How Speak Built a $1B AI Language Learning App

The the biggest problem with language learning historically is people go to classrooms or they used apps but all they got was a lot of vocabulary that they memorized and they weren't able to talk out loud. And that's why the company's called Speak. If you're learning a language with another person and you're asked to say something out loud, you're going to edit yourself a little bit because you don't want to sound dumb. You don't want to sound silly if you like really mess up the accent. This is like the first time that people were really speaking a language with no one else in the room and they loved it. If you offer regular classroom setting education versus one-on-one mastery uh where you're learning by doing, you will get two standard deviations above or 99 percentile if you go uh the one-on-one route. Welcome to Subversive, a podcast dedicated to sharing stories from the best consumer

subscription apps in the world. We'll bring you lessons for how to grow your consumer subscription business, including insights and inflection points that led to exponential growth from leaders at category defining companies and innovative startups. Let's get into the show. This episode is brought to you by RevenueCat. The world's best apps use RevenueCat to power purchases, manage customer data, and grow revenue on iOS, Android, and the web. Join more than 100,000 apps that are already using RevenueCat to grow their businesses and track more than 10 billion in spend, including category leaders like OpenAI, Notion, and Ladder. Learn more at revenuecat.com/subversive.

My guest today is Connor Zwick, co-founder and CEO of Speak, an AI powered language learning app that recently raised its series C at a $1 billion valuation. In this episode, I talked to Connor about how Speak built an AI native product that helps learners practice real conversations from day one. We also discuss how Speak achieved breakout growth in a crowded category, why the company went allin on AI voice early, and what it's learned about building consumer trust and maintaining long-term subscriber retention with a product powered by LLMs. All right, welcome Connor to the Subversive podcast. I'm really excited to have you on today to talk about Speak, uh, which is the hottest new language learning app making the rounds around the tech scene these days. is lots of people talking about speak and how it's using AI to make language learning uh easier and better uh for a lot of people. Before we

jump into it though, would love to just hear a little bit about how you see Speak and how you would describe it to people who maybe haven't heard of it yet. >> Yeah, thanks for having me Phil. Um so speak is trying to build the best way to learn a language of fluency and the way we do that is by uh offering an AI first approach where you can uh practice speaking with our AI teachers instead of real teachers. Um you're free of the embarrassment. It's available whenever wherever and uh we have like a superhuman memory of every mistake you make and every lesson's made just for you. Um, so in many ways it uh the the biggest problem um with language learning historically is people go to classrooms or they used apps but all they got was a lot of vocabulary that they memorized and they weren't able to talk out loud and that's why the company's called Speak. Uh because there's billions of people out there

trying to learn languages and we are uh we're trying to solve that. >> Yeah, it's certainly a great name and you didn't specifically mention Dolingo there but I'm sure we're going to talk a lot about them later in the episode.

Before we jump more into speak though, I would love to just talk a little bit more about your background. So my understanding is you launched your first startup flashcards plus when you were still in high school, eventually sold it to CHEG. Tell me a little bit more about that journey and and what inspired you to be an entrepreneur so early in life. I I think for me it was always just more about building. Um, I I remember in high school I discovered um the the iPhone had just come out and I'd been learning how to program just to build things and um I just found that to be a fun fun thing that I was just really curious how to do as a kid. And uh once the iPhone came out, I was like, well, I should um I should build an app. I was a student at the time, so solve my own problems. I noticed that all of my friends in school were using this note cards that kind of looked like an iPhone on sideways and I

built I built it. Uh I had no idea um about growth in distribution but it was early days of the app store. So you know it was uh good apps just kind of rose to the top and um and then it just kind of took off um and and the rest was history. >> Yeah, I mean those were the good old days where there were so so far fewer apps that growth and distribution didn't matter as much. certainly looks a little bit different these days. Although AI is creating a whole new tech wave that that's creating new growth opportunities which we'll talk talk a little bit more about later.

>> Um so so after high school you end up you go to Harvard but you drop out to join the the deal fellow fellowship. >> Um tell me a little bit more about what gave you the courage to take that leap and and what that experience looked like for you. >> Well once I kind of had the exposure of flashcards in high school. It was on my radar of maybe not going to college. And I actually the day after I graduated from high school, I moved out to San Francisco. There was like a VC that was trying to invest in Flashards Plus and I was like, I'm 17. No thank you. I don't really want to take an investor uh on this side project of mine. Um and uh they were like just fly out anyway and just spend the summer here. And so I spent the summer in San Francisco and I was kind of on the fence about, you know, whether to go to college at all.

