Config 2025: Advancing animation with engineering and design with Natalie Glance & Abhishek Mathur

Summary

In the "Config 2025: Advancing Animation with Engineering and Design" session, Natalie Glance and Abhishek Mathur from Figma discuss Duolingo's innovative use of engineering and design to enhance their educational platform. Natalie shares insights into Duolingo's culture, their principles, and the delicate balance between raising quality and maintaining a sense of urgency. Duolingo's focus on experimentation, AI integration, and cross-functional collaboration between designers and engineers is highlighted, showcasing how these elements contribute to their app's success and user engagement. The conversation provides an engaging look into the technological advancements shaping Duolingo's offerings.

Highlights

• Natalie Glance shares her journey of helping establish Duolingo's close-knit culture and core principles 🌟.
• The introduction of the 'energy' system by engineer Moses showcases innovation and the shift from punitive to more engaging game mechanics 🎯.
• Duolingo's adoption of AI for personalized learning and its partnership with OpenAI demonstrates their forward-thinking approach 🚀.
• The creation of interactive features using LLMs and the experimental AI team highlights Duolingo's commitment to innovation 🌐.
• The development of video call features with in-app characters like Lily, offering language practice in a natural setting, shows technological advancement 📱.

Key Takeaways

• Duolingo emphasizes a strong company culture with five core principles, including 'take the long view' and 'make it fun' 🎉.
• Experimentation is a key part of Duolingo's strategy, with hundreds of AB tests running simultaneously to optimize their app 📈.
• AI has always been part of Duolingo's DNA, enhancing personalized lessons and driving new innovations like chatbots and conversational characters 🤖.
• Collaboration between engineers and designers is crucial at Duolingo, utilizing tools like Figma and AI for more efficient development and creative problem solving 🤝.
• Duolingo's move from punitive game mechanics to more engaging features, like the 'energy' system, reflects their commitment to user-friendly design 🎮.

Overview

The session featured Natalie Glance and Abhishek Mathur discussing innovative approaches at Duolingo. Natalie delved into the importance of establishing a strong, fun, and focused company culture that has fueled Duolingo's growth since its early startup days. The environment they nurtured continues to shape their path forward as they expand globally.

Central to Duolingo's successful operation is their commitment to experimentation and strategic iteration. By running numerous AB tests and iterating on existing features like the Duolingo streak, they effectively drive engagement and refine user experience. The conversation highlighted how strategic risk and cautious creativity balance innovation with user-friendly design.

A significant part of the discussion centered around Duolingo's integration of AI. From personalized lessons to charming conversational characters like Lily, AI plays a pivotal role in their current and future plans. The team leverages AI to enhance educational capabilities and streamline collaborations between engineers and designers, pushing the boundaries of digital learning.

