Ep 40: CEO of Speak.com Connor Zwick on How AI Will Change the Way we Learn

We’re excited to bring you Connor Zwick, CEO and cofounder of Speak, on the podcast this week! Speak helps people learning a language have conversations with an AI speaking partner, which is critical to gaining fluency. It’s backed by OpenAI and most recently raised at a $500M valuation. Since launching in its inaugural market of South Korea in 2019, Speak has grown to over 10 million users and now has customers in more than 40 countries.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Jacob Effron?

  This is unsupervised learning, Red Points AI podcast. And today we have on Connor Zwick, the CEO of Speak. Speak.com is an English language learning platform backed by OpenAI and most recently raised at a $500 million valuation. Since launching in its inaugural market of South Korea in 2019, Speak has grown to over 10 million users and now has customers in more than 40 countries. We had a wide ranging conversation with Connor about moats and defensibility, what GPT-4O has unlocked in the audio side, what the biggest lessons have been in building Speak to date and more. This was a really fun conversation with a super thoughtful second time founder that's been building an AI long before it was cool. Don't forget to hang around for Pat and my debrief at the end. Now here's Connor. Connor, thanks so much for coming on the podcast.

• 0:36Speaker 01

  Yeah, thanks for having me.

• 0:37Jacob Effron?

  So I was thinking about where to start and I think, you know, there's many interesting things about your background, but one of them is I believe that in high school, you started this like flashcard app that you sold. And that was kind of the beginning of your entrepreneurial journey.

• 0:46Speaker 01

  Yes.

• 0:47Jacob Effron?

  It's funny because I actually think if you were, I'm curious if you were running that today. I imagine there was definitely a Gen AI play on top of these flashcards.

• 0:54Speaker 01

  Oh, yeah, for sure.

• 0:55Jacob Effron?

  And so what would you be doing if you were still running that company?

• 0:57Speaker 01

  Wow, that's a really, and no one's ever asked me that question before.

• 1:02Speaker 01

  So I think you're right. Like, I remember thinking, so for context, this app was essentially like pretty early days of the iPhone and the idea was, you know, replicate like flashcard studying, put on your iPhone instead of like the actual index cards. And it was really popular, had like several million users. They were making, you know, hundreds of millions of these decks, probably billions of these like cards, which were like front and back. You know, something that you were trying to memorize and the relationship between the two things was interesting because it was essentially like this, like these, like, I remember thinking we had hundreds of millions of these like pairs of knowledge that were like linked. And then they're also linked by like the deck. And this was probably in 2013, 2014. But the like, the thing that I always got really excited about was like,

• 1:47Speaker 01

  what if we could just like somehow aggregate this into some sort of graph? And then use this to like generate anything and teach anything that anyone wants to learn and make this like, you know, omniscient tutor who's got all this information. Well, fast forward, it turns out we kind of have that now. And we actually have the technology to do that. And so I think if I were still working on it, hopefully it would be even bigger, even more data. And, you know, I would be trying to build like a, some sort of. Omniscient tutor that can, you know, teach you anything. And so in an ironic way, it's kind of circuitous, but I feel like I'm kind of in that point anyway, but with Speak. So it's kind of a funny question that way.

• 2:25Jacob Effron?

  Yeah. Where would you stack rank flashcard data on like all the data you could have, you know, to build the omniscient tutor? Like actually, it's probably honestly pretty good.

• 2:32Speaker 01

  I think it would be really good data, specifically for learning, because it's structured in a way that is really centered around like the information that someone's trying to learn versus like all other information out there. That being said, like with hindsight and now knowing how LLMs work, it seems like probably the better idea was to just crawl the entire internet and try to compress it into a really big model. So, and they probably scraped, you know, all of Quizlet's data and everything else too, you know, in the trough. So probably needed way more data.

• 3:00Jacob Effron?

  You'd get really good user feedback from the flashcards and you could iterate on that. And so I'd imagine that would, you could build some sort of moat over time.

• 3:08Speaker 01

  Yeah. I mean, I'd be curious to see, like, I know Quizlet's, I've tracked from afar, but it seems like they're definitely doing a bunch of stuff in this space. So I think they've got a really interesting data set there to work with and.

• 3:20Jacob Effron?

  Yeah. I wonder if we'll still have flashcards. I feel like that was a key, key part of, of learning back in the day.

• 3:25Speaker 01

  Yeah. I mean, it's, it's definitely one method to memorize things and it's pretty simple. And, you know, you could argue that a lot of even what we do in Speak today is a form of, of a flashcard. So.

• 3:37Jacob Effron?

  A little more engaging.

• 3:39Speaker 01

  A little more.

• 3:39Jacob Effron?

  Well, you know, before we get into Speak specifically, the other thing in your background that I thought that we had to ask about, just because it's super interesting is I understand, like, you started your kind of crash course in AI in like 2015, but I think you literally crashed like a Berkeley course, right? And that's like how you got exposed to it in the first place. And you kind of became convinced that like these underlying models would get a lot better. One, very prescient of you, you should probably be in our job, that I wish we'd done the same in 2015. But, you know, you obviously kind of from there thought about language learning. When you were all like considering the different options, like what else were you thinking about building there? And, you know, what other, what other things were, were top of mind then?

• 4:14Speaker 01

  Yeah, I mean, this was back then, you know, the things that we were paying attention to were like RNNs and convolutional neural nets and stuff like that. I mean, the transformer wasn't invented yet.

• 4:23Speaker 02

  Two years, two years.

• 4:24Speaker 01

  Yeah, 2017 or whatever. Yeah, exactly. BT, that's a good. That's, that's, that's funny.

• 4:29Speaker 02

  Did you point that?

• 4:31Speaker 01

  It's it right now.

• 4:32Speaker 02

  Okay, nice.

• 4:33Speaker 01

  Right now, but we haven't, we haven't officially reported. Exactly. Yeah, so it was, it was two BT and, you know, we were paying attention to mostly like either things around computer vision and then speech. And we had a lot of ideas around computer vision, a lot of crazy ones, a lot of ones that would have been really boring to work on too. Like, I mean, just as a few examples, we were thinking about things like, you know, a little bit dystopic, but like, could you use cameras on like buses and other city vehicles and do like automated like meter maids? And then we were like, that actually seems like horrible for the world, so maybe we won't do that. We, we thought about a lot of things around vision for like measuring people's bodies for custom clothing

• 5:16Speaker 01

  or like medical imaging or things like that. We thought about, one, one crazy idea we had was, could we predict like weather better using deep learning? And everyone we talked to thought that that was crazy, but I definitely think we're starting to see some things there.

• 5:31Speaker 02

  Yeah, I guess every idea has its time. I feel like we've seen, we've seen a few pitches around that.

• 5:34Speaker 01

  Yeah, yeah, exactly. But in the end, I think we, we were just so much more drawn to like things around speech recognition because I think what drew myself and Andrew to the like problem in the first place was this idea that we could build some form of technology that felt like you were building an actual relationship with it. And, and there was some almost form of like a persona behind it. That, that was what really captivated us about, about the opportunity around speak. And so that I think was ultimately why we went down, down this path.

• 6:02Speaker 02

  Awesome. Well, that's a good segue into what you're doing with Speak Now. Maybe for our listeners, can you give a little bit of background about kind of the product that you have and, and what you do at Speak?

• 6:12Speaker 01

  Yeah, definitely. So Speak is basically like a, a full fluency solution for learning another language, specifically like learning how to speak it and have real conversations and communicate with other people. I would draw that as opposed to like other methods of learning that are maybe more grammar focused or, you know, memorization of vocab.

• 6:30Speaker 02

  Flashcard focus.

• 6:30Speaker 01

  Flashcard focus even. We saw a lot of language learning on, with Flashcards Plus, actually. And so. That's really like the idea in a nutshell. And we have like a whole kind of pedagogy and methodology behind like the way that the actual learning works. And obviously, there's technology as well that goes into it. But fundamentally, what we do is we teach people, even from the very beginning, we teach them like chunks of words and then have them that are like high-frequency words that always come together in common everyday speech. And then we teach you how to say that and then have you just say it a ton until it becomes automatic. And then you have these like patterns and chunks of words that you can pull out and we have you then, you know, practice them in simulated conversations where you're

• 7:11Speaker 01

  actually trying to achieve some sort of real goal that's tied to however and whatever is the motivation for you learning a language in the first place. So if I'm trying to learn Spanish to go to Mexico City and have fun with friends, maybe I'm practicing something around that and I can actually go apply the specific language. And the idea is that all of this is like extremely individuated to the individual user. So whatever their motivation is, whatever their interests are, whatever their level is, everything is the course and everything is kind of being molded to whatever those goals are.

