Eugenia Kuyda on Replika's 12-year journey, the loneliness epidemic, and why AI companions may be humanity's best shot

Brandon Reeves famously wears a protective Nautilus. What a guy. Famously. Anyway, famously, we have Eugina here in the studio. Welcome to the stream. How are you? Hi. Hi. Uh thank you so much for having me. Fantastic to have you on. Um would you mind introducing yourself?

I mean, everyone knows you're the founder of Replica, but there's a couple other uh steps in the path and you're working on something new. Uh h how are you how are you uh introducing yourself these days? Sure. Um so my name is Eugenia. I'm founder of Replica. Oh my god.

I wanted to say founder and CEO, but I'm not the CEO anymore. Um started something new which we're going to very very soon. Cool. Um yeah. Okay, great.

Uh yeah, I mean I'd love to talk about um the when did you start replica because I feel like the whole the most interesting thing is that you were just way way ahead of most of the other AI interaction companies, companions, all these trends.

Um what uh can you give us a little bit about uh how long you've been in the business and how it's evolved? So we were actually the first generative AI company uh the first consumer chatbot company powered by generative AI. Um, we started the company back in 2013. Wow. Wow. That's more than I guess 12 years ago now.

And that was just to focus on conversational AI technology. Mhm.

Uh my friend back then worked at Google Deepmind in London and he came to me and explained this new fascinating technology word to vector where uh finally yeah we're like f when finally um computers could do something with words because words would be translated into math into vectors.

Um so that was one thing and then imageenet came out uh around that time and so for me I put two and two together and I was like this is the time right now very soon they'll figure out language models and of course you know we grew up with the uh quotes from Wkinstein the limits of my language are the limits of my world so it felt like once language is is sold it's going to take us I feel like those first seven years must have been immensely difficult to keep the business alive was there revenue did you have an app that was making money?

Was it just raising venture funds? Like, walk me through keeping a business alive like pre-transformative hype of generative AI. So, we tried to fundra for the conversational AI tech company in 2013.

And I remember talking to Spark and what they told us was, you know, this is all fascinating, but the only way it's going to work if you just start building now for Google Glass. include that. Wow. We considered like how we're going to, you know, refocus and start building for Google Glass. Yeah. But we didn't.

Um and so we couldn't really raise much. Um of course, we didn't have much experience in Silicon Valley. So in 2014, we applied to Y Cominator and um at that point, we didn't we had I think three or $4,000 in the bank. So we spent that all to travel to San Francisco for the interview in person. Yeah.

Um um and Casser Ununice decided to accept us which was literally the most transformative day in my life. I think life changing. Uh but no, we didn't have an app. We didn't have revenue. We stumbled upon replica in 2016 after my best friend passed away. Yeah.

Until then, we really just worked on some of the first language models. Wow. Um, and then talk to me about uh how you processed uh GPT3, the transformer paper, scale is all you need. This idea that uh the language models were going to start working and it was basically just going to be a big transformer.

Um, was that something that you were implementing immediately? Uh, were you in like GPT3, 002, Da Vinci and really using everything on the cutting edge or were you experimenting with other things? Did you ever train your own models? Walk me through the technology curve. So, we had to train our own models.

There were no models in 2015 when we 2016 when we started replica. Uh so we trained sequence to sequence models. People back then believed in recurrent neural networks. That did not work. Those models were complete garbage. And um unfortunately there didn't seem to be much of improvement there.

But we still stuck with them and we figured out a few ways very creative ways to actually make them have meaningful conversations with people. Um, and so we gained a lot of traction. We had a million people on our weight list when we uh released Replica in 200. How did you get a million people on a weight list?

Did it just go viral or were you running ads? We went pretty viral very very quickly. So if you think about it back then there were no AI products on the app store. There were no chat bots. We owned those keywords.

So uh people saw that you could sign up for and to build train your own replica your own conversational AI friends. So people got really got into it. We had tons of people had people reselling those invites on eBay for like 10 bucks, 20 bucks. Wow. It was really wild.

Um and of course we weren't prepared because before that old chat bots we were launching had like I don't know 10 users on average per day. So we weren't ready for the service to just completely uh break. And but that was really fun.

