Duolingo CEO: AI Won't Take Your Job — But Someone Using It Will (And Here's What to Do Now) | Luis von Ahn

0:00This episode is brought to you by Indeed. Stop waiting around for the perfect candidate. Instead, use Indeed Sponsored Jobs to find the right people with the right skills fast. It's a simple way to make sure your listing is the first candidate C. According to Indeed data, Sponsored Jobs have four times more applicants than non-sponsored jobs. So go build your dream team today with Indeed. Get a$75 Sponsored Job Credit at Indeed.com slash podcast. Terms and conditions apply. K-pop Demon Hunters, Saja Boys Breakfast Meal, and Huntrix Meal have just dropped at McDonald's. They're calling this a battle for the fans.

0:36What do you say to that, Rumi? It's not a battle. So glad the Saja Boys could take breakfast and give our meal the rest of the day. It is an honor to share. No, it's our honor. It is our larger honor. No, really. Stop. You can really feel the respect in this battle. Pick a meal to pick a side. Ba-da-ba-ba-ba. Hey guys, welcome to Silicon Valley Girl. We are about to interview the founder of Duolingo. Oh. Oh, no, no, no. No, Duo, not you. No, no, no. Do we have the actual founder? Hi. Hi. Your employees are getting a little creative. We pay that guy and he has his own personality. He just does what he wants.

1:25AI is not going to take your job. Somebody using AI is going to take your job. This is Luis von Ahn, CEO of Duolingo. Recently, two of his employees built a chess course with AI in six months. No engineering background, no knowledge of the subject. It became the fastest growing course in the company. There are a lot of rumors that large companies are firing people and they say it's AI. We have never done a layoff. Despite what the internet may think, it is important to continue hiring people because a single employee is just way more productive now than they used to be. I'm going to name five professions and you're going to make some predictions.

1:58Gone in five years or not going anywhere. There will be fewer and fewer. Luis, you told your team that nobody gets hired unless he improves AI can do the work first. Can you tell me how you actually track that as a founder? Our goal here is to use AI to benefit our users. Internally, we have this golden rule. We're only going to use AI to benefit our learners. Some people may imagine that, you know, some companies may be firing people and having AI do their job. That's not what we're doing whatsoever. Our team has gotten significantly better at using AI over the last couple of years. And that has allowed us to do a lot more.

2:42It has allowed us to put out a lot more content, a lot more learning content, etc. We, as a management team, are not necessarily tracking like, oh, are you doing something that AI can do or not? We're just really trying to tell everyone, try to be as efficient as possible with AI and our employees are doing that. Can you give me some of the best examples from the team? For someone who's watching this and they're like, okay, my manager tells me to start using AI, but I have no idea how to do it. Do you have like best case examples? It depends on what your job role is. I think most of our engineers have basically really changed their workflows.

3:19They're using AI coding tools. A lot of our product managers, what they've decided to do is using AI to make prototypes of things. So the product manager may not be implementing the thing in the full production app, but rather than coming to us with a written document, they come with a prototype now. That's way better because it allows for much better decision making. So if somebody comes to me with a written proposal and says, you know, I'm going to do a way to teach Spanish better, it's hard for me to know what that actually means. But if they just show me the prototype and I can see that it really does seem to teach Spanish better, it's much easier to give approval to that type of thing.

3:57So it depends a lot in the role. Do you give them any guidance like, OK, instead of doing this, do that next time? Or how do they even find out they can do these things with AI? We try to do some things in the whole company. We had, for example, a few months ago, we had a day where everybody in the company had to vibe code something. Not just engineers, every single person, people from HR, people from the finance team, everybody had to vibe code something. So everybody could see the power of it. We also have a lot of documentation about best practices. But generally, people here in Duolingo are pretty smart.

4:36They're always finding new things. And I think what happens is rather than management telling them what to do, they tell each other what to do. We have a lot of Slack channels. You know, one of them is called Best AI Practices. And so people are just saying stuff. We also have another one that's called AI Fails, which is all the times, all the things that people try. That's such a great one. It is an incredible thing. very empowering for people who are like, oh my God, I made, you know, usually they're very small apps, but it's like I made an app. One of the things that has happened inside this company is everybody has made their own dashboard.

