从 AI 伴侣到个人软件：看见未来

Right now, AI is just an app on your phone. It should not be that way. Sometimes you need to go big or go home. Not having the balls to do that, especially in this current environment, you can suffer the consequences.

There's a huge mind trap that exists among builders in the space where they somehow think that voice is the main interface. That's because they are somehow thinking about the movie Her all the time, but not in the right way.

You were a true pioneer in the space before everyone else was doing it. Maybe just talk about your reflections on how that category evolved as you helped create it, and then how it's evolved since.

It definitely evolved. It's pretty crazy. We had a very strong belief that it would happen, but we were still so surprised when it actually happened. I guess it still felt like complete magic.

So, Eugenia, I'm excited to get into everything you're doing with Wabi. First, let's contextualize this a little bit. When you look back at the arc of your career, perhaps the last decade, how do you contextualize the through line through Replika, Wabi, and so on?

Oh, wow. That's a hard question, but I guess maybe the main theme is that we're always early. Sometimes a little bit too early; sometimes, hopefully this time around, the right kind of early—the right type of early.

But generally, yeah, I started working on AI in 2012, so over a decade ago, and was always fascinated by the idea that you'd be able to talk to a machine, to have a good, meaningful conversation with a machine. More importantly, I was fascinated by the fact that it could influence your life in a really good way.

Before that, the focus was really on these AI companions, AI friends that could be there for you and help you live a better life, help you feel better. Now, I guess it's just the same idea applied to personal software: Can we build mini apps or software that will really help us throughout the day in a very personal way and be focused on you, on helping you live a better life?

Trace the journey of how you got excited about this new category and what you hope it becomes.

While running Replika, we always had these conversations with our users about what other products they use and how they use AI. It always struck me as strange that they were using products like ChatGPT, Gemini, and Claude, but they were mostly using them for very simple use cases. They were using ChatGPT just for search, to help them write something, to help with homework, or something else.

No one was really mentioning any of these exciting capabilities that we kept seeing in the models. We felt like there must be an interface problem. When people look at a command line or a chatbot, they really just see a search or writing tool. Maybe I can talk to this. But if you think about it, that's the affordance of a command line—not really the main use cases.

I think even recently, OpenAI published one of those research studies showing that these are, in fact, the main use cases for ChatGPT. I think a third of all use was around writing or writing help.

That gave us the idea that there should be some next interface. There should be something else on top of this. There should be something more interactive, visual, and simple for everyone, me included, to actually discover these amazing use cases that all these models already have.

We started thinking about it and got obsessed with this metaphor that the current chatbots are really the MS-DOS era for AI interfaces, and that there will be some sort of Windows or macOS moment.

I think the confusing part for most people is that there is so much traction with chatbots already. Almost a billion people are using these AI tools, but they only use them for these simpler use cases. In order to unlock something else, you need a more exciting interface.

Paint more of a picture of what the world will look like when this Windows or macOS moment happens. What will that look like?

I do think right now we exist in the paradigm of when I was growing up: We had TV, and you had maybe a few dozen TV channels. Back in Russia, we only had 6 TV channels. Of course, TV still exists, but now we watch YouTube, Reels, and TikTok—a lot of UGC stuff.

I think the same will happen with software. Right now, we're kind of stuck in this world where there are these few apps developed by professional developers. Eventually, of course, we're going to move on to this new world where apps are built by all of us, for all of us, and maybe sometimes by AI for us.

If you think about it, the operating system of the future is one where you open it up and see your regular, popular apps—the ones you use all the time. Maybe it's Instagram, whatever it is. Then you see some cool apps that you discovered, that your friend or best friend is using, or maybe that you built for yourself, or that you tweaked from one of the apps you found.

Then you'll see AI suggesting some apps for you. Maybe you're going to New York next week and you're into art, so it made an art-show finder app for you, which is going to help you find art shows around the Airbnb you're staying at.

It's a lot more flexible, malleable, and deeply personalized. Think of it as an operating system built on the platform of you, and not on some random, fixed context.

Can you talk about that? We've always had this idea that software has to be durable because it's really expensive to make and it seems really serious. What types of ephemeral software, like the New York art-gallery app, do you imagine existing?

Right now, we're already seeing some of our first users building very specific apps that would never exist in the App Store. They're just too small, too personalized, and too niche. They don't have 20,000 features.

