Duolingo Co-Founder, Severin Hacker: How AI Impacts the Future of Work and Education

Summary

• “AI-first” at Duolingo is about content velocity. It took 12 years to build the first ~100 courses; with AI-generated sentence content (humans still design curricula), the company shipped another 148 courses in one year — “that was just not possible… it would have taken us decades.” The endgame Severin sketches is personalization on the fly: “why do we even create this batch on the back end?” — every exercise custom-generated when you open the app, replicating the one-on-one tutor only kings and billionaires’ daughters could afford.
• Severin calls “Silicon Valley hyperbole” on rapid engineer replacement. AI tools are great at 0→80% of a simple app and single-file transformations, but “the larger the code base… the worse they get,” and they generate tech debt they can’t fix. A year ago he’d have said “100% more” engineers in five years; now he’s less certain — but a cost-down/demand-up paradox may mean far more software and more people creating it, with product, engineering, and design (70-80% of Duolingo’s workforce) potentially converging into one “product engineer designer” role. AGI-and-UBI: “not next five years… maybe 10 years… more likely 20.”
• The OpenAI threat is framed as a motivation problem. ChatGPT devastated “the homework cheating business,” and Duolingo was mistakenly bucketed as an AI loser before flipping to winner. Severin’s defense centers on motivation: Duolingo is “a motivation engine,” and users asked what they’d do without it “never mention any of our language learning competitors — they say oh, I would spend more time on social media.” Harry’s pushback stands: language learning is OpenAI’s fourth obvious vertical after chat, support, and coding — “LLM is kind of in the name.”
• Retention is the only thing that matters in consumer learning — the space is winner-take-all. Every year for 12 years, one or two venture-funded competitors grow fast on paid acquisition and “inevitably they go down again… if you don’t have retention, it’s a leaky bucket.” The one company he fears is one with higher retention than Duolingo. The secret sauce investors don’t want to hear: “thousands and thousands of A/B experiments” — the streak alone has had ~300 fine-tuning experiments.
• Fundraising signal over terms: “It’s harder to raise 3 million than it is to raise 100 million.” Duolingo’s 2011 Series A was $3M at $15M with, in his recollection, one offer — “it was USV or back to university” — after every Silicon Valley VC demanded they relocate. “Always try to raise money from the best tier one VCs, even if the terms are slightly worse, because so much signaling.” There are only three deal types: no deal, the standard deal, and the once-a-generation outlier — Duolingo pretty much got the standard one.
• Europe verdict is brutal: to a young European AI founder, “100% go to Silicon Valley.” “I can guarantee you Duolingo would not have been able to raise any money in Europe — we had zero revenue for the first five years.” His dark twist: the best thing that could happen to Europe is the US restricting founder immigration, because today the ambitious simply vote with their feet. The EU AI Act is “some of the dumbest regulation ever written” — the Ottoman Empire banning the printing press.
• AI margin math, spelled out: content generation is one-off and “phenomenal for margin”; conversational features like Video Call with Lily carry per-API-call unit costs and sit gated in the top-tier Max subscription until inference prices fall. Decagon’s AI can do 70-80% of support tickets — and the cost-down/demand-up logic again suggests that at 10-100x lower cost, support could extend from paying subscribers to everyone. The model layer stays competitive except specialized audio/voice, where 11 Labs “might be harder to just replace.”
• Severin feels public markets care about predictability “even more than profit margins or growth” — which Harry flags as a weakness in an unpredictable AI era; Severin concedes the private case is stronger right now but still backs going public: Duolingo’s IPO minted millionaires who recycle into the Pittsburgh ecosystem, “that’s what we need in Europe.” Postscript on wealth: the post-exit founder study’s “how much is enough” answer is $100 million.

Deep dive

1. “AI-first” means 148 new courses in one year, not a pivot

• Severin’s reframe: Duolingo was always about giving everyone the one-on-one tutor only elites could afford — “the best education was only available to the richest people, the kings,” and it’s still true: a US Open final he recalls featured two daughters of billionaires, products of the 1:1 coaching model. “Duolingo was from day one technology first… back then we called it software, now we call it AI. In a way not much has changed except AI has become a lot better.”
• The proof point: it took ~12 years to build the first 100 courses; in one year with AI, Duolingo built another 148. “That was just not possible. It would have taken us decades.”
• Humans stay in the loop where it matters: curriculum design is “all human-made.” AI generates the short in-lesson sentences under tight constraints — introduce new words one at a time, “create a sentence that uses these words and doesn’t use any of the other words and only uses these grammar concepts.” Having had early access as an OpenAI launch partner for what was likely GPT-4, his reaction was “an iPhone moment — you just realize this is the future.”