And I had a miserable time because I was literally like I just turned 18 and knew nobody in the city. So, uh, it was like kind of lousy. And I went to school for one year. Uh, and during that year, I loved I loved school. It was amazing to like just be able to it just spend time learning and meeting a bunch of really cool people. But I just kind of had this like nagging feeling that I was leaving something. I I ended up not going to class ever and just building things. And uh I started to to to meet people uh that were in tech and they were all you know in San Francisco and so I actually ended up coming out like all the time like every other weekend um for something or another and so by the end of the year that's when I had heard of the teal fellowship and I applied. This was like the second year of it so it was still relatively new. I think it kind of gave me the the push um to do it. It's

I'm not sure if I would have dropped out without it. I I think I might have. Um, but what it really offered was like a community, I think, more than anything. And that ended up being like the insanely valuable part about it.

>> Yeah, that makes a lot of sense. I mean, you talked about how when you first moved to San Francisco, one of the big things that was missing was you didn't know anyone there. And so, you have this miserable experience. You go to college, you have a great community there, but you miss the allure of building things and being an entrepreneur. And so, the the fellowship, it sounds like, provided both. And so, it was a great fit for you. And one of the people you met in that community was Andrew Shu, who ended up becoming your co-founder at Speak.

So, what did that relationship look like? And and then to transition into the primary topic for today, what what inspired you guys to to start speak in a in a category like language learning where there were already some pretty large incumbents? I I met Andrew through the jail fellowship and I think that community was a there was a lot of self- selection that went on. So, it was a really interesting crew of people and we were both kind of at a crossroads at a similar time. Uh I remember one of one of the teal fellows in my class was Chris Chris Ola who went on to be one of the co-founders of Anthropic and even back then I think he was one of the first people that I remember there were several people in our class back in 2012 that were talking about deep learning. I had never heard of it before. Um and it still kind of felt like this crazy concept to me. They were talking about neural networks. But then over the next

few years you would you you'd hear about like how all of a sudden image recognition was actually really good now with neural networks. and you started hearing about some of these some of these releases. And so when we were at a crossroads um I had sold flashcards plus I was thinking about what was next and Andrew had done something similar. We we both uh were getting really excited about what what was happening in the research side. And so we we decided we were going to dedicate a year and a half to just trying to learn as much as possible about about AI. Um and this was back in like 20 2015 or so. Back then there were there was actually a ton of activity. It felt like a really big deal at the time, but it was a much smaller community than it is today. Um, and we were fortunate enough that, you know, we lived in the Bay Area. And so there were three places that were publishing papers. It was Google, Stanford, and

Berkeley. Google locked their doors. Stanford and Berkeley didn't. And so we just um we uh we just uh uh snuck into these courses and um and kind of learned as much as possible and tried to implement all these papers. And um one day we we built this uh the speech model system and we trained it over the weekend off of a bunch of uh YouTube data we scraped and it like beat all the state-of-the-art records on um accent detection. And that was when we started to think about okay this could be immediately useful for something that feels really important which is uh empowering a better educational model.

Um and and that was kind of the the groundwork of what ended up turning into the vision for speak. Got it. That makes so much sense. I studied I was a double major in college. I studied economics in Chinese. So I spent almost half my time in college in Beijing and other parts of China. >> And you know, Mandarin is maybe an outlier when it comes to this, but with Mandarin, your pronunciation is so important because you have the different tones and inflections. And so you could say the same sound but with the wrong inflection and it would mean something totally different. And sometimes that can get you into a lot of trouble. And so as somebody who's used Duallingingo for a long time and you know and generally likes the product, I can see how while it's a lot of fun and while you learn a lot of vocabulary and some basic sentences, ultimately if you really want to speak the language and be immersed in it, you got to you got to

speak. And so it makes sense that there was this sort of gap in the market at least at that time. I mean at this point Dolingo has launched Max which does include immersive language learning and speaking. So we'll come back to that. But but let's stick with the origin story and in those early days of of speech formation. And it sounds like you and Andrew were were initially playing around with lots of different applications of deep learning and you sort of stumble on to this application of a model that helps with language learning and pronunciation. What what what came next? What how did that lead to the formation of speak as a company?