5:07For their KPIs or something? Yeah, for whatever it is, whatever it is they're tracking. I mean, I see a few product managers that have vibe-coded a whole thing with like our users in every country and what they're doing. That's super impressive. And is that how you track AI proficiency? Because now it's part of performance reviews, right? At Duolingo. For a while, it was part of performance reviews. We decided not to do that and i'll tell you why i sent a memo to the company that said you know part of your performance review is going to be usage of ai and we found that people were i don't know if they were doing that but they were kind of asking like do you just want us to use ai for ai's sake and at the end we backtracked and we said no look the most important thing in your performance is that you are doing you know whatever your job is as well as possible a lot of times ai can help you with that.

5:51But if it can't, I'm not going to force you to do that. So I think we backtracked from that because it really, it felt like rather than being held accountable for the actual outcome, we're trying to just push something that in some cases did not fit. And do you have a specific example where somebody did something in a week without AI and then with AI, they kind of multiplied the output? We now teach chess on Duolingo. So, you know, for a long time, we only taught languages. Now we teach a few other things. Chess is the latest course we added. This course got started by two people, neither of whom knew chess, neither of whom knew how to program.

6:28They basically vibe-coded the first prototype of it. Now, the final version that is actually in the app, of course, we put some engineers in there, etc., but they really got very far. In a span of about six months, they created the whole curriculum for chess. They created a prototype of the app all entirely with AI. And again, these people did not know any chess. And whose idea was that? Was it their idea? It was their idea to do that. Well, they're the ones who wanted to add chess. By the way, they came to me a year earlier to say, we want to add chess. I said, I don't want to add chess because it's just a game and we're an education app.

7:04But what happened was that a few months later, I talked to the minister of education of my country. I'm from Guatemala. And she said to me, our education system, our public education system in Guatemala is so broken that I'm considering sending every student a chess board so that at least they'll learn logical thinking. And when she said that to me, I thought, oh, wow, okay, this is actually useful. Part of education. Yes. So then this is when I told them, okay, you can add the chess course. But I said to them, but I don't have any engineers to give you, so go ahead. And they figured it out. Six months, right?

7:37About six months, yeah. And now it's your fastest growing course? Yeah. I mean, at this point, we have 7 million daily active users that are learning chess. Wow, this is fascinating. Can you tell me step by step what was their process? So if somebody is watching and they're like, wow, if Duolingo is able to do such a spinoff, which is kind of different from languages, if I want to start something with AI and build a fast growing product, what are the five steps they need to take? These two guys, the first thing they did was probably learn chess. Probably because they didn't know any chess. And that's one of the reasons they wanted to add it because they themselves wanted to learn chess.

8:11but after that what they did is they really started looking at you know kind of the different tools that are out there for learning chess that they're basically doing market research try to figure out what's out there for learning chess they found that really what was out there was not all that great and then they decided to start vibe coding something to be fair this person does have some technical knowledge they're not an engineer but they have some technical knowledge And so they downloaded Cursor. And at first they made just chess puzzles. Then they realized that the AI was not very good at making chess puzzles.

8:45So then they decided to train it with some... There's an online database of a lot of different chess puzzles. So they trained the AI with that. And it got a lot better. And then after that, what they started doing was just more and more mobile prototypes for me to play with until I told them that it was kind of good enough to really put it in the app. And then you put it in the app, and then did you give it an external push, or people just started discovering it? By putting it in the app, very quickly a lot of people came. But chess really has a big draw. We put other courses in the app that have not grown as much.

9:18For example, it turns out chess is more fun than math. So someone, a student, who just heard about your chess course that was spun out so quickly, and they want to start building something with AI today, what advice would you give them? Well, the biggest advice I can give them is to start. You know, a lot of people talk a lot about starting. They're like, oh, I have an idea. But the biggest thing is just sit down and do it. You will learn a lot by just trying to do it. So that's the biggest thing. Other than that, you know, try to learn how some of the best tools work. Certainly, Vibe Coding will help you a lot.

9:54But it's not just Vibe Coding. I mean, trying to make, you know, the initial designs for it. now you can have AI tools to make the initial screens and everything. So use it all and try to make the thing. I mean, I haven't quite yet seen somebody who knows nothing about programming to really make a good app, but I have seen people who know a little bit about programming make apps. So I would tell you, it's still worthwhile learning kind of the structure of programs. That seems important. Even though you may not actually need to program like word by word, just knowing how things work, for example, knowing the difference between the server and the client, like just these very basic things.