2. The endgame is a course generated on the fly, per user

• Pressed by Harry on whether AI should design curricula given completion-rate data from “hundreds of millions” of learners (“the largest school” versus education studies run on 20-30 students), Severin goes further: “why do we even create this batch on the back end? Why don’t we create it on the fly?” — you’re into rugby, you have a trip to France, so every exercise is custom-designed the moment you open the app.
• Why personalization wins: it’s what makes 1:1 tutors work — “they’re really good at assessment. They know where you are, they know where you need to go,” and they know your interests. His view: personalization is the future of education, and it will be multimodal “100%.”
• The live version is Video Call with Lily — the purple-haired character who is “not a one-on-one tutor, she’s more like a friend, but she remembers stuff” (that he likes cooking, where he lives). Full conversation was the top user-requested gap — “I want to get better at speaking” — and “two or three years ago you couldn’t build” it.

3. Three internal AI lanes, and no tool mandates

• Severin’s taxonomy: (1) content generation — the biggest success, “completely changed how we produce content”; (2) AI features you previously couldn’t build, like Lily; (3) company-wide productivity — engineers on Cursor, AI customer support.
• On tooling policy, Duolingo is “surprisingly egalitarian and bottom-up”: no mandated tools — “use whatever makes you more productive and we pay for it” — Cursor, ChatGPT, Copilot — “and then we see what people actually use.”
• Hallucinations, the classic objection, simply don’t bite here: “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.”

4. Where AI coding tools actually stand: great at 0→80%, lost in large codebases

• His deflation of the hype, with hedges intact: “I think there’s a lot of Silicon Valley hyperbole around software engineering getting replaced very quickly. I’m less sure about that.” The tools excel at going “from zero to 80% for a simple app” — the simple-app-builder zone — and at isolated single-file transformations.
• The failure mode: “the larger the code base — even the one that generates itself — the worse they get.” The next 10% takes as long as the first 80%, “and they kind of create their own tech debt which they also cannot really solve.” What they can’t do yet: “adding a feature, even a simple one, to a large codebase.” He flags his own uncertainty twice — “maybe I’m wrong, maybe this will not age well.”

5. More engineers or fewer? Cost-down logic and the converged role

• Asked if Duolingo has more or fewer engineers in five years: “if you had asked me a year ago I would have said 100% more. Now I’m a little bit less certain.” But two forces push toward more software: the entry bar drops (“your mom can maybe build an app with Lovable”) and cost-per-line collapses — a cost-down/demand-up paradox: cheaper software may mean more demand for it.
• The structural call: product, engineering, and design compose 70-80% of Duolingo’s workforce, and AI enables “a new role, kind of a product engineer designer — one person who can do all of these roles to a certain degree.” More software, more people creating it; “whether we’re going to call them software engineers or not, I’m not sure.”
• On the CS degree: universities don’t teach coding — “computer science is kind of applied mathematics” — and he thinks the Java/Python details “are going to go away,” but logical problem-solving endures. Verdict for an 18-year-old: “for the next five years, probably still a good investment.”
• On AGI-and-UBI he’s a believer with a long fuse: “not going to be next five years… maybe next 10 years… more likely 20.” The philosophical kicker: people close to the AI revolution derive purpose from work, which is exactly why “we are most afraid of it — we feel we might lose our purpose.” In the post-work world, the opening question shifts from “what do you do?” to “what are your hobbies?”

6. Keep hiring juniors — the consensus “senior engineers only” move is a mistake

• Harry relays a call from someone likely named Jason Lemkin that AI matches entry-level tech workers and “in 12 months we’re going to see huge unemployment” at that tier. Severin: “I’m not sure.” The consensus — AI does the intern’s job, so stop hiring interns — “is a mistake, and we’re actually not doing that at Duolingo.” People who grow up with these tools become far better at using them: “I bet the person who spent a thousand hours using ChatGPT is a lot better at using it than I am.”
• The Gen Z proof case: Duolingo’s social media presence exists because they hired from Gen Z — Zaria “just fundamentally understands it in a way that someone who’s not Gen Z could never understand.” Severin’s confession: “when I open TikTok and look at our own Duolingo stuff, I don’t get it. Why is this funny? Why do people watch this? But it’s super popular.”

