guess, worrying about it because it’s so much. I mean, so much.

Did you have a number when you started?

No. This is a mission-oriented company, and I tweeted this recently: I think, counterintuitively, the founders who start with a purpose and a mission and aren’t focused too much on money usually end up doing better. Maybe people don’t give them credit, but I do, because I’m in the same situation.

I think Bill Gates or Mark Zuckerberg wanted to connect people. Bill Gates wanted to put a computer on every desk, I think. Then he realized, “Okay, how do I do this?” One way of doing it—or maybe the only way of doing it—is starting a company.

They’re not like, “I’m going to start a company to become rich.” I think that’s kind of reversing it. You have to think, “What change do you want to see in the world? What do people want? How can I help other people?” Ultimately, in an ideal capitalist environment, that should drive value.

I do believe that’s why these companies got started. Mark Zuckerberg was maybe focused on connecting people.

What do you think was the biggest needle mover for your happiness, then? For me, there’s “pay rent, easy; I can live life like I kind of want to, but I’m not loaded. I kind of need to work, but I’m not loaded”—probably a couple of million.

There’s the $10 to $20 million, which is like, “I’m pretty good.” Adding another $80 million is not going to change much, actually. I still have a super-nice house and I’m good. Then there’s the $100 million: I can take jets when I want to.

Maybe answering a different question, but I do believe I’m happier now, but not by a lot. I was pretty happy in my 20s too, but a little bit less.

I think this sense of achievement will never go away, even if Duolingo collapses tomorrow. It’s still a good thing. It’s like, “I tried, and it worked out for a while, and maybe then it goes away.” But I would still have this sense of achievement.

Another thing is when people share their Duolingo stories. I was in Venice a year ago or 2 years ago, and the buses are boats, right? So I took the bus on the canal, and sitting literally next to me was this woman doing Duolingo.

I was just looking at her and thinking, “If she knew I was sitting next to her…” I didn’t say anything. For a 10-minute ride, I was watching her do Duolingo, and I saw how she used it. I was just like, “This is amazing.”

Does money change marriage?

Wow. I thought it was going to be the quick one. It is.

I think the answer depends on your answer to the question, “Does money change people?” Because if it changes people, it’ll probably change marriage. So I assume the answer is yes. But it depends on the people.

Did it change in your case a little bit?

We met before I was successful. In fact, my now wife—at the time, my girlfriend—for the first couple of years, she made more than me. So we kind of reversed that. But I think she knows me from when I wasn't very successful, when I was just a dude with some crazy ideas.

I think it's different than if I had to date someone now, because everything starts with this: “Oh, co-founder of Duolingo,” and then super successful. Everybody knows Duolingo. My co-founder is obviously way more famous, but it's hard to remove that. In fact, I resent that. I wish you could go into incognito mode, like in a browser. It's like, “You know what? I'm just going to put on glasses.”

People always ask you, “What do you do?” When I'm in a room where I know they don't know who I am, I sometimes give a wrong answer. I used to say, “An Uber driver,” but of course, nobody really believes that. Or I say, “I'm a software engineer at a tech company.” Or I say, “I'm a patron of the arts,” which isn't entirely untrue, just because I know how the conversations go when I tell the truth.

The next question is, “How many languages do you speak?” The next question is, “How did you start Duolingo?” And then, “Why has it been so successful? Was it Brazil?” The next question, if they're a user, is, “Oh, I love your—my streak is this long.” If they're actual users, that's much more interesting to me.

Can I ask a final one? I promise, but it's an important one. Everyone has said, including Warren Buffett, that the most important thing in life is partner selection—who you choose to spend your life with. I spoke to everyone on your cap table, and they all said that the relationship you and Luis have is special. I'm intrigued. We mentioned marriage there. Whether it's Luis or marriage, I don't mind, really. What would be your biggest advice on partner selection?

For founders, I think it's a little easier. Actually, let me see if I can answer this for founders. I think the number one thing is: have you worked previously with this person? The key is worked, not hung out at the bar. Have you worked with this person prior to starting this company?

Luis and I had worked together for 2 years prior, so we already knew how we worked, what we were good at, what we were bad at, and our priorities. Very early on, we had this situation because I went from him being my superior—he was my PhD adviser and I was his student—to an equal relationship. We were equal co-founders, and I thought, “Can I trust this guy? What if he just fires me after a year?”

So we wrote down a little contract between the two of us. I still have a copy at home where we outlined exactly: “This is how we're going to make decisions. This is Luis's responsibility. This is Severin's responsibility.” We both signed it, and I think that avoided a lot of conflict.

The other thing that's interesting is that I think a lot of the conflicts happen in the first 2 years. What's the biggest conflict? Probably what to build—the product itself. There were conflicts there, and then also on hiring, not hiring, firing, et cetera. There were a bunch of conflicts as well, and then I think we had it all worked out.

It's amazing: I only meet with Luis every other week, and sometimes the meeting lasts 10 minutes. We go into an executive team meeting, and I have a pretty good mental model of what he's going to say or how he's going to react to a certain idea, because we had been working together for the first 6, 7, 8 years in the same room every day, 50–60 hours a week.

That gives you this—I know him better than probably any other person other than my wife, and vice versa.

Your wife—biggest advice on partner selection, romantically?

I would say: do you enjoy talking to this person? That's it. You enjoy talking to many people, but I call it the Tuesday night test. In 5 years' time, will you be excited for just you and them to sit down and just talk?

There are very few people where, 5 years from now, on a Tuesday night, it's a cold February evening in Pittsburgh or London or wherever you are, and it's like, “Hell, yeah.” I think that's one thing. Then also mutual support: do they support you in what you do, or is there competition? You don't want to have competition in any of these relationships.

Dude, this has been one of the most wonderful wide-ranging discussions I've had. I'm so pleased we got a chance to do it. Thank you for being so open. You have been amazing. All right.