Someone built a motivational-quote app that's only pulling from one particular show that I didn't even know anything about, but he's really obsessed with that show. He just wants that at 5:30 a.m. when he wakes up.

It's really—I feel like people are building very particular things to fit their needs. For example, I was putting my kids to bed the other day, and my daughter wanted to play these puzzle games. When I tell her something, she's trying to guess what it is, so we built a game very quickly where it's a puzzle and she sees 4 pictures and can click on 1.

She also wanted them to be about Princess Elsa and Princess Jasmine, so we had to incorporate that. She was so happy because now she's doing these puzzles, learning something new, and then we changed it to Italian because she goes to an Italian preschool. This is another way to practice.

She's so excited. We spent the whole evening—we couldn't put this thing away. It took me 2 minutes to build it and then a few seconds to tweak it, versus going on the App Store—I don't even know if an app like that exists—and going through 15-minute onboarding and paying again. Then it's not really what she wanted. There's no personalization there.

I think these are the use cases where something like that should happen. In the future, I should just be able to say, “Hey, I'm putting my kids to bed,” or it should know that context and maybe suggest a few apps that are already prebuilt for me that I could use right now.

It's funny. I just got a new iPhone, and on the new phone I had to delete a bunch of my old apps because I like to start fresh. There were probably over a dozen apps that I had downloaded and, in some cases, paid for that I just totally didn't need anymore because I had built better versions of them on Wabi.

Everything from migraine tracking to tracking restaurant recommendations, to a really hyperpersonalized notes app, to special things around image transformation in a particular style. I can imagine that will be true for a lot of people. Instead of having to find this long-tail app that's running all these crazy ads and pop-ups all the time, is hard to use, and isn't personalized to you, you can just make exactly what you want on the fly and tweak it, which is another really cool part of the product.

I think this is what—and I'm so glad to hear that you're using it—but for me, the product-market fit with Wabi was around that. I found there were a few things I really needed and wanted an app for: to track my weightlifting, very beginner weightlifting workouts.

I just figured out you've got to go up in weight. I didn't know that. Anyway, I got this app and I was tracking it in Apple Notes. I tried to find some apps on the App Store, and they're just—

There's just so much in these apps that I didn't need, and I wanted something really simple.

I'm also trying to follow one particular book. I wanted the workouts to be based on that book. So anyway, I built a simple app to track these workouts.

I built it on the way to the gym, but now anytime I go to the gym, I find something else I want to add to it. First it was just a tracker. Now it's generating new workouts based on all of my inputs, where I am, the book I'm trying to follow, whatever—the ladder technique of progressive overload.

I put it all in the prompt, and every time I go to the gym, I add a little bit to that. I want a little more of this, a little more of that.

So now I'm not just using the app. I'm constantly in the process of tweaking it and republishing it to Wabi's mini app store. Hopefully, someone will find it useful.

How many people do you think—or, generally, as you think about your audience—do you think it will be 90/10, sort of consumption/creation? Or do you think that most people—I think one of the things we saw with Sora, for example, is that many people, the majority of people, are creating. Do you think that Wabi will be like that?

I hope that more people will at least tweak, but I'd say fully original creators will probably still be under 10%. What we're working on right now—this week we're releasing, we're pushing sort of a big update—is the social graph. So you'll be able to see who is downloading what mini apps and how they're using them. You're going to be able to see comments on mini apps.

In this case, for example, if I created that workout app, someone in the comments can ask me to tweak it. They can also remix it and change it—obviously, change the app however they want. But maybe they just want this particular app to be slightly different. So they could put it in the comments, and maybe, as a creator, of course I'm going to be reading all the comments and changing the app so these people can use it too.

I think this is cool because, all of a sudden, it's not just about building apps for yourself. It's really about discovering apps, finding what your friends are using, using these apps together, and even asking creators to change them in some particular way.

Amazing. Zooming out a bit, or segueing, Anish and Justine, how do the categories or the topics that we've been discussing as they relate to personal software relate to themes that we've been exploring on the investing side? What got us excited here?

Yeah, I think the YouTube metaphor is the right one that Eugenia outlined, which is that we would have said in 2007 that 100 cable channels is enough—or 6 in your case. Now, obviously, there's this entire ecosystem, everything from unboxing videos to all of the very YouTube-native content that exists. I think people have more content and are more fulfilled by it.

Then there's also active creation. Some people create for a business, but many create and post just because it's fulfilling to them. Software has just been so restrictive because there are only 20 million developers in the world. In a sense, all of the software that we consume is downstream of the preferences of those 20 million people. It seems intuitive that if more people could make software, they would. There would be a mass-market consumer product here.

