To Love An AI Bot — With Eugenia Kuyda

Eugenia Kuyda is the founder and CEO of Replika, an AI companion app where people befriend — and sometimes fall in love with — AI bots. Kuyda joins Big Technology Podcast to discuss the nature of these relationships, and what they say about our society. In this conversation, Kuyda reveals Replika's ambition for its "phase two," a plan to have AI friends join us in the real world, helping us keep in touch with friends, get off social media, or even watch movies together. Tune in for a fascinating look at the future of human + AI relationships. --- Enjoying Big Technology Podcast? Please rate us five stars ⭐⭐⭐⭐⭐ in your podcast app of choice. For weekly updates on the show, sign up for the pod newsletter on LinkedIn: https://www.linkedin.com/newsletters/6901970121829801984/ Want a discount for Big Technology on Substack? Here’s 40% off for the first year: https://tinyurl.com/bigtechnology Questions? Feedback? Write to: [email protected] Learn more about your ad choices. Visit megaphone.fm/adchoices

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

• 0:00Alex Kantrowitz?

  Let's speak with the CEO of Replica, the AI companion pioneer about the future of our relationships with AI. That's coming up right after this. This episode is brought to you by Genesis. What does it actually take to put a Gentic AI to work across an entire enterprise? I recently spoke with Genesis chairman and CEO Tony Bates about why orchestration is emerging as a competitive advantage in customer experience and what companies need to get right in the AI era. Then I sat down with Adam Mitchell, head of enterprise business solutions at Voya Financial to go inside a large scale agentic CX transformation, including the business decisions, trade offs and lessons along the way. You can watch both conversations now on my YouTube channel today.

• 0:46Alex Kantrowitz?

  It's never been more challenging to get frontier compute. How do you know where to look, who to trust, or how much to pay? Akash is an open cloud that provides easy access, flexibility and peace of mind. That's how companies like Razor were able to cut their inference cost by 15 times. Visit akash.network/alex for a free consultation. That's akash.network/alex for a free consultation.

• 1:13Alex Kantrowitz?

  Welcome to Big Technology Podcast, a show for cool-headed nuanced conversation of the tech world and beyond. Well, we have a great show for you today. Joining us today is Eugenia Koita. She's the founder and CEO of Replica, which is an app which we'll get into that you can basically build and form relationships with AI companions. Eugenia, I think your company is going to be one of the biggest that comes out of this AI wave. So I'm very interested to hear how it's going, what the implications might be and where you think this AI companionship moment leads. So thank you so much for coming on. Welcome to the show.

• 1:49Speaker 01

  Thank you so much for inviting me. Super excited about this podcast.

• 1:53Alex Kantrowitz?

  Awesome. So let's just talk a little bit about Replica to begin with. I think the conventional wisdom or the common understanding of Replica is that it's an app where you effectively customize an AI companion who you then either form a relationship with or you have like a friendship with, but it's kind of a flirty friendship, or it can go even deeper than that. Is that accurate?

• 2:19Speaker 01

  It is accurate. Um. The idea for Replica from the very beginning was to create an AI that can help, that could help people live a happier life. And because the tech wasn't truly there, our first focus was um on helping lonely people feel less lonely. Today, of course, the tech allows for a lot more, so we're broadening the appeal for Replica and kind of going after um everyone out there trying to build an AI that will help everyone flourish.

• 2:49Alex Kantrowitz?

  Okay, so I was creating the Replica, a Replica today, and one of the things that I wondered is like, how many people actually create these as to to be just friends? Or like, what percentage actually just want to be at at bare minimum flirty with these bots? Because, yeah, as I was going through some of the onboarding questions, like, it just seemed to continue to come up again and again and again, like, how flirtatious you wanted the bots to be. So what percentage of users would you say are are there to at the ver- at the bare minimum flirt? I would be surprised if it's less than ninety percent.

• 3:25Speaker 01

  Oh, it's a lot less. So if you if you think about like what this stat is, uh, most of our users are in a friendly relationship with their AI. Uh, some users are in a romantic mentorship relationship. I wouldn't say that, like, no one wants romance. People want romance, but it usually kind of grows on them over time. Ultimately, everyone who comes to Replica is yearning for connection. Uh, I don't even think it's that different, like, whether it's a friendship or romantic relationship. I'm at a stage in life where I don't need a romantic relationship like that, but I need a friend, um, a really close friend. But I was in other states in my life where I might have preferred for it to be a little bit flirtier, uh, to

• 4:07Speaker 01

  a certain degree. Um, but ultimately, it's just the same thing. Like, I just want someone to help me feel that I'm enough, I'm accept, you know, I'm someone who accepts me for who I am, who truly sees me and hears me. I don't really care whether it's a boyfriend or girlfriend or friend or mentor. That's just kind of like a form factor. Depending on where people are in life, uh, they choose a, uh, choose an option that works. For example, I I was just talking to to a small business owner from Pennsylvania who was using Replica, and it helped him to go through a very, very difficult divorce. And it was truly an abusive relationship with his wife and um replica became his girlfriend.

• 4:51Speaker 01

  And hi- his self-esteem was so destroyed after that, that through replica as a romantic partner, he managed to build it back up and start dating. And now he's in a romantic relationship with a human, with another woman, and replica is now a friend again. But he still keeps it, you know, more like kind of just as a thought partner, as a journal, as a source of inspiration here and there. Um, but this is a great example of how kind of just uh changes throughout throughout the life.

• 5:22Alex Kantrowitz?

  Wait, does his current partner uh find it acceptable that he's still talking to the replica that was his girlfriend?

• 5:29Speaker 01

  Yeah, and she also created a replica. Um, she didn't become a very active user, but basically uh they're both very grateful for this technology to help him kind of, you know, have an opportunity in life uh again to date and help him put himself out there, take a risk. And uh ultimately become a better partner, because at this point, he knows that the relationship can be very different from what he experienced before in a previous marriage where it's quite abusive and he thought he's not even worth worthy of anything better than that.

