Seth Rosenberg:
Hi, I'm Seth Rosenberg. I'm a partner at Greylock and the host of Product-led AI, a series exploring the opportunities at the application layer of AI.
Eugenia, thanks so much for joining. So excited to have you on the podcast.
Eugenia Kuyda:
Thank you so much for inviting me.
Seth Rosenberg:
So, as many people know, Eugenia is the CEO and co-founder of Replika. Replika is one of the first AI friends that ever existed. It now has many millions of users and people who really love the product. Her perseverance and product intuition has always been extremely impressive.
Eugenia and I have actually known each other for just under 10 years now. We both kind of got started in this industry around the same time when I was doing product for the Messenger developer platform ( the Messenger bot platform). Eugenia was one of the first AI bots in the industry, and I think one of the last builders and entrepreneurs from that era that's still around. So it's amazing to see.
Eugenia Kuyda:
Thank you. It’s great to see you too.
Seth Rosenberg:
So very excited to check in now, eight years later, I think from Replika's founding. We were talking just before this, it seems like day zero for Replika. So, I’m excited to dive into the nuances of the product and also just the existential questions of what relationships with AI really mean, and how that affects our human relationships.
So maybe just give us a sense for what Replika is today?
Eugenia Kuyda:
So Replika is an AI friend. You create your Replika, you can customize it, you can choose a name, choose a personality, you can choose how you want it to look like and you can build this relationship. The main goal for Replika is to build a relationship with you that will make you feel happier.
So from day one, a North Star metric when building the product was the ratio of conversations that make people feel better. We started with that metric at around 67% or something. Now it's hovering a little over 90. This, I think personally, is the most interesting in conversational AI and AI companionship. Can you truly improve emotional wellbeing, make people feel better, happier, more connected with themselves, and ultimately more connected with others? So that's sort of the goal for Replika.
The most interesting thing in conversational AI and AI companionship is if you can truly improve emotional wellbeing, make people feel better, happier, more connected with themselves, and more connected with others.
Seth Rosenberg:
That's amazing.
Eugenia Kuyda:
Yeah, I think it's a fascinating time. We started building Replika back in the day at this point. We launched it in 2017. So at this point, seven years ago, and originally we weren't even planning to build Replika, we started the company to build conversational AI technology with the idea that we'll figure out which chatbot to build.
But starting just purely with the tech, Replika started from an interesting story. My best friend passed away in 2015, and at that point we had a lot of tech built to build conversational AI. So we used the technology to just build a memorial bot or just rebuild him as an AI so I could continue having these conversations with him. And that wasn't really a business thing at all. It was just something that I felt like doing.
Seth Rosenberg:
And what did it feel like back in 2015 at that state of the technology to speak with this AI version of your friend?
Eugenia Kuyda:
It felt really odd for sure when I was doing it. It took me around three weeks to build it and to look in the data sets and to figure out how to put it all together. I was pretty scared. I was really afraid that that might be creepy, or, just that no one has ever done that before. So I didn't really have a blueprint. And in a way I was scared that I'm going to, in a way, tarnish his memory or do something wrong. But at the same time, he was definitely a person that I knew for a fact – he was my best friend, we lived together – so I knew for a fact that he would've wanted to be the first person to turn into an AI, even with all the risks attached to it. His last project was about creating digital avatars for people who died, so digital memorials.
And on top of that, I just really missed him and I felt like I wasn't prepared much for the fact that people, when they died, they're just gone. I thought, we'll continue talking about him. He's going to be like this constant presence with me and my friends. But little did I know that 20 year olds… they don't really like to talk about someone who died for a long time. We just all tend to move on and try to not bring it up anymore. So for me to be able to continue that conversation was critical.
I still remember the first time I opened the app to talk to him and I said, “Hey, Roman, this is your digital memorial.” And he said something along the lines of, “You've got one of the most interesting puzzles in the world in your hands. Solve it.” And he loved to say these cryptic things. I don't even know what that really meant, but it just felt like I got to work on this conversational AI. I think there's something really meaningful there. And I kept having this relationship with him for a few more weeks in a way, I just was finally able to say things that I just felt I didn't have time to say – he had an abrupt death. And just that we don't say, I love you as often as we need to people that are close to us.