Chapters

• 00:00 - 02:00: Introduction and Duolingo's Early Culture This chapter serves as an introduction to Duolingo, an educational app designed for language learning on mobile devices. It highlights Duolingo's mission and the culture within the company during its formative years. The chapter may also touch on some initial challenges and aspirations the company had during its early days.
• 02:00 - 05:00: Duolingo's Core Principles This chapter delves into the foundational principles and core culture of Duolingo as the company expanded its offerings to include learning in languages, math, music, and recently, chess. An early team member reflects on their significant role in shaping Duolingo's culture, emphasizing how a strong company culture can drive success. The narrative begins with a recount of joining Duolingo when it was a small team of just 40 individuals, underscoring the pivotal impact of early team members in embedding lasting cultural values.
• 05:00 - 08:00: Balancing Excellence and Urgency The chapter titled 'Balancing Excellence and Urgency' describes the early days of a startup where the team worked closely in a small setting. Everyone was familiar with each other's lives, creating a tight-knit community. The team's united focus was on delivering the best education and making it accessible to everyone, emphasizing a shared singular mission.
• 08:00 - 13:00: Duolingo's Experimentation Approach This chapter discusses the fun and engaging work culture at Duolingo and how it was intentionally maintained even as the company expanded from its original Pittsburgh office to multiple locations including New York, Seattle, China, and Berlin.
• 13:00 - 18:30: AI's Role and Shift to Interactive Features The chapter discusses the intentional strategies implemented around creating and sharing a culture at Duolingo. A key recent initiative is the publication of a Duolingo handbook that encapsulates and communicates the company's values and culture. This resource serves multiple audiences: Duolingo employees (referred to as Duos), potential job candidates, and the general public. The core of the handbook lies in its five guiding principles, with a focus on taking a long-term approach.
• 18:30 - 28:00: AI's Impact on Team Collaboration and 3D Features The chapter highlights Duolingo's core principles impacting team collaboration and product feature development, with a focus on the AI-driven approach. The primary principles include: 'Take the Long View', which emphasizes sustainability and longevity; 'Raise the Bar', promoting a commitment to excellence in product and user experience; and 'Ship It', reflecting a strong sense of urgency to continuously deliver and improve, with an understanding that incremental changes have compounding effects over time. These principles guide the team in integrating 3D features and AI to enhance learning outcomes.
• 28:00 - 30:00: Closing Remarks In 'Closing Remarks,' the focus is on continuous improvement, engaging feedback, and enjoying the process. The speaker emphasizes the mantra 'Show, don't tell,' highlighting the importance of presenting a prototype rather than a lengthy document during product reviews. This approach aligns with the principles of making the product fun and maintaining a strong focus on roots and excellence.

Config 2025: Advancing animation with engineering and design with Natalie Glance & Abhishek Mathur Transcription