• 7:41Jacob Effron?

  How is what you're like able to do in the product? You know, if we start at like 2BT when you know, you just had this idea to now at 7AT, I guess?

• 7:49Speaker 02

  You can't overuse it, you know?

• 7:50Jacob Effron?

  I'm definitely gonna milk the hell out of this new thing. Like, maybe talk a little bit about the evolution as we've gotten better and better models, like, you know, in terms of what you've been able to do on the product side?

• 8:01Speaker 01

  Yeah, I think one thing that we did that in retrospect was really smart was we were very long-term oriented. Like, we knew in the beginning that there was a long ways to go on the technology and we couldn't perfectly predict it, but the thing we did know was like, over the next five to 10 years, like with more data, more compute, models will get better and better and better and eventually they would like surpass humans on various tasks. And eventually, you know, that would that would mean that we could fully replace the human in the learning process. And we knew that like if we over, like, as long as we kept that as like the North Star, this is where we're going, we needed to make sure that all the product decisions we were making weren't optimizing for the short term, but like we're aligned to that long-term vision. So the way that we always thought about it was... Let's think about what we can do today and let's think about what we hope to be able

• 8:46Speaker 01

  to do tomorrow and make sure we're building in that direction. And we basically always thought about it mentally as like a series of steps, a flight of stairs. And so every year or two, we were climbing up one step at a time and the product was meaningfully evolving but along like a consistent and coherent vision. And so I think that was like what allowed us to be able to have such a head start in the like space of AI-based learning because the whole product, even in the beginning, the unlocks were more around really accurate speech recognition to just allow people to speak into the app and have a good learning experience that was actually working and accurately understood what they were trying to say as an example. And then adding like phoneme recognition and then adding basic language understanding and kind of going up from there.

• 9:29Jacob Effron?

  Yeah, and I'm curious because one thing I feel like a lot of founders are thinking through right now is to what extent you like build around some of the shortcomings in the models today versus like go to a beach and then in two years they just get so much better and like it's like, cool, we didn't need to build all that stuff around the edges in speech recognition because we've just gotten so much better at speech recognition. And so I'm curious, when as you reflect back on like, you know, where you've invested product time and maybe in capabilities have gotten better and then how you think about that now, like, how do you think about that question where obviously you can solve real problems for users now, but also you kind of get for free a lot of the stuff in the underlying model improvement?

• 10:00Speaker 01

  Yeah, I mean, that is really the question of, you know, the moment right now, I feel like for a lot of companies. I do think you need to have, if you're building on top of these technologies, you do need to have a deep technical intuition about how they work today and how they're going to work in the future and like, you know, what those timeframes are. That's something that we're constantly thinking about at a strategic level to understand where to invest. I do think also, though, a super important part of a business is actually being able to better understand and articulate what the problem is you're solving for people. And even if the technology isn't there today, like, understanding and making progress on that problem, even if you have to swap out entire chunks of your technology in a few years, might still be very worthwhile. Like, I think we've got a huge head start with speak because even though there's still a lot of things we can't

• 10:44Speaker 01

  do today that we want to be able to do tomorrow, we've been working on the like actual methodology of learning for a very long time. And we've been thinking about how to engage users and keep them motivated and things like that. And so I do think certain technologies definitely, it's a little bit of a, you know, a balance to strike there. But I do think overall,

• 11:04Speaker 01

  You know, the technologies will only get better and it's important to like make sure that you're, you know, you're building something of value outside of like the core technology that you don't think you, you can have a comparative advantage building in-house.

• 11:16Jacob Effron?

  And one of the cool things I think about the product you built is, you know, obviously in addition to just learning vocabulary and whatnot, you help with the way people speak. And I think part of that is like, you know, accents and helping people, you know, be understood and speak in certain ways. Like, how have you built that? That seems a very interesting, interesting product.

• 11:30Speaker 01

  Yeah, yeah. I mean, because I think, I think, you know, there is definitely maybe several different areas of kind of technological moats, if you will, that, that, you know, one can build. For speak in particular, you know, we have opportunities, at least in the short to medium term here, to continue to build and do actual modeling ourself, like develop our own in-house models for certain tasks. And so examples of this are things where we think we can do something that's a little bit more specialized and niche than the generalized models. I do think that the general, like, large foundational cognition model will continue to subsume things over the long term. But in the, in the short term, as in a concrete example, we've developed our own in-house speech recognition that

• 12:11Speaker 01

  is super, super good for people speaking with accents and understanding what they're trying to say, and also understanding when they're making specific kinds of mistakes, and making sure this is super, super fast and reliable and streams back to our client really quickly and reliably, in a way that enables our product experience. Or we have like a phoneme recognition system that's built off of all of our data to be able to detect the types of mistakes that learners make around pronunciation and other, you know, prosodic types of mistakes. And that's another example of like, that will probably eventually be something that like a really good multimodal speech, speech model will be able to do one day, but it's not going to happen tomorrow and it's fairly specialized. And so even if we're only going to use those models for a few years, it'll still be like

• 12:53Speaker 01

  a huge worthwhile in us building our business. And it's not necessarily the like central node of what our kind of core specialty is or point of leverage long term, but it's definitely really valuable in like, for the next few years at the very least.

• 13:07Jacob Effron?

  Because in building those models, I guess you get something that works today that then lets you learn all the different things you have to build around that.

• 13:12Speaker 01

  Totally. And build a bigger business with more users that allows us to collect more data, to then build other models on top of it, and have more resources to take bigger investments and all these other things. Like, I think at a certain point. It really is just about like, you know, the pushing the ball down the field. And that's like the main thing you need to focus on. You can't be too philosophical about these things.

• 13:33Speaker 02

  How big of an investment were those models for you? Either from compute or team or resources? And yeah, I just would be curious.

• 13:42Speaker 01

  It's definitely a really big investment, but it's not. It's one of several or many investments we've we've made, if that makes sense, right? Like, it's definitely not easy to build models that are best in the world at a at a task. And you need data, you need specialized knowledge to be able to do that. And I think at the same time, you know, as I think a really common thing that people talk about here is like, and you kind of touched on it earlier, Jacob, around like the idea of when to just like use these larger models, these L L Ms and these foundational models. And I think like a pretty good analogy here for me is the idea that like, you know, I think in the eighties, there were probably 10 or 12 different personal computers from various different companies, including companies like IBM or Atari even.

• 14:27Speaker 01

  And a lot of these companies like Apple is a great example where they didn't. They were using Intel and they weren't building their own processor back then. And there was a very good reason for it. There were companies that were specialized on this and they were spending a tremendous amount of money building these things, doing it better than anyone else in the world. And then it wasn't like Apple wasn't building something really valuable on top, even on the technology side. I think today, like I see the same thing here where like if back then that the it wasn't just about building Mac OS, but it was also about building all of the like the firmware that was like directly sitting on top of the the processor. I think in the same way, like I see the modern day like AI firmware being all the like what we call internally like the M L scaffolding.

• 15:10Speaker 01

  It's basically all of the technology that we're building that's actually really technically difficult to do, to actually like have these models be orchestrated and working really well for our tasks and working with our backend and working with our product and all these things. And so there's a huge amount of complexity and depth to that part of the technology as well. And I feel like people are always talking about modeling. And I don't I don't really understand why, because to me, I'm looking at like when we're talking about investment. And I think the modeling is definitely one investment, but we're making a much bigger investment on this piece. And I actually think that's like, if I were going to say what's our like long term technological mode, I would actually say that's probably a bigger one.

• 15:43Speaker 02

  What are some of the big?

• 15:45Jacob Effron?

  Pieces of infrastructure that you've built there, whether in the ML scaffolding, is that evaluation or chaining of models together or inference infrastructure?

• 15:54Speaker 01

  It's pretty, it's all those things and more. I mean, I think all of it.

• 16:00Speaker 01

  A lot of it, I mean, this is also, I think, a trend that we've seen is like, there's so many companies trying to build services for all those pieces. And oftentimes, the thing that we need to build is so specialized and unique and like a little bit more hardcore that like, it doesn't make sense to use a tool. We should just be building in-house internally. And so, yeah, a lot of it is that. It's like, how do we get these models to be really good at individual tasks? And then how do we orchestrate those things? And then how do we continuously collect new data? When do we fine tune? When do, like, how do we evaluate even? And then generally, like, starting to think about building these larger pieces of infrastructure around. How do we make an investment in, like, an actual, real representation, for example, of a language?