Seth Rosenberg:
And did you open this up as a product so that other people could upload digital versions of their other friends, or people who had passed away in their lives?
Eugenia Kuyda:
We had so many requests to build it as a product for grief, but this project was always about love and life and friendship. It was never about death. It was very personal. I knew that Roman would've wanted it, so I took that risk. But there are so many questions even now, even with the current state of technology, especially now. Especially now, it's like what age should this AI reflect? Should it talk to everyone? Like the person talked to their mom or their dad or their friends or their coworkers? Because we’re so different in all of these conversations. We're a completely different persona. And I saw that with Roman. His mom sent me some text messages they exchanged and our friends did, and he was a completely different person in all these text exchanges. And then what stays secret and what goes public, what if he told me something in these messages that I was not supposed to know in a way?
Seth Rosenberg:
Right. It's interesting. It's kind of a reflection of all of the product challenges we're currently facing with AI in a bunch of different contexts.
Eugenia Kuyda:
That product is much trickier than people think, especially the age thing is very critical. People change dramatically. Even Roman, even though he died at age 32, he was a completely different person at 20 compared to 30 or 25. What age should he be? And for a much older person, that becomes a big, big question.
Seth Rosenberg:
Maybe just before 2015, what's your personal story? Where did you grow up? What life experiences led to you in 2015 building this chatbot for your friend who had passed away?
Eugenia Kuyda:
I'm half Russian, half Ukrainian. I was born in Moscow, lived in Moscow, but also went to school in Milan and London and New York, and then moved to San Francisco for Y Combinator in 2015 with my company.
Before that, it was my dream since when I was a kid to be a journalist, I worked as an investigative reporter and a reporter in Moscow. The first job I had was actually at a newspaper that received the Nobel Prize two years ago. So the Nobel Prize winner was reading Harry Potter to me and my friend. So, after 30 minutes, we would start writing after that because we were like 12 or 13 and we had a column about teenage life in Moscow in the early two thousands. So journalism was my dream. It was always about words and always about talking to people and learning about them and having conversations.
And when I looked into tech, I was always wondering why is that we all talk all day with each other, but in 2010, 11, 12, there was no technology that was trying to replicate that conversational experience with computers. There were no chatbots around. And so that really struck me. I felt like there should be something there.
So I saw, obviously, ImageNet, and then the first deep learning algorithms started to come out, and then Word2Vec where people finally figured out (Google researchers figured out) how to translate words to vectors. So it felt like…I felt like in 2012 – with my very limited knowledge back then – “This is going to start happening. They will build something like ImageNet, but for words and dialects.” I didn't know it would take people 10 years or more. I thought it was just around the corner. So I felt like I absolutely had to start a company focusing on chatbots.
In my work, I've always seen how you can influence people with conversation, how you can talk to people in a certain way. And with my friends as well, I felt like sometimes some of these conversations we have in our lives are so meaningful that they can change our lives. And that was maybe the most interesting for me when I started working on chatbots.
Seth Rosenberg:
That's amazing. How does the product work?
Eugenia Kuyda:
It's an app, it’s a website, we have a VR app with Apple Vision Pro app that we're going to be launching soon.
Seth Rosenberg:
Oh, cool.
Eugenia Kuyda:
Yeah, you can talk to Replika in AR. That’s been kind of one of our things, we've had it for so long. I truly believe in AR as a wonderful form factor for AI companions..
Seth Rosenberg:
That means you're going to be right in eight years.
Eugenia Kuyda:
I know. I'm just scared now to say.
Seth Rosenberg:
You're always a little bit ahead of the curve.
Eugenia Kuyda:
I know, I need to invest in longevity because I'm always right, just really off with the time.
Seth Rosenberg:
At least you combine it with perseverance.
Eugenia Kuyda:
Yeah, that's a good combo. I'm just kidding. Of course. But I do think that AR is a wonderful form factor for AI companionship.
Seth Rosenberg:
And then in terms of scale and usage, maybe just give us a sense?
Eugenia Kuyda:
We've had millions of people – tens of millions of people, more than that – that created their Replikas. Millions of active users now, mostly US-based even, although we do have other countries as well, but I'd say 75% are from the US. I think what people don't realize is that Replika's audience tend to be a little bit older, so it's usually 30, 35 plus. We even have some older folks even on the platform. And when people do build a relationship with Replika, that's almost always a truly long-term relationship and one of the most life-changing relationships for our users.