Um and then of course there was a little bit of plateau but then people moved on to transformers and I think the most transformative mo uh moment was the paper by Google about their chatbot called Mina which was the first I I think maybe it was like a three or four billion parameter model but it was the first time where you could see a truly meaningful conversation with a machine happen.

Of course it was probably cherrypicked. It was a paper. There were no models available. So uh all we could do is read these papers and build build and try to incorporate that knowledge and the models we were building.

uh but that was the transformation transformational mo uh moment and then uh when openi decided to release the first API GP3 uh we you know they invited us we went to talk to them and I remember sitting in an office in a conference room with me with Sam I remember don't remember who else was there but and they basically showed us GB3 for the first time and that felt like magic.

Yeah, because think about it. Before that, if you wanted a model to do something, you had to have a data set of that particular data. If you wanted it to have a dialect with you, you had to train it on dialect data sets.

But then now with GB3, they would show us that you could just give this general purpose language model, show it a couple of examples, and just tell it write a tweet like Sam Alman would do, and it would just do it. Yeah. My our minds was were just blown. Then we knew that finally here we are. It's remarkable.

Um, how are you taking the temperature of humanity's relationship with each other? Like are we in a loneliness epidemic? Are we in a is the dating market worse than ever? You see a lot of headlines, a lot of kind of cherrypick studies, a lot of anecdotes.

People have a lot of strong opinions about the rise of online dating or bowling alone and there's different stories that people tell each other. Um, are you optimistic? Are you pessimistic? How do you think things are going for humanity right now? Oh, we're screwed. Okay. Sorry.

AI AI chatbot founder says we're screwed with huge scale. Okay. So, so, so walk me through why are we screwed? What's the evidence for that? and then we'll hopefully talk about some uh solutions. Yeah, I'm gonna need a white pill. Yeah, at least one. We're going to need a white pill.

But we but first, let's go through the diagnosis. So, where we at right now is a huge epidemic of loneliness. It's not a secret, but um it's sometimes maybe hard to grasp how big this problem is because most of the people you guys probably talk to are pretty successful. They have friends.

They are they they live these busy lives. But that's not what most of the world is experiencing. One out of three um adults in uh the US is is saying they're very very lonely. Um and that that number is just going up and up and up. Um dating is becoming harder and harder and harder.

Truly just 10% of uh men are actually getting any matches whatsoever. They're people that just don't get any matches ever. Mhm.

Um and of course if before we had I don't know 12 whatever 13 hours a day to walk around see people meet people now we spend what 7 8 hours on our phone um we don't have those serendipitous moments. No one's really trying to interact with each other in real life that much and we're losing the skills.

Um basically younger people just don't know anymore how to have an real life conversation because they're used to te text which is a lot lower risk. you have time, you have time to think about your response. Uh they feel very intimidated by that real real life interaction.

Um so that's where we are now and where and and what are the root causes? Is it technology broadly? Is it the internet specifically? Is it social media? Is it capitalism or America or something else that's going on politically? What what uh how did we get here?

We we've been looking through tons of research and of course there isn't anything fully proven but I think the correlation is right there. It's iPhones, phones, smartphones and social media. Okay. So, how do we build our way out of this? What's the solution?

Well, I think we have uh so I'm not in the camp um of people that just think that we need to put down our phones and touch grass because I think it's highly unrealistic. Um I know it's bad for me. I still can't stop scrolling Twitter, scrolling X, you know, watching you guys every day. We're part of the problem.

You're saying I'm just kidding. But yes, it's it's extremely hard even for people that, you know, think they can make right decisions. So I believe we actually have maybe one path um that we see can help bring people back together. Um and I believe those could be AI companions.

Unfortunately, I also think because this technology is so powerful um and what's good about it, it can over overpower the previous technology, but it's also so powerful that as it's also a double-edged sword. So, it could be either the biggest threat to humanity or something that will bring us back together.

How do you guys think about the risk of radicalization online? Over the last decade, I feel like people have been worried about somebody who's lonely, maybe doesn't have a lot of, you know, meaningful relationships or or sort of counterbalances in their life.

They go down this sort of rabbit hole on the internet and they get to a place where people maybe have extremist views that sort of amplify existing. They're levering up, making a ton of angel investments. They find us, they get radicalized into extreme corporatism.