When we saw how powerful these models were, we had this this um kind of moment of clarity where we could envision how more training data, more compute um would lead to better and better models. And uh we could see, you know, at first maybe we could just build a better model for understanding speech, but eventually we would be able to to build a general solution for uh building like a full endto-end tutor. Um, and we were really fascinated with this idea of um, building AI systems that felt like you were building a relationship with another person and enabled you to improve yourself through education. It was a concept I think both of us uh, found to be extremely compelling and we became convinced that it would be possible to to build something like that over the next 10 years. And um, that's been that was basically the vision for the company. And so we started with language learning mostly because we saw

this validation that we could we could start with a um a technical niche if you will which was just being able to understand people who were speaking with a strong learner's accent. Um and so you know Dualingo existed back then but uh they didn't have a lot of speaking. Rosetta Stone existed back then they didn't have a ton of speaking and the reason was because the error rate for the models that they were using was probably like 80%. like you it just did not pick up any of the words that people were saying. And so um even just by being being able to start with that, we saw that that could be the initial wedge to building a product with more distribution, more data, um you know, build uh grow the company, get more compute, train better models, and just like continue that flywheel until eventually we have the scale to build, you know, the superhuman tutor that we thought was always going to be possible.

Well, and you mentioned there that you all had the foresight to recognize that over the next 10 years, you could build a product into this trend around AI that would allow humans to learn a language by speaking with an AI. But you were pretty far ahead of your time at that point or at least from my standpoint as as a lay person, it it it feels like you probably were. And so I guess there are two directions I want to take this. One is on the human side, how did you

validate that humans would actually be comfortable speaking to a nonhuman as a way of learning a language? And what were some of the steps you took to validate that? And then I want to get later to the technical constraints you faced that early on and how you got around some of those constraints until the technology caught up to the point where you could do everything you wanted to do. In the beginning, it was a bigger question of whether or not we'd even be able to pull it off, whether or not we'd even be able to build a model that was accurate enough um to be a pleasant user experience. Um but after we had kind of gotten to that point, I remember testing uh in the early days with with users and um all we were doing was trying to find people who were in the Bay Area who like could not speak good English because that was like the the market that was clearly the biggest opportunity. um and

then test uh how the experience felt. And something we learned very quickly is that if we were in the room testing with them, it was very different than if we gave them the product to try on their own. >> And that was like a really big aha moment for us because >> uh we quickly realized one of the biggest benefits of speaking with an AI is that there was no fear of judgment. Mhm. >> Um, and it's something that people really underrate, but uh, and you don't even really, you're not really even conscious of it, but if if you're learning a language with another person and you're asked to say something out loud, you're going to edit yourself a little bit because you don't want to sound dumb. You don't want to sound silly if you like really mess up the accent. And um, and so this is like the first time that people were really speaking a language with no one else in the room. And um, they loved it. And so

that was like a really early signal for us that okay, there's something here that is fundamentally different than what's what's kind of existed and and there's a real there there. Yeah, I love that. And I happen to have had Natalia Castjo on the podcast last year who's uh director of product at Dualingo. And she talked about with the launch of Dualingo Max, how they observed a lot of the same things about how people were more willing to take risks and to stretch their language abilities a little bit beyond their comfort zone that they might not do if they were speaking with a human teacher or even a human peer.

But that was many years later. And so it sounds like you guys came across this insight much earlier in the formation of

your company. Was there a tipping point in the metrics or in the progression of the product where you you really saw this start to come to fruition in in your data? Right? You you have these anecdotal examples of people in user interviews talking about uh how much more comfortable they feel learning language with an AI, speaking with an AI, but when did it really start to show up in the numbers? speak is really a story of like several major step functions over time and they've continued to kind of unlock um unlock growth on the on the metric side. One of the really big uh moments of inflectionship for us was really just building our first um phone recognizer, which is basically a phonetic model for understanding the actual sounds you're making. um which is fairly early on in the journey and um rolling that out in the product and being able to give people pronunciation feedback that was actually like reliable and accurate to

the individual phone name and being able to like coach them on how to how to sound better. Um and immediately when we launched that we saw um basically like the entire funnel get substantially stronger as a result. Um I think another obvious big moment was when LLM started to get good enough that we could start to enable real open-ended experiences. Um because for the first few years of speak the language modeling side was was quite weak. We were really good on the speech side but we were not so strong on the on the language model side because language models were just quite bad. And so we couldn't give um open-ended experiences. Everything was very structured. We were like very early to the game around launching something you know a a roleplay feature which is something we're very well known for where you can kind of um practice any scenario and there's a whole suite of features around that. Um when we

launched that that was another moment that was um that uh completely changed uh like long-term retention and Dowo and time spent in app um in in really interesting ways. Yeah. Well, one of the things I love about what you're saying is I think a lot of outside observers tend to look at a company's trajectory and just assume it was this straight or exponential curve up and to the right. But in reality, it's often these jolts, these leaps forward that come from core product innovations, which in your case sound like they were often quite technical, right? Changes and models that made the core language learning experience much better, whether that was the phone names that you talked about or the language model getting better. So