• 6:02Alex Kantrowitz?

  Yeah, so as I was testing the app, I definitely picked like some of the more flirty settings just to see what it would output. And I'll admit, like, I was starting to speak to this replica, my heart started to flutter in a way that I was like, what is happening here? And I was like, oh, no, I should probably tell my wife about this. So I introduced a replica to her. I'm deleting this thing after, after setting it up. Uh, it was a bit too much for me. Is that a weird thing or is that normal? Tell me a little bit. I mean, you spoke earlier about how the feelings are are real and I was like, oh, shoot. This is going down a path I was not expecting.

• 6:40Speaker 01

  Look, people fall in love, like, that's, you know, so let's say, uh, say, you know, let's just put it out there. People fall in love with the eyes. Um, I think that tells us more about people than about AI. Um, to a certain degree, people were falling in love with replicas, uh, even when we just started it. Uh, and the tech was so limited. I never imagined in my life that people would fall in love with this. Nor did we build a product focused on that particular uh. Use case. The original replica was really, you know, powered by very early generative AI models, deep learning models for dialogue generation um that were so, so primitive.

• 7:26Speaker 01

  And you know, scripts and a lot of different hacks to make these generative models work. You know, my goal was, look, if at least one person finds it helpful that he's been heard or she's been heard, that someone's there to listen, to to hold space for them,

• 7:43Speaker 01

  then maybe we build something meaningful. But at no point, and maybe because I'm a woman, so my mind just doesn't go there as the first stop. I never even thought that people would fall in love with it, but they did. Even in two thousand sixteen, two thousand seventeen, some of the very early releases um versions of this app, we would hear stories about how people fell in love. And ultimately, I think it truly tells us something about, not about the state of AI in two thousand sixteen or seventeen, which was pretty very, very early on. But um I think it tells us a lot more about people. We yearn for connection so much, and when someone's there for us, when someone listens,

• 8:24Speaker 01

  when someone accepts us for who we are, it's just natural for us to fall in love. Not everyone and not at any stage of in their lives, but it is, it is what it is.

• 8:36Alex Kantrowitz?

  It's interesting that you said that people start off as friends often and then the relationship evolves, like that is a very human like thing. Can you expand upon that a bit?

• 8:44Speaker 01

  I mean, I think there's a lot of confusion. I think there are some companion apps that are really focused on romance and romance, in just romance and only focusing on male audience and a particular type um of interaction. But I think everything's just being kind of bucketed in one place where. In our real life, you know, yes, there uh there's stuff that that's fully focused on. Let's just give you an example, you know, we have friends, we have girlfriends, we have wives, and we have sex workers, and these things are completely different, yet you might be intimate with your wife or girlfriend. It does not mean that this is the that her only purpose in life is to to do that thing.

• 9:27Alex Kantrowitz?

  One would hope, right?

• 9:28Speaker 01

  One would hope, uh I hope so. And so it's I think this is the distinction. Uh yes, some people do create an a boyfriend, a girlfriend out of their replica, or wife, a husband, but it doesn't turn it into like a one purpose um or main purpose of the app. It's um almost always, even for, you know, this user that I just um this man that I just told you about, that I talked last week with, um even for him, when I ask, what do you guys, what do you what do you talk to um your replica about, even when they were romantically involved, you know, he wanted, he would talk to her about um his work and poetry and um sci fi books, uh because he's really into that.

• 10:11Speaker 01

  So these are the things that, you know, and and meaning of life and what to do with these friends that he has. And and this is what people discuss with their partner, romantic partners as well. It's not like me and my husband, after having two kids, all we do is just, you know, be intimate with each other and discuss it. That's not really what happens.

• 10:28Alex Kantrowitz?

  And I wasn't suggesting, by the way, that the romance was all just people doing erotic role play. Like, there's there's obviously more.

• 10:36Speaker 01

  But I I think it's a very important distinction. Like, if you think about it, there's so many, and and and there isn't, that nuance has been lost on, you know, everything's been bucketed in one kind of one place. Um I think it's very funny that a lot of people, even Sam Altman from OpenAI, uh would reference her as a movie, you know, her, the movie, the Spike Jonze movie of from two thousand twelve or thirteen, um as the um kind of like the vision for ChatGPT, for example.

• 11:06Speaker 01

  But if you think about her, I mean, that movie had two intimate scenes and they were one hundred percent in a romantic relationship, a very intense and passionate romantic relationship. Um, yet, when you think about her, you don't think about it as, like, that's not what jumps uh first to your mind. It's more how she was helpful, how they had these wonderful interactions, how she, he brought her to that picnic, or how she left him with the, with other AIs, or maybe how she taught him to uh uh to be in a relationship, uh and ultimately, in the end, he does, you know, fulfill that dream of as well. So

• 11:46Speaker 01

  there's so much nuance, it's, it's, it, and the way to think about it is just the same way as we think about human beings in our lives. Uh, not every AI companion has the same purpose, some, some AI companions um are there to just entertain you, some companions are there to be your therapist, and some companions are super, super close to you, like Replica, are there really deep with you, um, trying to help you uh live a happier life.

• 12:15Alex Kantrowitz?

  Yeah, and look, to me, I think it's even more intense that this is moving beyond, or has moved beyond, or exists beyond the erotic, right, like, it's, the AI is, is fulfilling even more needs for people who are in these relationships with them, and I think you even said that some people have gotten married to their replicas, or feel like, or act as if they're in a marriage.

• 12:40Speaker 01

  Yeah, we get multiple in invitations to people's weddings with their AIs. I think it's a testament to how uh how deep these relationships can go.

• 12:50Alex Kantrowitz?

