20VC: Duolingo Co-Founder on Why $3M is Harder than $100M to Raise | Why You Should Always Take Tier 1 VCs Even at Worse Terms | Why Europe Can't Win Unless the US Screws Up | How AI Impacts the Future of Work and Education with Severin Hacker

"You should always try to raise money from the best tier one VCs, even if the terms are slightly worse because of so much signaling. It's harder to raise 3 million than it is to raise 100 million. But I can guarantee you, Duolingo would not have been able to raise any money in Europe. You know, we had zero dollar revenue for the first five years."

"Duolingo came out as being AI first the other day. What does that mean? And what does that not mean? So people have a clear understanding. Take me to that discussion then. We're going to use this everywhere. And then it's like the second question is like, how do you use this at Duolingo?"

"When we started Duolingo, we saw the potential of technology to do this, but for everyone. We can't all afford these one-on-one tutors. It's too expensive. But with technology, there's a way that anyone can have access to a one-on-one tutor with the same efficacy. And that's why we started Duolingo. From day one was technology first. Back then it was software. Now we call it AI. So in a way, not much has changed except AI has become a lot better."

"One of the most amazing things about Duolingo is we have this massive user base. We're by far the largest learning platform out there. By far, orders of magnitude. We see how people learn. It's like the largest school. Usually in education, the studies were done with, you know, 20 to 30 students. We have hundreds of millions. That allows us to look at, okay, does this curriculum design work? The industry is definitely moving that way."

""So, you know, right now, we still have these courses. And, you know, your curriculum is the same as mine. But really, I think the future is going to be quite different where you say, like, you might be interested in rugby and you have a trip planned to, I don't know, France. I think France also is big into rugby. And you could then custom design a course just for you. It is exactly the vocabulary you want to use when you go to France. You know, it could be completely different from mine.""

""We use AI in three ways at Duolingo. Number one is content generation. And that's where we've seen, you know, the biggest success. Number two is AI features that you previously couldn't build. So this is, again, this video call with Lily. It's an interactive conversation with an AI bot. Two years ago or three years ago, you couldn't build. Now you can do it. And it's actually the thing that, you know, was lacking in the product. If you ask our users, like what's missing in Duolingo? It's like this conversational piece, like speaking. I want to get better at speaking. And now we can do this. And it's a really, really, really powerful feature. Great adoption with this video call with Lily feature. And I think there's more to be done there, both making this feature better, but then also maybe there's new AI features that we can have built. And the third is overall productivity improvements across the company. Software engineers using cursor, using AI for customer support.""

""Duolingo is surprisingly egalitarian. And we just say, use 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.""

""On the software engineering side, I think what they're really good at is going from zero to 80% for a simple app. The repli, the lovable. And then the larger code base gets, the worse they get. Adding 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. And then the other thing they're really good at is just isolated transformation. So in a single file, add all the comments or change a function to use different input parameters or something like that. Really good at that. What they're not good at is large code bases and adding a feature, even a simple one, to a large code base.""

"So in a way, what these AI tools do is they lower the entry level to get into software engineering, right? So, you know, your mom can maybe build an app with lovable or something like that. She has to. Right, right. So there you go."

"People think like, oh, CS means coding. What they teach is kind of the fundamentals of not even software engineering, but like how computers work. And then computer science is kind of applied mathematics. Okay. So I do believe there's still a lot of value in thinking logically, which is something that the best CS courses teach you. They don't teach you Java or Python and they shouldn't. I think that part is going to go away."

"The general consensus is that these AI tools are really good for senior software engineers or staff level software engineers, right? Yeah, they have been trained and now they can use AI and they become much more productive. They're now a 10x engineer."

"Maybe it's going to be next 10 years if you really see a broad adoption of these technologies. More likely, maybe 20 years or so. Today in the US, if you run into a stranger at the conference summit, the first question is usually what do you do? But in this future world, the first question might be like, what are your hobbies? You know, because we don't have jobs anymore. We have UBI, et cetera. It's like the paradise. Do you actually buy that?"

"I think we still hire from universities. One of the reasons is we believe that people that grow up with these tools, and, you know, start using these tools early will become much better at using them. So in a way, yes, they're... Because young people have more plastic minds to new tools and new processes? I think it's just like, I bet the person who spent a thousand hours using GPT is a lot better at using it than I am. And I think that's the case for the new grads."

""It's like, you need to know your target demographic. And, you know, Zaria, who was on your show before, she's Gen Z. She just fundamentally understands it in a way that someone who's not Gen Z could never understand it. In fact, when I open TikTok and look at our own Duolingo, I don't get it. It's like, why is this funny? Why do people watch this? But it's super, super popular.""

""For a lot of the, you know, AI can do 70 to 80% of the tickets that we get. And then people always say, oh, well, okay, you guys, you don't need human customer support agents anymore. That's not true because you still have the other 30% or so. But then not only that, now that, you know, you have these kind of like an unlimited AI customer support agent, you can actually give customer support to a much broader base.""

Why is it harder to raise $3 million than $100 million, and why does Severin advise prioritizing tier one VCs despite worse terms?

Severin explains that early-stage fundraising is harder because there are fewer investors willing to take risk on unproven startups with no or minimal revenue. At this stage, investor quality serves as an important signal to the market, future investors, partners, and customers, which can enable subsequent fundraising and growth. Therefore, even if the financial terms are slightly worse with tier one VCs, the signaling, expertise, and access they bring outweigh the costs. This approach can facilitate better company trajectory and valuation in the long term.

What does 'AI-first' mean at Duolingo, and how has the adoption of GPT-4 transformed their operations?

Being 'AI-first' for Duolingo means fundamentally changing how content is produced and operations are structured to harness AI generative capabilities. Severin describes that after partnering with OpenAI and accessing GPT-4, they accelerated course production from 12 years to one year for a similar volume. Instead of building everything manually, AI models generate content within human-defined constraints, which increases speed, scalability, and adaptability. This shift represents a major strategic pivot, making AI central rather than peripheral in product development.

How does Duolingo balance AI content creation with human oversight in curriculum design?

Though AI generates sentences and course content, humans still design curriculum constraints and verify quality — ensuring new words are introduced one at a time, grammar concepts are correctly applied, and viability is maintained. This human-in-the-loop model preserves educational integrity and efficacy, given AI’s limitations. It reflects a pragmatic approach where AI empowers humans rather than replaces them, maintaining control over learning outcomes and user experience.

What advantages does Duolingo’s massive user base provide in terms of learning data and product development?

With over 100 million monthly active users, Duolingo operates like the world’s largest school, generating extensive learning data. This data enables precise measurement of curriculum effectiveness using real-world metrics such as completion rates and user satisfaction, which small traditional education studies cannot provide. The scale allows continuous optimization of courses and personalized learning experiences powered by AI. This data advantage forms a barrier to entry for competitors and legitimizes data-driven education innovation as fundamental to their success.