• 16:40Speaker 01

  In a way that like, we can then, you know, retrieve on top and build like, you know, knowledge graphics, for example, to understand like, where is someone proficient and where are they not proficient, you know? That's like one example of like, but, but probably at least 50% of our product development and time right now is spent on things that are like touching these systems. So it's like, it's definitely a huge investment for us.

• 17:01Jacob Effron?

  What's the most like painful part today? I always think like, you know, you're working on early technology. You think about in 10 years when you're, you're going to tell some set of engineers, oh, I used to have to do things like this way. And they'll laugh. And I'm sure there'll be many things like that. But what's, what's kind of like top of your list right now?

• 17:14Speaker 01

  You mean from a technology standpoint?

• 17:15Jacob Effron?

  From just like all the stuff you have, like hacky stuff you have to do in the back to like get this, this, this to work?

• 17:20Speaker 01

  Yeah. Huh. That's an interesting question. I, I do think

• 17:25Speaker 01

  it feels like there's a lot of things that are rather silly, like the fact that we're still doing any amount of kind of prompt optimization. Feels kind of silly.

• 17:34Jacob Effron?

  Pretend you're very friendly when you get.

• 17:36Speaker 01

  Yeah, like, like, that's definitely diminishing as the models get more intelligent.

• 17:39Jacob Effron?

  Is that how you start all your prompts? It is, yeah. Otherwise, the models are too mean to me, you know?

• 17:44Speaker 01

  So maybe you should use Grok.

• 17:45Jacob Effron?

  Yeah.

• 17:46Speaker 01

  Yeah, totally. I think like that's like a great example, though, of like, I feel like that's going to be a short window of time and people will be like, what? Why did that? Why was that necessary?

• 17:56Jacob Effron?

  Part of what you're building, obviously, is a different way to learn language, right? And so it's, you know, there's something, I think we all kind of, there's been behavior in the past of folks, you know, using tutors and talking out loud, but it's not what people think when they pull up an app for the first time. Was there like a, is user education hard here? Or do people just kind of intuitively get how to use this from the start? And what have you learned about that over the years?

• 18:16Speaker 01

  So a strong tenet for me always when I'm working with our product and design teams here is like, any sort of tooltip or user education or explanation of how a feature should work or config panel, that just is a sign like we haven't done a good enough design yet and we haven't like actually built this in an intuitive way. But that's really challenging because something I talk about a lot with like when we're recruiting designers in particular is like, we really are inventing new interface paradigms around like audio first experiences, right? Like people are talking into speak in a way that is kind of fundamentally unfamiliar when it comes to technology. Like you don't really interface with technology in this way historically. Or same with like this kind of open-ended

• 19:01Speaker 01

  nature. Like, for example, in our onboarding now, when you open up the app, there's just a microphone button and we have a simple question for you, which is like, why are you learning English in this case? And you literally just press this button and then you're just supposed to start talking. But people are like, what am I supposed to say? What language am I supposed to talk in? Like, am I supposed to give you like a minute-long answer or like five words? Like, I don't know. And they're also just fundamentally kind of like, not exactly creeped out, but they're taken aback by this. And this is a good example of like, how do you make an experience like that powerful? Because we want to know in their own words, like what their motivation is. And then we want to then empower highly personalized experiences on the backend after they give us that information. And we want

• 19:44Speaker 01

  to demonstrate how the app works, but at the same time, we know this is like a highly unfamiliar thing. And so how do you design this in a way that's still intuitive, but still will be like very future-proof? So that's like a good example of a challenge. But I think that as minimal of user education as possible is always the goal here. And it is interesting to think about, like, this feature is one example. We launched the first version of this a few quarters ago, maybe. And the amount of familiarity, because apps like ChatGPT are so prevalent now across the world, there's already been a meaningful shift in the average user's understanding of these paradigms. And so I do think all this stuff is rapidly evolving and that's another interesting design challenge.

• 20:23Speaker 02

  How do you think about the UI over time? Like, will more and more of it move into audio and to the point where maybe it's just an agent that you're talking to? Or do you think there's always a place for this kind of UI? Maybe you're selecting things from menu and then having structured conversations around that?

• 20:41Speaker 01

  I think my best guess is it will just become more fluid and it'll become intuitive how you can maybe do both. Like, a thing we're often thinking about is, and we use the word hybrid to describe this, but how do you build something that's hybrid where you can kind of at any point choose to talk or type or tap? And this is, I think we're really in the early days again of how that paradigm works. But like, I don't think speech is necessarily always better, but it's definitely better some of the time. And that's going to be a huge shift, especially as these audio speech-to-speech models get better and better. At the same time, though, there are definitely going to be many times where you just prefer to type, right? Or tap, depending on the interface. Like, if you have access to a keyboard, that's actually much faster of an input in some cases.

• 21:26Speaker 01

  And so that's my guess, and that's definitely the way that we're thinking about it. But I think we're going to develop a lot more here.

• 21:34Speaker 02

  Totally.

• 21:35Speaker 01

  And it'll be interesting to see where it goes.

• 21:36Speaker 02

  Yeah, it's been really impressive how quickly there's been progress made in this voice area, right? Where even where you can now interrupt the models and they sound a lot more lifelike. And I think there was kind of, for a while, we were in the mode of, for decades, we were in the mode of humans kind of bend to technology. And now I think there's this like, technology will bend to how we interact with each other and learning. So it's cool. Speech, I think, is one of the most interesting spaces for that, so.

• 22:04Speaker 01

  Yeah, totally. Another interface that I feel like is highly relevant to this as well is the idea that like, you know, a little bit better at like push versus pull. Right, like historically, like most things that people, most of the time we get value, it's from us querying our computer or a piece of technology. We have push notifications, we have emails, things like that, but like it's mostly 99% spam and it's not very useful. And it's for the very basic kind of send updates about something that's happening somewhere else.

• 22:35Speaker 02

  Yeah.

• 22:37Speaker 01

  It does seem like a huge other kind of interface unlock here is like the fact that there's now this GPU that's like thinking about you in the background.

• 22:44Speaker 02

  Yeah.

• 22:45Speaker 01

  And maybe observing pieces of information or data about you and then doing things for you in the background. And that's like another big area that we've been thinking a lot about at Speak that I think is like going to be very unintuitive how it eventually pans out. But we'll have like totally different interfaces there.

• 23:02Speaker 02

  What would that mean in the Speak world? Like you're traveling in Paris and it says, hey, you know, have you thought about? Or do you need to order food at this cafe or something like that or?

• 23:12Speaker 01

  Yeah, I think it could be proactive stuff like that for sure. I also think, you know, things that we're thinking about here are just like. Sometimes there's a lot of processes that actually take a fair amount of compute and latency. And aren't actually necessarily something we want to surface to user immediately. For example, if you've been using Speak for an hour today. Maybe it makes sense for us to like on, you know, when Korea, where that user is, maybe when that user base is kind of asleep. Like maybe we run those GPUs overnight and we start thinking about the users and saying, okay, they've been using, you know, they did all these things today. Like what would be like a really good set of like distilled analysis and lessons that like would be interesting to send them tomorrow to start out the next lesson.

• 23:57Speaker 02

  Oh, cool.

• 23:58Speaker 01

  As like an example, right?

• 23:59Speaker 02

  Yeah, it's just waiting for you to pull up the app.

• 24:01Speaker 01

  Exactly, right. And it's just kind of like doing that in advance and, and really like, I think the mental model is like literally it's thinking about you. When you're not using it. And I think like the more information we have about the user, the like more you can do there, but like. This is like a totally new area, I think, that like not many people are playing in yet. But it will, it'll be like a profound shift, I think.

• 24:23Jacob Effron?

  How like is the curriculum kind of like very explicitly mapped? Or and at what point do you imagine like, I mean, do you ever find lesson plans or something? You're like, oh, that's kind of weird or something or like, is there, uh, is there ever going to be? You know, as you think about how this evolves over time, how much of it is kind of like you're dramatic about this is, you know? There's a bunch of paths we go down versus like, I don't know, there's, there's many different ways we could go. And I wouldn't have expected that, but that's actually pretty interesting as a, as a kind of subject for today.