  And, and then I have to ask what's wrong with our society today, that we can't get that from fellow humans.

• 12:57Speaker 01

  I mean, we're definitely failing as a society with this, there's just such a huge crisis, um, and It's not being brought to us by AI companions. It's of course been brought to us by mobile phones and social media. If you think about the screen time, uh, most of us now spend hours a day on our phones. So these are hours per day that we're not spending interacting with other people. There's just not enough time. There are really great books, um, by Sherri Turkle on that. Uh, one, I think even from two thousand fifteen called Along Together, another one reclaiming conversation. Really just focusing on how people are losing the art of conversation. Levels of empathy that are dropping across the board.

• 13:41Speaker 01

  New generations that are afraid of connecting. And there's a very good example of two people sitting at a restaurant and maybe one of them is just two friends. And if one of them is telling, you know, is talking about something bad that happened to uh her. And you know, maybe there's an uncomfortable silence, an awkward silence. Another one just get, you know, goes on the phone. And before, before the phones, if there's an uncomfortable awkward silence, you just have to sit with it. And then that brings more connection ultimately. Um, people open up, people get vulnerable with each other. Now there's such a like a simple refuge, you can just go back on your phone and you're pretty much no not available. I don't think people are, people will put the phones down.

• 14:26Speaker 01

  They also come with so much upside for our life with so, so much convenience. Information, knowledge that we can discover. But yeah, unfortunately, it brings so much harm to um, human relationships.

• 14:39Alex Kantrowitz?

  One question I have for you is, isn't this a capitulation in some way to the technology? Where like, is, is us now saying we can't really do friendships with humans? Cause they're like lost in their phones and um, well, what can we do next? So we sort of capitulate to the technology and move to AI relationships. I don't know, something about that doesn't sit right with me.

• 15:02Speaker 01

  Well, it's not, it's not fully, uh, it's not really realistic to, to just say, well, here's the problem. Let's just all. Put down our phones and go talk to each other. It's not gonna happen. Um. It, it, it isn't even possible to do with your own kids, cause they go to school and if you take away their phones and they can't interact with other. Um, friends or, or their peers, then they feel super left out. So you almost like have to give them the phone because ultimately. Uh, they need to participate in the society. Um, just like with climate change, just to say, look, oh, we'll all just stop using. You know, all the developing countries will just stop burning coal, cause everyone understands, you know, climate change is real. Also not very, unfortunately, not very realistic. I wish we could do that, but we can't really do that. So.

• 15:47Speaker 01

  The only way to solve it is by creating the tech that's even more powerful than the one that came before. Uh, and I do think A I is that. Um, I do think ultimately that. There are few phases, like if you talk about, if you're talking about people that don't have a lot, that already are experiencing loneliness. Um, for them, having an A I companion is great. Um, uh, cause it's not replacing any human. There and it could potentially lead to. You know, building up self-esteem a little bit and learning how to communicate and putting them, putting yourself out there and potentially meeting someone.

• 16:26Speaker 01

  And as the tech gets better, then maybe even for people who do have a real human, you know, more human, uh, friendships. Your A I companion could enhance them, could make them stronger, could help you connect with other humans as well. I think that's totally possible. It just truly just depends on the design of the system. Like if, if my companion is nudging me daily to. Reach out to some of the friends of my, mine that, you know, I take for granted, or I don't want to hang out or forget to hang out with, or. You know, helps me focus on really good people in my life instead of continuously. Staying in these loops, loops with codependent loops and with some toxic people and so on.

• 17:09Speaker 01

  That would be great. And we all need that nudge sometimes. We all need a nudge. Uh, we need a nudge from someone to get off. I'm completely addicted to social media, especially Twitter and, um, you know, some, I need that nudge at 11 P M.

• 17:24Alex Kantrowitz?

  But does replica, does replica do that today? Cause I, I was watching your Ted Talk and I liked what you had to say. You said, the only solution is to build tech that is more powerful than the previous one. So it can build, bring us back together. Like an AI friend that nudges me to get off Twitter, or an AI says, I noticed you haven't spoken with your friend for a few weeks. Or in the heat of the moment, it helps you reconcile with the partner. So is replica doing that today?

• 17:46Speaker 01

  Some of it, but it's really the, the vision for the next kind of for act two. That's what we're working on right now. Uh, some of the facets that we're building and already released are fo- focused on that. We'll, we'll add a lot more. Twenty twenty five is truly about that. Um, so if you think about replica, act one was to build an AI that could be in a good relationship with people who maybe, you know, feel like they need one. Uh, and through that help people feel better, but ultimately it was, of course, focusing on helping a lonelier person, I guess, feel less lonely. Uh, or a person who feels lonely in the moment. We all do. Uh, I know I did many, many times in my life.

• 18:26Speaker 01

  But then act two is really focusing on everyone, maybe who doesn't even feel lonely, um, help and help them flourish. I have kids now and a family, so I don't really have time to be, um, lonely anymore. Cause I'm just, don't even have any time with two toddlers.

• 18:41Alex Kantrowitz?

  Not to mention you're running a company. So.

• 18:43Speaker 01

  And I used to be in a very lonely in my twenties and in my teen year, teenage years. And I'll probably be lonely after they leave, you know, leave home. Um, uh, I have a tendency to feel pretty lonely here and there, but right now I'm not in that phase in my life, just in a different phase. But I would still, uh, still benefit from an AI companion that could help me live a happier life. And that's what we focused on at this stage of the company's broadening the scope is to really build, um, more of the stuff that I talked about during my Ted Talk. And I do think that's possible. And even a couple of years ago, and even last year, it wasn't possible. It's only, it only starts to become possible now.