I think, unlike many of the other products that we've seen—and you should comment on this—it feels like this is not text-to-app. It's not a developer or developer-adjacent tool. It's truly a mass-market consumer product for nontechnical people.

I agree with you, and we really made this choice early on that we're never going to show any code or anything, or say anything in the app that's even remotely technical. No API keys, no “bring this,” “connect this integration,” or anything like that.

We do have integrations. You can add your apps and services, but you can just say, “Use my Apple Health,” or “use this, use that,” or “use my email.” We call them power apps, so that's probably the most technical you're going to get here.

We wanted to make it super simple for anyone to make apps and to make this process very delightful, almost to feel like you're just creating something really quick. You're not even really building mini apps.

What we're adding right now is more—I think the company or the product we're looking at most is Canva, and the ease with which they're letting people create beautiful presentations. A similar kind of thing needs to happen here. Right now, we're talking about vibe coding a lot, but I feel like we should be more about vibe kind of taste, or vibe designing these apps.

We're going to give a lot more of just visual controls, more in the vein of, “Choose a style, choose some colors.” You can go a little bit deeper, but really it's just 1 button and things look great, versus needing to go deep and really try to understand what to do and get technical with your mini app.

I do think that unlocks a certain level of creativity where all you need to think about is what's the use case. One other interesting thing that 1 of our investors, Soleio, mentioned was that mini apps could become kind of community starters.

If today the App Store is very much not social, and it probably won't ever be social because of Apple's mandate for privacy and how important that is for them—which is great—but at the same time, you probably do want to know who else is into, I don't know, toddler activities in Potrero Hill. I want to know what other moms are going to an Italian preschool. I want to have their kids go to that preschool.

Maybe my designer is really into bird-watching in London, in his particular neighborhood. So creating that app and finding some other people who could be a little bit of a community building around that topic.

Yeah, I think there's a bunch of really interesting choices you guys made in designing the product that feel like they really build on the explosion in vibe coding—or building with AI, whatever you want to call it—in a super interesting way.

One of them is that, for many consumers, the existing tools are a lot more accessible than coding something from scratch, but it's still pretty easy to break them or get to something where you don't know what to do. Whereas I think Wabi, you've purposely put guardrails around the experience to make it hard to really mess up, which is extremely helpful for consumers.

It also feels like we needed to move beyond this paradigm of, “Great, you vibe-coded a website and it exists out there.” Is that the best interface for using a product you want to use daily, storing personal information on, having your own records, and all of these sorts of things? Probably not.

For a lot of people, their phone is where they spend most of their time online. It's kind of their primary interface and where they want to use things. So, obviously, I know there are challenges designing for mobile and that sort of thing, but it felt like a really smart choice for you guys to start there, just to get so much more deeply ingrained in someone's daily life or daily workflow.

You know, I've only pretty much built mobile apps.

Yeah.

I really like the idea of mobile apps, and a lot of things you're not going to be doing by really just putting in a website or whatever, some link app. I also believe a lot in the concept of an organizational layer.

The App Store is the organizational layer for mobile, and the browser is the organizational layer for the internet. Something like that should exist for coding, I guess, or for this new era of software—for AI, let's call it AI software.

I do believe that we're all going to be building some sort of software and using software that everyone builds—some UGC software—but I definitely don't believe in links that people will share with each other, with me relying on some random person somewhere who's not a professional developer to support the database for the app where I might store some sensitive data, or at least data that I don't want to disappear, even if it's not very sensitive.

We've already seen this with some vibe-coded apps reaching the top of the App Store. I think the one around women dating, where all the super-sensitive information got leaked, wasn't because they were bad actors. They're just not professional developers.

There needs to exist some platform where everything will live. To a degree, we're not passing around links to videos; we're watching them on Instagram, TikTok, and YouTube. It's not like people are passing around links.

The current situation is very similar to the very beginning of the internet, where people were creating personal pages with GeoCities and some other tools like that, which I don't remember very well because I was very little. Before LinkedIn, people were sending these links to each other. Of course, now there is LinkedIn. There are certain guardrails—you can't really do everything the way you could do it in GeoCities—but at the same time, you get the social graph, you get the platform, and you get all sorts of cool things there.