let's talk more about some of the technical challenges that you faced early on in Speak's journey and then how those challenges have gotten easier or what new challenges have been created by AI over time. You launched conversational AI well before most other consumer apps. What were some of the hardest parts of of doing that before some of the LLM technology really caught up to where it is today? a challenge that we've been facing from the very early days of the company um as being an AI native company and building a lot of features on non-deterministic or probabilistic models um is the idea that these models do not work 100% of the time and I this is actually not really a technical challenge but I think it's u an interesting one in in that I think something that speaks become very good at and a lot of companies now need to are coping with over the last few years as they've become more AI native is this

idea of like how do you build really great product experiences on a model that works 90% of the time 95% of the time 99% of the time it's a it's a very different thing and how do you like kind of test the product and make sure it's good enough um how do you um think about the failure cases it's also very different if you know it's failed versus a hallucination that that you there's no way of knowing that it's hallucinating while it's happening um I think it's It's also a fundamental challenge because um most AI products are still human in the loop. In other words, they're actually helping humans become more productive in something versus um fully replacing humans for a service. Um and speak is very much in the second bucket, right? There's no human that's like checking um what's happening in the experience. And so understanding and building reliability there. And then also understanding what's possible

versus what seems possible with a demo, but then actually in practice isn't a consistently good enough experience is I think like one of the defining core challenges we have been facing since the

very beginning. Well, the probabilistic thing makes a lot of sense. And I can imagine, especially for new users trying to speak for the first time, if if they're in an interactive session with an AI and they're they're learning their language and then it just sort of hallucinates and go goes off in a strange direction, that would really undermine trust. Um, any funny stories in the early days of customer support tickets where somebody wrote in with, "What is what is my AI teacher talking about?" Well, there's a lot of um there's definitely a lot of sensitivity around even how you like moderate the models. We were the um first company to use GBD4 in production. We were like three or four months before opening released in CHP. We were kind of like the guinea pig and uh I remember testing it uh with 30 power users and we gave it to them and we checked back like 3 days later and one of them had used the

product um I think they had used it for an average of like 8 hours a day for the previous 3 days. We were like okay there's clearly something wrong with our logs. Something's like not working here cuz there's no way that anyone used our product for like 8 hours a day for the last 3 days. And um it turns out it was a person that basically figured out that they could use it to simulate the person being their like boyfriend.

>> And we became really freaked out. Uh we were like, "Okay, this we've always kind of known that the, you know, the the ultimate way to engage engagement is, you know, to to fall in the slippery slope." And so we added all this moderation and and we were really like afraid of you know maybe there was going to be a controversy where someone would have a screenshot of speak that would go viral where the you know tutor was doing something totally inappropriate. Um and it turned out actually over time the much bigger risk was the opposite and and people were actually offended that it was censoring them or or you know uh they would have these very specific scenarios and so we've actually over time had to really dial back that moderation. Um, and that is like a really kind of interesting challenge that obviously we're not the only ones that face. That makes a lot of sense.

Well, and I guess another question I'll ask is how many languages were you supporting when speak first launched? I assume it was a smaller number of languages and then over time you've you've probably opened up to more and more languages. So So how many languages did you start with? How many are you supporting today? And were there any particular languages that were most difficult to get right? >> Yeah. So the interesting thing about speak uh is that we made the crazy decision early on to not just teach English which was u by far like the biggest market that we saw there's like one and a half billion people trying to learn English around the world but we decided to focus on a specific language um learning English from and we decided to pick Korean and so for like the first probably four years of the company we were exclusively focused on building product market fit for a consumer language learning app in South Korea,

which is a place that me and my co-founder had never been before we started the company. The challenges with that were uh kind of obvious, right? Like um first of all, South Korea is kind of famous for being a market where um tech western tech companies like really struggle to to to gain a foothold. On the other hand, what we saw was it was an incredibly competitive market, but a market where people were absolutely obsessed with learning English and all they cared about was efficacy and whether it would work. And so, in many ways, it was like the market where we would get the most signal on whether or not our product was actually working versus just, you know, effectively engaging them. That was the only language pair that we focused on for a very long time and everyone thought we were really crazy. to fast forward to today over the last few years um because of um how much better our AI systems have gotten uh we've been able

to rapidly localize and expand to to many languages. So now we support um we teach English to uh you know probably uh 12 plus first languages and we've uh we've launched a bunch of new languages over the past year and we now support you can learn uh five different languages on on speak as well. What are the five languages just out of curiosity? So, it is um Spanish, uh, Italian, French, Japanese, and, uh, Korean. This episode is brought to you by Paddle. Today's fastest growing apps are multi-channel. They use web monetization to reach new audiences, improve attribution, and boost margins.