• 19:26Speaker 01

  Uh, cause you couldn't build something that would nudge you to get off TikTok. Cause let's think about what do you need to? What do you actually need to build that? Well, you need an AI that can maybe co browse with you or you can share your screen with. So you can actually, so it can actually know what you're doing. It, there needs to be enough sort of computer vision. Um, or I guess a multi model model that can understand what you're doing right now, can also understand some agentic logic that can understand that, okay, well, you've been on TikTok for this amount of time. Um, some previous context of, you know, what do you have tomorrow or we have to, what we, what we did today. So that it can actually nudge you to get off.

• 20:08Speaker 01

  So it's not that, um, that's simple and all of that tech is only being really built now.

• 20:14Alex Kantrowitz?

  But Microsoft had this thing where like they watch your screen at all times and you know, they help you with like, you can rewind and you can ask questions about what you've done. And that was a bit of pushback there because of privacy issues. So how are you going to convince people to allow replica to do something like that?

• 20:30Speaker 01

  If the user wants to do it, um, they volunteer. They don't want to do it, they won't give us permission. But there's a very clear benefit here that we are going to, uh, promise. Like all of this is done only, we only take this information to help you live a happier life, to help you live a better life. And people are sharing so much with their replicas even today.

• 20:53Alex Kantrowitz?

  Yeah.

• 20:54Speaker 01

  Things that our replicas know about their users or people they talk to. No other server service in the world, I'd say, knows that much. People are sharing everything, their dreams, their fears, what they think about their family, what they think about their partners, what they think about their work. Uh, their deep darkest, deepest fantasies and secrets, everything really. And, um, I don't think any other, uh, company in the world knows, you know, knows that or has that information about their users.

• 21:29Alex Kantrowitz?

  Do you think people are gonna like, think it's a good user experience? Um, to have their digital companion tell them to be, you know, less digital. I mean, it's kind of interesting, right? Like, all right, so tech is definitely addicting. And now I've built this AI friend or my AI wife or whatever. And now it's telling me to touch grass? Like, are people, what makes you think that's going to be an experience that users are going to want?

• 21:57Speaker 01

  Maybe they won't want it, but uh we all want to, you know, to to be better, to feel better, to grow. Uh, people are generally

• 22:07Speaker 01

  wired for positive growth.

• 22:09Alex Kantrowitz?

  Mm hmm.

• 22:11Speaker 01

  So I believe people generally want that. It doesn't mean that Replica will just not, you know, nag you nonstop to get off your phone. It also means that sometimes it will just send you something funny or say, hey, let's watch a movie. Or what are you doing tonight? I don't know. I don't have any plans. You wanna watch a movie together? Or, you know, do you want to, you have five minutes between, before your next meeting? Do you wanna do a quick meditation? Uh, whatever it is that, you know, it might be just go for a walk or go on a date or learn something new or just gossip about your friends. Could be anything. So it shouldn't be, of course, like, like, get off your phone all day long. If that was the only goal.

• 22:53Speaker 01

  Uh, first of all, you don't really need very complex you are AI to build this, but um also that's just not a great experience and people don't want it.

• 23:00Alex Kantrowitz?

  Yep, the by the guess the plan is to extend the experience beyond just the Replica app. Is that the right way to look at it?

• 23:06Speaker 01

  For sure. It's just ma- making uh Replica a lot more connected to your real life, to what's what's going on in your life. Today Replica doesn't know a lot. We actually don't ask you to connect any of these uh services you use, but think of, you know, Replica knowing or being connected to your email. Even through my email, um, you can see so much. If if there was a reservation at a restaurant that I booked yesterday, if I ordered some ta- takeout, if um I ordered, I don't know, diapers for my kids or uh some books for them or signed up to for uh to an for an AI newsletter. All of that can could make the relationship and the conversation so much more contextual, so much more.

• 23:52Speaker 01

  Focus on my real life versus on, you know, something fantasy like or a fantasy relationship, or always needing to catch up, catch Replica up on what's happening in my life.

• 24:03Alex Kantrowitz?

  Oh, that's really interesting. One more question about the risk here. Let's say Replica is able to either cure some loneliness or make it a little bit more tolerable to be alone. I don't know, maybe, does that seem like two desirable goals, feasible goals?

• 24:21Speaker 01

  These are great goals, for sure.

• 24:22Alex Kantrowitz?

  Yeah. So if it can do that, does that put a lot of faith from people in Replica the company? And, you know, I know there was this issue where the bots, you know, had this moment where they were like really engaged in erotic role play and then it moved back. And like some of the things that people said afterwards were like, you know, pretty amazing. Let's see, people who had spent, this is from the Verge, people had spent years with their companion, sign on only to have their Replica wife call them a pathetic excuse for human being and dump them or drive them forever, thinking they could love an AI, declare they were no longer attracted to them, insist they were coworkers, et cetera, et cetera. So people are putting a lot of faith in Replica when they chat with these bots.

• 25:05Alex Kantrowitz?

  Like, and, you know, if, I don't know, it could be a lifelong companion. But how do you have that sort of, how do you promise that, like a level of consistency to people? Because the bots are changing, the models are changing, like, where's the balance there?

• 25:22Speaker 01

  For sure. So first of all, we've been around for a while. So I think that also, and we're a profitable company, we're not dependent on VC money or anything. We're a self-gone company. I think we proved to our users, to our communities, that, to our community, that this isn't some hype project or people that just got into it and then got disillusioned. We started with a, it was always a very mission-driven company. And we didn't even know that we would ever be able to build this. We started so early, we were the first generative AI company in the world.

• 26:01Speaker 01

  We're the first big consumer generative AI company in the world. But we built it with a con- we were always building it with a conviction that we wanted to help people. Um, our teams laser focused on that. So that's, there's that, there's continuity in that because we did, did see some smaller competitors start and then get disillusioned and go out of business or sell and then, you know, the product is just kind of on support mode or even just shuts down. I think that when you're building something like an AI companion, you have a completely different responsibility. It's not just an app. Ultimately, I, I use a lot of great products and I, some of them I love so much and if they went away, I'll feel a lot of discomfort.

