Duolingo’s Severin Hacker on AI Tutors, Learning and Building a Company

Duolingo co-founder and CTO Severin Hacker joins Harry Stebbings to discuss AI, education, company-building and life as a founder. Hacker speaks from the perspective of an executive and owner of a company that sells educational products and uses AI suppliers.

Hacker says Duolingo began with the ambition to make excellent education universally available. He sees one-to-one tutoring as a powerful but historically expensive model, and believes technology could eventually bring a comparable experience to far more people.

Duolingo was an early partner for OpenAI’s GPT-4, and Hacker says the technology immediately impressed him. The first practical opportunity was to accelerate the production of course material, a task that had previously taken substantial time and human effort.

He says Duolingo created roughly its first 100 courses over 12 years and then another 148 within one year. Curriculum design remains human-led, he says, while AI produces short lesson sentences under vocabulary and grammar constraints.

Hacker expects learning content eventually to become much more personalized. Instead of giving everyone the same sequence, a future system might generate exercises around an individual’s interests, goals and current level; this is his vision, not a demonstrated outcome for every learner.

He compares that vision to a good tutor who assesses a student, understands what motivates them and gives timely feedback. Duolingo’s video-call feature with its character Lily already enables AI conversation, although Hacker describes Lily as a friend rather than a one-to-one tutor.

Hacker groups the company’s AI use into three categories: creating learning content, building features such as conversational practice, and improving productivity across the organization. Employees can choose among tools rather than being required to use one particular coding assistant.

On hallucinations, Hacker argues that some invented details matter less in a language exercise when the language itself is correct. That is his assessment of a particular use case, not a general guarantee of accuracy or educational effectiveness.

He is skeptical of claims that software engineers will be replaced immediately. In his experience, AI can produce much of a simple app or perform localized code transformations, but struggles to add features reliably to large codebases and may create technical debt.

Asked whether Duolingo will employ more or fewer engineers in five years, Hacker says he is less certain than he was a year earlier. He expects cheaper software production to increase demand and sees product, engineering and design responsibilities potentially converging.

Hacker still sees value in studying computer science, particularly logical thinking and problem-solving rather than memorizing a particular programming language. He calls a CS degree probably a good investment for the next five years, while acknowledging uncertainty about longer horizons.

He speculates that a future with advanced AI could eventually mean much less compulsory work and a basic income, perhaps on a decades-long timeline. Stebbings challenges this, arguing that past technological shifts have also created new work; neither presents the outcome as settled.

The two consider the value people place on work beyond income. Hacker acknowledges that the social and personal consequences of a world with much less work are difficult to predict, even if productivity improves.

They discuss entry-level hiring and the danger of assuming AI removes the need to train the next generation. Junior employees become experienced people over time, so the near-term economics of a role cannot be the only consideration.

Duolingo uses Decagon for customer support, and Hacker says AI handles a large share of requests while human agents continue to address the remainder. He argues that lower costs could make support available to more users; he is also an investor in Decagon.

AI has different cost implications depending on the feature. Hacker says generated course material is largely a one-time production expense, whereas real-time interaction with a model creates a cost for each use.

Video Call with Lily is mostly in Duolingo’s higher-priced Max subscription because of that ongoing cost. Hacker hopes competition will bring model prices down so the company can extend such features to many more users, but that is a forecast.

Duolingo also works with ElevenLabs on voices. Hacker expects competition among model providers but says specialized audio and multimodal capabilities may be harder to substitute than more widely available models.

Stebbings asks whether OpenAI could enter language learning directly. Hacker acknowledges some competitive concern, while arguing that Duolingo’s focus on teaching and motivating people differs from using a general assistant to answer questions.

Hacker calls motivation the hardest part of learning a language. In his view, streaks, points and other game mechanics encourage users to return and do difficult work; this is a product claim from the company’s co-founder, not independent evidence that all users achieve fluency.

He argues that a dedicated learning experience has an advantage if a competing feature is buried inside a broader app. Stebbings pushes back that language is naturally suited to large language models, and Hacker concedes that the outcome remains to be seen.

Hacker says details in design affect retention and whether people recommend a product. Co-founder Luis von Ahn is particularly attentive to small imperfections, and Hacker sees those details as part of the difference between a good app and a great one.

Duolingo initially built a separate math app. The team later realized it was recreating streaks and other infrastructure while asking users to discover, download and configure another product, so it brought the educational content into the main app.

Chess raised a question about the boundary between education and entertainment. Hacker says two employees who cared about chess built an AI-assisted prototype, which made the idea tangible enough to persuade the team to pursue it.

He says AI helped the company build the chess course in about nine months. For Hacker, the prototype demonstrated both that the experience could fit Duolingo’s design and that a person with limited recent coding experience could make an idea concrete.

Asked about a physical school, Hacker separates instruction from other functions of education, including social interaction and credentials. He believes AI tutors could transform instruction, but does not claim that software has solved the social side of schooling.

He discusses Duolingo’s interest in measuring and certifying learning, alongside the possibility that humans would still play important roles in a physical educational setting. His vision of highly capable AI tutors is a prediction tied to his company’s mission.

The conversation turns to how Duolingo organizes its work and reviews products. Hacker describes close involvement from leadership and the importance of building teams that can turn detailed feedback into a better experience.