• 24:48Speaker 01

  I, I try to resist the urge to make it mutually exclusive when I'm thinking about that. Because I don't think it needs to be. And I think both can kind of be true at the same time. Um, and in the sense of as an example for like learning a language is, is a, is a really good, uh, example of this. There is just a certain right sequence of ways to learn a language. Like you need to start with like the very basic words in vocabulary because there are, you know, 100 words that we use 20% of all, all the time. And, you know, the next 500 words are 80% of the time, right? So like, there's a, there's a certain set of words that you should learn first. But like, maybe the specific ordering of those things within that, you know, first concentric circle of language.

• 25:31Speaker 01

  Maybe we can like make that really individual and bespoke for the user, right? Um, and maybe like the individual 500 words that like you should learn first is very different than someone else. And so maybe we can change that. But I do think that there is for a very long time, going to be a need for, um, humans to be in the loop here. To empower like very, very thoughtful, like artistic almost creation of the actual curriculum. At least at a high level, like strategy and like the, the way that we take our methodology. But it starts to look a lot more like, a lot more of that is actually being done by instead of our like curriculum team. It's also being done by like our machine learning team. Um, and that's like an interesting challenge because the machine learning team now needs to understand how the methodology works and how we teach people language.

• 26:16Speaker 01

  And that's like super cross-functional. So, yeah, I think, but like, get back to your question. I do think that it's ideally both. And the really cool, powerful thing is when you get, when you kind of start venturing off of the path a little bit. And you start learning in a way that's like, kind of. Really kind of weird or like very unique to an individual.

• 26:38Jacob Effron?

  Yeah, I wonder if you had like an AlphaGo moment where you know, there's some sort of move and then

• 26:41Speaker 01

  Right.

• 26:42Jacob Effron?

  Language learning, like I didn't expect that, like.

• 26:43Speaker 01

  Right.

• 26:43Jacob Effron?

  Uh, but it's actually is, uh, is quite effective. I wonder if we'll still be doing the, the hundred vocab words in, in 10 years.

• 26:49Speaker 01

  Yeah. I don't know if you guys have ever read any books by Neil Stephenson, the sci-fi author, but he, uh, wrote a book called Diamond Age, which is essentially all about this. It's about like, uh, a girl who, uh, finds a book and the book is like basically powered by AI. And it's like, uh, you know, it's a, an all encompassing primer that can teach her anything. And the, the whole story is about this. And it's actually a huge inspiration for Speak and like a big part of it is like how the kind of the book can actually go like full, like fully in it, in a really unique and creative direction for this, for this individual. So, um, that's like something that we often talk about internally. And I, I think it's like a very cool feature to imagine.

• 27:25Jacob Effron?

  I mean, I guess, you know, the present today, I'm, I'm curious, uh, one, one thing that seems to be happening is model costs just keep coming down. Like every, you know, every few months it feels like. And so, uh, I'm always curious if folks are, feel like you're constrained at all. And like, are there things you'd want to do in the product, but you're like, oh man, that just costs like a little too much for, for what we charge today. Or is it like literally anything you think about the, the model inference costs are kind of, uh, are, are not super important.

• 27:49Speaker 01

  We don't feel super constrained by that. I think, and, and, you know, obviously we're, uh, more kind of at scale company, lower cost per user compared to enterprise services. Um, we don't really have a free tier, so that would be an interesting challenge. But for our tier where people are subscribing, we don't really feel constrained. And I think if we did feel constrained, we would probably elect to just build it anyway and eat the costs in the short term, because we believe that the costs will go down over time. It does feel like the lower, I mean, I think this is clearly OpenAI strategy, but like the more the cost goes down, the more there will just be more demand, like it will one for one increase more on the demand side. And, and these companies will make more money if they lower the cost.

• 28:32Speaker 02

  Flip side of that, how have you thought about, um, uh, pricing for the product and, um, and charging users? We kind of, uh, VCs love to talk about modes and pricing. We hit on modes. So, you know.

• 28:43Jacob Effron?

  Yeah, exactly. We gotta get our contract.

• 28:44Speaker 01

  I gotta get the other one.

• 28:45Jacob Effron?

  Yeah. Tried to coin.

• 28:47Speaker 01

  Check the box.

• 28:47Jacob Effron?

  We tried to coin a new term too.

• 28:48Speaker 02

  Yeah, exactly.

• 28:49Jacob Effron?

  I feel like we're, we're doing a good job.

• 28:50Speaker 01

  Yeah. Nice work. Yeah, pricing is a great one. I think it's definitely something we haven't had the time to like really be as thoughtful as I would like to be about. But in general, we're thinking about it kind of on both extremes. Like, how do we make Speak radically accessible to anyone who wants it? Because in the end of the day, like, we are offering a software solution to a problem that traditionally has not been available with such little marginal costs. And so, we can go actually, we have an opportunity to like, go provide value for like, literally hundreds of millions of people. And I think the way to build a really big business is to offer something that scales to all those people and, you know, charge overtime for them. But at the same time, that's on one end. On the other

• 29:35Speaker 01

  end, I think there's actually a really interesting opportunity here to charge dramatically more for a consumer product. Because I think truly, we are offering something that currently, there are millions of people that are paying, as consumers, hundreds of dollars per month for, which is offline tutoring or classrooms, education outside of school. And so, I think there is a question here of like, is there a way to also offer that high-end experience and be able to charge for it as well? Not necessarily a dollar for dollar for the current substitution, but, you know, dramatically more than what consumers are paying for it. You know, I think that if we can build something differentiated and truly that valuable, then I think we will have built something where you're not competed down to a lower

• 30:17Speaker 01

  price. But still very early days there. And like, you know, obviously, it's, pricing is one of those things that is often very counterintuitive. So. Yeah. And always in flux. Always in flux. Yeah.

• 30:28Jacob Effron?

  What about like, on the model evaluation side? Like, how do you know, when a new model comes out, like, obviously, you have the high-level goal of making it easier to learn languages. But how do you know if like, the, you know, new models you're testing are actually good?

• 30:38Speaker 01

  Yeah, I think evaluation is one of those things that people really underrate how difficult it is and how important it is. Like, for our ML teams... I often like to try to say that the like maybe the most important thing is evaluation. Because if you can actually distill an evaluation, especially when we're talking about open-ended tasks that LLMs often are doing.

• 31:03Speaker 01

  But even for speech, if you can distill the perfect evaluation, you've basically distilled the problem you're optimizing for. And then the optimization is almost straightforward. Like as a concrete example, even for speech, let's forget about like the gray area of LLM tasks. For speech, it's not just about like what percentage of word error rate we have and like how many words are we mislabeling. But there's also like an idea of, you know, what the, are we catching individual mistakes? Are we, sometimes users say a word, but it's like substantially kind of unintelligible. And we could actually now train a model that can understand words that humans wouldn't understand in communication. So like, how are we evaluating that? Like what's correct, what's not correct?

• 31:45Speaker 01

  And, and we actually like, the evaluations can get quite convoluted and complex.

• 31:49Speaker 02

  You have to dumb down your understanding model to like human level?

• 31:51Speaker 01

  Yeah, exactly, literally, yeah. And so there's all this weird clarity you need to achieve with your evals. But once you've got a good eval framework, I think that it really drives a ton of like execution clarity that makes the team just make better decisions. And so I always like to really like, bring it back to that, that conversation first, if I don't feel like we've got a good framework for it. And then from there, it's, you know, it's fairly straightforward, like the technology to actually like go run an eval loop on something like, that's, that's something that's like the much more straightforward piece.

• 32:24Speaker 02

  Yeah. Related to evals, I've always been super curious about this. So when GPT-4O comes out, what happens inside of SPEAK? Is that something where you guys run it against all of your evals and decide, okay, we want to ship this or?

• 32:41Jacob Effron?

  Or

• 32:42Speaker 02

  Siren go off?

• 32:43Jacob Effron?

  Yeah.

• 32:45Jacob Effron?

  Yeah, I guess like, how do you, how do you, you know, or do you assume there's great evals out there, you know, we've heard that it's amazing and we're gonna kind of open it up to our users today?

• 32:55Speaker 01

  Yeah, no, we, we have a whole process for this and thankfully, I think we've usually, we've got a pretty close relationship with them and we can, we can usually get a pretty good sense of it pretty quickly. But absolutely, it's like something that we've got, we've had to get really good at over time is like, what are all the like, what are all the, we have a ton of internal tools now and technology we've built in-house for this and it's literally like, here are like the, you know, 40 different major tasks that we're doing, here are all the different eval loops that we run for these, here are the things that we need to go run human loop, like evaluations for, like every single time we need kind of, we have a new run of a, of a model. And all that is like basically been distilled down to like a playbook now, which I think

• 33:40Speaker 01

  it needs to be because otherwise you have so much organizational thrash every single time something changes. So that's definitely like the, like something we've gotten kind of dialed in over the last year that's been really beneficial.