7. Support economics: Decagon can do 70-80% of tickets, and cost-down logic strikes again

• Duolingo uses Decagon for AI customer support, which can do 70-80% of tickets. Human agents stay for the rest — but the real shift is expansion: support currently goes only to paying subscribers, and “if we can reduce the cost of customer support by 10x or 100x, then we can actually give it to everyone.” The framing is a cost-down/demand-up paradox: “as cost goes down, demand goes up.”
• The margin structure, use case by use case: content generation is a one-off — “phenomenal for margin.” Conversational features like Lily carry a real unit cost per API call, which is why Lily sits in Max, the highest of three tiers (free, Super, Max). “We would like to give this to everyone, but the reality is today there’s a unit cost.” The bet: batch costs have already fallen on competition; real-time multimodal will follow, “eventually… almost everyone.”
• On model relationships: competition is expected in general models, but in specialized lanes — vision, and especially audio — “there’s less competition and you can see companies that keep a lead.” 11 Labs’ voices “might be harder to just replace with one or the other model.”

8. The OpenAI question: motivation is the core defense

• Harry takes him to the stock hit when markets feared ChatGPT would eat language learning. Severin’s structural answer: ChatGPT’s first killer use case was “homework cheating, for lack of a better term,” which was “devastating” for companies that “call themselves education but are really more in the homework cheating business.” Duolingo got lumped in by mistake — the market’s “bifurcation into is this an AI winner or is this an AI loser” first filed Duolingo as loser, then flipped to winner with AI tailwinds.
• The core insight he says nobody believes after ten years of repetition: the hardest part of learning a language is motivation. “Duolingo is a motivation engine… if you stick to it, you will actually learn a language.” The tell: users asked what they’d do if Duolingo vanished “never mention any of our language learning competitors — they say, oh, I would spend more time on social media.”
• He concedes Harry’s scenario — ChatGPT skimming casual users with a simple chat-with-a-friend mode is “definitely somewhat of a concern” — but counters twice: “if I was Sam Altman… I would go after the big trillion dollar companies like search or social media,” and Luis’s frame that gamification buried inside another app is a second-class citizen — “it’s hard to make that experience as sticky as if you build an app just for this.” Harry’s rebuttal is worth keeping: OpenAI’s obvious verticals are chat (owned), support (will buy), coding (just bought), and language learning fourth — “LLM is kind of in the name… It’s not a stretch.” Severin: “Well, we’ll see.”

9. Details are the product — and the math-app detour that taught the super-app strategy

• Channeling someone likely named Gustav Söderström from Spotify (“the details are not the details, they are the product”), Severin agrees flatly: “ultimately product quality is details.” Luis “can spot something being not properly aligned within half a second.” Details drive retention — the main metric — and referrals: “that’s the difference between a good app and a great app.”
• The wasted-time confession: expanding to math, they built a completely standalone Duolingo Math app — then caught themselves re-implementing leaderboards, leagues, and streaks, while users faced a gauntlet of discover-download-install-permissions and a confusing second streak. The rule that came out of it: “no more new apps. If it’s education, it goes into the main app” — internally, a super-app strategy. Content transferred in, so nothing was thrown away.
• Chess shows the new build velocity: two internally passionate employees — a PM with some engineering background but little coding experience, plus a designer — built the first version with Cursor and AI tools, and “the prototype is what convinced us… if it had just been a Figma design, we probably would have said maybe, maybe not.” Full course shipped in nine months, the fastest course ever created — versus years, normally.