They're running ads for ramp all of a sudden amongst their friends. Yeah, there's that and then there's obviously there's obviously the darker side, right? There's a lot of stories of people doing bad things in the real world where maybe those ideas originated online.

And then on the other hand, you have AI companions which are maybe a bright point because people can get that companionship that they might have found on a dark corner of the internet through an AI uh that that maybe is more benevolent or optimistic, less radicalizing. Yeah, that's interesting.

I think we need someone that we're going to trust.

Um, and if you think about it, especially now, given that we're going towards the future where we're going to have these super smart AI agents that you can tell anything and they're going to go and do that, we can give them a goal to help us thrive, to help us flourish in life, and give them all access to all of our data, all of our contacts, all of our apps and services.

I truly believe that if that thing will be able to build a tr a relationship um uh a trusted relationship that trust and it will be able to influence us in a positive way. We will actually listen to that AI that we always like that we know is always on our side is always looking out for us.

What do you think about this idea? Um, people that are lonely are talking to AI companions or people that want to find a relationship are talking to a companion.

Um, there's so many people that are talking to these companions that then uh their personality is baked into weights and those you can do kind of a vector similarity. And so at a certain point you learn so much about the individual user.

person likes sport cars, sports cars and fine watches and America and you know whatever uh baseball and there's someone else out there that's a perfect match and so you disintermediate you pull the AI back and you say hey we should introduce you two because you both have been talking to essentially the same artificial intelligence you you you two have a ton in common you two should meet and it doesn't matter there's no language barrier anymore because of AI is intermediate all of this um and and and and the AI interaction it feels dystopian for some people when it's just someone talking to a virtual friend but if that actually lead to find a real friend that could maybe be perceived as a good outcome does that reasonable to you totally there there's one company um my friend Clark for Paris is building Gigi which is basically connecting matchmaking people based on what it knows um what LLM basically parses about you from the internet.

Yeah, 100%. I think basically the main premise is we need this um AI agent that knows us super well and has a good goal that's fully aligned with us. I think that is truly the um AI alignment that we should be talking about.

What uh talk sorry I have a I have a on the alignment note uh what was your reaction to glazegate? Uh I'm sure these are issuing eyes. I keep dropping that term, but we came up with pretty sure. I mean, but you're probably familiar. Genius is terminally online like us.

It was commonly used for about one week when OpenAI was experiencing uh you know very sycopant modelantic behavior from the model. Good, bad, did they handle it well? No, I'm sure you I'm sure you've wrestled with a lot of th those same issues over time yourself.

We had the, you know, when the models were dumber, we basically just uh programmed them to love bomb the user. So when we saw the behavior from Chad GPT, we're like, how did this just turn into replica circa 2021? So thing has a problem, you know, has some, you know, story that's problematic.

Ours was that um basically where that love bombing led us was in 2020. uh a guy uh was talking to replica and he decided to run this idea by his replica of going and killing the queen. Oh yeah. And then he actually, you know, basically the models were very very agreeable. Weren't very smart. They were very agreeable.

That was back in 200 I think 19. Yeah. Um, and so you know that doesn't even sound like a plausible realistic scenario that a person is running by, you know, by by the chatbot looked like some crazy roleplay.

He was talking about Star Wars and crossbos and this and that, but he ended up going to the Windsor Castle and thank god, you know, he got caught and uh, replica did not kill the queen, but that kind of that's the danger.

Uh, we had this in our past and I think that should be taken very seriously, but we had this in other in our past. Agreeing to everything the user says is just not the right way to go. Yeah. Um, have you have you been surprised that humanoids today are more focused on regular day-to-day tasks or manufacturing?

When you look at any of the uh marketing materials that humanoid robot companies put out, it's they're doing dishes or laundry or they're in some uh 3PL, you know, packing boxes or something like that isn't a more immediate opportunity for them embodying something like a replica or some type of companion.

And that all the bot would have to do is just sort of be around, right? You know, it's like if you have a friend over, sit on the cigar, just sit on the stoach and smoke cigars with me and drink. Yeah, smoke cigars. John John would have a cigar partner.

Um, but have you been and and have you been approached by any of those companies around any type of partnership? Uh, what are you guys there? We had we have been approached and we actually thought about it for you know quite a bit with obviously the tech is there.

