Why Speak's Founders Picked South Korea as Their First Market, and What They Learned About Building an AI Tutor
EO Korea
Speak is an AI-powered language-learning app founded by Connor Zwick (CEO) and Andrew Hsu (CTO). Its founders describe it as "the best way to learn a language to fluency," using AI to create a hyper-personalized learning experience. According to the founders, roughly one in every ten South Koreans has signed up for the app. About a year and a half before this interview, the company reached unicorn status in a funding round led by Accel, with existing investors such as the OpenAI Startup Fund participating.
Two things stand out in their account. The first is that a San Francisco team, neither of whose founders had ever been to Korea, chose South Korea as its first market. The second is their argument that AI is finally making personalized one-on-one tutoring economically possible for everyone. The stated vision is to "reinvent the way that people learn, starting with language," and eventually to expand the product and mission to support any form of learning.
Where to Be Versus Where to Start
Connor separates two questions that founders often treat as one: where the team should be based, and where the company should launch its first market. Speak answered them differently. The team was in San Francisco, and the first market was South Korea. He calls this combination particularly unusual.
For the team's location, he thinks "ecosystem" is the key word. A startup wants investors, prospective employees and talent to hire, and a community to learn from and form partnerships with. He says hubs like San Francisco show that power most clearly. He considers the first-market decision just as important, though. He points out that many Y Combinator companies that stay in San Francisco pick first markets very far away from it. He argues this matters especially for teams based in San Francisco, which he calls "a complete bubble" and "a very weird place in many ways." Getting out of that bubble, in his view, is how a company gets real information from the real world.
Why Korea, Despite the Obvious Drawbacks
Connor says the founders came to Korea with a clear-eyed view of the market's problems. When they visited and talked to users, they found it hyper-competitive and not growing. He says there were "many reasons to not enter the market." They also knew that many founders from outside the United States would rather start with the American market.
What convinced them, he explains, was that despite those difficulties, Korea was where people were most motivated to find something that truly worked. Speak's whole thesis is that AI can make learning more effective. So the founders wanted a first market that would give them the most useful signal on whether they were building the right thing. Their bet was that entering Korea would be harder in the short term but more likely to lead to long-term success. Connor suggests there is a lesson in this for any company, and adds with a laugh that maybe people just think they're crazy.
The First Technical Problem: Understanding Accented Speech
Andrew describes the early engineering challenge. Speak had to accurately understand very heavily accented non-native speakers of English. Many components had to be developed in-house. That work is what unlocked the first version of the app, which could understand what users said and have them repeat a line.
He says the lessons from fine-tuning speech recognition models for particular accents turned out to generalize well. They helped when Speak began expanding beyond Korea several years ago to Japan, Europe, the US and elsewhere. He describes Korea as the place where the team learned "foundational truths" about the problem, about who their learner is, and about how best to serve them with the product.
A Learned Fear of Making Mistakes
The founders say they developed a deep understanding of the psychological angle from which people in Korea approach English. They observed a huge fear of making mistakes and describe it as "learned," which they attribute to the education system. They see this fear as a trap that causes people's ability to speak English fluently to plateau.
To address it, Speak ran a pop-up event called "English Exorcism." Users were invited to come and get rid of their fear of speaking English.
According to the founders, the same pattern appeared when they expanded into Japan and then other markets. It was far more prevalent than they had originally anticipated. The way they built this insight into the product has tended to work across many markets. This leads Connor to a broader conclusion. They believe in localizing for every individual market, but he thinks that "at the end of the day, the way that you solve the problem of learning and teaching is universal." Whatever culture people come from, he says, the solution looks pretty much the same. He finds it somewhat surprising how well the core product has worked in very different markets.
Two Unconventional Educations
Both founders connect the company's mission to their own schooling.
One founder remembers always wanting to learn in their own way and getting into trouble as a kid for not following the rules. A fifth-grade teacher recognized this. Instead of forcing conventional classroom learning, the teacher handed over a textbook and let the student learn alone, awarding extra credit whenever the student answered their own questions. The founder says this was exactly what they needed. What made it so impactful was that the teacher accommodated one student's particular needs in a classroom of probably 30. The founder hopes nearly everyone has had at least one teacher who changed their life, and sees that as evidence of what high-quality education can do.