10. Disrupting Harvard means solving credentials and community, not just instruction

• Severin’s decomposition of higher ed: instruction, credentials (“otherwise you would always have to prove again that you know something”), and the social layer where people meet partners and networks. “Most edtech companies just focus on number one. If you really want to disrupt the Harvards of the world, you have to solve all three.” Asked if achieving Duolingo’s mission disrupts Harvard: “Yeah, we would be.”
• On credentials, the wedge is the Duolingo Score — “my Duolingo score is 61 in French” — which they want to become the standard vocabulary for language proficiency. The social layer is his current obsession and admitted blank: massive user base, huge brand love, “an untapped potential of connecting the people that are on Duolingo — I don’t know how to do it.” (Harry’s Duo Dates pitch: “not a terrible idea.”)
• K-12 realism: “a lot of K-12 is childcare” — even a superb AI tutor doesn’t dissolve schools, and humans keep guidance, coaching, and the why-should-I-learn-this role. Still, the categorical bet: “AI tutors will become really, really good and it’s going to be the way most people learn most things in the future. I do believe that.”
• Mechanics still matter more than anyone thinks: push notifications “still make a huge difference — whenever the service that sends push notifications goes down, you can see the dip.” The streak is pure loss aversion; users email from Everest expeditions for streak restores, and Duolingo proactively extended streaks during the Portugal-Spain power outage.

11. Winner-take-all: retention is the only scoreboard

• Twelve years of pattern recognition: “every single year there was one or two companies that grow really fast in language learning… and inevitably they go down again.” They buy growth with venture money on Facebook and Google ads, but “the only thing that matters in our space is retention. If you don’t have retention, it’s a leaky bucket.” His fear function: “the one company I’m scared of is one that has higher retention than Duolingo.” The daily metric: of everyone using Duolingo today, how many come back tomorrow — all retention curves correlate off it.
• The secret sauce investors are “disappointed” to hear: not the streak, not leaderboards, not cracking Brazil — “running thousands and thousands of A/B experiments.” The streak was one experiment; “we’ve probably run 300 experiments fine-tuning the streak mechanic, and that’s where you get the gains in retention.” Same process (“the green machine”) powers marketing: try things on TikTok and Instagram, double down on what works.
• Harry’s caveat — compressed test timelines kill compounding bets, like the gym (“you don’t gain muscle day one… but 3 months in it shows”) — draws a real concession: run a portfolio of changes, small low-risk tests plus big swings like chess and math, “otherwise you end up in a local maxima where you optimize yourself into it.”
• Related brand math: marketing is ~40 people and was built on a zero-dollar budget (“we can’t spend money that we don’t have”) — “Duolingo marketing is not powered by money, it’s powered by ideas.” At some point Duolingo was bigger on TikTok than Nike.

12. Europe: “100% go to Silicon Valley” — and the SF-office trap

• He’s “very conflicted internally,” but the advice to a young European AI founder is unhedged: “100% go to Silicon Valley. You absolutely maximize your chances.” The receipts: “I can guarantee you Duolingo would not have been able to raise any money in Europe. We had zero dollar revenue for the first five years. Not in Switzerland, not anywhere.” Of European founders who stayed: “best maybe half as successful, or maybe some of them 0%.” The cultural rot: “there’s this suspicion almost — you can’t be too successful, you shouldn’t be too ambitious.”
• His most contrarian mechanism: “the best thing that could happen to Europe is if the US made it really hard to immigrate for founders” — because right now the ambitious vote with their feet, so there’s no domestic pressure to fix regulation. On the EU AI Act: “some of the dumbest regulation ever written — it’s literally like the Ottoman Empire banning the printing press… let’s not do AI because there could be negative side effects. It’s that dumb.” Against Harry’s “no great entrepreneur was stopped by regulation”: Uber and Airbnb faced local rules with 200 other cities as alternatives — founders facing EU-wide rules may leave.
• The non-obvious operating insight: Duolingo nearly opened an SF office until they polled founders of non-SV-HQ’d companies who had. “All of them said: worst single decision ever” — it creates an internal funnel of your best people to SF, where “they get recruited away by whatever the hot Silicon Valley company is at the time… you basically create a hiring funnel for Silicon Valley.” Duolingo will never open one. (For Harry personally: as a founder he’d move; as an investor, being the top name in a less crowded Europe “is not clear” a loss — that was exactly Severin’s own Pittsburgh angel thesis.)