He reflects on the evolution of the product organization and the tension between founder involvement and specialist managers. A company can learn from conventional structures, but needs to decide which practices actually serve its own work.

Hacker describes Duolingo’s marketing as a small team that experiments with social content and repeats what works. The distinctive public-facing style is the product of iteration as well as creative choices.

On conventions for building a company, Hacker urges founders to question inherited rules during a major technological transition. The existence of a standard process does not by itself prove that it is needed or should be performed in the same way.

His personal framework is to ask first whether a task needs doing at all, then whether it can be automated, and finally whether it can be delegated. He has handed much of the day-to-day engineering management to the company’s engineering leader.

Hacker says he spends much of his time on AI’s implications for Duolingo and on acquisitions and external investments. He suggests prior experience working with a company can make an acquisition easier to assess; Duolingo has also made a small number of minority investments.

In language learning, Hacker sees retention as the central competitive measure. Paid acquisition can produce fast growth, he argues, but a company that cannot keep users has a leaky bucket. He watches how many people return the next day.

Stebbings asks whether Europe risks falling behind in technology. Hacker says he would advise many ambitious AI founders to consider Silicon Valley because of its ecosystem, although he notes Duolingo itself was built in Pittsburgh.

Hacker argues that Duolingo, which he says had no revenue for its first five years, would have struggled to raise its early funding in Europe. Stebbings challenges the claim and argues that the European funding environment has changed.

Hacker says he invests in European venture funds and is involved with Project Europe, giving him an interest in the ecosystem he wants to improve. His views about founder migration, regulation and European competitiveness are opinions, not a comprehensive comparison of all companies.

Duolingo considered opening a San Francisco office but decided against it after other founders warned that employees could move there and then be recruited away. Hacker presents this as feedback from those founders and a decision suited to Duolingo’s situation.

He argues that easier movement to the United States can draw European founders away instead of forcing improvements at home. He criticizes European AI regulation; Stebbings challenges whether regulation can really stop the most determined entrepreneurs.

Looking back, Hacker identifies two major mistakes: waiting too long to take monetization seriously and waiting too long to hire experienced managers. He says the company initially concentrated on user growth and was unsure how to charge without undermining its accessibility mission.

The company later developed a model with free and paid tiers. Hacker says most users do not pay and that people can learn without a bank account; these are the co-founder’s descriptions of the product and its mission.

Duolingo began as a Carnegie Mellon research project involving Hacker and his PhD adviser Luis von Ahn. It had research grants before its first institutional financing, so the Series A was its first venture round rather than following a seed round.

Hacker recalls the Series A as approximately $3 million at a $15 million valuation and says Union Square Ventures supplied what he remembers as the only offer. Other interested investors wanted the founders to move to Silicon Valley, he says, while Union Square accepted Pittsburgh.

He recommends seeking supportive, highly regarded investors even at somewhat worse terms because their participation can help signal credibility to later investors. That is his advice based on Duolingo’s experience, not a universal calculation that brand always outweighs economics.

Hacker says the founders also tried to avoid raising at valuations that would create expectations they could not meet or increase the risk of a down round. He describes most of Duolingo’s financings as relatively standard rather than exceptional outlier deals.

Asked whether Duolingo would still go public if it were private today, Hacker says yes. He describes the IPO as a personal career highlight and believes public listings can return value to the wider startup ecosystem.

He also describes the demands of public investors, including the importance of explaining and forecasting the business. Stebbings questions whether the availability of private capital makes public listing less necessary; the two emphasize different trade-offs.

Hacker says the company’s growth has no single secret. Its team runs thousands of experiments, including hundreds of refinements to the streak mechanic, and applies a similar test-and-repeat approach to marketing.

Stebbings warns that some investments take time to show results. Hacker agrees that teams need a portfolio of small experiments and larger initiatives such as math or chess; optimizing only tiny details risks becoming stuck at a local maximum.

Hacker says AI is the subject on which his views have changed most. He considers higher productivity potentially valuable in aging societies, while acknowledging that the scale of those gains and their economic consequences are uncertain.

He predicts that human-to-human interaction and verifiably human-created work may become more valuable as AI improves. Asked about an underrated AI company, he names Decagon, in which he has invested.

Hacker says he wants an identity beyond being Duolingo’s co-founder. Stebbings relates to the problem: when a founder’s identity is closely tied to a company, the company’s performance can have an outsized effect on personal happiness.

Despite those concerns, Hacker says he still enjoys his job and believes AI could help Duolingo pursue its educational mission more effectively. His claim that AI tutoring could eventually match the best human tutors is an aspiration, not a proven result.

The discussion shifts to money and happiness. Hacker says it is difficult to separate the effects of money, achievement and self-knowledge; he values the sense of having built something useful and seeing people use it in everyday life.

Hacker reflects on how success changes social interactions and his marriage. He says his wife knew him before the company’s success and describes occasionally avoiding his usual introduction so he can meet people without being reduced to his company.

On co-founder selection, Hacker recommends having worked together before starting a company. He and von Ahn had collaborated for two years and wrote down decision-making responsibilities as their relationship changed from adviser and student to equal co-founders.

He says they worked through disagreements over product and personnel early and developed an understanding of how the other would respond. For a romantic partner, he values enjoying conversation and offering mutual support rather than competition.