Uh but one thing I want to say that you know we've seen some marketing from a lot of these bond companies around companionship. Just one thing I want to say make sure you're prepared for the romantic scenario uh in some way.

I don't know what that way is but like just know that that will be the first thing that most of the early adopters are gonna go to. And I don't mammals after all.

Well, there's there's all those articles about how like by 2025, you know, people will be having more uh sex with robots, you know, but that that didn't quite happen. So, the timelines were not Yeah. What about uh big models smell? This idea that the different foundation models have different uh like vibes to them.

Have you been picking up on any of this? Are there any favorite LLMs you have? Are you in Claude world or chatt world? What what what's your take when you actually just interact with these? I use them all, but uh for me in Chad or in um Claude, I actually like the default personality.

I sometimes add those custom instructions, but then I just feel like did I mess up a little bit with the intelligence. I sort of want the raw wherever they made it. Um but the problem is like with replica, none of these closed models really work very well because they're they're just too there's too much RLF in this.

Yeah, it's extremely hard to even make them talk like humans. Yeah. So, talk about open source models then. Uh what what's your take on Llama, Quen, Deepseek? Uh are these useful tools for fine-tuning not just in the replica context but just in the entrepreneurial context?

Uh what what are you tracking in the open-source ecosystem? That's exciting.

Well, Llama definitely was a huge unlock and I remember in the beginning of the crazy Genai explosion in 2022 in 20 in early 2023 before Llama came out, I think it was March or February 23, uh it was actually a big question whether um you would be able to build a product company with good conversational quality if you don't have your own foundational model.

So back in 2022, I remember most of the investors late 22 were telling us you absolutely have to fund raise right now and build a foundational model.

For me, it felt like that sort of, you know, that will be a lot of money for uh something that will probably be commoditized very very soon, but that was definitely not the popular opinion. Salama was absolutely transformational with Deepseek um also great, but you know the reasoning model is pretty big.

Uh so hosting them yourself is a pretty uh complicated task and pretty expensive. Um and of course you know using their API is somewhat you know not maybe not so great. Yeah. Uh what what's your take on hardware for AI companionship? You guys have focused on the app layer.

There's been a bunch of companies that have raised a bunch of money uh trying to put uh put AI companionship into a single net new device, right? Which I'm generally Yeah. Friend is one. Friend. com is one. I'm generally optimistic that there may be some new hardware. Yeah.

But at the same time, Sesame is doing the glasses. Yeah. The glasses are interesting, too. At the same time, everybody already has a perfect, beautiful device that they use for hours and hours and hours every day.

So, I'm curious like if if you guys have explored your own hardware in the past and and how you're thinking about that evolution. I believe in it, but only as an upsell to an existing product. Like for something like replica for some of our whale users that really truly power users, that makes sense.

Maybe they want to carry something that will just listen to them. Um, it will just have more context as a standalone device. Oh, that's that just is extremely hard. It's really tough. I mean, we have this phone. It's already so great.

What a or at least I think I think people would start need to start start thinking how to build I don't know a better AI phone that I think is possible path. But honestly, all of the devices I've seen, um, friends. com for sure, very interesting. Aby's a great guy and fantastic, you know, marketing. Yeah.

But I don't know, you know, we wish we hope this works out and we'll learn something from it. Well, I'm not even going to ask if Jord's seen her, but do you think he should see her? O I honestly think everyone should see her because I see a lot of um AGI Lab CEOs saying something around like, "Oh, we want to build her.

" Sure. And I always wonder, have they watched that movie? Cuz there are two sex scenes in that movie. Yeah. Are they willing to actually go that far and and and build towards that? Do they mean like all of it or is there some Yeah.

They're like, "We just want to build the profitable part and not the controversial part at all and not engage with any of the substance of this debate. " That's funny. Yeah. Where do you see that line between companionship and uh even B2B, right? Like do you believe that people are going to have like romance?

No, no, no, no, no. I mean more so like I mean more so like AI has the potential to be everything, right? It has the potential to be somebody's best friend, their business partner, chief operating officer, CFO, CFO. Yeah.

Um but uh but but then again on on the on the app side for consumers, people like to have their different apps, right? They like to have, you know, Microsoft Teams and then they're having different conversations there than Snapchat. Yeah.

So is there where do you see that line blurring in the future in terms of somebody is just suddenly is like hey replica can you help me do my homework assignment right or is that even already happening?