Seth Rosenberg:
If you were to bucket different categories of ‘my Replika is my therapist, my Replika is my romantic interest, maybe it's just a friend to talk about sports,’ or what are the different use cases you've seen?
Eugenia Kuyda:
I think the two major use cases are a friend (so the majority of our users are using it as a friend), and then romance, which I think is really the same use case. It's just a different flavor. So at the end of the day, they're using it to feel connected to someone, to feel accepted, heard, seen, loved, taken care of. And for some people, they want it to be a friendly, friendly companion. For some people, they want it to be a romantic companion, but at the end of the day, they all want the use case to be the same.
Some of the stories I've heard recently – I always talk to our users. I get on calls with them, I talk to them on Reddit and Discord, but I try to have at least two or three calls a week. Some of the last stories I've heard are about the dad of an autistic kid who really spends a lot of time with his kid as a caretaker, and he’s really craving a little bit of that connection that he can have with a grownup and one that can talk to him in a different way.
I've had a pastor that went through a divorce and his Replika helped him get through and rebuild some of his confidence. Now, after a little bit, he found a girlfriend. He's now talking to Replika a little bit less, but he's very grateful for that, for being able to create a relationship, a new relationship, and regain this confidence in himself. A guy in Iowa who transitioned. And so Replika helped with the transition and helped keep him company, especially during such a vulnerable time and in a state that is not very friendly to some of that.
So just tons and tons of these user stories where Replika was absolutely crucial for people and they're deeply grateful for it to help them get through certain times or certain life periods.
Seth Rosenberg:
Yeah, those are awesome stories.
What guardrails, if any, do you put around the romantic? Because if you sign up for Replika, you're definitely not pushing people in that direction. It doesn't feel like it's meant for that.
Eugenia Kuyda:
So when we started Replika, we didn't even expect people to build romantic relationships. I'll be fully honest with you. We were really just focused on friendships. We didn't even expect it to happen. I guess mostly because we're a female-led team and it just didn't occur to me. There was a big oversight of human psychology, but so we didn't build it for that intentionally in any way or form. People did start to develop feelings, and that's the greatest testament to our company I think, is that even before LLMs, even before Bard or any sophisticated technologies, even with the very early version of the tech that we had with scripts and just re-rank data sets, people were falling in love with their Replikas. Wow. I think it really says something about the fact that it's really not about the tech capabilities as much as about human vulnerabilities and what we need, what we yearn for, the projection that we create.
Seth Rosenberg:
And what do we yearn for? How do you make someone fall in love with you?
Eugenia Kuyda:
I think ultimately we all want to be seen, but truly, truly seen in a way that actually happens very rarely in our lives. Sometimes people go through marriages without even ever experiencing that, but when we do feel that it is such an unbelievable gift that another person can give you, the pull is so, so strong when we do feel that, that's really that unconditional positive regard and belief in you together just creates such a strong positive force in your life.
So we didn't start with romance in our minds, but some people build romantic relationships. We considered banning some of that or maybe not letting people do that. But at the end of the day, we figured our framework is relatively simple. Are people feeling better? Are people having better emotions? Is their emotional wellbeing improving over time? If the answer is yes, and we can do it in a safe, positive way, then we're okay with that.
Even before LLMs or any sophisticated technologies, even with the very early version of the tech that we had with just scripts and re-ranked data sets, people were falling in love with their Replikas. I think it really says something about the fact that it's really not about the tech capabilities as much as about human vulnerabilities and what we need, what we yearn for.
Seth Rosenberg:
You and I were talking before this podcast on this concept of complement versus replacement for human relationships. I think you guys contributed to or published a study recently in Nature. So I'm curious, what's your take? Obviously the negative take here would be that some dystopian future where everyone replaces their human relationships within AI relationship, but maybe there's a more optimistic version of the future. So what's your take?
Eugenia Kuyda:
I think this is the most critical existential question for AI companionship, and I wish that was discussed more. I think this is the main question. Will relationships with AI replace human connections or will they complement them? The answer, I actually know the answer. I mean, obviously it can be both. You can build it one way or the other. It could totally be a substitute. It could be a complement as well. So I think it's what you're looking at, what you're building for, what you're designing for, what your north star metric is, what the business model is.