The other founder describes a very unusual path. As a curious child, they chafed at the public school system, and their parents eventually took them out of school to homeschool them. Learning at an accelerated pace, they went to college at 12 and started a PhD at 16. They say this experience instilled a core belief: everyone can learn ten times faster than they think possible. They acknowledge the controversy and turmoil around AI's effects on the world. Still, they argue that frontier AI models can bring a meaningful part of that kind of experience to anyone on the globe, and that Speak is trying to use AI to "make humans stronger and better."
Bloom's Two Sigma Problem
The founders cite Bloom's two sigma problem as one of Speak's original inspirations. They describe it as a famous pedagogical study by Benjamin Bloom in the 1980s. In it, a typical classroom of 30 students to one teacher was compared with a treatment group of students given one-on-one, mastery-based tutoring. As the founders tell it, the tutored students performed two standard deviations above their results in the normal classroom model, wherever they had stood on the bell curve.
The name, they explain, comes from the question that followed immediately: how could this be given to every person in the world? Forty years later, no one had solved it, because giving everyone a tutor was not economically viable. The founder speaking argues that AI changes this. In their view, it has become "incredibly obvious" in 2026 that this will be possible and is almost inevitable. They predict that within the next few years we will live in a "two sigma world," where anyone who wants it can have a 99th-percentile outcome.
Riding a Secular Trend
Connor also frames the company's timing as a strategic choice. He argues that almost every great company became great partly by riding a massive secular trend. In those moments, he says, incumbents and competitors don't matter. His example is Google, which he says rode the wave of the internet and of information coming online. The founders believed AI would be "the single greatest technology that humans had ever created" over the next ten years, and they wanted to ride that wave with a very smart tutoring system.
A Secret Test From OpenAI
The founders recount one turning point. At the time they were partnered with OpenAI but had not yet received its investment. Sam had angel-invested in Speak years earlier, so the relationship was close. He knew Speak faced unusual speech problems. One day, the founder recalls, he called and asked whether they could keep a secret, because OpenAI wanted to test something with them. It was a speech recognition model, and OpenAI described it as so accurate that they didn't measure error rate in the usual sense. Instead, they measured whether it made any errors at all.
The founder remembers that Connor and Andrew looked at each other and thought there was no way it would work. When they tested it on their own data, they were "completely blown away." They call it an absolute transformation and a massive "aha" moment. They could finally trust their transcription of what users said, and then use a language model to build a much more open-ended experience.
"The Model Is Not the Tutor"
Looking ahead, the founders say that as frontier models grow smarter, capability thresholds unlock entirely new products. Two capabilities they see as very close are agentic intelligence and real-time models. In their view, these will enable an AI tutor that can own and guide a learner's entire journey and provably help them reach their goals as efficiently as possible.
They also address a pressure they see across the industry. Application-layer companies feel squeezed from two sides. On one side are the foundation model companies. On the other are new entrants, where one or two people "vibe coding" can build overnight what would have taken years only a year or two ago.
Their answer is the phrase "the model is not the tutor." The model supplies a huge amount of raw, general intelligence. The tutor, in their framing, is the system built around the model to direct that intelligence, which they describe as a deep and complex product and technical system. As an example, a good language tutor must hold an extremely immersive conversation. They say today's real-time voice models are close but not quite there.
They argue the fundamentals still matter most. Their learners pursue an outcome that takes years and requires daily dedicated work, so they look for the most effective, purpose-built solution. The founders even see foundation models as helpful. They improve Speak's product, and many people use an app like ChatGPT, realize AI is useful for language learning, find it falls short because it is not a purpose-built tutoring product, and then look for something like Speak. Speak's focus, they say, is building foundational technology, using its data as a flywheel to improve its models, and layering scaffolded machine-learning systems on top.