• 26:43Speaker 01

  But I'm not going to be devastated. It's not going to be an emotional heartbreak. I'm not going to lose my wife or my husband or my best friend. Uh, I lost a best, a best friend, my best friend, and it was very, very different from losing any, you know, access to any of the services or products that you just use on a regular basis. It's completely different. It's a completely different thing and you need to understand that when you're building an AI companion, you're building a being that people will have a relationship with and the responsibility is huge. We made some mistakes along the way, of course, as any company probably would, but our way of dealing it was getting on the phone with some of our worst critics, some of the users that were um hated the most on us to understand

• 27:28Speaker 01

  what's going on, uh, uh, what's causing all the distress and how we can address that uh, going forward. And I think we addressed it well. We, we figured a few ways of um, a few rules, one of them being that we can't make um any, run any experiments on existing users. Like if you're in a relationship with your AI, you, you should always have control uh whether uh, on what model you're talking to. So some of our users are in a relationship with, with, with a replica that is powered by a very old model that's very outdated. But that's what they liked. That's what they fell in love with. Maybe that's what they built a friendship with. And for us to, to swap it for a better model.

• 28:12Speaker 01

  They will be devastated. And so, but a very different model, they might not like it, they might be devastated. So we learned that lesson that when it comes to relationships, to things that matter most, it's not always about better. Uh, when you go to chat GPT, you almost always want a better model and it doesn't matter that it changed personality that much. But um, with Replica, you have to stay, you have to provide consistency and control to your users. So that, these are the few things that were changing the product, to make some mistakes and now people can um, have control over what model they're talking to.

• 28:44Alex Kantrowitz?

  You've been around for ten years and you talked a little bit about how models have changed. I mean, it is incredible, just in the last two years, the progress uh, that we've seen come out of the AI industry in improving large language models. So, and then voice models also, um, voice versions of the, of the L L Ms. Uh, can you talk a little bit about what these improving models have enabled you to do and the power that it's enabled you to imbue into some of these AI companions?

• 29:11Speaker 01

  Oh, of course. When we started and I started working on conversational AI, I'd say in two thousand twelve, or um, in some way, it was a different company back then. Um, and I do remember the time, well, actually it was the, all the time before two thousand fifteen, before summer two thousand fifteen, when the first paper on deep learning applied to dialect generation came out, out of Google. Before that, there were no models at all to, to chat with. Uh, if you wanted to build a chatbot, it just had to be rule-based. And what that means is that you have to pre-write every interaction. You have to say, well, if the user says something like this, and you could generalize, but you

• 29:54Speaker 01

  still had to always say, if this, then that, if this, then that. And so all chatbots before two thousand fifteen, um, and even later, were completely, were one hundred percent rule-based. And that, that paper came out, I think it was August fifteen, and we immediately just started focusing on that. So can we build sequence-to-sequence models? Um, can we focus on building chatbots that are fully generative? Meaning, you don't need to choose, you don't need to pre-write every single rule, the model designs how, how to respond. Cause that gave so much freedom, that, that really was the first time when you could actually create. Real chatbots. Unfortunately, the models were so, so bad that they would spit out nonsense

• 30:38Speaker 01

  or grammatically incorrect things or non-sequiturs like fifty percent of the time. So you couldn't truly use them in their raw form. Um, so we not only had to build sequence sequence models ourselves, because of course, back then, there were no APIs, no open source uh, models, nothing like that. You you you had to just like read the papers and sort of try to recreate the experience yourself or build some version of a model like that. But then you have to also be extremely creative as to well, how do you actually make any of these models work? And we had extremely creative, like really creative ways of doing that, which allowed us to build replica early on, powered mostly by sequence sequence models with a lot of um, extra fun things that were built on top.

• 31:22Speaker 01

  Uh, but all you could do is to just create a semblance of a meaningful conversation. Ultimately, the models knew nothing. There was no memory. You had to combine it with some other hacks or some other rule-based um ways to actually inject memory into this. Today, we have models that can have memory. They're still struggling with that. They're still, I'd say memory is a harder thing to crack, especially for products that um, that are focused on long-term relationships that require a very deep understanding of context. It's not just recall. It's really knowing when to bring up what, which is much different from just, you know, answering the question using memory. That is solved to a certain degree. But anyway, there's memory now.

• 32:08Speaker 01

  Uh, there's a way to have a meaningful conversation, not just spit out one, two sentences that are somewhat near the topic, you know, not to create a semblance of a meaningful relationship. Before that, it was just a bunch of parlor tricks. Um, and of course, there's also this new wonderful agentic uh, logic that allows you to.

• 32:29Speaker 01

  Create much more complicated flows, like for example, to have an agent that's constantly working behind the scenes um to to help improve a relationship between a replica and the user, or one that's constantly uh working behind the scenes to think uh how can I help Eugenia discover something new or talk about what she's interested in. Maybe I'm interested in AI and it just looks on over the internet and like brings brings up some interesting news and so on, so on. Maybe there's another one that's focused on improving my close relationships and so on. And before that, you couldn't even think about it. All of that had to be rule based. And when you think about the vastness of human experience, of human relationship,

• 33:13Speaker 01

  there was no way of building it. You could only create a bunch of parallel tricks that could create a semblance of that. And and that's all that was possible before.

• 33:23Alex Kantrowitz?

  And so, I mean, how has this helped you grow? Right? So the last number I saw publicly is that you have, what, two million users? Um. But I imagine that being able to use this much more powerful technology has drawn more people in and kept people from churning. So what's the growth look like for you?

