The Hardest Part of Learning Is Getting Yourself to Start

Sarah Guo opens No Priors with Luis von Ahn, the Carnegie Mellon computer-science PhD who founded reCAPTCHA, sold it to Google in 2009, and now runs Duolingo. The app has more than 116 million monthly users and a market cap around $17 billion, plus an owl mascot that recently faked its own death. Guo frames the hour around AI in education, motivation as the hard problem in learning, brand risk, and the 16,000 A/B tests behind the product. Von Ahn is there to describe not only the language app people know, but what he wants it to become.

His first answer is straightforward: Duolingo is already the most popular way to learn languages in the world. Over the last couple of years it has also started teaching math and music, and chess is about to land. The through-line is not school-on-a-phone as a slogan. It is an app for the things a lot of people want to learn that also take a long time to learn.

He was still a professor when the company started in 2011, an odd origin for a gamified consumer product. He did not plan that. He had a PhD student named Severin, now co-founder and CTO, and they needed a thesis topic. They agreed to work on computers teaching people something, then settled on languages—especially English, because in most countries it raises income potential, and about two billion people are trying to learn it.

Von Ahn built the first Spanish course; Severin, a native German speaker, built German. They promised to learn each other’s languages and immediately failed. “I did not want to learn German. He did not want to learn Spanish.” If the founders could not sit through their own product, they could not expect anyone else to. The fix was to turn it into a game as much as possible, and by launch it was pretty fun, largely because neither of them likes learning languages. That, von Ahn says, is why they built something for the average person instead of for language obsessives.

Guo asks why this counted as a computer-science thesis. Von Ahn’s list is less about linguistics than about systems: how to get a computer to teach, how to adapt to a student with data, motivation, and human-computer interaction. At the time, “AI” meant training classifiers to teach a bit better, not large language models. There was, he says, plenty of thesis material in that mix.

Once they decided ordinary humans do not have infinite motivation, the first change that worked was shrinking lessons from 30 minutes to 2. People did not spend less time. Two minutes is a commitment you can make at any moment; thirty minutes is a decision. You may still stay for half an hour, but you start because you believe you can leave whenever you want. He calls that the first real game changer among many later bets.

The next surprise was streaks. A streak is only a counter of consecutive days you opened the app to learn, and von Ahn did not expect it to be this powerful. People talk about their numbers in public. Duolingo now has 10 million active users with a streak longer than 365 days, meaning they have not missed a day in at least a year.

Notifications were the other accident. The app used to ping lapsed learners with a bland “come back to learn Spanish,” then stopped after five days of silence. Von Ahn thought that if they were going to stop, they should say so. The fifth-day note became: these reminders do not seem to be working, so they would stop sending them for now. It is one of the strongest re-engagement messages they have, because people feel the app has given up on them. They did not even write it to get users back.

Guo puts the objection on the table. Learning is hard; a friend compared it to the gym for the mind: you just have to do it. Von Ahn says he would be delighted if every education-app builder thought that way, because then Duolingo would have no competitors. “The hardest thing about learning is motivation.” He thinks the same is true of the gym. People can debate elliptical versus treadmill all year; 90 percent of what matters is that you show up.

The same logic applies to books versus apps versus tutors. Effectiveness differs, but what matters most is that you actually do the work. To get people to do it, you have to make starting as easy as possible. That, he says, is why Duolingo has grown so much. New language apps often pitch themselves as “like Duolingo, but without the gamification.” His reaction is simple: great, carry on.

“99% of the world’s population just is not that motivated for any activity.” The highly motivated few can use whatever they want. The product is for everyone else. Language lovers building for language lovers, he implies, keep mistaking their own taste for a market. Duolingo’s edge is that the founders were never in that club.

Guo then raises flow. Knowledge workers treat deep, uninterrupted attention as almost holy, and two minutes of Duolingo on the subway does not look like that. Von Ahn understands the idea but does not optimize for it. For an English speaker, getting to a pretty good spot in Spanish takes about 500 hours. “We just got to clock those 500 hours.” Whether distraction turns that into 600, or deep focus into 400, the order of magnitude does not change. The same number for Chinese is 2,000 hours. Guo, who holds a master’s in Chinese literature, deadpans that it does not seem like a language meant to be learned.

She finds his view of learners a little dark and still useful as an end user. She has three kids and a full day of work; she will not carve out a painful hour and a half at night for chess, Chinese, or math. Two minutes at a time over years, adding up toward hundreds of hours, actually seems doable. After enough consumer work, von Ahn says, a few blunt facts stop being optional: people do not read, people are lazy, and given a choice they will scroll Instagram or TikTok, whatever they tell you in a survey.

The optimistic half is that you can point some of that behavior at something useful. That is the crux of Duolingo: take the tricks mobile games use to hold attention and spend them on learning. You stop arguing with how people actually spend their time and start designing for it. The app is not trying to make users more virtuous. It is trying to win the minutes they would otherwise give to a feed.