10:38I think that's important. Did it give you more ideas to add more courses that are non-language related? Yeah, it's given us a lot of ideas. We are not yet working on other stuff, but we have a long list of things that we want to teach all kinds of things. K through 12 science, we want to teach. Again, we're not yet working on that. We want to teach K through 12 science. We want to teach how to draw. There's all kinds of things, but at the moment, we're not really working on them because we want to continue working on chess, math, music, and languages. Got it. But then any employee can just go and vibe code one of these courses and show you and maybe they can make it to the other.

11:08Yeah, that's what I tell the company, by the way. A lot of times people come to me and they're like, what course are you going to add next? And one of the things I tell them is, I wanted to add K-12 science and then we added math. I wanted to add K-12 science and then we added music. I wanted to add K-12 science and then we added chess. At this point, I also want to add K-12 science. But it turns out that what people do here is somebody comes up with a really good idea. And if it's good and they're passionate about it, we let them do it. Yeah, they do it. Let's go back to the actually the AI failure, Chad.

11:37Can you recall any examples of AI actually failing in a task? Oh, yeah. I mean, there's a lot of things where it's failing. I'll tell you the biggest one, which I think we're starting to see a difference here. But it wasn't the case a year ago. If you were to look at Twitter and just read what people are saying on Twitter. two years ago I should have fired all our engineers because on Twitter they're like AI is better at coding than engineers this is kind of for the last two years this is what they've been saying on Twitter and it was an interesting thing because I would come here and I would not really see a speed up on engineering and I'm like well what's the disconnect why you know there's all these people saying that AI is better at coding than humans and the reality is it's not yet the case that AI is better at coding than humans I think you still really need engineers and you're going to need them for a long time I think but we've just seen a lot of cases where you tell the AI to program something and sometimes it works, but the other, I don't know what fraction of the time it doesn't work, but when it doesn't work, there's a real problem in that because you don't really know what it did, it's really hard to debug it.

12:39And that's kind of what has happened. You get into, the happy path is really fast. Okay, it worked. But the unhappy path makes it so that it takes so long that you end up spending more effort on that than the time you saved on the other thing. So we've seen that quite a bit. we've also seen ai not be able to generate things like narrative like stories sometimes it does a good job sometimes it doesn't do a great job and it just comes up with things that don't make any sense you know it's interesting because when you see a demo and you ask it to come up with a story somehow those demos always look amazing but then when you try to come up with a hundred stories you realize that only like 30 are good and the other 70 are so humans still need to check and Yeah, all our content needs to be...

13:22There's a lot of steps to try to either check it or spot check it or something to make sure that the quality is high. So if we compare today and a year ago, how much more productive the company is with AI? I don't know. It's more in pockets. And I don't know if any larger company that has seen like a 10x speed up in like... I don't think so. I think startups really see it. Because when you're a one-person team, then you can work as true. A one person can do a lot of stuff. With a larger company, it's harder. it just turns out for example most engineers don't spend eight hours a day coding they have to go to meetings they have to so there's a part of it that you just cannot speed up then there's the part that they spend coding okay maybe you can speed that up but you cannot speed it up at least at the moment you cannot speed it up by a thousand x you can just maybe i don't know a little faster so overall we're not seeing like we're putting out 10 times as many features we're not we're seeing some speed ups here and there in different pockets of the company but again it's what you said.

14:16One-person companies are a lot faster, but it's because you don't have to interface with all the other parts of the company, etc. Also, AI is not as good with existing code bases rather than with a brand new code base. We've been talking a lot about how AI is changing the way people learn and retain information. And that got me thinking about something I changed in how I organize my own daily life and work. Someone told me once, if you're not capturing everything that happens in your calls with your team, you're actually losing to other companies. you're losing decisions that were made you're losing context that gets dropped things you're committed to and then you just forgot team thanks strategy sessions call with investors even a quick conversation with a contractor it all disappears the moment the call ends and i don't know about you but i'm on calls every single day sometimes i forget to press record sometimes i don't think the conversation is worth recording but then after the conversation i realized i had to actually record that.