Mission as the One Constant
In closing, the founders reflect on what a founder can rely on when the world changes so quickly. One says there are very few truths to hold on to. That is why a company's core mission and long-term vision matter: if you have conviction in them, they can stay constant while the path to reach them shifts dramatically as the landscape changes.
They recall that in Speak's early days, every version of the app seemed not to work, no one cared, and the situation felt "highly existential." What got them through was the mission. They stress that it mattered for more than motivation. A strongly held vision helped them keep making long-term decisions that built the compounding advantages they would need. They say this was important because it is always tempting to work on short-term things when times are hard.
The final reflection frames company-building as a sequence of judgment calls. One founder suggests that a startup's chance of success can be approximated by the percentage of decisions it gets right. If a team keeps seeking the truth, keeps adapting, gets more of those calls right, and is correct about the long-term direction of the world, then, in their words, "you've got a shot at succeeding."
So we've been live in Korea for many years at this point. We have roughly one out of every 10 South Koreans have signed up for Speak at this point. And when we visited South Korea and we talked to users here, we saw it was hyper competitive market and it wasn't a growing market. There were many reasons to not enter the market. We started with South Korea though because we believe going into this market while it would be more difficult in the short term would lead to a higher likelihood of succeeding in the long run. We developed a very deep understanding of the psychological angle upon which people in Korea approach English.
Yes. Yes, I'm fine.
There's a huge fear, a learned fear of making mistakes. And we see that be this huge pattern and a trap plateauing people's ability to actually fluently speak English. We are running a pop-up event called English Exorcism where we're inviting a bunch of our users to come and get rid of their fear of speaking English.
He there or is it brown? Brown hair.
Yeah,
that's not the guy.
Somewhere like San Francisco that's a complete bubble. It's incredibly important to make sure that you're getting out of that bubble and getting like real information from the real world. I think there's a lesson in there for any company. Maybe people just think we're crazy. A year and a half ago, we reached a unicorn status with a fundraise led by Accel with a bunch of existing investors like the OpenAI Startup Fund. Hi, I'm Connor, founder and CEO of Speak.
And I'm Andrew, the other co-founder and CTO of Speak. And Speak is the best way to learn a language to fluency and it uses AI to create a hyperpersonalized learning experience. And so the vision for the company has been reinvent the way that people learn starting with language. Eventually we believe that the product and mission of the company can be expanded and generalized to support any form of learning.
I think one of the most unique parts of the Speak story up until now is that we were this San Francisco based team that somehow decided to start in South Korea of all markets despite neither of us having ever been to Korea before in our lives. There's fundamentally two different questions around when you're starting a company in terms of the location that you pick. One question is where should you be as a team and a second question is where should you start as your first market.
In the case of Speak we started in San Francisco but our first market was South Korea, which is a particularly unusual combination. When I think about where startups should be formed I think ecosystem is the key word because you want to find a place where there's going to be an ecosystem of investors, of prospective employees and talent that you can hire and a community of people that you can learn from or create partnerships with. That's where you see the power of technology startup hubs like San Francisco really show the impact. But I think just as important of a choice is where your first market is.
And I think a lot of the companies for example that stay in San Francisco from Y Combinator, they pick places that are very far away from San Francisco as their first market. And I think that that's a very important distinction. And especially if you are based somewhere like San Francisco that's a complete bubble, it's incredibly important to make sure that you're getting out of that bubble and getting like real information from the real world because it's a very weird place in many ways. The reason we started with South Korea though, despite many people from outside of the United States wanting to start with the American market, was because we believed in the long-term vision of the company and we wanted to make sure that wherever we started, we were going to get the best and most useful signal on whether or not we were building the right thing. And when we visited South Korea and we talked to users here, we saw it was hyper competitive market and it wasn't a growing market. There are many reasons to not enter the market, but the thing that convinced us it was the correct market to start with was despite all those difficulties, it was the market where the people that were the most motivated to actually find something that truly worked. It was truly effective. And so for our use case and our product category, the whole thesis is that you can use AI to make learning more effective, which meant that going into this market, while it would be more difficult in the short term, would lead to a higher likelihood of succeeding in the long run. I think there's a lesson in there for any company. Maybe people just think we're crazy.