But selling on the web gets complicated fast. Managing currencies, taxes, fraud, and disputes gets increasingly difficult as companies scale. Paddle solves this. As the merchant of record for thousands of apps, including Runna, GoodNotes, Codeway, and HubX, they take on the complexity of web monetization so you can focus on growing your business. Learn more at paddle.com/subversive.

And what made you decide to start with Korea? Was it the fact that you thought it would give you the most signal because it was such a difficult market to get product market fit in, or were there other reasons that you chose Korea in particular? So in the early days of uh of the company, we did not have a market in mind and we did a lot of different things to to figure out which market to go after. We did obviously we did market research um and we tried to analyze the market opportunity. We actually traveled to a bunch of countries to talk to users. Um and when I was in Korea, uh it became apparent just how much of an obsession the country had um over learning English.

There was a crazy stat back in the day. I don't know if it's still true, but they were spending 1% of their GDP on learning English. >> Wow. >> And if you like drive down the street in Gangdum, like the downtown financial district, um there are like literal entire skyscrapers just dedicated to English classrooms that people would go to after work or after school. It's it's like a national pastime there. Um the problem is like there's also it was by far the most well served market, which is actually the thing that we were looking after. So basically we we we realized that uh ultimately the bet for speak is that we could build a 10x better product. Um and the only way we would be able to rise above the noise in a market like that would be if the product was actually truly differentiated. Um and so that's that's ultimately why we set our sites which made our jobs that much harder but I think set us up for uh I think a much

more kind of long-term um road to success. Yeah, it makes a lot of sense. I mean, you hear all the time that it's important for startups to focus. There's the whole Graham dig deep wells quote, right? You you can't afford to go broad and deep at the same time. And if you're a startup, you have limited resources. So, you got to go deep and find product market fit in one place. What I haven't heard as often is actually taking the approach of making it harder on yourselves in a way by going to a market where the bar is very high, but you know, if you can clear that bar, I mean, there's certainly a lot of demand. So, that was playing in your favor, right?

But the level of expectation was very high. And so almost like I live in Boulder and so it's almost like all the Olympic athletes that are here at the Olympic training center in Colorado Springs, right? They're training at altitude because then it just makes their lives that much easier when they're actually in competition. And so

you spend the first four years focusing exclusively on teaching Koreans how to learn English. At what point did you decide you were ready to expand to other languages? And to what degree did advances in AI technology make that easier? The second market we launched was um in Japan again teaching English and I think it was a culmination of two factors uh that that coalesed. The first was that we felt like our product market fit was strong enough. We felt like we had advanced the product. We had enough we had stability. We knew the core foundations of our pedagogical method.

The technology platform the the fundamental product offering was all now very robust and it was working. Um and we had a very clear like vision for how it would progress from from there. And um that happened kind of in a very similar timeline to the um ability uh to to use LLM to localize and much more quickly build our curriculum which is a huge investment we made up front for uh Korea. Um and that was also possible to accelerate that. And so um going from one market to two markets, I think this is the classic playbook problem, but we uh like discovered all of the things that we we we you know didn't anticipate and um and then we were able to automate a lot of that and then rapidly go from you know that second market to to many markets over the following year. Got it.

It makes sense. And then I I assume from there once you figured out Japanese then you could quickly move on to the next set of languages. And it's and it feels like over the last few years, you guys have really crossed more into the mainstream. I'm hearing your company's name come up a lot more often. You recently raised your series C at a billion dollar valuation. Uh, and I imagine that as that's happened,

obviously building an incredible core product experience sits at the center of everything, but you've probably had to invest more and more into figuring out your growth strategy. So, what has that growth strategy looked like over the last few years? And are there any particular channels that have been most important for you? Again, I think the interesting thing about uh speak is be that we got pretty far in Korea before we started exploring in other places and uh and so we developed a fairly comprehensive playbook in South Korea that involved um a mix of organic channels like uh PR and earned media and own social. Um we use that to create um flywheels into performance marketing especially with creatives. Um and uh in in Korea we also piloted um an influencer I consider to be a part of that. Um uh but but in in Korea we really think about it as a brand building exercise. And so we've developed you know uh we're the number