• 33:51Jacob Effron?

  It's true, we've had a few folks on the pod lately that have been like, look, rather than you figuring out you can come up with a perfect eval data set, just release it to like some subset of your customers and track the metrics you care about on the product and like, your customers will tell you very quickly whether

• 34:04Speaker 01

  This is very true. This is like a, a big piece that I forgot. It's absolutely a big part of anything is like tracking, tracking experiment and like literally just seeing like, do the metrics we care about and the kind of Garterell metrics, do they all look okay? And that's like definitely the, a very important piece as well.

• 34:21Jacob Effron?

  I guess, you know, to, to switch to like another favorite VC topic, so we have to, we have to hit them all. I feel like there's a lot of people that are like, look, you know, Gen AI, very cool technology, benefit really accrues to the incumbents, right? And I think in your space, that would be, you know, Duolingo gets to just now incorporate a bunch of this stuff in. And so I'm curious like how you, how you think about that broadly and then in the context of, of ed tech specifically?

• 34:44Speaker 01

  Yeah, I think like, I do think that AI broadly does help incumbents and it's a sustaining technology. If you're solving, if the problem that you're solving

• 34:58Speaker 01

  Is just made better and it's the same problem that you can now solve with AI. And that is sometimes the case, but sometimes what the case is, is actually like, for example, for maybe like customer support. If the problem you're solving is you make it really efficient to manage a fleet of customer support agents and you have a really good software to like manage and evaluate those individuals. And then there's now L L M's that offer like a fully automated solution. I would argue that's not really going to help. It might help you better evaluate those individual agents, but the disruption there isn't making agents more efficient. It's like literally fully automating this loop perhaps. And then in that case, I would say that's highly disruptive. Right?

• 35:41Speaker 01

  Just as an illustration, I think like the, you know, to, to, to talk about this in the context of language learning and Duolingo, I think that like we're fundamentally solving very different problems, Duolingo and Speak.

• 35:55Speaker 01

  Duolingo's, you know, vast majority of their subscribers are folks primarily in America, UK, Australia, native English speakers who actually, one of the most mind blowing things I've ever heard is that most Duolingo subscribers were not learning a language before they were using Duolingo and which is really cool. They've gotten all of these people to learn a language that otherwise wouldn't, wouldn't be. Must be that really good TikTok they have, I assume. It's, it's gotta be mostly the TikTok for sure. And, uh, and the point though is like, they've built something really good as like a casual, almost like a brain training app that makes you feel like you're productively doing something versus scrolling on Instagram or whatever. Um, and that's, I think, an amazing achievement, to be clear. Um, and they definitely care about fluency as well, but their core thing that they built

• 36:39Speaker 01

  is like a very casual app experience. And I'm not totally sure. I'm, I obviously don't know all the details, but I'm not sure that AI necessarily. Helps you build a better casual learning experience? It might, it might. But I do think that the like for us, we the our customer base is actually very different. And our users typically are not using Duolingo. They're people that are like, our specialty is learning English and teaching English for folks at home that don't know that. And we're really focused on like helping people who previously have spent 10 plus years learning a language, learning English, trying to get the conversational fluency, but they lack access to human speakers.

• 37:21Speaker 01

  And so it's a very different kind of experience even using speak. And the I think that AI clearly, really, really helps with our use case. And in this case, I think it's like, it's again, it's it's all about understanding what the actual core product market fit you have is and what problem you're solving for what users. And then I like, that's the kind of the way that I think about that in terms of sustaining versus disruptive.

• 37:44Jacob Effron?

  If we have like real time translation and like accent, you know, changing like, does that obviate the need for some of your users to learn English?

• 37:52Speaker 01

  Probably some. I think there's maybe a few things here, though. So one, I think the best translator in the world, right? When when, you know, you have two world leaders talking to each other. If you watch like what actually happens in not edited version of that, the amount of latency and lag between one person talking and the other is massive. And that's because languages are fundamentally different, right? Like the order of words are different. And so there's just an inherent latency to translation and there's an imperfection of that. That's like maybe a little bit more of a tactical point. I think the real thing, though, it's like, what if I distill that, like, what is speak? And what are we offering to to to like our users? What our users fundamentally are after is like human connection.

• 38:37Speaker 01

  They're trying to like connect with more humans around the world, professionally, culturally, socially, whatever it is, they want to connect with more people than the people that just speak their language. And I would argue that like for most of them, they've spent so much time learning this language. The reason they're doing that is for something fundamental that doesn't shift. And the best AI kind of live babblefish, if you will, in the world doesn't really solve that problem. I do think it solves the problem of casually going to a random country as a tourist. I think that'll be really cool. But I think that for our user base, the people that fundamentally want fluency, the reason that they want to learn is super important.

• 39:18Jacob Effron?

  I thought, I mean, one thing that was funny, I remember when G P four O came out, like Duolingo stock just went down immediately. I don't know if that was just like, you know, too much AI buzz or like whether that was rational or not.

• 39:28Speaker 01

  Yeah, I mean, I don't know, these, the markets feel really noisy right now, so I don't really know for sure. I don't have a super strong view there, but I do think that the four O speech to speech stuff is super, super exciting for us. I think there's probably the speculation there is like, okay, people are going to learn languages with Chat G P T now. If I kind of understand the sentiment correctly, I may be, I may be off base. And for us, we are extremely excited about, one, multimodality in audio in general. We think it's a holy grail for our use case. I mean, our app's called Speak.

• 40:05Jacob Effron?

  But at a minimum, speak.com is going to be a very valuable URL.

• 40:09Speaker 01

  It is, it is. We could have just, maybe that will be the full business at the end of the day. But, but, but I think the like, the broader point here is like, actually, more people using Chat G P T, I think more, a lot of people will start to learn languages with Chat G P T and practice learning languages with Chat G P T. I think it'll be a really interesting, cool, cool tool. But I think that those users are, I think it's fundamentally a good thing for AI language learning products, for people to use Chat G P T and realize they can use AI to learn a language. And, and then if they're serious, go look for a specialized solution and actually go look for like something that's going to offer them something a little bit more effective.

• 40:54Speaker 01

  Because at the end of the day, if you're willing to spend the amount of time and money required to learn a language, which It's not something that happens in 10 hours. It's like sustained habit over many months, if not years. You're going to want to have the most effective solution

• 41:11Jacob Effron?

  Yeah.

• 41:11Speaker 01

  out there, right? And clearly, there's room to specialize and build something more effective. And so for us, we're really focused on owning that category. Like Airbnb owned house sharing and Uber owned ride sharing. That's kind of the, we see it as a similar opportunity. And we actually think that like, in many ways, this is going to increase the number of people using AI to learn languages, which is quite exciting.

• 41:38Jacob Effron?

  They won't be as confused when they come to your first prompt.

• 41:40Speaker 01

  Exactly. They'll know the paradox, exactly.

• 41:43Jacob Effron?

  Yeah, I guess there are other things, you think about audio all the time. So I'm curious, like are there other, now that GPT-4 is out and more people are thinking about it, like completely independent of speech, are there other like, you know, audio-first applications that you think are like really interesting or?

• 41:55Speaker 01

  Oh, man, I wish I had a more interesting answer. I feel like I pay very little attention right now to anything outside of speech, probably to my own detriment.

• 42:02Jacob Effron?

  That feels right.

• 42:03Speaker 01

  Yeah, yeah, so maybe that's, maybe that's my full answer. I do think like, it's been really cool to see the dev tools, like even 11 Labs is an example, like, it's really cool. Like, obviously, OpenAI needs to be way more conservative with how they let people use their tools, because they have a giant target on their back. But it's cool to see these, like, other players release tools that are just unleashing a torrent of interesting, creative applications. So, I don't know, I think 11 Labs is really cool.

• 42:31Jacob Effron?

  I mean, I feel like with the release of GPT-4, there's always been this question of like, you know, will an audio-only model kind of still have a place? Or will they're just, you know, will something that's multimodal can do, you know, and have the most complex reasoning engine behind it just went out? But it seems like, at a minimum, there is certainly an arbitrage on the fringes of like voice cloning and like some of these other areas that OpenAI won't go into. But it seems like, do you think there's still room for an audio-only model to win?