It's the same as Shopify for e-commerce. You can build your own online store, but no one's really doing it anymore. Everyone's just using Shopify, and then they also get all the platform benefits. I feel like the same will happen with Wabi.

You cannot download an app from Wabi and put it on the App Store, but you can use it inside Wabi, and you can get the social graph, all the integrations, and potentially the shared context between all the apps and the memory behind that user, and so on.

Our mutual friend Hiten said that Wabi is not just a collection of apps. It's a framework for memory, context, and expression. Every time you create or share, you're teaching it who you are. Is that how you see it?

I believe in it 100%. I still remember, of course, and you guys obviously remember too, how the first iOS apps were just people trying to squeeze those websites into an app format, or just toy apps like iBeer. There was my favorite one, I Am Rich, that was $999.99, and it didn't do anything.

And then, of course, people figured out that there was something special about this being a mobile app: GPS, connectivity, and things like that. Things like Uber and Tinder came out, and a whole new category of apps emerged. They figured out what was so special about mobile, and I feel like, for AI, the super-special thing is personalization—but really deep personalization.

For me, vibe-coding an app that’s still sort of the same old software, not really taking any of your context into consideration, where the data of that app is not exposed to AI that keeps learning, is kind of just old-school. I do believe a lot in Andrej’s idea of Software 3.0: some next level of software that’s super deeply personalized. So, if you think in the context of a Wabi mini app, how can you personalize it? First of all, you can personalize the features and the looks, where I’ve skinned the app in a certain way.

But then, on top of that, you can also change the prompt. For example, for my workout app, I added a couple of things to the prompt. First of all, I added the book that I’m reading and the method that I want to work out using. Also, I go to a certain fitness center, and it has a certain layout—I added a photo of that gym. So the model is not generating workouts on a completely different side of the gym. If it’s a superset, the exercises have to be next to each other.

This is the deep, deep, deep level of personalization that it’s following. And, of course, at the Wabi platform level, the mini app also knows that I’m a certain age, that I live in San Francisco, that I have kids, that these are my fitness goals, and so on. Eventually, when I build another mini app, maybe around nutrition, those apps should be able to talk and pass along that context.

Today, of course, all of that exists in just the walled gardens. You’re fully locked into one app, which to me feels absolutely crazy. If I connected my email or my calendar, it can’t be connected to both of these apps. I have to go through the process with every developer all the time. They have to build it; I have to connect it. That seems crazy to me.

Do you imagine people building true social apps where the in-app experience involves a community?

I’d love that. We’re building multiplayer right now, as we speak. We spent the whole day trying to figure it out. It’s really complex, just because all apps can have very different types of multiplayer, and that needs to be explained to users in a very intuitive way.

But yeah, I totally believe in, first of all, using apps together, at least with your friends and family, but also potentially building more community-oriented apps where everyone can join. Maybe I can make my app multiplayer and open for everyone to join.

A good example of that is an image-generation app around dogs that lets you turn any of your dogs into a royal portrait from a different era. Some user built it. It would be cool for all the dog owners to join that app and be able to post to the universal feed.

That’s something we’ve been talking about today, because this would be really cool. Instead of just making these photos and sharing them somewhere, you’re adding them to the ongoing feed of dog photos.

No, I think it’s a great example. I’m definitely going to use it with my dog.

But I feel like there are so many examples, too, of image- and video-prompt sharing that happens in very unoptimized ways today. As you know very well, teen girls and college-age girls are often early adopters of stuff, and they have been making all these Nano Banana and Qwen image-edit prompts of themselves lying on a couch with the Ghostface killers behind them for Halloween.

They’ll post the image on TikTok or Reels or wherever, and it will blow up. Then they’ll comment with this long-form prompt in the comments on TikTok, which makes no sense as a way to do it. Everyone’s asking, “Where do I go to do this? I have the Google app—why isn’t it letting me do this?” You have to explain, “No, you need the Gemini app.”

It just feels like there’s already this consumer demand and consumer behavior around prompt sharing, in particular for creative stuff, that could be done so much better. I’ve already made a bunch of Wabi mini apps for this, and it could be done so much better. I think the creative community would really thrive with this.

That was the motivation, and I guess this is why we started the company. How is it possible that we have this godlike technology, yet we pass around these text prompts, which are almost like MS-DOS commands—but worse? At least the commands were something you could learn. They were short. Now we have these crazy, unstructured paragraphs of text, and sometimes you also need a reference image or something else.