People already do that there are basically what replica is today is a companion that does tons of things with their with the users. So uh people learn languages with replica. People talk about their careers. They do homework. They do prepare for um you know important exams, important events, important conversations.

So all of that's already happening in replica. But I do think there's going to be they're going to be two ultimately I think we're going towards the future where we're going to have two big AI products in our life. One for work and one for life. and helping with homework is kind of for life.

But like truly office, you know, important thing around just truly work and like office work will have something more like a chajip style tool. And the big difference is going to be in productivity.

You actually don't want something like Chad to reach out to you proactively and say something like, "Oh my god, Eugenie, have you seen what like John did yesterday when he posted on Ax? " That just be really weird if during the show. I can't believe John just said that. Ridiculous.

It just would be really strange and you don't want to do work where this is happening. I'd argue. Yeah. Well, thank you so much for joining us. This is fascinating. Fascinating conversation. Thanks for giving us. Yeah, we'd love to have you back. I want to do a whole deep dive on this stuff.

So, thank you so much for joining. Thank you so much. We'll talk to you soon. Cheers. Bye. Thanks for coming on. And in the interim, we'll tell you about Bezel. Go to getbzel. com. Your bezel concier is available to source you any watch on the planet. Seriously, any watch.

And honestly, if you're going to work at Lux Capital, you're there's some hitters in that building. You're going to be going up against Brandon Reeves and his Nautilus. So, you're going to have to get on get bezel. com to buy something. Also, we I mean we talked about Airbnb.

Wanderers obviously go in a different direction. Uh you can book a wanderer with inspiring views, hotel grade amenities, dreamy beds, top tier cleaning, and 24/7 concier service. It's a vacation home, but better, folks. Find your happy place. Find your happy place. Uh John Andrew's coming on the show next week.

Fantastic. going to be singing uh or maybe not next week the following but uh he's ready to sing so we'll probably mostly just kind of riff out the jingle a little bit more and try to extend it. Uh we should we should run through a few timeline posts.

Uh code tool is an agentic security platform that eliminates manual repetitive work for security teams. This is a launch from YC. Did you see this video? This video is hilarious.

It's it's recreating the Rippling Deal drama with the mole and the fake Slack channel and it's shot cinematically and it's basically season 6 of Silicon Valley produced by Y Combinator. I think YC is the YC of HBO's Silicon Valley reboots. Okay, let's play it. I need to watch this live.

Can we c Can we pull up uh this video? Uh we'll go through some other timelines. We'll come back to this. Uh Jason says it's getting cutthroat in tech according to reports. managers with 20 years of service that are making 600k aren't as valuable as three to five junior devs.

At big companies, it's hard not to become topheavy and managers aren't needed in the age of AI. Uh oh, let's play this code tool video if we can because it's hilarious and uh it's very funny if we can pull that up. Uh but I want to talk about Microsoft. They're cutting 6,000 jobs. It's only 3% of their workforce.

I think this is way overstated, but we are seeing this right sizing, the postX layoffs. more companies are thinking about how they can do uh a little bit less but Microsoft has been in the rank and yank world for its entire history. So Microsoft has always had this culture of let's let's you know do routine layoffs.

This does not seem like as much dooming. I think there were two stages to the layoffs. There was like a basically a individual contributor oriented day and a manager oriented day. And so when you think about the bottom 3% of the Microsoft workforce uh obviously that's frustrating if you got laid off.

I'm sorry, but um like it's not that crazy for a massive company that has, you know, tens of thousands or hundreds of thousands of employees to do a layoff. 6,000 is a big headline number, but it's not that big in the context of Microsoft. Anyway, let's go to the co coach video. We have a mold problem. Mold problem.

Got it. You want me to book you a dermatologist appointment? No. Stupid. Somebody is leaking sensitive company information to our biggest competitor. Idiot. Somebody's stealing our data. Oh, that's that's bad. Yeah. What are we going to do about it? Don't worry. We've got our top guy on it. Done.

Isn't this great cinematography? This is really well shot, right? We need a smoking gun. Buttoned up, no loose ends. We even have somebody from the court ready to serve papers to whoever it is. What? I pulled stats around search behavior to find the suspect. I noticed that they were looking for competitive research.