• 42:56Speaker 01

  I bet there is. I think like, I don't have high conviction here, but I would say like, even for a speak, we're building our, we're investing in our own speech recognition technology because we have a niche use case that we believe will not be fully solved by these large cognition audio models anytime soon. So I think that that's going to happen for general, general purpose things. Like people need certain security with their, with their speech data, or they need it on prem, or they need a really weird set of vocabulary that, that, you know, isn't in the, on the internet. So I suspect there's still an interesting opportunity there. And if nothing else, they've got the advantage of they're willing to make

• 43:39Speaker 01

  a lot more risks as a, as a scrappy startup than, than the, than the big players.

• 43:44Speaker 02

  Are there specific capabilities on the audio model side that, obviously you guys are building a bunch of this in-house, but are there capabilities where you are looking to the market and say, okay, if that comes out, then we could do all this, this new set of things? Or are there specific features that you're kind of waiting for to come from the model layer?

• 44:03Speaker 01

  Yeah, absolutely. I mean, I think there's still, we're really early days on these general like cognition models, these LLMs, and what I would categorize even as like multimodal audio that's connected to an LLM essentially. So I think there's a lot there that, again, going back to the earlier point, like, we don't think we should build in-house because these models will literally cost hundreds of millions, if not billions of dollars to train and develop. But, so, so I would say that's the main thing. And I think it's all, for us, like, the main thing we're paying attention to is this, like, multimodal audio. And there's, for us, it's really like, I think, the holy grail. It'll take some time to get it really good. And I think there's a lot of opportunity there to build something

• 44:45Speaker 01

  specialized on top of it. But there's, I think, endless possibility as that continues to get better.

• 44:51Speaker 02

  What are like the specific things you're thinking about? Like, oh, man, you know, hopefully by the time we're at GPT-6, like, I can do X in speaker.

• 44:57Speaker 01

  Yeah, I mean, there's definitely a ton around just generally better multimodal capability, multilingual capability around the audio side and being able to do generally produce something that feels a lot closer to like a real tutor, which we don't always want in our experience, but sometimes is very useful. And what that means is literally just having something that's a lot more natural and lower latency, because instead of having a speech recognition model and LLM and then speech synthesis, you've got one continuous model that can do all that in one turn. And also, it's not so lossy because it's not reducing the complexity of your speech into a little string of text and then putting that into the LLM and then trying to expand

• 45:38Speaker 01

  that back into synthesis that kind of has the right nuance and tone in the response. Not to mention the like more kind of complex understanding of what you were saying, how you were saying it, whether or not you felt confident, your emotions, all that stuff, what mistakes you were making. So that's definitely like, I think, a really, really big thing that is starting to comment. We're still early days, like, there's a lot more that needs to be proven out there. But then generally, I think for the like actual kind of cognition piece, clearly there's a lot more room for improvement around reasoning and general ability to like actually follow a task through completion and do that well and reliably.

• 46:21Speaker 01

  I do think there's a big open question of how that happens and we'll see what happens.

• 46:25Jacob Effron?

  What would you do with that if you had it?

• 46:27Speaker 01

  I think it would generally mean much smarter planning around the curriculum. I think that's like, that's like the biggest thing that kind of is the missing piece right now for us is like doing that really, really well and reliably.

• 46:42Jacob Effron?

  And so it still requires some sort of like non-LLM intervention?

• 46:46Speaker 01

  Yeah. I do think that broadly speaking for us, we're much more excited. I don't think we're that constrained by the like reasoning piece compared to most use cases. Like, like one really exciting and special thing about the language learning space in general is that we can We can actually like build something that's pretty much fully useful and disruptive with the current set of technologies and maybe a little bit more on the audio side. And we don't really need that. And we can, we can, we can basically fully take the human out of the loop and build something cool and useful. Whereas I think right now, a lot of industries are still finding themselves where they still kind of need a human because it's just not quite good enough. So like any, anywhere that's a little bit higher stakes than language learning, you

• 47:30Speaker 01

  just need that higher level of reasoning and consistency with lower hallucinations. And so that's like, I think maybe a particularly noteworthy thing about our industry that's an advantage versus other areas.

• 47:40Jacob Effron?

  What about like public speaking or other kind of like speech related things people try to get better at? Like is, could we build product around that today? Like, is that something you guys think about?

• 47:48Speaker 01

  Yeah, absolutely. We've got like a really interesting, very like fast growing part of our business, which is like, our main market right now is South Korea and we've got pretty good brand recognition there. And every company wants their employee or employer to pay for, pay for, speak for them. And so we're building out a like enterprise version of our product where, you know, Samsung and SK and all these companies can go offer it to all their employees. And this is a really big deal because it's like, it's like certification of skill that is, and it's a very valuable skill that is a true professional skill, unlike, you know, in other, in the English speaking world, like in other parts of the world, English speaking is a professional skill.

• 48:31Speaker 01

  And so, so there's like certainly a huge kind of level of adjacent things to that as well. Like it's not just about learning how to give a presentation in English, but it's also like give a good presentation in English. And so there's, there's, there's definitely a lot in that and many other spaces, I think, that are, are very interesting down the road.

• 48:49Jacob Effron?

  Yeah, I guess the, the data you gather that for that will be fascinating, like how you actually start training models to even back to the evaluation problem, like what are you actually, what are you actually evaluating?

• 48:58Speaker 01

  Totally. It's, once you start thinking about the kind of expansion from language, you start to realize like a lot of these things are very similar patterns, slightly different, but similar. And so I do think one really exciting thing long-term is the opportunity we have to expand beyond language. But all the things that we've built and and kind of developed in terms of even just like our our expertise and how you assess and teach will generalize quite well.

• 49:26Jacob Effron?

  How do you think about like those expansion areas? I mean, I guess what you're watching like those opening I demos and they have the Khan Academy one with like, you know, the sign and cosign. Are you like, I'm gonna, we'll do that one day. Or are there like areas that are not, you know, not within scope? Are you going to build the kind of sci fi, you know, learn everything from from one place?

• 49:42Speaker 01

  I think there's a huge opportunity there. I think that we'll see exactly where we play, but I think there's a huge opportunity outside of just language learning. I think there's, you know, going to be multiple players. There's so many other complexities to these spaces, but like fundamentally, there are three major sectors we're paying attention to. One is like, you know, schools in general. Obviously, people will spend a ton of time learning in schools. The second is, as I was talking about earlier, businesses and professional skills. Huge opportunity to like be able to actually certify and assess and do skill development for companies. And then the third is personal. And I think personal learning is something that people, it's almost invisible to people right now, but it's a massive, massive.

• 50:26Jacob Effron?

  Most people are, a lot of people are using perplexity, chat GPT, a lot of these tools.

• 50:30Speaker 01

  I think that like learning, personal learning will be one of the biggest areas that changes of human activity. And people don't realize how big it is, but like so many things that we do on a daily basis are could be categorized as reading books, listening to podcasts, watching videos on YouTube, like all these things are reading articles, like this is all learning or learning adjacent. It's all this kind of desire to know more about certain topics to basically become like a better version of yourself. And I kind of think, I can't help but think about like early internet and like all the bad predictions about how people were going to use time on the internet going forward. And it turned out that like there were things like YouTube or like searching, search engines that turned out to be this really interesting behavior that people were doing in other ways before.

• 51:13Speaker 01

  But you didn't really realize there was such a thing necessarily as a search engine in the early days and what a search engine could be in in 2020 or 2025. And I think personal learning will be the same thing. So I think there's an opportunity just merely on the consumer side to be able to build like a more specialized premium version of that. But there will be a lot of different players there and it will be crowded and competitive and we'll see which angle is the correct angle.

• 51:37Jacob Effron?

  Yeah, if you had to predict at this point, you just said all the predictions, early predictions are usually wrong, so we'll caveat with that. But like, what do you, or what is one vision of what personal learning might look like in like 10, 15 years?

• 51:48Speaker 01

  I think it's, I think it's really something that is highly individuated to the, to the person. So it's really like that vision of, you know, like the movie Her and the, the idea that like this thing has long-term memory. And it's got like a pretty good mental mapping of everything you kind of know, what your interests are, what your personality is and what you kind of want to know. And, and then uses that to like actually be able to, be able to like provide you the right information when. And I, I would expect that like there will be different levels of this, like just like Google or YouTube are like these very, very wide, widely used platforms that have a very wide like growth in terms of what they can actually do.

• 52:32Speaker 01

  The actual, like there will be very like casual things like that platforms like chat, maybe it's chat GPT, maybe it's something else. But then I do think there will be more kind of specialized solutions for this as well. And I think the question will be like, how does that get monetized and what, what are the patterns of usage? Are there going to be more like siloed things or will it be that are specialized and how does that work for these areas?