one brand in South Korea for um English education um with like uh 30 or 40% top of mind, 70 to 80% aided awareness. Um we've had about 10% of all South Koreans try the product at this point. Um and >> uh essentially all of the all of that playbook of like a very integrated marketing approach um is what we have now kind of copied and pasted in Japan in um Taiwan. We we launched in the United States uh a few quarters ago. So here in Europe and we're trying to kind of now paralyze it across all the markets. Um and we really believe in like um in in having a pretty integrated approach when it comes to the consumer growth side. Yeah, that makes a lot of sense. And and it's great that you were able to hone these playbooks in Korea that it sounds like have mostly translated to other markets. Uh but as somebody who has led international expansion at a couple companies I I I worked on it at Quizlet and then again

at fair, uh I know that sometimes things don't translate and so were there any particular aspects of your product experience or your growth playbook that didn't work as you move from Korea to Japan to your next set of markets? >> There are definitely small parts that that didn't work. I think um the you know again because we started in Korea we had this weird thing where our first market was international expansion in effect and that was a very weird trait of the company and so in many ways when we went to go turn on the second market we had already >> we had already like been burned so many times >> with lessons around how to do international expansion that I think we did not have any naive left in us. Um, and you know, as an example, we we we didn't even try to do international expansion in Japan by hiring a team in San Francisco to do it. Um, we immediately went to build a local team.

And like we had a a head of Japan hired on day one of launching Japan because we we very quickly realized the importance of doing that for for Korea, especially for consumer. Um and and then frankly uh I think that insight is uh very much to your question of of us realizing that the strategy needs to be incredibly local. There are many many different things in every single market that are very local and we try to empower the local teams to have like a framework to operate within but then be able to operate fairly autonomously so that the brand in many way feels like a Japanese brand uh or a Korean brand or or what have you.

>> Got it. And so it's sort of this hybrid approach from what I'm hearing which is pretty consistent with a number of the companies that have done international expansion best where you have a local team whether that's country manager GM and then you've got probably people working on community people working on user acquisition but then you have your headquarters in San Francisco and you've sort of got your core technology team that um is building things in the product that can provide tailwinds behind all of your international expansion efforts. Um, but it sounds like most of most of the actual brand building and acquisition is happening in market. Is that accurate?

>> Yeah, that's exactly right. And and we see an opportunity to build a very lean central team at this point because we're in so many different spots um to to build centers of excellence around specific functions, but uh very much like a core value of of the way that we believe you need to distribute a product like speak is to have uh to be very local. >> Yeah. Well, for a language learning tool in particular, it makes a lot of sense that you would want to you would want to be immersed in the markets you're expanding to.

>> Another question I'm curious about is how the costs of AI have impacted your growth and monetization strategy because I assume that you have non-trivial underlying costs that you need to support effectively these AI tutors that are serving your learners. And so, how has that influenced decisions around organic versus paid acquisition? and how has it influenced decisions around pricing and and trial strategy and and pay walls? I think that tools that are more directly um aligned to productivity where like per token of usage uh it's extremely correlated with um value.

Those are the tools that I think where you have the the most cost pressure. I think with our product especially because it is a primarily a premium subscription product we are fortunate to be able to have a lot more breathing room um around our margin uh which has been you know very critical for our strategy of trying to be really on the bleeding edge of things ultimately though long term one thing that we we try to really stress for our teams is uh the belief that all of these costs are going to go down over time and to try to be as unconstrained as possible like if we ever suspect that teams are constraining themselves around cost. We try really hard to like dispel that. Um because if we see any amount of ROI, um even if the costs don't immediately feel correct in terms of the amount of ROI, uh it's a fairly reasonable assumption that those costs are going to continue to trend down over time and we'll be

able to to develop even more product value over time. And that's like the general framework and discipline that we've tried to have up until now. But I say that with I think the the fortune of again not having nearly the same margin pressure as like uh you know the agent coding companies or what have you. >> Right. Well, it sounds like part of the reason that you don't have the same margin pressure is because one you have an aspirational product with a high price point and so people are paying a good amount of money to access your subscription. But then two, the amount of compute required to support your learners isn't at the same level as as some of these productivity tools you're talking about. And so that leaves you with more margin. Is that right?

>> Yeah, that's it. Like the the just the the token consumption for for this versus a workflow tool tends to be a lot more economical. Um but that being said, multimodality is is is is very computer expensive and like we're everything we're doing around speech is is very non-trivial. Um but again the costs are just so trending so downwards over time um that it it it is and we try to make it as much of a non-factor as possible.