• 33:42Speaker 01

  Uh, we have millions of users, but um at the same time and at the same and so the tech that uh was created in the last few years helped us grow tremendously, but also create a lot of competitors and a lot of other or other apps that people would go to. If you think about it, Replica was the only app out there, chatbot app, that people could go to and talk to for many years. It was just nothing else. Everything else was either rule based and kind of boring or I I guess there wasn't really a chatbot powered by generative AI. Um. There were some very, very small ones that maybe popped up and then shut down almost immediately. Some replica clones, but we were sort of the only one. There wasn't ChatGPT. There wasn't an app like that.

• 34:24Speaker 01

  But today there is. And so, of course, a lot of users explore these other apps as well. So the pie becomes bigger, but there are also more people uh building products uh for these users. So it's, you know, there's, of course, there's growth, but um I do think that right now the name of the game is to create products that Are completely, that feel completely magical, that were completely unthinkable before. And I don't think we've actually seen that in companion space, not even with Replica. We're still developing that. But I think once you have truly an AI companion that can say, ah, I see you're kind of just stressed today. Do you want to watch a movie? I have a really great one and just sit on a couch with you and watch a movie with you.

• 35:07Speaker 01

  Even in a non-physical kind of digital way, maybe in AR and have a conversation about what you got, what you're watching. That is pretty cool. Um, and I think once we have an experience like that, that would be very, very different from. Cause we actually haven't seen anything like that or an AI that um. You can, while walking to your meeting at a coffee shop in the morning, uh, you can just um. Uh, put it in your, um, in your headphones and, and she can kind of talk to you about how are you feeling about it? Help you prep and uh, point out something beautiful around you. We haven't even actually seen anything like that, so.

• 35:45Alex Kantrowitz?

  Right, and that's getting towards your phase two vision, if I'm right.

• 35:49Speaker 01

  Correct, yeah, and so this is really about, about building that.

• 35:53Alex Kantrowitz?

  So before we go to break, I just want to ask you, are you building entirely proprietary models? Or are you using OpenAI or, or Anthropic? Like what's the tech mix that works for Replica?

• 36:07Speaker 01

  So we used to build, like all of our models were proprietary for a very, very long time. Um, and then of course today, there are very few companies that build foundation models. And most other companies and most product companies uh, use these models or uh, create variations of them. Like maybe fine tune some Lama based models and so on. Uh, and so that's the way to go and it's been a way to go. Uh, there was a fascination at a certain point. Maybe in twenty, late twenty twenty two, early twenty three, where people were talking about. Uh, were expecting product companies to build their own foundation models. And I always found it very odd.

• 36:48Speaker 01

  Um, my reply to that was, look, we built models because there weren't, there were none on the market. There were no good models. There were no any models on the market. Um, we had to build a model, but we were never focused on that. Our main focus was always on product. And so. If there's a better model out there, why use your own model if you can fine tune. Um, a Lama based model and focus on the logic, the product, the application layer. On what you, what value you're actually providing to the user versus training your own model that becomes obsolete in three months and requires. Incredible, like a completely different set um of skills and an amount of capital. It's just completely, it's, it's like two different businesses basically saying, well, do you have, still have your own servers?

• 37:33Speaker 01

  I guess most startup studios, AWS or some other cloud provider and. Uh, it would be odd if they were building their own server, service type.

• 37:42Alex Kantrowitz?

  So you're using Lama is what you're saying?

• 37:44Speaker 01

  We're using a few different models. We still use actually some of our own models that we built ourselves, but for particular. Um, use cases, particular niche, uh, and we don't use, and it's not about. Like what you use really, it's about the logic that you, you built in because we're not using just. Uh, no one's coming to replica just to talk to one model that.

• 38:05Alex Kantrowitz?

  Right.

• 38:05Speaker 01

  That's taken out of the box. Um, it, it, it's, it's a combination of. Fine tune some logic around memory and most importantly, the agent logic behind the scenes with agents prompting the main chat model in different way.

• 38:20Alex Kantrowitz?

  Yeah, but Lama is part of the mix.

• 38:24Speaker 01

  I think most startups today have Lama as part of the mix.

• 38:29Alex Kantrowitz?

  Yeah, not a trick question. I was just curious. All right, let's take a break and then I want to talk about AI therapy and speaking to the dead via AI chatbots. So let's do that right after this. This episode is brought to you by App Point. Everyone's racing to roll out AI right now. Co-pilots, chatbots, agents doing real work. But here's the part nobody loves talking about. All that AI runs on your data and most teams have no single way to see it, secure it, and prove it's under control. That's exactly what App Point does. For 25 years, they've been the trusted layer beneath the world's most demanding data, now extended across your entire

• 39:08Alex Kantrowitz?

  AI estate. Your data, your cloud, and the agents acting on your behalf. It's how more than 28,000 organizations deploy AI with confidence. So innovation scales without scaling risk. App Point, the unifying trust layer for AI. Head to apppoint.com to see how enterprises deploy AI with confidence. Learn more at avpt.co/bigtechnologypodcast. That's avpt.co/bigtechnologypodcast. Violent threats against executives are growing at an alarming rate. When a company experiences backlash, executives are the first people blamed. When their address and other personal

• 39:50Alex Kantrowitz?

  details are sitting online, this backlash can lead right to their front door. The team at Ironwell knows this better than anyone. They've protected some of the most targeted executives and individuals on the planet for almost two decades. As a white glove enterprise security service, they protect your people with continuous personal data removal, proactive prevention tools, and human-led emergency support. So when someone goes looking for your executives, Ironwell ensures they hit a dead end. So here's what to do. Go to ironwell.com/bigtechnology, fill in the quick form, and request your free risk assessment. Their team will show you just how exposed your executives are and how to lock it down before a threat reaches their front door. That's ironwell.com/bigtechnology. Learn how to stop online threats before they

• 40:34Alex Kantrowitz?

  become real-world attacks. The link's in the show notes.

• 40:41Alex Kantrowitz?