15:13A lot of my friends told me, Marina, you should start using granola. You'll be blown away. And now I use it on every single call. Now it is an AI notepad for meetings, but it's not a bot that joins your Zoom and makes everyone uncomfortable. It transcribes your computer's audio in the background while you stay completely present in the conversation. It listens to every call automatically. And if it's a recurring call, it merges them by folders. So by the time the meeting ends you have clean structured notes ready to go and after the call i chat with my notes i'm like okay we just had this conversation can you create a follow-up email that i'm gonna send to monica my manager uh we're gonna proceed with this this and that creates the email and i just send it takes 15 seconds it works across zoom google meet and teams it's one of those tools i genuinely can't work without anymore and i don't think you'll be able to either once you try it try it on your very first meeting after this episode, head to granola.ai slash marina to get three months of granola for free or enter the code MARINA at checkout.

16:14I left a link in the description. And now back to my conversation with Luis. What about you as a founder? Does AI help you make decisions or any workflows that work for you? Research, sure. It used to be the case that whenever I had something that needed to be researched, like, I don't know, what is the chess landscape in India, for example? So, you know, it used to be the case that I needed to either spend a lot of time myself or get a team of people to help me. Whereas now I can get a pretty good idea by just asking Gemini or something like that. So research, I do a lot more by myself. That has helped me.

16:50But ultimately, the decisions are made by me. It's not like I asked the computer, what would you decide? Okay. I haven't. So you don't use it as a coach or you haven't bi-coded your own KPIs or whatever. I've done a bunch of iCoding for things, but yeah, certainly my KPIs, but it's not decisions I still make myself. Okay. Yeah. Okay. I interview a lot of cool people on this podcast. In the past three weeks, I had Reid Hoffman and Bill Gurley, who are legendary investors. And I asked them a question, what is the first market that's going to be completely changed by AI and market you should be looking at?

17:21And they told me language. Thank you so much. You know, it's been my thing for 10 years. What do you think? I don't know what that means, change. Like you don't have to learn a language because everything is translated. Oh, no, I just don't buy that. If you look at our users, we have more than 100 million active users. Half of them are learning as a hobby. It's a hobby. Whether AI can do it or not, it's a hobby. Actually, chess is a great example. You know, computers have been better at chess than humans since 1997. A lot more people are learning chess today than they were in 1997. It's a hobby.

18:00So for half of our learners, it's a hobby. And I don't think whether computers can do it or not, I don't think it'll matter. The other half are learning English. And I think anybody who tells you that people don't want to learn a language has not had to learn English. It's interesting how you say like 50 % they're probably learning other languages. Hobby. Yes. English necessity. Yes. But that's how it is in the world. If you're learning, I'm sure there are, I'm generalizing here. But generally, if you're learning French, it's a hobby. It's not always true. I'm sure there are people who actually need to move to France.

18:28But it's a small fraction The majority of people in the world that are learning French Will, you know, want to feel cool when they go to Paris So they can order a croissant And that's our uses And again, somebody who's lived their entire life in the US Has never had to learn English Because that's their native language, etc Will say something like Well, languages are unnecessary, etc But somebody who's had to learn English It's a different thing You just need to learn English And you may want to move to an English-speaking country You may need to do an educational like you may need to go to a university.

19:00Nobody's going to be allowing you to go to a university with like a phone that translates everything the professor says. Like that's just not. So I'm just not particularly worried about that. Is there a fraction of the market? So I'm thinking about translators, for example. Somebody spends like four years learning that. What would you say to those people? Depends on what it is. But there are fractions of the market that are probably going to change quite a bit. Also, in terms of the demand, I agree that some people may have wanted to learn a little bit of a language if they were going to, just visit Germany for two days, that maybe now they're just not going to do that because they can just...

19:33But at least when you look at Duolingo's users, that's a very small minority of our users. The majority really are hobby or English. So I'm just not... So you're not worried? I'm not. The other thing about language translation is that simultaneous language translation has been really good for... This is not an LLM thing. Like LLMs are good at it too, but Google Translate did not use LLMs and it was really good 10 years ago. So I just, you know, we've seen the demand to learn a language actually go up as opposed to go down. This is not something that we internally are worried about. But I understand people say that.

20:07Yeah, I'm also thinking like if you have glasses, for example, that are smart and they translate everything for you automatically. I haven't seen the case with the headphones. I don't think people really use the AirPods to translate. No. I'm thinking maybe it's glasses that translate the science. I don't know. It's because when we came to Silicon Valley, I think 2015, we were pitching our company LinguaTrip that does language travel, language courses. And I think 50%, but to your note, American investors are like, oh, this is not a market. Oh, this is going to go down. We've had this problem. And the entirety of the company of Duolingo, we're a company based in the United States.