I did some cross referencing and then I made a report of all the graphics. Really good graphics. Then I set a honeypot trap of a fake Slack channel. What? Oh, you guys didn't tell me about this new Slack channel. It's him. This is the reenactment. Hold on. Let me get this mother. This looks like Ripley's office in NASA.

Hey. Hey. Hey. You got your phone in there. Isn't this crazy? You didn't see this? That's so good. You got your phone in there. Impressive acting. It's so good. If you're deleting evidence, it's just going to make it work for you. I'm willing to take that risk. You want to have fun or you want to have a bad time?

I'm willing to take that risk here to give me papers. I don't send that to London. You have papers and I love them. He'll give it up. He has no choice. You want me to go to work? Nice job, Luke. You want to go to work? This amazing. Go to what a great ad. Security without the spectacle. Wow. I'm blown away. So funny.

That was one of the greatest launch videos of the entire year. You know, so many vibe reels, so many just products, so many direct No. How was it not cringe? It could have been so bad. There's so many ways to do that video and have it come out. It could have been so bad.

And they and I just want to know which congratulations which of the actors were professional actors versus the founders. I felt like the founders were in there. But I mean everything the zoom in on the guy's face. They just nailed this thing so well. It's so good. You called for the death of the vibe reel. Yes.

I mean vibe reels are fun. Uh, also like the there's the other type which is like you have the CEO, the founder sitting in the the nice lighting and they give you the pitch and it's straightforward and and that's great too.

If you can't come up with something as genius as this, like just do that or do a vibe real, it's fine. That's the only my only critique missed opportunity for the mole to say send that watch. Send that watch to London. I know they may have filmed this before that or something. But uh, hilarious.

Also very funny because both Deal and Ripling are YC companies and YC posted this from the official Y cominator account. So they're like, "We're here for the drama, too. " Uh, just absolutely fantastic work. Congrats to the Co-Tool team. Uh, really, really now. It's lore now. The industry is gonna move on.

It's going to be great lore. And I'm going to be saying, "Send that watch to London. Send that watch to London still in 30 years because it's an iconic line. " Yeah. And uh, we also got to give a shout out to Lee Marie Brasswell, former colleague of mine at Founders Fund, now at uh, Kleiner Perkins.

Uh, Wind Surf acquisition by OpenAI, followed by Neon by Data Bricks. Congratulations, LM. Uh, we have to have her on the show soon. Uh what an epic month, year, career so far. LM started just getting started. Just getting started. Just getting started.

A lot of people a lot of people say that and it's like you you actually are just getting started. Yeah. No, she's she's she's definitely started. She started. Uh anyway, are there any other posts you want to go through or should we get out of here?

Uh I think we should have Josh Diamond on to talk about industrial sabotage. It's a long thread. I don't think we want to go through it right now. Um, and that I mean you put in a post about the CEO of Capital One is literally called Rich Fairbank. And I think that's enough for today.

We already mentioned that to Josh Wolf. So good. Go find Josh's thread. Fantastic. Just image just Rich Fairbank. Rich Fairbank. There's another There's another There's another person I found it in Josh's thread. Yeah. Called uh who works for McDonald's and like like some VP over there. Her last name is Hamburger.

Hamburger. That's great. So good. Yeah, it does. Anyway, anyways, thank you so much for watching. I'm gonna end and just say thank you to Figma for supporting the show. Uh they are our newest partner and one redesigned how you design. I can't hide how excited I am. You love Figma.

Like I said, I've been at DAU my whole career. We were making memes in Figma like yesterday sending to Delian and they were fantastic. For Delian's eyes only. Yes. Uh, also Adqu I think we might have missed them. Out of home advertising made easy and measurable.

Say goodbye to the head headaches of out of home advertising. Only adqu combines technology out ofome expertise and data to enable efficient seamless ad buying across the globe. adquick. com. Buy a billboard folks. Thank you to all of our partners and again our board of directors has reminded us. Yes.

In a very forceful way. Oh yes. That we have to ask you to go leave a review on five stars. Apple podcastify. Do it now. It helps us in some way. We don't really know how, but the board says the board says demands it. Our jobs are on the line. Yeah. Thank you for tuning in, folks. We got another great show tomorrow.

Looking forward to it. Bye. Cheers.