• 00:00 - 00:30 [Music] Natalie, welcome to Config 2025. Welcome everyone. Um, Natalie, what what is Dolingo? What does it do? So, Duolingo, for those of you who might not know, is an education app that runs on your phone and you can learn
• 00:30 - 01:00 languages and math and music and as of just a couple days ago, also chess. Wow. Uh Andrew said that you were a very early member of Dolingo team. Uh I'm sure you had a really important part to play in establishing the culture. Culture is super important for success of any company. tell us a little bit about uh the Dualingo culture and the people. Sure, I'd be happy to do that. As Andrew said, I joined 10 years ago when the company was just 40 people, so it was
• 01:00 - 01:30 quite small. Uh it was still a startup. We were all basically all in the same room together um sitting next to each other and everybody knew everybody's name, knew the name of each other's significant others and children and even pets. So, it was really a very very close culture. Um, everybody was really single-mindedly focused on our mission to make the best education possible and make it universally uh available and accessible. Um, so very very single-minded focus on that. Um, and
• 01:30 - 02:00 also just I had never really experienced a a work environment that was that fun, so much fun. We had a lot of fun together every day. Um and so over the years as the company grew, we were very intentional about thinking about how could we maintain this culture even though we were growing um and expanding beyond our original office in Pittsburgh. So expanding to many offices, Pittsburgh, New York, Seattle, China, Berlin and so on. Um, so we
• 02:00 - 02:30 thought very intentionally about how we could do this and we we did a lot of different things. And then more recently, what we've done is we've published a Dualingo handbook so that we can help share what our values are and what our culture is both internally to Dualingo employees, we call them Duos, um, to candidates who are thinking about coming to Duolingo and and also to the world. Um the handbook is centered around our our five core principles and those principles are first take the long
• 02:30 - 03:00 view. That's um our CEO's favorite principle take the long view because he wants Dualingo to be a hundred-year company. The second principle is raise the bar, build a truly excellent product and learning experience for our learners. The third is ship it. uh we have a strong sense of urgency at Duolingo and one reason for that is that we know that as we ship changes for our learners those compound over time and
• 03:00 - 03:30 make the app better and better. Um the third is show don't tell. Um I'm sure a lot of you here are familiar with the idea of product review where people come to product review to get feedback on their ideas. Um and so we don't want people to come with a long dock. We want them to come with a prototype or you know something in Figma. Yeah. Um and then the last one is make it fun. So staying true to our roots to focus to have strong focus to have an excellent
• 03:30 - 04:00 product and to and to have fun doing it. Um sometimes people ask me what my favorite principles are or my favorite principle let's say and I have a hard time choosing just one. And what I'd like to talk about is um the two principles of raising the bar and shipping it. And the reason I like to talk about those two is because it can feel to people a little confusing. There's a tension between those two. If you're raising the bar and making
• 04:00 - 04:30 something really excellent, how do you also focus on that sense of urgency to push out changes quickly? Um, the way that that I think about it is that for raising the bar, well, one thing to say is that we've outlaw outlawed the term MVP at Duolingo. We're not allowed to use that word MVP. I wasn't expecting that. So, instead, we call it a V1. And the idea with the V1 is that we're not sacrificing on quality, but we are being
• 04:30 - 05:00 pragmatic in cutting the right corners. Um, and one thing to say is that everything we launch goes out as an AB test. So we can see what the impact is in terms of the user behavior with that new feature, that new product. So we push a V1 out that raises the bar and then we see what works and when we find something that works, we double down. Very interesting. Um, finding the right balance in intuition and data is a very
• 05:00 - 05:30 challenging thing. A lot of companies get it wrong and go either way too hard. How does Duolingo think about experimentation and what are some of the learnings that you've had? Thank you. I love that question. Experimentation is really at the heart of Dualingo. It's been one of our I'd say one of our superpowers from the very very early days. We were doing AB testing on everything. uh at any given time we're running hundreds of AB tests in parallel which
• 05:30 - 06:00 especially in the earlier days was very unusual for a company our size. Um I'd like to talk about um our approach to how we do experimentation as a portfolio approach. So any we we organized around pillars which are groups of teams and areas and uh so for example our growth pillar their northstar metric is is daily active users or DAUs and within this pillar which is about 100 people or so 150 people. Um we have this portfolio
• 06:00 - 06:30 approach where they have a mix of bets a mix of things that they do. uh on the one hand there are some known growth levers growth levers for daily active users that we can iterate on through incremental AB testing. So some of those growth levers are things like the Dualingo streak. Those of you who use Dualingo are probably very familiar with the notion of the streak. Um so and you would you would you might think that okay the streak is there it works. What