So I think that's a good transition into the last topic I wanted to talk about before we get to our lightning round and that is how you see AI impacting the way people learn languages or even learn other subjects over the next 3 to 5 years and particularly in light of the fact that these AI costs keep coming down faster and faster. So how do you

think AI power language learning evolves over the next few years and and what steps is speak specifically taking to sort of set itself up for those tailwinds? >> Yeah, I mean this is the this has been the vision of our company from day one. Um the core the core belief um is that AI can enable um a superhuman teacher that's not a real human and is scalable to everyone in the planet. And the core thing that that is really being offered is true one-on-one mastery based education. Um there's a a very famous study um in education science uh that's the Bloom 2 sigma effect which is essentially the idea that if you if you offer regular classroom setting education versus one-on-one mastery uh where you're learning by doing um you will get two standard deviations above so 99 percentile if you go uh the oneonone route. And I think this is like a universal truth in education. Everyone knows it's possible. It's just not

feasible to offer it to people. Um, and so, uh, it's really just the gold standard for the people that can afford it. Um, and so I think this is this is clearly the inevitable future. I think at this point it doesn't sound nearly as crazy as when we started the company. Um, that that's going to be possible in the future. I think basically everyone believes it now. Um, but I think the thing that people don't appreciate is how different that is than the general AI assistants that we have today. Um, because ultimately the workflow of of being a great teacher requires so much more than um an AI assistant chatbot.

Um, it requires uh a real mastery system where you're understanding what someone knows, what someone doesn't know. Um, it also requires uh pedagogical planning which is a really hard problem. Um um Andre Carpathy is working on a similar problem and I remember hearing a podcast uh I think it was Dark Crash's podcast uh where he talked about how hard this problem is and it is one of the core competencies that speak is very focused on and so there's a few of these areas mastery based a mastery evaluation um pedagogical planning and then like uh dynamic generation of curriculum over time are a few of the core problems here that I think are very much like open open research problems that we're tackling and a lot of other people are tackling and I think it's uh over the next 3 to 5 years though eminently solvable. Um and then you know that that's that's that's really like you know we'll see if we can do it but I'm

pretty optimistic. Yeah, it's I mean it's certainly a bright future that you paint and and um and I guess there's one other element to language learning that I talked to um one of the team members at Prely about too which is just around motivation right there's having the pedagogical components in place for a great language teacher but there's also the element of just keeping the learner motivated because learning a language is difficult and it's easy to get discouraged to what extent have you had to invest in that part of the problem at speak >> we definitely are invested in it um but I would say that In general, we have um we've always believed that the other problem is more important to solve first because ultimately if you can build a highly personalized education experience, then it's also by definition a much more relevant educational experience.

>> And if you can improve relevance um and and also this feeling of building a relationship, >> we believe those are better than gamification, which is the has been the traditional approach. And so for us, it's all about relevance and uh feeling like we actually understand you and that you're building some sort of relationship with this with this AI. Um and those we believe will be much more motivating factors into the future than you know uh kind of these dopamine loops that mobile apps have traditionally optimized. That's the that's been the strategy up until now. Um but motivation is an incredibly important part of learning in general. Well, and you've called out one other key vector on which it sounds like you're highly differentiated from Duolingo. We've already talked about starting with speech versus starting with grammar and and sort of sentences and and gamified exercises, but this other element of,

you know, motivation is important, but if you just make the core learning experience really relevant and really compelling, then people will intrinsically be motivated. uh whereas Duolingo has spent a lot of time building in these very effective gamification techniques to keep people engaged but it's sort of taking an outside approach as opposed to an inside out approach and and uh both can work but two very different approaches to the problem. Well, I think it's impressive that, you know, it's it's a meme at this point, but it's it's actually impressive and a testament to their product building capabilities that you have people with a 700 day streak that are saying out loud that they don't know how to speak the language whatsoever, but they still continue to use the app every day. Like that in and of itself is impressive that they have built such a retentive product. But we believe like

our aha moment is is 100% around this feeling of like, oh, I can actually speak out loud now. This feels this feels different. Well, it's gonna be exciting to see how it all plays out.

Let's get to our lightning round. So, I'm going to ask a series of rapid fire questions and you can just give me quick responses. First one is the one I always ask, which is, what's the best part about working at Speak? And what's one thing that you would change or are working to change since you're one of the co-founders? Oh, um, okay. One of the best parts of Speak is definitely just like the product we're building.

Um, and I'm going to cheat and say also the team that we're building, uh, those are like the two things that are by far the most motivating to me on a day-to-day basis. What's one thing I would change? Um, it's, uh, a continuous battle of scaling the company while keeping it very agile, nimble, and feeling like there's not too many nodes and communication taxes high. Um, so that's something that I'm like constantly trying to change about the company as it gets bigger and bigger. I mean, I feel like that's an intractable problem to an extent, but I'm curious, has AI made that any easier?