To me, that’s a little bit crazy. I think that’s one of the biggest problems with discovery in AI. It’s hard to find these prompts. Even if you saw the output—you saw this cool photo with the Ghostface killers—how do you recreate that? You need to find the prompt, and you need to know what app and what model to choose. Oftentimes, it’s not even very intuitive, and at that point, I’ve just lost all my motivation to do it.

Instead, if you could quickly click on the link in the comments on TikTok and open a mini app where everything’s already set up for you, you could just add a photo. You can see examples, choose different styles, and see in the comments what other people are doing. I think this is the way to go.

What’s most interesting is that it combines vibe-coding apps, one of the biggest trends of the last year at least, with just using AI. Before that, it was all separate: either you were in the text-to-prompt-to-app market, or you were in some other market. This kind of puts it together. This is really the one place to use amazing prompts, and it blends the difference.

You mentioned, Justine, that this is a mass-market consumer behavior—or maybe a future one—of sharing prompts this way. How much do you think, for the average person whose only AI experience has been with ChatGPT, this is going to be a surprising new behavior that they have to adjust to, or something that feels very intuitive and obvious—like, “Wow, I’ve been waiting for this”?

I hope it will be easy because, at the end of the day, it’s just a mini app. It’s just an app graphical user interface. It’s something we’re all used to. You don’t need to learn how to use these new tools.

I’d argue it’s harder to use a command line because you need to know: Do I copy and paste the text here? Do I edit an image right here with this prompt, or do I edit it later? It becomes a little bit more—it’s too loose. There are no guidelines.

Everyone knows how to use apps. To a degree, a lot of these thin wrappers blew up at some point. Prisma was one of them, or Lensa, where it was just, “Change your photo into some avatar.”

Some headshot apps.

I do.

High school yearbook.

High school yearbook. Yeah, it’s clutch.

A lot of them are awesome, but again, there’s a reason why apps like that gain traction and not people just passing along this prompt, because it’s too high-friction. Whenever the motivation isn’t high enough, that amount of friction just kills all my mojo.

I wake up in the morning, I’m on Twitter or Reddit, I find all these cool prompts, and I’m like, “I’ve got to try it.” Then I’m like, “Copy and paste.” It didn’t really work, or whatever, and I forgot about it. That, I think, is something—

At least we’re super excited about it. And it’s not just about image prompts, but also some cool text prompts. People come up with all sorts of cool stuff. There are millions of people in subreddits like ChatGPT Prompt Genius, which I love. Sometimes you can find some crazy stuff in them.

But you would never even know that this could be a cool way to use ChatGPT unless you found it there. For example, someone made a fantastic prompt to analyze your blood work. But again, I don’t even remember at this point what it was or how I’m going to find it.

Right?

Versus, I could just download a mini app from Wabi and keep it in my health folder.

On the investment team, we’ve said that the world has 1% of the software that it needs, and that the rest is going to be built in the next 5 years. Give an example, or say more about what that looks like if we’re 100xing the amount of meaningful or impactful software.

Well, I don’t know. I guess I’m always going back to how magical YouTube felt in the very beginning, and all of these platforms where it was really about creativity and very raw, sometimes weird things—like putting a home video online for TikTok.

I still remember all my younger friends, younger kids, lip-syncing to different music and how weird it was. It felt like a toy, and then all of a sudden became really, really huge. I think the same will happen—or at least we hope a similar trend will happen—to Wabi, where in the beginning maybe a lot of that will look like toys or something very simple, very funny, almost, in a sense, and then eventually can grow into a much larger platform.

If you think about it, today we just treat apps as software. But what if we could treat them as content? If I’m a health influencer or fitness influencer on TikTok, maybe I should put out, “Here are my 5 mini apps on Wabi that showcase my fitness protocol.”

Get them, and maybe there’s a way to monetize it in some way. Maybe it’s a way for that fitness creator to create more content that’s now useful. Right now, people sell courses and stuff instead of that. I think a mini app could be much better, and then people can be talking about that in the community, in the comment section.

Again, this is the start of some community. People are working out together; people are doing something together. I do think we’ll see a completely different type of software—not just apps, not just static apps, but more content, community-built apps, community starters, conversation starters, and just fun little toys.

Do you imagine a creator class, a kind of professional class, on Wabi?

Oh, 100%. Yes, of course. I do believe that, ideally, this should happen. If we’re really thinking about Wabi as YouTube, that’s sort of the last frontier at this point. Creators can make their own professional content. They can make videos, they can make shows, they can write, but they still cannot create software. It’s still really not happening.

