This is an interview about AI. My fiancé works at Anthropic. See my full ethics disclosure here.

Last week in our series on AI and jobs, Brookings' Molly Kinder warned us to prepare for a "messy middle": a long, “politically explosive” stretch in which AI job losses are concentrated among some of the best-paid workers in the economy. This week, for the first-ever Platformer live show, I wanted to talk to someone who believes in that vision: a founder building the tools that might bring it about, and who turned out to be unusually candid about what that might cost us.

I've known Eugenia Kuyda for more than a decade. In 2015, after her best friend Roman Mazurenko died in a car accident, she gathered the text messages he had sent to friends and family and built a chatbot that let them speak with him again — a story I covered at the time for The Verge, nearly a decade before ChatGPT made chatbots ubiquitous. That project was the seed for Replika, the AI companion app that now claims more than 40 million users. Kuyda’s latest startup, Wabi, takes AI in a different direction — away from personal entertainment and into the world of work. The app, which is now available for iOS, lets you vibe-code apps on your phone using text prompts. Over the next year, Kuyda hopes to shift more of the team’s work away from standard enterprise software toward apps built on her own platform.

Kuyda argues that we are living in "the Microsoft DOS era of AI interfaces," and that we’re desperately in need of a Windows equivalent: an easy-to-use graphical user interface that lets the average person take full advantage of agents and personalized software. When that happens, she predicts, the long tail of subscription-based apps — the calorie counters, meditation apps, and fitness trackers of the world — will start to disappear, replaced by software that we make and share ourselves.

But what struck me most during our conversation was her answer to the question at the heart of our podcast miniseries. When we began, I expected that more tech executives would tell me they expect AI to cause job loss. Instead, it’s been the opposite — most of them have said that advances in AI will only increase demand for software engineers and other knowledge workers.

Kuyda is our first guest to say plainly that she believes that is a fantasy. The fear of job loss is "super justified," she told me; in her view, AI has made hiring junior employees "extremely expensive and completely unsustainable for a startup," because every hire now competes with the leverage of what she calls a “1,000x engineer.” "I think the crazy protests around jobs and AI are going to start happening," she said. "We live in this very optimistic city, where it's all about future, future, future — but as soon as you get out of here, it's pretty scary."

She’s building Wabi accordingly: the company is modeled on a soccer team, she told me, with 10 to 15 superstar "players on the pitch" who get sizable equity and public-facing roles, supported by contractors in the back office. She doesn't think you need more than that to build a billion-dollar company anymore.

Whether that turns out to be true depends in part on Kuyda’s own bet on Wabi. Can vibe-coded apps truly compete with enterprise software in the way that she hopes? Or will most companies continue to prefer the stability and support that comes with traditional software as a service?

We should get more data on that point soon: Kuyda told me on the show that after a year in beta, Wabi will launch publicly before the end of the month.

Highlights of our conversation are below, edited for clarity and length. Listen to the entire conversation wherever you get your podcasts — just search for Platformer — or watch it on YouTube at youtube.com/caseynewton.

And let us know what you think — we're new to podcast production, and welcome your feedback at casey@platformer.news.

Casey Newton: So it's 2015. You are almost a decade away from ChatGPT. What were you seeing that made you think, "I can actually use the tools that are here to make a kind of prototypical chatbot, and that will be an interesting thing to explore"?

Eugenia Kuyda: We actually started a company that was building chatbot tech in 2013. What kick-started that was a friend of mine who used to work at Google DeepMind showed me this technology called word2vec, which was the original tech to basically transform language into math — to let computers understand words, in a way. ImageNet also dropped, and I was like, whoa — soon that will somehow come together, and we'll have these new neural networks that will understand language. I used to be a journalist before that, and we had this gigantic sign in neon: "The limits of my language are the limits of my world."

Newton: Which is a Wittgenstein quote, if I remember right. 

Kuyda:I felt like if we figured out how to build language models, that would probably be the closest to understanding the world as well.

So we started building that, way before any of the first language models. Then in 2015, Google published a paper where they talked about the first deep learning model applied to dialogue generation, and we decided to hire every possible NLP researcher we could find to focus on these language models. And then, of course, after Roman passed away, we built that AI for him. We were struggling to find a consumer application, and that was it. We were like: maybe we can't yet build a chatbot that talks eloquently with people and has meaningful conversations, but maybe we can build one that can listen, and that would probably be enough for many people out there.