• 52:55Jacob Effron?

  Right, and how do they share like memory and knowledge about you?

• 52:57Speaker 01

  Exactly. How, you know, if the web 2.0 is any, you know.

• 53:04Jacob Effron?

  It'll all be walled off.

• 53:05Speaker 01

  It'll all be walled off.

• 53:06Jacob Effron?

  I thought crypto was gonna fix all this.

• 53:08Speaker 01

  Yeah, right. We'll see. But maybe we'll get something closer to email. I don't know.

• 53:13Speaker 01

  But I, I do think like broadly speaking.

• 53:18Speaker 01

  When we think about everyone now, it's not, it's not, it's definitely not controversial to say that like AI is going to change everything. People are, you know, if anything, over optimistic about how fast things will change.

• 53:27Speaker 02

  Certainly, we are.

• 53:29Speaker 01

  It's our job to be over optimistic. That's, that's good. You need to believe and be excited. Yeah, the, the, I think that for me, though, education truly is going to be one of the biggest and most exciting areas of change and disruption, though. Like, I think just like zooming out, I, what I often think about is the fact that like, you know, to, to, you know, use the, the expression of software eating the world. Software did eat a lot of the world over the last, I don't know, call it 20 years. But education, like, the way that people learn, every, now we have like Chromebooks in every classroom, but like, has the quality of education changed?

• 54:11Speaker 01

  Like fundamentally, people are still basically taking quizzes, but they're taking them on a laptop instead of a piece of paper. They're still learning. They might be watching Sal Khan as a video teacher who's teaching a million people.

• 54:22Speaker 02

  Don't forget flashcard apps instead of flashcards.

• 54:25Speaker 01

  They're using digital flashcards instead of physical flashcards, but like, the quality of all that, the efficacy, like, I don't really think it's changed very much. Like, the, you know, the way that, the best way to learn 2,000 years ago, like Alexander the Great famously learned, you know, from Socrates, and that's like, still the best way to learn. That's going to change, obviously, with AI. And education, people learning things, is literally one of the biggest categories of human behavior out there. And so, we'll see how exactly it happens, but I do think it broadly speaking, it's going to be like, one of the areas where I feel very confident that there will be a tremendous amount of change in a way that like is fundamentally hard to even predict or grok, to full extent of it.

• 55:05Speaker 02

  How do you think about the timeline for that? Like, do you think it will be five years, 10 years, or sooner than that?

• 55:12Speaker 01

  Yeah, what's the, like, I forget who said it, but the, the expression of like, you know, nothing's, not much happens in a few years, but everything will happen, like, way more than you would expect will happen in a decade. I think that's generally like, going to be the case. Like, I think people are way too overhyped on everything right now.

• 55:28Speaker 02

  Yeah.

• 55:29Speaker 01

  There will be a ton of disappointments. But over the long term, like clearly, I think lots of things will change. But also, I, what worry I have is that the- On the research side, on the technology side, like, people are so excited about the transformer. I hope that that's not like a local maximum we get overly obsessed with and like we're we're still spending a lot of time on research for other things because I think it's very TBD if like that really will get you all the way. But it really does feel like we have a huge obsession over that now and no one's really like, people aren't thinking enough about other things, but you know, that's not an original take. A lot of people think that, so hopefully people are spending time doing other things.

• 56:11Jacob Effron?

  What other areas of like education, AI education, do you pay attention to like, I mean, you know, or like, are we kind of getting somewhere with like teaching math and other subjects or like, where do you kind of, you know, both where do you like think other areas you think about and then kind of where are we in some of those other spaces? It feels like language works and I'm curious if like the, if any of the other stuff has really been unlocked yet.

• 56:31Speaker 01

  I think there's a little bit less of an advantage with other subjects until it's really, really good. Because the problem with language learning is it really does require like a classroom model of one to, one teacher, thirty students doesn't work well and it really doesn't work well compared to.

• 56:46Jacob Effron?

  You can speak from personal experience on that.

• 56:48Speaker 01

  Right, right. And so I think the like, almost like the delta of efficacy needs to be that much better, which just means that it will take longer. I'm not convinced we don't already have the technology to do other subjects really well.

• 57:04Speaker 01

  It may, we may need more reasoning for certain subjects like math. I'm not actually totally sure yet, but I think regardless, the like bar to kind of like having a thing that people will really enjoy using is much higher because the kind of current solutions are, are better than, than the status quo for language learning. So I think it's more of a timing thing and like we've, even if there was no more AI progress like today and going forward, I still think there would be a huge amount to build on top of it and build better experiences just off the current technology. But it will take time to do.

• 57:42Speaker 02

  Right, yeah. I don't know if you saw the, the GPT four O math demo, but that was super impressive where you were kind of, they were streaming just the video of them doing math problems and then they were talking about it. But to your point, it's like, there's still a delta between kind of.

• 58:01Jacob Effron?

  Having a human there teaching someone and then having the kind of AI type experience where, you know, it knows what you're doing and can comment on the math in real time.

• 58:10Speaker 01

  Yeah, I mean, it's like, it's easy to imagine that many millions of people are going to be using ChatGPT for homework help. So I'm not sure I would build like a homework helper tool right now. But like, if we're trying to fundamentally improve the way that math is taught versus like getting help with homework, that's like a much deeper solution. That may not be technology gated, but it's more gated around can you truly build something that's substantially better such that people will like actually go adopt it. And like, you know, people aren't learning math casually as like in their personal time. So like the bar is even higher because you need to go sell to schools for math, right? And maybe

• 58:52Speaker 01

  you go for go after parents for like tutoring or whatever. But like, yeah, that I think like the the oftentimes it's like actually just like building a really good product and finding the right market are like the harder parts here. And that's the thing that it's not really a technology problem. It's a product and a market problem.

• 59:08Jacob Effron?

  What's the next thing you think you build after language learning?

• 59:11Speaker 01

  I think that's a great question. I we're going to be focused on language learning for quite a while. We've got a lot more to do there. But yeah, I think it'll be like adjacent adjacent things to that. And I think we have to be thoughtful about where we want to invest across the sectors of schools, personal and and businesses.

• 59:30Jacob Effron?

  Well, we always like to end our interviews with a quickfire round where we ask you some questions and get your takes. I thought we've already alluded to some of this, but I'll ask it anyway. You know, what's one thing that's like overhyped and one thing that's underhyped in the AI world these days?

• 59:42Speaker 01

  Underhyped? Well, overhyped, probably everything. I don't know. I don't know what you guys see, but it does feel like a lot of funding and not that much actual true product market fit and not really real activity usage. But and then but I do think over the long term, clearly there will be really big winners. And I don't know which those are, but there are definitely certain things that are underhyped right now. And maybe just like technology that isn't the transformer is underhyped right now. And like the amount of research going into that.

• 1:00:13Jacob Effron?

  What's been the biggest surprise in building AI features? Maybe one thing that you thought would work that didn't, or one thing that you thought wouldn't work but ended up working?

• 1:00:22Speaker 01

  I think the constant thing that I struggle with, or everyone struggles with, is like you see a new technology or capability, you get extremely excited about this is going to change everything. And it turns out like it's always never as good as you think it will be. And it's not a panacea. And basically, building something that will actually change user behavior is really, really hard. So I feel like that's like a continuous loop for us.

• 1:00:44Jacob Effron?

  Was there one you thought, you're like, oh, my God, this is going to just be a game changer for engagement? And it's like, OK, it like keeps on the gradual path of improvement?

• 1:00:50Speaker 01

  I think like maybe just like our general like open-ended lessons when we had like human-level transcription and then we're able to tie that into GPT-4.

• 1:01:01Speaker 01

  I was like, it was, it was good, but it wasn't, it wasn't like a game changer.

• 1:01:04Jacob Effron?

  What's something you've changed your mind on since like building Speak? Something you thought strongly before and maybe less strongly now?

• 1:01:09Speaker 01

  Oh, man. That's a good question. There's always a million little things and basically constantly learning. I'm, I don't know if there's something that stands out as like a, like, this is, this was like a fundamental.

• 1:01:24Jacob Effron?

  There's not like never build a model and now you're like, oh, we like to build models.

• 1:01:27Speaker 01

  Yeah, I don't know. Maybe, maybe in the very beginning, we thought we were going to be doing all the modeling ourselves. And now we realized, oh, no, actually, some models are going to cost a lot of money to build.

• 1:01:38Jacob Effron?

  Or maybe not as much if you, if we find alternatives beyond the transformers.

• 1:01:42Speaker 01

  Yeah, true, true.