Guo notes that Duolingo already had scale when machine learning started accelerating. Von Ahn says the timing has been “really positive,” in part because they teach languages and the new models are language models. Teaching requires a huge amount of content, and the old pipeline was half hand-built, half automated—faster than pure craft, still slow. Over the last couple of years the company retooled that pipeline to run entirely on large language models. Humans are barely involved. The result is a volume of courses that was not possible before.

Duolingo already taught 40 languages, but each course is directional: Spanish for English speakers is not Spanish for German speakers. English speakers could take all 40; a native German speaker could take about four. Because AI made authoring so much faster, the company can now teach those 40 languages from every base language—a major jump in the catalog, and something they were launching around the time of the interview. Guo later calls it “140-some” new language offerings.

The second new capability is conversation. Vocabulary and reading were always teachable; spoken practice meant another human. Most people do not want to talk to a person in a language they are bad at. It feels like shame. With a model, they can practice without being judged, and usage of that feature is rising fast. Von Ahn is also excited about math. Tutors have strong learning outcomes and one big problem: they are boring. Most people would rather play Candy Crush than sit with a tutor.

The goal is something as effective as a tutor and as fun as Candy Crush. The realistic version, he admits, is about 90 percent of each—and that combo is still more useful than today’s options. New subjects have to clear a few bars: hundreds of millions of potential learners, because the app charges about $10 rather than $30,000; hundreds of hours to mastery, because a two-hour skill belongs on YouTube; a subject that is good for the world and teachable on a phone; and someone inside the company who is genuinely excited, which is how music and chess got in. History is the counterexample: Duolingo is good at drills, and history wants well-produced video. Even so, he does not think there is much that computers will eventually fail to teach well.

Guo asks whether large language models could threaten Duolingo itself. Von Ahn does not dismiss it. The industry is in a platform shift, and nobody knows the other side. He is not super worried, but the same uncertainty hangs over Netflix: a model might one day press a button and generate the perfect movie for you. At the moment it does not look like a major threat. The company’s bet is distribution, unique data on how people learn languages, and a brand people already associate with that job.

Guo turns to that brand. Duolingo takes more risk and more distinctive voice than most public companies. Why did the owl die? “He didn’t die. He just faked his death.” It was not a day-one plan to be unhinged. In the product the owl was already a little pushy about lessons. The internet invented the darker memes—kidnapping your family if you skip a session. Duolingo did not create those; it leaned in, and leaning in worked.

A few years ago a very junior marketing hire asked to put an old owl suit from company events on TikTok. Von Ahn was against it; he did not think anyone would care. The videos were just the owl doing dumb things, twerking and falling down, with no “learn a language on Duolingo” caption. They went viral. The character took on a life of its own. The company has since formalized which risks it will take, but it still takes more than most public companies.

Two structural facts help. They have a mascot, so the cartoon can behave badly instead of the corporation. They are also an education company, and it is hard to argue that education is bad. The same marketing from a vaping company, or a business that just wanted to extract money, would get more pushback. Education gives them leeway other consumer brands do not have.

With more than 100 million monthly learners, Guo asks what Duolingo knows about learning that others do not. A lot of it lives in the algorithms and is hard to verbalize. Distance from your native language makes a language harder. The company can predict with high accuracy whether you will get the next exercise right, because it watches what each user is good and bad at—this person is weak on the past tense, for example. Dumping only weak spots on someone would make horrible lessons. The science is when to serve the hard stuff.

Whenever they give you an exercise, the right target is about an 83 percent chance of getting it correct. That rate maximizes enjoyment, and enjoyment is how you reach five or six hundred hours. One hundred percent is too easy; 50 percent is too punishing. You have to mostly win. Guo then asks what this means for school. Von Ahn thinks education will change in less than 20 years, because teaching with AI scales in a way teachers do not. Teachers will not vanish: someone still has to take care of students. Schools will not vanish either, because parents still need childcare.

Guo summarizes the picture as schools becoming childcare while everyone Duolingoes. Something like that, he says—maybe not Duolingo, but one teacher with about 30 students, while a computer knows each child’s strengths and weaknesses at a granularity no human can. There are extremely good teachers; there are not very many, and most people never get one. Formal school can demand 20 minutes instead of 2, because attendance is not optional. The change will still be slow. Go to a real school, he says, and they are doing stuff from 30 years ago. The drag is government and regulation; Texas trying not to teach evolution is his example of how weird school systems can be.

Private schools may move first, though a $50,000-a-year campus will struggle to tell parents the pedagogy is an app. Some countries that are currently behind may leapfrog, because AI teaching is the only way they can scale. When Duolingo went public, investors treated it as a COVID bump; it was not. The deeper miss, even among other education companies, is how central motivation is, and how much testing sits under the cute art: more than 16,000 A/B tests, including when to show an animation. Guo talks about dropouts who wrote wireless textbooks as sophomores; von Ahn expects 15- and 20-year-old experts in five or ten years, amplified by AI. Inside the company, artists are still there, but a month of illustration can now take a day, which Guo dubs “vibe cartooning.”