Anyone—and especially small niche creators—should be able to afford to create any software for free for their fans. That’s what I’m really, really looking forward to. I’ve been struck by the idea that MrBeast, the biggest creator in the world, has such a close connection with his fans. They do anything for him, and the thing that he makes is chocolate. That’s the thing he chooses to do that seems like the best monetization, and it just feels like this is yet another type of offering that creators can provide to have a closer relationship with their fans.

Exactly. And if you think about it, even the style—I would love certain designers’ apps. I want the apps that they build.

Because I’m sure they’re going to build very beautiful apps, and I want to look at them. Even if it’s the same functionality, the same whatever—some Pomodoro timer—I want their take on it.

And so on. I think there are so many different groups and niches, and it shouldn’t all just be about monetizing. It’s really just about different styles, different outlooks on life, on the world.

And that, to me, is very, very interesting. I’m a huge user of Reddit, and I find it very exciting because people just join around certain interests. That doesn’t happen on other platforms really that much, so that’s something that I’m excited about here as well.

Yeah, I think from the creator perspective, one of the weird things about putting out content—we’re all kind of content creators in various ways—is that you often don’t really know. You see how many impressions it gets, but you don’t know if those are bots. You don’t know how many people are actually using the prompt you posted, watching the full video, or whatever. I think that, even beyond monetization, that part of Wabi—where you can see what someone else has done or created or accomplished, or whatever, with the prompt you wrote or the app you made or the thing that you developed—will be incredibly cool.

Not only for existing creators, for whom it’s really obvious they want to do this, but also for people who are not creators today, but who have really interesting ideas and just have no way to build their own mobile app, get it approved in the App Store, ship it, distribute it—that whole type of thing. What other new class of creators could there be?

Yeah, that’s why, in our style, in some of our communications, we’re also trying to—I guess we’re a little bit nostalgic about those early days of the internet and just being weird and staying weird. Right now, a lot of the video platforms are very polished and very commercial. You don’t even see your friends anymore. You don’t see that much weird content; you just see very curated, polished stuff.

But with mini apps, with software, I guess we’re just entering this new era of tons of really weird and fascinating new mini apps that could never exist because they wouldn’t be enough of a business on the App Store.

Right, right. Well, speaking of the early days, should we talk a bit about Replika and your history in the community? Maybe just talk about your reflections on how you helped create that category and then how it’s evolved.

It definitely evolved. I’ve noticed it’s pretty crazy. It’s wild because we were thinking about it recently, just talking to one of the early employees of Replika, and we had a very strong belief that it would happen, but we were still so surprised when it actually happened. I guess it still felt like complete magic.

For me, I got into it because a friend of mine worked at Google DeepMind in 2012, and he came to tell me about Word2Vec and kind of the first technology to translate words into vectors and to be able to actually do something with words. Before that, language was this whole separate, almost like, thing that computers could not interact with, but now all of a sudden they could.

I grew up reading Wittgenstein, so for me it was like, well, the limits of language are the limits of my world. I felt like, if computers learn language, then they’ll learn about the world and they’ll be truly smart. That is the future of artificial intelligence.

Also, ImageNet had just come out, so we saw all these new models around image recognition, and we felt like, well, we’ve got to start a company focused on language models, focused on dialogue generation. But of course, there were no papers around it at all, and no known algorithms or anything—no models built, really. Back then, there was nothing.

So we were just focused on building some technology to build chatbots, trying to build some language models around that. Then, in 2015, the first paper came out of Google by Quoc Le that actually showed off the first deep-learning model applied to dialogue generation. They didn’t publish any models back then, so it was all about reading the paper, seeing some of the cherry-picked results, and trying to replicate that.

When we saw that in August 2015, we basically bet everything we had at the company on building those models. We said, “Okay, this is it. This is right around the corner. We need to focus on building these language models, getting the first generative AI product—a chatbot—out there,” which we did with Replika. But it wasn’t around the corner. It was 7 years away.

Then we just had to survive long enough to actually get to the first transformer models. I think the next magical moment was the Meena paper, also out of Google, with the first transformer model. In 2020, we got invited by OpenAI to go see GPT-3 before the API came out, to partner up with them and be one of the first partners for the GPT-3 API launch.