• 06:30 - 07:00 is there to do? It turns out that we can just keep on iterating on the streak through AB tests. Uh, and we can keep driving DAUs by doing that. Other examples of known levers for DAU growth are leaderboards, friend streaks, daily quests, monthly goals, and more. And then um, to complement the approach of of these working on these levers iteratively inside of our portfolio, we also have some bigger bets. So every
• 07:00 - 07:30 year we have a few big bets that we that we work on. Some of which succeed and some of which with which don't. Um so I'll give you an example. Uh if again for those of you who are familiar with Duolingo um you'll be familiar with the concept of hearts within the app. I see our animation isn't working here. Let's try it again. You'll be familiar with the concept of hearts in the app. Uh, so as a free user of Duolingo, you
• 07:30 - 08:00 start off the day with five hearts. This is a mechanic that we put into place years ago. Um, so you start off with five hearts and every time you make a mistake, you lose a heart. And when you run out of hearts, well, you're a little bit out of luck. You can either wait for the hearts to regenerate or there other mechanisms to bring them back. So hearts worked as a mechanism, but we were never entirely happy with it. um we felt like hearts were just a little bit too punitive. Um and so recently in the past
• 08:00 - 08:30 year um there's a an engineer with the company his name is Moses who came up with the idea of could we replace hearts with a new mechanism called energy. This is a mechanism that a lot of games use. And what's nice about energy is that um instead of being punitive, it's it's more of a pacing mechanism that as you do exercises in a lesson, you gradually run out of energy. And then there's fun gamification elements that we can add to
• 08:30 - 09:00 the app. For example, when you get five in a row, you might get an energy boost. Now, this was kind of so this is a big bet and it was pretty interesting because um it came out of Moses and a and a designer Vicki who worked together. Um, and initially they didn't actually have that much leadership support, but they p they persevered and they were very scrappy and pragmatic in how they built their first prototypes. Uh, and eventually they they won us over. They won leadership over and we ran a first AB test uh in Q1 this
• 09:00 - 09:30 year. Um, and when we have a new bet like this, our our philosophy about new bets is they don't actually have to improve metrics. They should just at least be neutral. And the surprising thing with energy is that it was actually positive not just for monetization but also for DAU growth and that is really very unusual for us. All right. That is incredible like uh being
• 09:30 - 10:00 able to measure it and see the actual outcomes come through is really really interesting. Um you know the other thing you talked about was uh sole engineer coming up with an idea and seeing this in the into a product. Um we talked about it last time the slides on which this uh this presentation is being broadcasted was a hackathon project in one of our events and now it's a full-blown product which hundreds and thousands of people use every day. So I
• 10:00 - 10:30 I love that power. Um the other thing Natalie which is on everybody's mind here is AI. Do you mind telling us how is Dolingo thinking about AI in your products? Uh where is Dolingo going with that? AI is definitely very very top of mind. Uh I'll start by saying that AI has always been part of Dualingo's DNA well before LLM's burst onto the scene. We use we've been using AI for years to
• 10:30 - 11:00 personalize lessons. Um, and the way we do that is to try to get the sweet spot in terms of difficulty. We also use AI extensively in the Dualingo English test, which is a test of English language proficiency. Um, but now of course we've we've picked up the use of AI quite a lot. Um, we were an early partner with OpenAI uh back before they launched um the the 4.0 GP2 4.0. Hey, how's it going?
• 11:00 - 11:30 Hey, Lily. It's uh it's Did I do that? I think I might have done that. That was a little early. A little teaser to the teaser. Uh a little too anxious with a thumb there. Um, so one thing that our learners were often telling us with Duallingingo is that, okay, it works great, but we really want more opportunity for spoken conversation. We want to practice speaking. And so before LLMs came into the scene, um, we had a couple iterations on that years ago, not
• 11:30 - 12:00 that long after I started, we worked on an early chatbot that was really heruristic based, kind of like choose your own adventure. and and we we had that at live in the app for a while and we weren't super pleased with it. So we ended up unlaunching it and then in the next iteration we thought well what if we brought human tutors into the app. Um so we did that. We actually built a platform where paid users could practice conversation with human tutors and we also that wasn't working all that well for a couple reasons. One is well the
• 12:00 - 12:30 best human tutor is really really good but the a the average human tutor wasn't that great. And then probably more importantly, um, our learners have a lot of social anxiety around practicing language with another human. You know, if you've been in that situation, unless you're really outgoing, that's very hard to do. And so then when we saw um what OpenAI had done with with Chat GBT and with their with GPT 3.5 and then 4.0 know because we were an early partner we