• 1:01:43Jacob Effron?

  Which AI startup outside of your space are you most excited about right now?

• 1:01:47Speaker 01

  Oh, man, we kind of covered this earlier. I don't track this space well enough, so focused on Speak.

• 1:01:54Speaker 01

  Yeah, I wish I had a good answer. I definitely feel like I'm, I pay extra close attention to all the startups that the OpenAI startup fund is investing in now. Because I just think they have incredible deal flow. And there's been a lot of really interesting stuff coming out of there. But I don't track the space close enough.

• 1:02:11Jacob Effron?

  Yeah, what's it been like working with that?

• 1:02:13Speaker 01

  It's awesome. Highly recommend.

• 1:02:15Jacob Effron?

  That seems fun. Well, we want to leave the last word to you. You know, where can folks go to learn more about you, the AI work you're doing at Speak? For folks that want to get better English, speak itself as an application. I'll turn the floor over to you.

• 1:02:26Speaker 01

  Yeah, I mean, you can just visit us at speak.com. And we're hiring across the board for pretty much all roles. And so you can visit speak.com/careers and learn more there.

• 1:02:39Jacob Effron?

  Awesome. Was it hard to get speak.com as your URL?

• 1:02:41Speaker 01

  Took a few years. Took a few years, the negotiation.

• 1:02:44Speaker 02

  Good you have it now, because yeah, the prices only go up.

• 1:02:47Jacob Effron?

  Do you have speak.ai too? Or has someone else going to come in and you know?

• 1:02:51Speaker 02

  We should corner the market.

• 1:02:53Jacob Effron?

  We did look at that, yeah.

• 1:02:54Speaker 02

  Yeah, well, let's get it right after this and then.

• 1:02:57Jacob Effron?

  Yeah, exactly. We'll sell it back.

• 1:02:58Speaker 02

  Yeah.

• 1:02:58Jacob Effron?

  Right, right, right. Podcast for a few years.

• 1:03:00Speaker 02

  Yeah.

• 1:03:01Jacob Effron?

  No, that's awesome. That was great. Thanks so much for coming on.

• 1:03:04Speaker 01

  Yeah, thanks, guys. It was great.

• 1:03:06Jacob Effron?

  That was a fun episode.

• 1:03:07Speaker 02

  Super fun.

• 1:03:08Jacob Effron?

  I feel like it's just, you know, one of those products that clearly has product market fit.

• 1:03:12Speaker 02

  Totally.

• 1:03:13Jacob Effron?

  Just from your own experience in language learning, you know that, one, the current system isn't great. And two, the only way you really learn is speaking. It's why people study abroad or kind of find themselves in, you know, in environments that can actually learn a language. And so it's just such a good use case.

• 1:03:29Speaker 02

  It really is an incredible use case of AI, of something that is really hard to deliver this solution without AI. And then now everyone can kind of have this personalized experience and this, you know, AI tutor for, for learning a new language. It really makes so much sense. So it's, it's cool to see the traction that.

• 1:03:49Jacob Effron?

  And it feels like you're on the precipice. You know, this is maybe the first one because it's so, to his point, it's so much better than, than the status quo. But it feels like this is coming for, for other subject areas too.

• 1:03:58Speaker 02

  Yeah, absolutely.

• 1:04:00Jacob Effron?

  Which will be, to your, to your point on the pod, I think, just incredibly impactful. So it's also exciting from that perspective.

• 1:04:05Speaker 02

  For sure. The other thing, a few things that stood out to me. One, I loved his take on moats and what he was saying about the, how he thinks about his moats at, at Speak. And there were kind of three, if I were to summarize, at least the ones that, that I distilled down. One, specialized models. So they're building these kind of specialized models for. Kinds of speech synthesis or understanding speech. Second was around the scaffolding. That was, he put a lot more emphasis around that than I feel like we've heard in the past of how hard it is to build those systems that, that scale. And then the third is one that we hear a lot around the end to end experience and kind of really solving the problem for their, their customers.

• 1:04:49Speaker 02

  But it was just interesting to hear about all of the things that, that they have built on top of the L L Ms.

• 1:04:57Jacob Effron?

  Yeah, I love the way he articulated it. And I think it's funny, we talk about this so much internally, you know, everybody wants there to be some like crazy moat, like, you know, data partnership that nobody else has or like model that's a hundred times better. And I think the reality of what we're seeing in the application space is it's often like a hundred smaller things that you come together just to make a really good product.

• 1:05:15Speaker 02

  And, you know, like you said, the models are amazing today, but like maybe in two

• 1:05:18Jacob Effron?

  or three years, that's actually not where they're gonna end up differentiating. And, you know, I think the, the orchestration and the scaffolding just feels really important. And I feel like there's, it's kind of one of those things you know, when you see it, when you use one of these products, it just works really well.

• 1:05:31Speaker 02

  Yeah.

• 1:05:32Jacob Effron?

  It's kind of like a delightful experience, kind of like the traditional SaaS world.

• 1:05:35Speaker 02

  Yeah, that's, that was the, the thought that I had hearing that back was it's not that different than traditional SaaS, right? Where, you know, maybe in the limit, you have a moat around distribution or scale, but in the early days, it's mostly around UX or workflow, something that's hard to build, technology, defensibility, all of those things.

• 1:05:55Jacob Effron?

  I think of all these investors that are like, where's the moat, where's the moat, where's the moat? And if you ask the same question of early stage SaaS companies, you know, you'd get a lot of the same answers.

• 1:06:03Speaker 02

  Yeah, totally. Any other takeaways for you?

• 1:06:05Jacob Effron?

  I thought it was really interesting, this idea that, you know, that they still do develop their own models for, for some of these areas. And to his point, like these use cases for, you know, the person was trying to say a word, but like, maybe the machine, maybe like, there's just different.

• 1:06:19Speaker 02

  The AI is too smart.

• 1:06:20Jacob Effron?

  There's different end use cases you're optimizing for. And so obviously, if you're just trying to make a automated support agent, you'll just correct it for the user, but here you're trying to help the person learn and how that's actually like a different model you have to build.

• 1:06:31Speaker 02

  I, I thought his take on audio and UIs and how there'll be this fluid UI that they're looking to build where you can. Talk to it in some instances, but then also use the use the actual UI and others. I thought that was was really interesting and they're just at a really cool space where they are shipping so much with voice. So there is, you know, I think they're one of the first companies that's kind of pushing the envelope of what can be done with voice because it is so natural to be in that medium, obviously, when you're learning a language.

• 1:07:05Jacob Effron?

  I love this point on chat GPT two or like, you know, some people think, you know, GPT four O and people learning language through audio could be a big threat and Duolingo stocks down. And he's like, actually, like, you know, when you hear him talk, he's saying, look, so much of my time is just how do I convince people they can learn language by doing this? And like this kind of rising tide of more people like believing they can use AI to learn language, doing it via speaking is just net good for us because we'll always build the most specialized version of that. And so I do feel like these early pioneers are doing a lot of work on behalf of the ecosystem of just treating teaching people new ways of doing things. And then it becomes, you know, and if there's so much power in the same way that chat GPT having become really popular now, everybody is like building a UI that just like lets you do the same thing because it's you know how to use chat GPT. You know, maybe as we see more speaks and people using GPT chat or using GPT four

• 1:07:50Jacob Effron?

  O, they'll just be, yeah, like there'll be a million other people that are copying that UI and then there'll be another one next year. It's a it's cool to see.

• 1:07:58Speaker 02

  For sure.

• 1:07:58Jacob Effron?

  And I also really liked, you know, what he was saying about a lot of people that have problems with evals are really actually struggling with what's the problem they're trying to solve. You know, Linus from Notion said the same thing on many episodes. That's a long call back.

• 1:08:14Speaker 02

  Deep track.

• 1:08:14Jacob Effron?

  Yeah, deep, deep cut. But I was like, you know, I think this idea of actually doing good eval is doing really good product management, too, of like, what exactly are you trying to solve? And if you thought that all out beforehand, really in a really throwaway, then you should have some really good kind of examples and test cases.

• 1:08:30Speaker 02

  Yeah, defining the the problem and the solution and what good looks like and being able to run all of that.

• 1:08:36Jacob Effron?

  So maybe even everyone thinks product management is dead, you know, maybe maybe still important.

• 1:08:42Jacob Effron?

  But yeah, Pat, that was a that was a really fun episode. Yeah, super fun. Folks should definitely stay tuned. Next time we'll have another AI founder leader in the space talking about the future where everything's going.