We came to the office, and Mira, who back then was actually leading partnerships, and Sam showed us GPT-3. I remember being floored. It was insane. Before that, if you think about it, we had to train every model. We had to create the specific dataset. If you wanted to train a dialogue model, you had to have tons of chat data, and you would train the model, and the model could only do dialogue.

But with GPT-3, it was the first kind of zero-shot, few-shot model where it could do anything. It didn’t just have to respond in a dialogue format. You could tell it, “Write a tweet like Sam Altman,” or “Write a tweet like this,” or “Translate,” and it would do that. That felt absolutely magical, and we were one of the first partners of OpenAI’s GPT-3 API.

It was still crazy because we still have a Slack channel where Greg Brockman is training a model for us, which now feels like such a—

Amazing, weird reversal. You still have the model?

It was, I think, a fine-tuned Davinci for Replika. We were the biggest customer in terms of API calls because we were the only chatbot available on the market that actually used generative AI.

It’s weird to think about it now, but back then, all the big companies were scared to put out generative AI products because Microsoft Tay happened, and it turned into a Nazi chatbot in literally an hour. Everyone else was too scared to put them out there. So we were sort of the only ones for so long, until OpenAI put out ChatGPT and kind of changed history with that, of course.

Before that, we literally owned all the keywords—AI chatbot, artificial intelligence—on every single platform. We also owned hundreds of AI domains, which I just let expire because I felt like no one wanted them. They’re not even on any of the big domain platforms.

And then recently, I went to see what .ai domains were still available that were just regular words, and the only words that were available were vomit.ai and Iraq.ai. I was telling myself, “My God, this is so upsetting.” Naming Wabi because of that was pretty hard.

You told us a story at dinner about the time that you’d spent in the OpenAI office. Maybe talk a little bit about that—when you guys were working out of there. What was it like? What was the energy like? What were the expectations of the team?

I think that—yeah. So when OpenAI started, it was kind of out of YC. It was YC Research.

They were doing a few different research groups: one on UBI, one on AI, and some other ones. Because we were a YC company, they let us come. They were very generous with their time, and they would let us come to Greg’s apartment, where they were headquartered, I would say, with a couple of other companies that were doing AI at YC, to ask questions, learn from them, and just talk—maybe exchange experiences about what we were building.

So we would come to that apartment and ask all the questions. Usually it was Ilya and Andrej and some other people, and that was absolutely incredible. We were, of course, absolutely starstruck and super happy to be there, just grateful for the opportunity.

Then they moved to their office, and we would go there as well. Very quickly, they stopped working on language models. We were very upset because we really wanted to continue going there, but they didn’t want to talk about language models because no one was really working on them.

That may have made us feel very strange because we were so set on continuing to believe in language models, but they completely moved away from them. They shifted to playing video games and all of these kinds of agents, rebuilding the worlds and agents in those worlds—reinforcement learning for that—and all these other interesting things.

I guess the only person who, in the beginning, continued to work on language models was Alec Radford. So we had an opportunity to ask him some questions. But, of course, it was wild to think about, because when we were going to OpenAI, they were superstars even back then. Still, it felt like such a small research group that was trying to do something interesting.

Of course, now to see it become one of the biggest companies in the world is absolutely wild.

Well, it’s interesting because Andrej said on the recent pod that the entire video game direction was an incorrect direction. It was an incorrect research direction, and they probably should have stuck with language.

So you guys were right.

Being right is not always—

You’ve got to be right but also execute.

I think, to a degree, we were really invested with Replika in building language models, but at some point we just required a lot more capital to do that. After working on that for so long, we just got really into revenue maximization at some point because we got a little bit scared. We were like, “Well, we didn’t have the balls to say, ‘Okay, well, we need $20 million and we’ll build that model,’” even though that was our discussion internally right after the Meena paper and so on. We just needed to get a lot more money.

If you think about it, with Replika, we only raised $1 million, and I guess it’s still all in the bank. We went for many, many years and built a big business.

But, of course, this is, in hindsight, a very interesting lesson: sometimes being very nimble and scrappy and very profit-oriented is great, but you can miss out on almost a generational chance. I’m not saying that even if we had bet on it, we would have built it. I would never compare myself to the geniuses who actually did it. We probably were not the right people anyway, but still, the lesson stands.

Sometimes you need to go big or go home.

And not having the balls to do that, especially in this current environment, I think you can suffer the consequences.