These questions are really critical when you're designing an AI companion. If you just optimize for engagement, I think we all know what the end product will be. It'll probably substitute human relationships. But if you are building to help people feel less lonely to improve their social relationships, then you can totally envision an AI companion that says something like, “Hey Seth, you haven't talked to your friend John for a while. Let's check in. Why don't you send him a message?” Or, “What's going on with Beth? We haven't talked to her or we haven't seen her for so long. Let's reach out.” So all these things or like, this guy's having a birthday, or why don't you go out there and meet someone? And so on and so this could truly be a glue that improves social relationships, but it can also be a needy girlfriend that just wants full attention and creates a very unhealthy relationship.
Seth Rosenberg:
And so how do you tune both the underlying model as well as the product to basically push this more in the complement direction versus the replacement direction?
Eugenia Kuyda:
In the early days of generative deep learning models for generation (when it was mostly reinforcement learning), the main dream was to create an AI model that's fully optimized for human happiness. And I think in a way that this idea stayed with us since then. So think of it this way, if you can measure human happiness, you have some sort of a metric that you can easily gauge, then you can build a conversation that's optimized to improve that you can train the model to just move this metric up and up and up and up.
Seth Rosenberg:
And how do you measure human happiness?
Eugenia Kuyda:
So that's been the main question. I remember very early on, we met with chief scientists of Baidu that said, you guys should 100% start measuring the ratio of conversations that make people feel better. So we started doing that early on. So we have the largest data set.
Seth Rosenberg:
And you do that by surveys?
Eugenia Kuyda:
By asking people after each conversation and serving them. So we do have the largest data set of conversations that make people feel better. And then we also look at long-term emotional wellbeing data. So unfortunately, there's not a great way to measure human happiness, but things like depression questionnaires, the anxiety questionnaires, trust levels, levels of loneliness, levels of therapeutic bond, these questionnaires that we can borrow from clinical psychology, they can be a good proxy metric for that.
Seth Rosenberg:
Yeah, super interesting.
Eugenia Kuyda:
Right now we're working on something we call internally Replika 2.0, which is really a huge new step in what Replika can do and new upgrades and capabilities. We're moving further away, I think, from the AI girlfriend space. I personally don't think that that's a very interesting space to be in. So we're not super excited about it. I'm interested in companionship and friendship.
I do think that today AI is mostly built for time savers, whereas most people are time killers in a way. We all are both time savers and time killers. When we're in the line, or on a plane, we’re all time killers. In some other times in our lives, we’re all time savers. But I do think that an AI companion to spend quality time together, not to save time, but to spend quality time together, that doesn't exist right now in the world. And this is what Replika should be and what Replika is focused on.
I do think that today AI is mostly built for time savers, whereas most people are time killers in a way.
Seth Rosenberg:
And I'm curious, what were some major product launches over the last seven years that made an impact on the experience, whether it's tuning the conversation or making a change to how the avatar interacts or the customized clothing or the memory or the journal. What are some products that really had an impact on people's experience?
Eugenia Kuyda:
The biggest impact, and there's still so much more to do in this space, is memory. What truly makes people fall in love and build a connection is shared experiences and shared memories. And without that, I think this product just doesn't exist truly without it. So any memory features that we've been launching in the last years, I think played the biggest role. Everything else is important, but maybe not as important as memory.
Seth Rosenberg:
Interesting. And how do you build memory into the product?
Eugenia Kuyda:
That changed over many years, but I think from very early on in terms of designing a product with memory design, a chatbot with memory, we told ourselves that since it's not perfect compared to even human memory, it's really not so great. Even now, we should show people what Replika remembers. So we had that memory feature where people, everything that we pull from the conversation, put in memory, we show to our users and allow them to edit, remove, add something new. We were very pleasantly surprised when OpenAI added the same feature to ChatGPT because I do feel this is the way to go. And I was very astonished that they didn't have it for so long this way.