13. One offer or back to university — and why tier-1 signaling beats terms

• The origin: a Carnegie Mellon research project (Luis the professor, Severin the PhD student) funded by NSF grants — no seed, no pre-seed. The 2011 Series A: $3M at $15M, raised on a website prototype, with what he recalls as one offer: “it was USV or back to university.” Silicon Valley VCs were interested but all conditioned on relocating; USV “said we’re the only investor that doesn’t care that you’re in Pittsburgh.” What got it done: Luis’s two prior exits — investors were “basically buying a ticket for the next company.”
• The generalized lesson: “you should always try to raise money from the best tier one VCs, even if the terms are slightly worse, because so much signaling.” After USV, “every other round was a lot easier. It’s harder to raise 3 million than it is to raise 100 million.”
• His demystification of terms: every founder thinks their deal is custom, but “there’s only three possible deals” — no deal (the default), the standard deal (~15-20% dilution at the going price), and the extreme outlier deal “one or two companies of a generation get” (Facebook, Uber). Duolingo pretty much got standard deals, deliberately avoided peak-valuation raises to dodge down-rounds and investors with “extreme expectations we cannot live up to” — and never had a layoff.
• The two admitted mistakes, “not fatal but”: monetizing too late — the business model answer for five years was literally “venture capital,” investors pattern-matched to Twitter’s grow-users-first, and the mission felt incompatible with charging until they reframed it as inclusion (“you should be able to learn everything on Duolingo without a bank account” — still true; most users don’t pay). And hiring senior managers too late — flat and chaotic to ~30 employees. Against the podcast-circuit orthodoxy of promoting youngsters: “you cannot run a thousand-person company with all new grads… the oldest organization in the world, the Catholic Church, is very hierarchical — there’s probably some reason for it.”

14. Go public anyway — even though Severin feels the market pays for predictability

• What surprised him post-IPO: finance went from 2 people to ~25, and public markets “really care about predictability — I feel like they care even more about that than about actual profit margins or growth,” reading unpredictability as management not understanding its own business. Harry’s pushback lands: in an era of new-model shocks and model turnover, “I don’t think we can predict very well — that’s a negative to place a premium on predictability.” Severin concedes “right now there’s a stronger argument for staying private,” but holds the line for mature businesses like Stripe.
• The pro-IPO case is ecosystem-level: liquidity “made lots of millionaires” whose money flows back into Pittsburgh; “if you had more European companies like Spotify go public and employees make some money and start their own companies — that’s what we need in Europe.” Ringing the Nasdaq bell was his “career highlight by far.” Harry’s counter-data point: Deliveroo priced at $1.5B on the UK exchange, sold for $2.9B — “a $1.4 billion delta.”
• Quick-fire adjacencies: the near-death experience wasn’t a day but “the first five years — we just didn’t know how to find a business model that aligns with our mission.” Founder mode at Duolingo: Luis, the detail-obsessive, still sits in most product reviews at a $20B public company. Severin’s own operating principle is “reduce, automate, delegate” — he’s handed off day-to-day engineering and now spends “80% of my day” on the AI question plus M&A (4-5 acquisitions to date; the number-one M&A predictor, same as co-founders: “have you worked with them before”).

15. Money, identity, and the partner-selection contract

• What he’s changed his mind on most in 12 months: “AI and the impact of it — I’ve gone back and forth on basically any of the fundamental questions.” His macro argument is that even hypothetical 20-30% productivity gains would help aging demographics — “if you can grow the economy 10% year on year… all these problems will go away. It will really be an age of abundance.” But no to AI friends outnumbering human ones: “the value of human-to-human interaction will go up significantly… art made before 2022 is going to increase in value because it’s verified human. Picasso — we know it was Picasso.”
• On wealth: a study of post-exit founders asked “how much is enough” — the answer was $100 million (Harry guessed $20M: “Jesus, inflation is real”). Severin never had a number; his tweet-thesis is that mission-first founders — Gates putting a computer on every desk — “end up usually doing better… starting a company to become rich is the reversing of it.” His honest 20% gap, and why he took the podcast: “I feel like I’m reduced to just Duolingo… there’s much more to my identity.” He sometimes tells strangers he’s an Uber driver or “a patron of the arts, which is not entirely untrue.”
• Partner selection, the closing keeper: for co-founders, the number-one predictor is prior work together — “worked, not hung out at the bar.” He and Luis had two years of research collaboration, and when the professor-student hierarchy flipped to equals, they signed a short written contract — decision rights and responsibilities — “I still have a copy at home… I think that avoided a lot of conflict.” Today they meet every other week, sometimes for ten minutes: “I have a pretty good language model of what he’s going to say.” Romantically: “do you enjoy talking to this person?” — plus mutual support, and “you don’t want to have competition in any of these relationships.” Are Duolingo’s best days ahead? “Now I think we can really do it — an AI tutor as good as the best human tutors, for everyone, not just the rich… this AI wave is going to really enable us to do this.”