I think a lot of the initial challenges in the early days that we had to solve were really around how do you accurately understand a very heavily accented non-native speaker of English. And there were many things that we had to develop custom. And that was what really unlocked the first version of the app where we could actually understand what users were saying and then have them repeat a line. A lot of the lessons that we learned around how do you fine-tune speech recognition models for certain types of accents have ended up being very generalizable and helping us as we started expanding several years ago outside of Korea to Japan, Europe, the US and so on. I think that in many ways I think of the Korea market as the place where we learned so many foundational truths about the problem that we're solving, who our learner is, and also the best way to serve them with the product. We developed a very deep understanding of the kind of psychological angle upon which people in Korea approach English.
Yes. Yes, I'm fine.
There's a huge fear, a learned fear of making mistakes through I think the education system here. And we see that be this huge pattern and a trap in terms of plateauing people's ability to actually fluently speak English.
So that's a great example of something that we saw when we expanded into Japan and it's the same thing we saw when we expanded into other markets as well and it was actually way more prevalent of a thing than we originally anticipated and I think the way that we've translated that into the product experience as well, it tends to work across many markets. I would say actually in fact that even though we really believe in the importance of localizing to every single individual market, I think at the end of the day the way that you solve the problem of learning and teaching is universal. Like I think all humans, no matter what culture you came from, the solution looks pretty much the same. And I think that's actually been something that maybe has been surprising, how much the core product just works in many very different markets.
When I was a student in school, I think the thing that I remember the most is that I always wanted to learn in my own way. I got into a lot of trouble as a kid because I would often not be following the rules. And I remember I had a specific teacher in fifth grade who I think recognized that in me and instead of forcing me to learn in a conventional manner in the classroom, he just gave me a textbook and let me go learn by myself. And every time I had questions, he would just give me extra credit to go answer them myself. That was exactly what I needed as a student. And I think the thing that was so impactful is that he recognized the unique things about what I needed as a student and figured out a way to accommodate that despite it being a classroom of, you know, probably 30 students. I think almost anyone or hopefully everyone has been fortunate enough to have at least one teacher in their lives at some point that changed their lives. And I think that really shows the power of what really high-quality education can do for somebody.
So I had a very unusual educational journey as a child. I was a very curious kid. I loved learning about the world and I sort of chafed a lot at the public school system. Long story short, my parents actually ended up pulling me out and homeschooling me and I was able to learn on my own at a much more accelerated pace and I ended up going to college when I was 12 and starting my PhD when I was 16. I think it really instilled this fundamental philosophy that everyone is really capable of learning 10 times faster than they think possible. There's obviously a lot of controversy and turmoil around the effects of AI on the world, but with frontier AI models, it's actually possible to bring a meaningful part of that type of experience to any human on the globe. And we are trying to use the power of AI to very directly make humans stronger and better and unlock that type of human potential.
One of the original inspirations is something called Bloom's two sigma problem. This was a famous pedagogical study that was done back in the 80s by a guy named Benjamin Bloom. And what he showed was an experiment where there was a typical classroom of 30 students to one teacher. And then as the treatment group, there were individual students that were given one-on-one mastery based tutoring. The results were profound. What he learned is that the one-on-one students, no matter where they were on the bell curve in the classroom, they performed two standard deviations above what they did in the normal classroom model. And so it showed the power of learning in the way that you can really accelerate how fast someone can learn something by giving them one-on-one tutoring. And the reason it's called Bloom's two sigma problem is that as soon as everyone realized that this is possible, the immediate next question is how could we give this to every single person in the world, and this was back in the '80s. So it's been 40 years since then and no one has figured it out because the reality is that giving every single person a tutor is just not viable economically until now. I think that everything changes with AI and I think it's become incredibly obvious now in today's age of 2026 that this is going to be super possible to do and is almost inevitable and I think that we'll be living in a two sigma world where every single person will have that 99th percentile outcome if they want it in the next few years.