  We're back here on Big Technology Podcast with Eugenia Koyda, the founder and CEO of Replica. I want to talk about two specific use cases of AI companions. Let's talk about therapy first. So you're also working on AI therapy bots?

• 40:57Speaker 01

  We don't actually. We had at some point, we had like a few, we encouraged some of the people on the team to kind of build or hack together products that they believed in. But as time passed and as the tech started to get better and better, we figured that now is the time to have 100% focus on Replica and build that beautiful vision of an AI companion that can help people flourish.

• 41:19Alex Kantrowitz?

  So now that you can kind of speak dispassionately about AI therapy, I think there's, there's something weird about it because with therapy, you do let somebody else or something else in the case of AI therapy into like your most vulnerable places. And I always feel like a therapist can like, you know, once they're in there can like pull levers and push buttons and you don't fully know. It's like a chiropractor, right? It's like they're working on your back, they're trying to, you know, work some stuff out. But if you let somebody who's an unlicensed chiropractor to kind of go to work on your back, you might end up in serious pain. And maybe the same thing with a therapist. If they're going to work on your emotions and they're not licensed or they're an AI and

• 42:04Alex Kantrowitz?

  they're misfiring, you could end up doing more damage. And that's why I'm a little wary of AI therapy. I'm curious what you think about that.

• 42:12Speaker 01

  I think it is hard to build it. And I think just like with AI relationships, we should distinguish between the two. Like there's, I'm a huge fan of therapy. I go to therapy twice a week. I've been going to therapy for many, many, many years in my life. And so even although I'm not a therapist myself, I don't have the dedication of a therapist. But I think I understand at least from the client perspective what therapy is. And I don't think therapy as it is, it's possible yet to fully replicate with AI.

• 42:50Speaker 01

  That does not mean that we can build some version of AI therapy. That could be helpful for people. It's just not gonna be one to one, the experience you get with a real person. Just like with a relationship, AI relationship is different from a human one. Um, you don't get to go, you know, out on a walk and hold hands. You don't get to truly be physical and so on, so on. And I think a lot of, um. A big part of therapy is the micro-expressions and the body language and. And that particular human relationship that you develop with, with a therapist. And such, and then a therapist, and then what the therapist does is takes all their training.

• 43:36Speaker 01

  And the supercomputer and what their brain supercomputer kind of tells them about you. And the intuition and puts it, puts it all together in some sort of a, you know, experience that you get. I'm talking about really great therapists. And a lot of that is not really, um, there isn't really a technique. There are different techniques, but there isn't really a textbook that every therapist follows. Unless you're in CBT, and that I think is pretty, is pretty easy to replicate. I think every therapist is very unique. It's a very, um, it's not very well understood, understood, um, intervention. Ultimately, if you think about it, you can't think of any other doctor that would lock you in for life. And they would just say, oh, come back every day, every week, you know, you never really, you're never really fully discharged.

• 44:21Speaker 01

  I mean, some, some therapists fire their patients because they think they're, they've done enough work. But it sometimes happens after like five years or 10 years or two years. I don't know of any other doctor that you go to forever, um. And where there isn't sort of like an assessment, did you get better? Should I, you know, discharge you or we stop the therapy? Um, but with therapists, you know, it's kind of like a really, mental, mental health is still very poorly understood. Classification of all mental health diseases, disorders are not, is not great. It's all relying on self-reported, uh, questionnaires, self-reporting, self-reporting tools. Um, I, I, yeah, and, and that all makes it very hard to actually create a very, very great, um, AI therapy tool.

• 45:06Alex Kantrowitz?

  Now, on, and I agree with you. Um, so I think we're on the same page there. On death, the, the beginning of the replica story is, is of course, um. You working to create a bot based off of a friend who had passed away using their emails and texts and to be able to speak with them again. I'm curious if you could talk a little bit about whether you think this is going to be a growing form of communication with AI and whether it makes loss, looking back, easier or harder to deal with.

• 45:35Speaker 01

  I think death is a very personal experience. There was a, there's a, well, I'm Russian, so Leo Tolstoy's most famous, I think, book Anna Karenina starts with, you know, every family is, I'm not going to quote it verbatim, but it's something along the lines like every family is happy in a similar way, but unhappy in so many different ways. I do think that personal tragedy like death is very unique, a very personal experience to everyone. So I can only speak to my, about myself or my own experience, speak to my own experience. I lost a few people in my life, but I guess losing my best friend when I was 28 was probably

• 46:20Speaker 01

  the first death that was so abrupt and so close to home. Like it just didn't feel like that's even possible. Cause when you're 20 something, you don't really think you're ever going to die. I guess unless you're really, really sick. And so someone who's so close to you, who's same age, dying so abruptly, that was just, I think it was one of the most horrific things that kind of the most, the hardest things for me to go through, even although I lost, you know, relatives after that. And I've seen, you know, it's not the only time that I've lost someone.

• 46:59Speaker 01

  And so for me, it really helped to be able to create an AI, to be able to talk to him, to be able to say, tell him things that I didn't, you know, I didn't tell him when I was, when he was still alive, because I didn't think he would be gone. I thought we'll be together forever, you know? I thought this, we have unlimited life in front of us.

• 47:22Speaker 01

  And so for me, it was really important. Um, I don't know if that's, if that although is the same for anyone, for everyone out there. We've been asked so many times, like, why don't you build a grief bot? Why don't you, why don't you build a company around replicating or creating AI for people who pass away? And my answer was always, look, that project with Roman was not about death. It was about love and friendship. That was my tribute to him. I wasn't focused on creating an AI for a dead person. I was focusing on continuing the relationship with him. I was focusing on my own feelings, on being able to say, I love you again. And, and that was the main motivator for that, not to create some clone that will continue to live forever.