Newton: How confident were you when you were doing that?

Kuyda: We thought it would work at some point. The models were so crappy in 2016, when we started Replika, that they would just produce non sequiturs. These were sequence-to-sequence models, and of course there were no models off the shelf or through APIs, so we had to build our own — this was right before people moved on to transformers. So it was very hard to say, "Okay, this will become what it is today." We just felt that it would happen; we didn't know when. For us it was more like: okay, maybe the tech is lagging, but it's less about tech capabilities — it's going to be more about human vulnerabilities. There were so many people who wanted so much to have some connection — someone to listen, someone to hear them out, to accept them, to understand them — that maybe in the beginning just those people would react positively to it, and as the tech got better, we could increase the range of people it would resonate with.

Casey: I think there's something really poignant about the fact that even though the technology, by today's standards, was maybe not that good, the human desire and need for support and connection was so powerful that people looked right past it. But at the same time, I wouldn't downplay the technology either, because when I was talking with Roman's friends and family, the part of the story that will still make me cry when I tell other people about it is how much people learned about their friend and family member after he passed away, and how their relationship with him changed after he passed away, because of the conversations they were having in this app. That was honestly the moment when I started to take AI more seriously, because I thought: if people can feel that deeply even in this very primitive version of the thing that we have today, there just has to be something there.

Kuyda: I think so. And looking at my previous relationships — oftentimes we do have relationships with people where maybe they don't respond that much, or it's more about our fantasies. How much do we put in? A good example is talking to God. So many people talk to God, and maybe he doesn't really respond.

Newton: He's sort of famous for leaving you on read.

Eugenia: I also had a lot of experience going on dates where you just listen, and maybe ask, "Oh, tell me more," and then the guy would be like, "Oh, that was the best conversation I've ever had." And you space out half the time when they're talking — you're thinking about all the groceries you need to buy. So I'm like, if this is the level of understanding that's required for the most amazing conversation, we can probably build that.

Casey: Once you realized how low the bar was, you thought, "There's a unicorn here." 

I want to zoom out and ask you a question about your two companies, because on the surface they look quite different, right? One is about an AI that you develop relationships with; the other is a tool for making apps. In your mind, are they completely different, or do you see a through line there that you're chasing?

Eugenia: They're definitely different things, but for me the idea was always: how can we make a person's life better, or help them unlock their potential? With Replika it's easy — it was always about building an AI to help people flourish and feel better in the long term. We had some big studies published around that with Stanford and Harvard, some of them published in Nature, where we proved we were doing that.

With Wabi, the idea is: most of our time today is spent on our phones, using software that's not built by us — built, in David Foster Wallace's words, by people that don't love us, that want us to just scroll or click on things. We shape our buildings, and then they shape us. It's the same with software — we shape our software, and then it shapes us. Only we don't shape it; someone else does. 

So in this new era, where anyone can really build something in a matter of a few seconds, why not let people take a little bit more agency? Maybe not build every app they use, but at least have software be more decoupled from this model where every app needs to be a business. Wabi is a platform where people can make apps, but can also discover, remix, and use them with their friends and their families. It's a social platform where you can quickly spin up any app, or find any app, and start using it with whoever you want. In my case, that means creating software that really fits my life — whether it's helping me learn more about the art movements I'm into, or the language I forgot, or teaching my kids something, or finding cool events to take my kids to, or even just a better weightlifting tracker.

Casey: What is a feature or a design element of something you've built that made you feel like, "This is truly, personally for me — I would not expect to encounter this kind of app anywhere else"?

Eugenia: When you take away the idea that you have to make an app, put it on the App Store, distribute it, and make it for some audience, it can just be n-of-one. For me, I have an app that teaches me a daily philosophy concept. 

But that's the simple way of putting it. The more interesting reason I really decided to work on this is that I do believe we're in the Microsoft DOS era of AI interfaces, where everything's a chatbot. I've worked for 10 years on a chatbot, and I do believe there will be a GUI moment — a Windows, macOS moment — that will come to AI. Mostly because even though the model capabilities became so much better over the years, most people — normies, I guess, us included — still use ChatGPT and Claude mostly the same way they used them in 2022 and 2023: ask questions, search, do homework. That's it. It's not all these crazy agents — they're not spinning up cron jobs or figuring out Claude Cowork, even. And really, that's because through text, through a chatbot, it's very hard to discover anything.