• 12:30 - 13:00 we saw that this could really revolutionize the ability to build something like a chatbot. Um so we were very early on. So what happened then is that we went in and just immediately pivoted that team working on human tutors and said working work on using this instead work on using generative AI instead. uh and they pretty quickly launched a couple interactive features using LLMs and then at the same time we set up this new team called experimental AI uh just a couple people a PM and an
• 13:00 - 13:30 engineer and we told them just go just go and try stuff let's just go try a lot of stuff and we'll see what sticks and so they tried a lot of things um and then eventually they came up um with this idea that well what if you could basically FaceTime with Lily, who's one of the characters in the app. And that metaphor of FaceTiming with Lily just was really strong. Of course, we called it internally FaceTime, but externally we can't use that. So, we called it video call with Lily. And and
• 13:30 - 14:00 here we'll we'll I'll give you the teaser for for real now. Hey, how's it going? Hey, Lily. It's uh it's going it's going well. How's it going? Oh, you know the usual. I wanted to ask you, got any travel plans coming up? Uh, yes, actually I'm going to Mexico. I'm going to Mexico City in October. Nice. You seem to love that play.
• 14:00 - 14:30 So, that's a little bit of a teaser. Um, as you can see, the goal here is to have a natural conversation um with Lily in the language you're learning. So imagine that you were doing this in Spanish or French or Chinese or whatnot. So a very natural sounding conversation with Lily and and what's interesting here is unlike the other, you know, chat bots you may be trying, you're doing it in the language that you're learning. And so this means that Lily um what's special about her is that she adapts her language to be at the
• 14:30 - 15:00 level that you're at. You of course you're learning with Dualingo, so we know what level you're at and we can use that to guide the conversation. Um, and there's just a lot of technical complexity here in terms of turn taking and managing interruptions. Um, and then also she has some memory. So, she remembers the past conversations that she's had with you and weaves that into the conversations that she's having with you. That is really cool. I really love to
• 15:00 - 15:30 talk to Lily again. That and such an awesome way to learn a new language. Um the other part I wanted to learn about was how has AI changed the way PMs, designers, engineers work together. How are how has that evolved? Right. So just you know before I answer directly the question about the AI piece um our engineers and designers work very very closely together. You can see that
• 15:30 - 16:00 in the video call with Lily example that I gave you. Um the example that I showed before, Lily is in 2D um using a 2D rig that a creative technologist on the design team builds uh with a a a system called Rive. Um so Lily is a 2D rig that's like a puppet um uh that software engineers uh control using code. Um and there there's really really close
• 16:00 - 16:30 collaboration here. It's quite a it's quite an interesting technical problem. There's a lot going on. So, for example, Lily's mouth moves of course in sync with what she's saying, but um even more than that, her expressions, her body movements also move to be coordinated with what she's saying and to respond to what the learner has said to her. And then here, what you see on the screen um is that we envision the future of Lily to be actually in 3D.
• 16:30 - 17:00 And what's really interesting about 3D for us is it adds a lot more ability to have context around the con the contextual clues around the conversation. She can talk about the guitar that's in the background in the corner. She can talk about um what she's doing on a computer, what's in her bookcase. And 3D for us has been a really interesting challenge um for both design and engineering to add to the app. One thing to know about Duolingo is it's coded natively on both iOS and and Android. It's not inside of a game engine. So, what we've had to do for 3D
• 17:00 - 17:30 is embed a game engine inside of the app. We're working with GDAU, which is an open source 3D game engine that came out just a little over a year ago. And that requires a whole new set of workflow for both our designers and our engineers. Um, and then beyond that, in terms of how we collaborate together, uh, of course we collaborate a lot through Figma. Um we also collaborate by uh through videos shared over Slack. So
• 17:30 - 18:00 for instance an engineer will implement what the designer has put into Figma and create a video of let's say the animation and send it back over Slack to the designer who will give feedback and this way we can go back and forth really really quickly and get to um our raising the bar and the level of of uh delight and engagement that we want. And then in terms of AI, um what's really interesting now too is that our designers are starting to lean in lean into AI. They're doing their own kind of vibe coding. Um so in some cases we have
• 18:00 - 18:30 some designers that are are vibe coding the more complex animations interactions so they can send it over to the engineer uh closer to what giving giving a much better sense of what they're looking for. Well uh this is really awesome. Thank you so much Natalie for sharing all your deep insights on this one. I'm really happy to see all the great work. Looks like a sci-fi world to talk to an avdar and learn a language. Thank you for sharing your thoughts on this. Thank you everyone.
• 18:30 - 19:00 [Music]