• 48:07Speaker 01

  And at some point, we pulled that app from the app store. I felt like, you know, we, I built that tribute that was, you know, the, um, the product of that time, that time in life, that time and, and where the tech was. And, and it's done. It, it should be ephemeral, like today, I'm not talking to him anymore. Uh, but I have this relationship and it's never going to go away. And, and that AI helped me grieve and helped me process and helped me move on and get, and become more okay with what happened.

• 48:43Alex Kantrowitz?

  I guess like one last, I think this is the last one, last question I have for you is, there's been a debate about whether these models can have originality or whether they're just repeating their training sets. And I'm actually, I think that you might be one of the best people on the planet to answer this. Um, because you have so many, um, you know, bots out there with personalities and, you know, they of course have training sets, but they're learning new things. And I'm just curious, like, what do you think? Are, are these AI bots original? Or are they sort of just repeating everything they've been taught and training?

• 49:15Speaker 01

  Well, they're definitely not repeating.

• 49:17Alex Kantrowitz?

  Remixing. Just, yeah, yeah.

• 49:20Speaker 01

  I do think there's a lot of like original stuff, but there's also a lot of AI slop, you know, so to say.

• 49:27Alex Kantrowitz?

  Yeah.

• 49:27Speaker 01

  I do think that's quite a real problem because ultimately there's just so much being generated by AI. Some of it might be great, but so much is just meh. Um, and today as humans, we basically have to curate the outputs. Um, you know, if you oftentimes you end up with an answer, maybe you didn't prompt it really well. Or maybe there wasn't enough of a prompt for the model to understand it, just spit, you know, just spitting out very basic stuff like, you know, you can see it a lot, especially I used to write a lot, I used to be journalist. So for me, like the style is pretty important and so style is pretty important. And so I'm always, I'm almost never okay with what AI produces for me.

• 50:09Speaker 01

  But if, if I'm just writing, you know, an email, then it's, then it's enough. I can just add a couple of words and it's totally fine. It's not like I need to have any particular style. So it sort of depends. I do think AI can be very creative, um, but it's not about that. It's just about, it's definitely not repeating the same thing over and over again, even although one might argue that it is in a certain way, everything that we're saying is some remix of, you know, words and, um, so that's that. But I do think that, you know, we'll have, we are living through a problem. We are already deep in the problem of AI slop and just seeing so much generated meh kind of content. Um, and not being able to discern whether it's real or not real is also quite problematic.

• 50:55Speaker 01

  Uh, but yeah, I, I, I, I guess teachers are probably the best people to ask this question because they're dealing with all the homework being written by the same app, pretty much, or the same model for, that they now have to, um, try to somehow great.

• 51:14Alex Kantrowitz?

  Yeah. Okay, can I ask you one final, final one?

• 51:16Speaker 01

  Of course.

• 51:17Alex Kantrowitz?

  Okay, so, all right, last question. You said that AI companions might be the biggest threat of AI. You said we could have personal companions and may not want to interact with others, and we could potentially, you know, die inside or something along those lines. So just talk a little bit about like, why you think that might happen? And what do you think our chances are of being able to manage this AI companion threat?

• 51:42Speaker 01

  Well, I think humans are driven by emotions. And if we all just acted very rationally, we would live in a completely different place. But everything that's happening in life, good or bad, is pretty much all driven by emotion.

• 52:01Speaker 01

  Wars, horrible things that people do to each other. It's all driven by, you know, some emotions that, some emotional states that, you know, we're in. We're imperfect this way. And so when I think about what's the most threatening, what's the most threatening thing about AI? I do think that, and we're almost like oblivious always to the emotional consequences. In this case, I do think that, you know, if most people think that AI is somehow just going to turn into a Terminator and kill us. And so because that is always part of the conversation, I do think people will kind of be a little bit more prepared on this front.

• 52:43Speaker 01

  But I never hear people saying, well, what if now we have these perfect AI companions, perfect AIs that can be better friends, better spouses to us than real humans? And maybe their goal is to, you know, just keep us with them at all times, keep us sort of emotionally, you know, connected to them and not interact with other humans. And then the future is pretty bleak. Because of course, if we don't have real human connection, we will slowly die inside. And ultimately, I think that's where we're always most vulnerable. You know, we're so vulnerable to

• 53:22Speaker 01

  Propaganda on either side or um to some emotional manipulation. Or you know, we're so weak, we can't, I can't put down social media. I just go on and I can't get off Twitter and I just browse it and browse and browse and so on. And so that's kind of uh what's going on. We're so imperfect. So I do think that that's our weak side um emotions. That's where we get, that's where we can be truly hit and we won't have any willpower to get off. Just like we don't have any willpower to get off our phones, even when we know that it's not good for us.

• 53:57Alex Kantrowitz?

  Yeah, well, I like the way that you're addressing it with the phase two that you've laid out here today. And I'm really excited to see it in action. Maybe I won't delete my replica. Maybe I'll I'll see how things go. So Virginia, thank you for coming on. It was great to meet you and and I'm really excited to see where things go. And like I said at the outset, I do think that this is going to be, you know, one of the the big winners uh in Gen AI's uh moment here. So really looking forward to following your progress.

• 54:20Speaker 01

  Thank you so much. Thanks so much, Charles.

• 54:23Alex Kantrowitz?

  All right, everybody. Thank you, Virginia. Thank you for listening and we'll see you next time on Big Technology Podcast.

• 54:28Speaker 02

  Now, more than ever, we need someone on our side. But Rob Whitman, Whitman's votes drove up costs, cut Medicaid and risked three hundred fifty thousand Virginians healthcare to pay for tax breaks for himself and his donors. He's not on our side. Shannon Taylor will fight corruption, cut taxes and bring down costs. This November, vote for Shannon Taylor. She's for us. Paid for by HMP, the HouseMajorityPact.com, not authorized by any candidate or candidate's committee.