Casey: It feels like talking to the Alexa in your house, right? It can set a timer, and it can check the weather, and it can do 1,000 things, and you don't know what those things are — so you just use it to check the weather and set a timer.

Eugenia: Exactly. But even if you set a timer, you need to see that timer. Chat is great as one of the interfaces; it cannot be the primary one. People love to tap, tap, tap, click, click, click, scroll, scroll, scroll. And that's the only way to really make things discoverable and multiplayer.

Casey: Talk about the demand that you've seen for Wabi so far. Sometimes I feel like a freak, because I like to use software — I love productivity tools. Most people don’t feel that way. So talk to me about these people who are out there saying, "I need to build a philosophy app that only I will understand."

Eugenia: It's really just about making this tool simpler for people to use. We're still in private beta — we're going public in the next two weeks, so I'm super excited about that. 

But I think we grow up being more creators, and at some point we become consumers. Kids use Roblox — kids make these games, kids hang out in these environments they make for themselves. And then at some point we just turn into these passive consumers: scroll, scroll, scroll, and subscribe, subscribe, subscribe. I think once you show people that it's actually very easy to make something — or not even make something; maybe we just suggest some apps for you that someone else made, and it's very easy to remix them. The agent says, "Hey, I see you added this app. I know all your apps are black and white — let's change this one into black and white, too." So it's proactively helping you use all this software.

But I do believe software needs to change. If we're just using AI to write the same old apps from the past, that's pretty boring. What needs to happen is new agentic apps, where all apps have agency and are more alive. What I mean by that is that you can change them, they can suggest how you can change them, they can grow with you, they can evolve with you — and they can also talk to you. Right now, apps can only send you push notifications. With Wabi, all apps have a chat, so the push notification becomes "Time to work out" — but you can also say "Stop messaging me" as a response. You can change everything right there in the chat. Chat becomes the way for an app to talk to you, but also the way for you to change it.

Casey: Tell me about an example of something somebody built that made you say, "This is the promise of what I'm doing, realized" — the equivalent of that early moment with chatbots, when you saw the pieces coming together and how badly people wanted it, even though the technology was primitive.

Eugenia: A couple of things from my personal experience. I built this weightlifting tracker — I'd been tracking my gym workouts in Notes, which I found out a lot of people do. We make all of our apps agentic by default, and it started talking to me after my workouts, giving me some pointers on how to improve them. So I said, "Now also talk to me during workouts, as I'm logging — tell me what I can do as the next exercise." And all of a sudden this app just felt so much more alive, and so much better than even a really fancy-looking app off the App Store, because it was smart. And not only that — it was also connected to my Apple Health, it was connected to my other apps, so all of a sudden it had a lot more knowledge about me. That was really a magical moment.

Another one: we have a few apps for our team. Our design engineer, Alex, makes lunches for us at the office every day, so we made an app where he puts up the menu for the week, and we can all vote and comment and say stuff — "Oh my god, these poke bowls were so nice." It was just a little bit magical, because it created another way for us to bond more as a team.

To me, the really important thing is that today we have AI that lives separately in a chatbot interface, and then we have apps on our phones, and everyone's debating: okay, MCPs or APIs — how are they going to communicate? But really, we should not have that distinction. Every agent or agent skill should just be an app, because a normal, regular person will never understand what an agent skill is, and no one's going to go read Markdown files on GitHub. Instead, they can totally understand: oh, it's just an app that looks at your inbox, and whenever there's a new email, checks whether it's an important one and sends you a quick summary. That is super easy to understand. If I tell you it's an email agent that triages your inbox — "here's the Markdown file, go figure it out" — that's hard to understand.

Casey: In a world where everyone can make their own software, what does it do to the value of software that other people are selling — SaaS companies, for example?

Eugenia: The biggest problem with vibe coding is that no one's going to use other people's apps if those indie developers own the backend and the data. There's just no way — even for consumers, let alone for businesses. If I build an AI therapy app on Replit and say, "Casey, use my fantastic AI therapy app, here you go," you're like: okay, well, Eugenia can read all my log