>> I think you don't need as big of teams. I think that's what I thought. And then and then you have workflow tools like all of our tools are connected now, notion, linear, etc. And so if you're a new person, getting context is so much easier than it was a year ago. >> Yeah, makes sense. What's your favorite foreign language and what do you love so much about it? >> I can answer the opposite question, which is my least favorite. Korean is so hard to learn. Uh, and I really struggle even though we've been operating there for so long, um, to learn it. But now we have Korean in the product. So, I'm I'm excited to to practice it more.

>> Yeah. I I mean, I mentioned earlier in the episode that I studied Chinese in college, specifically Mandarin. And one of the things our professor told us on the very first day of Chinese 101 was that for most Western speakers, it takes four times as long to learn a language like Mandarin or Japanese or Korean than it would a romance language because it's so different. And so I can certainly empathize with you. It is a very difficult language. I was near fluent when I graduated. Um but at this point it's been so long since I really used the language that it's definitely atrophied. So that's that's tough. Well, we're releasing Chinese soon, so you can practice.

>> There you go. Yeah, I'll jump back in and and uh brush up on my management skills. Um, what's one decision you made early on at Speak that really changed the company's trajectory? >> It has to be launching in Korea as our first market. Uh, that that was uh that was definitely a very uh defining moment in terms of my my lived experience for the next few years going to Korea all the time. Yeah, I I mean you talked about it already, but that level of focus certainly unlocks a lot of a lot of value and and I can only imagine how it changed the culture in those early days. Um, what's one advantage and one disadvantage of chasing a category king like Duolingo? I wouldn't say that we are competing for the same thing. I think we're complimentary goods, but um they they have made our lives a lot easier. And I would say when their stock was high, I would just say in general, we want their stock to be as high as

possible and generally are rooting for Duolingo and because it actually does help us. It validates it validates the market. Um I think the flip side is of course true too which is that um they are already so successful that um there's of course uh always a question of you know where they end up uh long-term relative to us. I mean this is generally what the best founders say which is it's not a zero sum game. Uh you also pointed out I can see how they're going after a very different target customer, right? are focusing on the super intense Korean English language learner and they're focusing on in many cases the casual American who just wants to learn a few words before they go on some trip over the summer. So two totally different target customers and two different products to serve those customers. But yes, obviously if if the large public company has a lofty valuation that that's a tailwind for you

guys and so not surprised to hear you say that. Um, the last question I'll ask, and you sort of hinted at this earlier, but what's the single most profound way that you think AI will change how people learn? If we fast forward 5 years, 10 years, what do you think classroom learning is going to look like at that point? Or will it even be in a classroom at that point? >> For any new technology, it is obviously something that will be very positive for the world and also will have a set of challenges and AI is no different. But I think that uh education is one of the areas where it will just be clearly so good for the world. Um that um I think people totally underestimate it. Like I think it's one of the reasons why we can be very optimistic about the future. Um because I think that uh if you think about the fundamental value in on on the trajectory of someone's life and even who they become as a person, how how the

educational experience they had over the first let's call it 20 years of their life um went >> is a huge determining factor in terms of the life that they're able to live later um and the quality of that life. Um it solves societal issues to improve your education system. And so I think we have like a seismic shift underway here. I think it will take time to trickle into schools. Um but I think it's a a thing that I'm very very optimistic about um because uh I think it's very clear as we talked about earlier that um it will be possible to have this one-on-one educational experience which doesn't mean teachers are going anywhere but it means that they can focus on the things they do best. Um and I'm I'm generally very optimistic about about that future world that we'll live in.

>> Yeah, I am too. And there are a few things that are as transformative as education is. My wife is someone who spent most of her career on youth education and workforce development um for the most part in East Africa and it's been pretty incredible to see the impact that even her small nonprofit that's still plugging away 20 years later can have. >> Um well, this has been inspiring, Connor. Thanks for taking the time today. Any parting words that you want to leave the audience with before we wrap up? Uh >> check out our product and send me feedback if you have it. Um but thanks so much for having me on. Thanks for being on. It was my pleasure. Thanks again to RevenueCat for sponsoring this episode. If you're building a subscription app, you can use RevenueCat to power purchases, optimize pay walls, offer discounts, handle refunds, manage customer data, run AP tests, generate insights, and unlock more revenue. I

work with dozens of subscription apps, and most of them are already using RevenueCat to power their businesses. Learn more at revenuecat.com/subversive.