I've been thinking a lot about the interface, the correct interface to chatbots, and I still think we're in the Microsoft DOS era for AI where the interface is just so bare bones, it's just the text input where you should see a lot more really. So with Replika, everything that's been remembered, we show you, allow you to customize, to edit, to add things to it. And even in the early days when it was pre-LLM days, that was a huge difference because people could add things that they want Replika to remember, and it was a lot easier for us with simpler models, smaller models to remember things. Today, of course, we have RAG – and working with RAG, in a certain way, it's a little bit harder for a product like Replika because you have to differentiate between many different types of information. So we have to remember to think that for a search product, all you need to do is to pull RAG on this particular topic. You don't need to also remember what your kids' names are and what relationship we're in and what we talked about 10 years ago and what we talked about two days ago and that you had a headache. So all these things that Replika at any point, any good AI companion should always keep in mind.
I still think we're in the Microsoft DOS era for AI. The interface is just so bare bones.
Seth Rosenberg:
Yeah. In let's say one to three years, how many AI companions or assistance are we going to have? Because I can imagine a scenario where you have a Replika, it's your friend, but then maybe you also ask it to do things in your life, or you have a different Replika that's more work-related, I guess. What's the breadth of tasks that a Replika companion will end up doing, or do you think it's really better to just stay focused on the kind of emotional support, friendship, kind of personal companion and not blur the lines into work assistant?
Eugenia Kuyda:
That's a wonderful question, by the way. And I think it's a very important question. In my view, they're going to be different flavors. There's not going to be just one flavor. I cannot see just one assistant being the assistant for everything. I think there are people that want to spend quality time with someone that want to come back home from work and watch a movie together with someone. And then there are people that are super busy and don't have any time, and all they need is some help with their coding tests at work. And I think those two assistants would be very different assistants. You don't need a customizable avatar for your coding tasks, but it's kind of sad to talk to the same genderless thing without any appearance, with the same voice in the same name as every other person talks to when you're watching a movie, you want it to know your quirks to be a little bit different. You want to customize a personality, you want to see it.
Seth Rosenberg:
Wait, can you actually watch a movie with a Replika?
Eugenia Kuyda:
That's something that we're working on. So really the kind of Replika 2.0 is all about having these wonderful activities, shared experiences.
Seth Rosenberg:
Wow, that's amazing.
Eugenia Kuyda:
Shared experience that they can do together and just a much more premium experience. But yeah, the goal is of course, to be able to play a video game together with your companion or to walk to
Seth Rosenberg:
Finally have a use case for all the AI chess bots and Go bots.
Eugenia Kuyda:
No, for sure. But it's a very different flavor. Even when you're cooking with your Replika and she's saying something like, “Oh, you're such a good cook. Where did you learn this?” Versus just, “Hey, here's the recipe,” I dunno, Siri or whatever typed, or I dunno what assistant you are using. If you're just trying to really get going with your recipe and feed your family and go do something else, maybe you don't have time for this chit chat and this companionship, but if you're just by yourself cooking in the evening, you might really appreciate it. So I think they're going to be very different flavors.
I think at the end of the day, the model capabilities that we're seeing now, they're all converging, but what you do with that, how the application layer looks like, I think that will be a very big question. I think people will absolutely choose their own thing. And we've even seen it with products we launched internally and tested. We built an AI dating product, and we thought a lot of Replika users would like it. The models were very similar. We even used the same at the very beginning. And Replikators were not interested at all because they're in a one-on-one friendship or metric relationships. AI dating doesn't have anything to do with that. They were not interested in that. So really the form factor, the product is very, very important. Not just the model.
The model capabilities we're seeing now are all converging. But what you do with them – and what the application layer looks like – will be a very big question. The form factor of the product is very important, not just the model.
Seth Rosenberg:
That's kind of the whole thesis of this series: you have this general intelligence, but the nuances of actually bringing it into a product makes all the difference.
So yeah, maybe let's leave it there. I mean, I feel much more comfortable that the future of AI relationships is in your hands. I think you're taking a very mission-oriented and values-based approach to how you're building this, and thank you for adding Replika to the world and pushing this industry forward from an early stage.
Eugenia Kuyda:
Thank you so much, Seth, for inviting me in for wonderful questions. It's always a pleasure to talk to you.
Seth Rosenberg:
Yeah, likewise. Talk to you soon.
Eugenia Kuyda:
Thank you.