Almost every single great company that's ever been created, part of the reason it was great was because it was riding a massive secular trend. And those are oftentimes the moments to create an incredible company. And those are the moments where I think incumbents and competitors don't matter. As an example, the reason Google is great is because they rode the wave of the internet and all of the information coming online. We saw an opportunity to use AI which we believed would be the single greatest technology that humans had ever created over the next 10 years and we saw an opportunity to ride that wave and so the vision for the company has been reinvent the way that people learn starting with language but eventually we believe that the product and mission of the company can be expanded and generalized to support any form of learning and I think that's the power of being able to use a really really smart tutoring system.
3.5. That was a really big moment for us. We were kind of partners with OpenAI at that point, but we hadn't actually had the investment. But Sam had angel invested in the company years earlier, so we were close. And he saw that we had all these interesting speech problems that were really unique. And one day he rang me up and he was like, "I know you have a bunch of really hard speech recognition problems. Can you keep a secret, and we want to test this thing with you." And they were like, so we have this speech recognition model and we don't measure error rate in the normal sense. We actually just measure whether or not it makes any errors at all because it's so good. Andrew and I both looked at each other and we were like, there's no way this is going to work. And I remember testing our data and being completely blown away. It was an absolute transformation and it was a massive aha moment for us because we could finally accurately trust our transcription of what the user was saying and then actually use a language model to generate a much more open-ended experience.
As the frontier models get smarter and smarter, there are thresholds of capability that unlock entirely new products. And the thing that is very close and is about to come is agentic intelligence and real-time models. These types of capabilities will actually unlock an AI tutor that can own and guide your entire learning journey. That can provably help a learner achieve their learning goals as efficiently as possible. I think a really common pattern that we're seeing now is that a lot of application layer companies feel like they're being competed against from both sides. On one side you have the foundation model companies and then on the other hand the new entrants where just one or two people vibe coding an app can build something overnight that would have taken years just a year or two ago.
The way that we think about it is that the AI model is not the tutor. The model is not the tutor. And what we mean by that is the model supplies a huge amount of raw general intelligence. It can do many things but the tutor itself is a system that we build around the model to direct that intelligence. So we think that there's actually a very deep and complex product and technical system. For example, a really good language tutor needs to have an extremely immersive conversation with you. And today's real-time voice models are close but not quite there.
Ultimately, the fundamentals are still what matters the most. When we think about our learners, they're motivated by an outcome that's going to take them years to get to and it's going to require dedicated work every single day. And ultimately what that means is that they're looking for a solution that is the most effective and purpose-built solution. The foundation models, they're actually really really great because they not only help our application get better, but we see a lot of people using an app like ChatGPT and they realize that AI is useful for language learning, but there's all sorts of reasons why it falls short because ultimately it's not an actual purpose-built tutoring product and then they go look for a solution like Speak. So for us, we're really focused on building the foundational technology, using our data as a flywheel to improve our models and build the scaffolded and machine learning systems on top of that to create a much more effective solution. Man, I think the world is changing so fast right now that there are very few things that a founder can hold on to as truths. This is why I think the core foundational mission of the company and the vision of the company long term are so important, because if you can have conviction around that, hopefully that's something you can hold constant at the highest level and then the pathway upon which you get there can shift tremendously as the landscape changes over time. I remember in the very early days of Speak it felt like every single version of the app that we were building just wasn't working and no one really cared and it was highly existential and what got us through in that moment was the mission. It wasn't just important from a motivational standpoint. Having a really strong vision that we had absolute conviction in also allowed us to continue to make sure that we've always made decisions for the long run that were compounding advantages we would need to succeed long term. I think that's really helped because it's also always very tempting to
work on short-term things when things get hard, and the vision and the mission have really helped us stick to what really matters.
In many ways, I think the chance of that startup succeeding could be approximated by the percentage of decisions that they can get correct. And operating a company is just a series of making a bunch of calls in a row. And if you can continue to truth-seek and adapt and make more calls correctly, and you're right about the long-term position of the world, then you've got a shot at succeeding.
Article published · Updated