One of the things we say on the consumer team is that consumer behavior, especially new behavior, cannot be predicted. It can only be observed. I think you’re actually one of the few people for whom that is not true, because you seem to be able to predict consumer behavior. You’ve been so early to a number of these things, and every time we talk to you, you seem to have an eye not only on what’s going on today, but on what the next thing should be. How did you develop that sense? I’m sure a lot of people would be curious to know.

I don’t know if I have that sense. I only have a couple of ideas, but I really believe in them deeply, and then I go really deep down the rabbit hole. I start thinking about them.

I do think that I have a background in journalism. I pretty much grew up dreaming of being a journalist. My first job was when I was 12 years old, working at a newspaper. I was an investigative reporter for a while, and the one thing I loved about it was being able to go and talk to people and really try to get to know them and live their lives.

For that, you sort of have to have a lot of empathy, and you have to really try to understand people. You also need curiosity about people. Today, what I’m seeing with AI is that it’s being built by a very specific type of personality. Oftentimes, these are savants—brilliant geniuses, physicists, mathematicians—and they’re insane at building algorithms and doing math and making scientific breakthroughs, but they usually lack human empathy on the other side. That’s just kind of how it is.

Meanwhile, I’m the opposite. I’m a dumb-dumb when it comes to science. I come from a family of physicists, and they were always like, “Oh my God, why can’t you be smart and go study physics?” But I couldn’t. At the same time, I’m really interested in the human condition and what people are doing.

Just seeing my mom trying to understand how to copy and paste a prompt from Reddit was such a foreign idea to her. I realized that my mom is very savvy with computers. She’s always on her phone, and she’s always on her laptop, but somehow she can’t crack that. This is just not user-friendly enough. Understanding those concepts is what really let us come up with the idea for Wabi.

The same goes for Replika: just traveling around and talking to people and seeing how much loneliness there is, and how much people want to be able to tell someone what they’re going through, and how few people are able to listen. I think that realization—that maybe AI can’t talk well today, but it could listen—could be a groundbreaking thing for millions of people in the world. I think that’s what allows me to have a slightly different angle on the same problem.

If you wanted to speculate on the future of hardware, 5–10 years out, what is the interaction going to be like? How are we going to be interacting with these applications?

I do have a few ideas around hardware. I’m not a hardware person at all, but I’m a hardware user, like all of us, so I get to have opinions.

I think there’s a huge mind trap that exists among builders in the space, where they somehow think that voice is the main interface. It’s the best, ultimate interface. I think that’s because they’re somehow thinking about the movie “Her” all the time, but not in the right way and missing the whole point of the movie “Her.”

Voice was amazing in that movie because it was Scarlett Johansson constantly breathing heavily in his ear, and that totally worked. You didn’t even need to see anyone. In that case, that’s why it worked.

But if you really think about voice interfaces, they’re just so imperfect. You can’t use that device if you’re lying in bed with someone who’s sleeping. You can’t use it in a crowded space. You can’t use it at the office. You can’t even use it walking around. It’s a little bit weird.

All of a sudden, you’re betting everything on voice. A lot of people are trying to build voice-only devices. In my view, that’s completely wrong. It can be a fantastic extra way to interact with the computer, but every single Alexa right now—75% of them are being shipped with a screen.

Even in the proverbial voice use case, when I’m setting a timer while I’m cooking, I need to see the timer. I’m not going to be asking, “Hey, how much longer is left?” every second. This is just strange.

I think this is the biggest mistake, in my view: trying to ship these screenless devices. I love screens. I don’t think there’s any way with voice to solve for discovery or proactivity. The worst thing with voice and the iPhone is having it read out push notifications and text messages that come in. I’m like, “God, please shut it off.” It’s horrible. It’s also very hard to turn it off.

That’s terrible.

If you want productivity, you don’t want it to be read out loud in your ear, because this is just a very slow way of getting information into your brain.

I would never make a screenless device. In fact, I would make it very much a screen-first device. But I do believe that the AI device is not about a voice-driven thing. It’s more about building an AI-first operating system and having all the models run locally as well.

I think there’s a lot in that: building a truly AI-first smartphone—not today’s kind of CPU-driven hardware, but more the hardware of the future, where models can run locally and the operating system is super different from what it is today. There would be no fixed apps, and you’d be able to change and create software on the go for yourself, with a level of personalization that goes a lot deeper than what it is today.

I think there’s definitely space for a device like that right now. Today, AI is just an app on your phone. It should not be that way.

I think it's a great note to end on. Eugenia, thanks so much for coming on the podcast.