20:40When we first pitched Duolingo to investors here, I don't know, however many years ago, the most common thing is like, nobody really wants to learn a language. Math though, people want to learn math. And I'm sure investors think that because they're good at math and they wanted to learn math. But the reality is there are more people learning languages in the world than there are people learning math. What was your mindset when the smartest people in the world tell you that? You're like, I don't care. I mean, I grew up in Guatemala. I could see what it was to have to learn English. Look, it is a huge thing to learn English.

21:11It changes people's lives. I know. That's my case as well. It changes people's lives. Absolutely. Do you have another worry that, so you have an app, right? But now we're in the era when anyone can vibe code an app for themselves. What if I go to, like I'm learning English, I go to Claude and ask, can you gamify my experience, create me an app that's personalized to me, you know, my interests, you know, my hobbies? Do you worry about that? A little bit, but not really. I mean, this, it's funny. Inside the company, we never talk about this. This is, you know, we see, again, we see Twitter talking about that or investors talking about that, but we internally don't talk about that.

21:49Look, ultimately, I think you can ask AI to make you an app, but making a really good app is not that easy. Ultimately, we, for example, have data from hundreds of millions of people about how they learn a language. Every single day, more than a billion exercises are answered on Duolingo, and we use that data to teach you better. And so we just have a lot of data on how to keep you motivated. There's probably 2 ,000, 3 ,000 language learning apps in the world. they've been there and now with vibe coding there now there will be probably 20 ,000 rather than two three thousand but at the moment we're just not particularly concerned you know something may happen is there anything concerned because i'm asking a question about it because i talk to people and they're like oh this is going to be eliminated there's going to be we're done and you're like it's fine no i wouldn't say that it's like all is fine i do think a lot of things are going to change what do you think is going to change what are you preparing for i think that But user expectations are going to change, and I think we have to stay ahead of it.

22:49So a good example is, you know, one of the things we have in the app is conversation practice with AI. When we first put that out there, the cost of it was high for us. So we put it behind the most expensive tier. So you have to pay a lot of subscription to get the conversation practice. At the moment, the cost of that has come down enough that we're going to start giving it to much cheaper tiers. And we are probably going to eventually give it away for free. We're doing that because I believe that customers are going to start expecting that. Like, I think there will be apps that start doing this for free.

23:26And if we don't do it now, we'll be forced to do it in a few years. So that's the type of stuff we're doing to prepare for this. Because, again, I think we just need to stay ahead. Yeah, so users will expect more. I think users will expect more. I think users will expect the apps to be a lot more intelligent, which they should. We're undergoing kind of a platform shift here. And what ends up happening in platform shifts is that the companies who were the winner before the platform shift may or may not remain the winner after that. So I hope that we can do that. Also talking about AI and workforce, I just talked to Gary Vaynerchuk and I really like what he said.

24:01There are a lot of rumors of people and news people firing, large companies firing people and they say it's AI. Then Gary said something that struck me because he said, if I fire 100 people, my competitor hires them and they still 10x their output, I'm dumb for firing them. How do you think about that? We have never laid anybody off here. We have never done a layoff, despite what the Internet may think. Yeah, I think they misunderstood you. We've never done a layoff here. I think that it is important to continue hiring people because now, you know, the way I see it is a single employee is just way more productive now than they used to be.

24:41So, you know, I get better return on investment by having another employee. So that's how I see it. I mean my sense I of course cannot speak about the very specific companies but my sense is that at least in some of the cases AI is just an easy PR reason yeah I mean usually what happens is you overhired and when you overhired you're like okay because of AI we're going to do that I don't at least at Duolingo the way we run the company I am surprised that there are companies doing that because I see no reason yeah I talked I was just at Davos and I was talking to somebody who releases their jobs report.

25:19All the layoffs are structural, like overhired during COVID. So it's not really... A lot of companies overhired during COVID. I understand that. And then, you know, if you overhired, you probably don't need these people. But, you know, blaming AI is an easy scapegoat. It is. It is. What he just said actually worries me the most. Everyone's talking about like AI replacing jobs and big companies firing people. This is not what's happening here, right? He's actually hiring people, but he's hiring those who know AI. Same for me, where we're just hiring a social media manager. And the questions I was asking them during the interview, how are you going to automate this?