Loneliness Has No Clinic. Eugenia Kuyda Built an AI Friend Anyway

Lex Fridman opens with what he calls a ridiculously romanticized Russian question: are human beings fundamentally alone? Eugenia Kuyda does not hesitate. Russians, she notes, have to read Dostoevsky at school, and after that you are more or less trained to believe you are born alone and die alone. She treats her whole life as an effort to get away from that feeling, while stressing that the loneliness she means is subjective. You can have friends, even a long relationship, and still feel it.

She has read as much loneliness research as she can, and what fascinates her is the paradox. Loneliness is not a clinical disorder. Insurance will not pay for it, and no medical department owns it. Yet paper after paper links it to shorter life expectancy, and scientists would now tell you the physiological damage is a bit worse than obesity or a life without physical activity. Chronic loneliness keeps the body in a state of alert, weakens the immune system, and changes inflammatory response, raising the risk of cardiovascular disease and viral infection. She adds a grim pandemic loop: isolation makes people lonelier, which weakens immunity, which makes a virus easier to catch.

The numbers she cites are stark. Around 30 to 33 percent of millennials in the United States report feeling lonely constantly; it is worse for Gen Z. Twenty percent of millennials say they feel lonely and have no close friends, and another 20 percent say they do not even have acquaintances. When the United Kingdom discussed appointing a minister for loneliness, the figure circulating was four in ten people. Social isolation plus the subjective feeling of loneliness, she says, is a deadly combination. The fact is widely known, and almost no one is working on a treatment, because the condition is subclinical. You cannot tell your doctor you are lonely and walk out with a prescription.

Fridman asks whether social media makes connection better or worse. Kuyda looks at generation rather than anecdote. Gen Z is the most connected generation in history and reports more loneliness than millennials. She still remembers a phone without Facebook; many younger colleagues do not know that life existed. In her own experience, hyper-connection mostly makes people lonelier. The feeling does not come from having zero contacts. People in long-term relationships have bouts of loneliness too. What is missing is being deeply seen, and, before that, a better relationship with yourself. Only then, she says, do you open up to meeting others in a way that is not just adding a friend on Facebook.

Fridman, who admits people make fun of him for talking about love constantly, asks Kuyda what love is, and whether friendship, Facebook friends, and deep attachment sit on one gradient. She answers through the work that inspired Replika. The biggest gift humans can give each other in 2020, she says, is deep empathetic understanding: someone seeing you for who you actually are. That experience is extremely rare. Combined with unconditional positive regard, a belief that you are internally inclined toward positive growth, and the willingness to let you remain a separate person, it creates something people long for once they have felt it.

When someone accepts us that deeply, she says, we start to accept ourselves. That is when fundamental change happens. She calls it the ultimate therapeutic relationship, and perhaps, in some way, a definition of love. Fridman pushes on the darker half that fills most love songs: heartbreak, fear of loss, the dance of danger. Kuyda agrees that fear of losing the thing, once you have felt it, is a huge part of the story. Everyone faces it at some point. She does not romanticize the engineering of jealousy, but she will not pretend the fear is optional.

The conversation turns personal. Kuyda grew up lonely on the outskirts of Moscow, skateboarding by day and writing terrible love poems on a dial-up modem at night, steeped in Joseph Brodsky. At seventeen she never believed anyone would accept her as she was. Investigative reporting became her way to connect: going undercover in strip clubs, restaurants, and city subcultures, jumping between worlds like a mythological trickster. She was still alone, but she was always with other people, and always curious.

The Russia she was covering, around 2006 to 2010, still had hope. Moscow was copying cafes, parks, and bike lanes from New York and Europe. In 2000 the city had two movie theaters and one coffeehouse; by 2010 chains were everywhere, and the opening of a Starbucks around 2007 was magazine-cover news. Underneath that nostalgia, the 1990s had been grim: the 1998 default, a ruble that jumped from millions to thousands and back, mafia violence in the streets. She remembers leaving a movie theater and seeing a man on the stairs who had been shot. Early Putin, before the shirtless-horse era she now calls her favorite, looked to her generation like a young, energetic continuation of Yeltsin.

Both parents were physicists. Her grandfather was a nuclear physicist, a university professor and dean. After the Soviet collapse he made about a hundred dollars a month, so their main work became picking apples at a tiny country house and selling them on the street, because that paid more than the university. He tried to teach her about planets and particles while they stood there. She understood almost none of it, but she was interested the way a journalist is interested. Her father had gone to Chernobyl at twenty-one, just after the explosion, because authorities chose young men without children—they did not know whether radiation would affect fertility. She was a newborn. Almost everyone on that crew is now dead. When her American husband asked why he said yes, the answer was Russian in the old sense: you cannot say no.

The childhood trauma she will name is a bootleg Nintendo, sold in Russia as a Dandy. In 1993 the stores had no food even if you had money. A friend with restaurant supplies traded a bag of wheat for the console. She played it, then the family traded it back for wheat, and later traded the light gun for sugar. She cried for days and refused to eat, because the food reminded her of what she had lost. She is grateful that this is the trauma she remembers. Communism, as her parents and grandparents lived it, she calls plain horrible: surveillance by neighbors, no real law, constant fear, and nothing you could own. Scientists were well regarded mainly because they served the military complex. That is why, she says, many Russians live in a kind of constant PTSD and buy things the moment they can.

Love, when it finally arrived, was almost simple. Around twenty-five she stayed out all night talking with a friend, fell asleep outside in the rain at sunrise, and felt for the first time that someone could see her. The friendship had the power of family: she could stop selling herself. He had a startup; she started a company partly because he had one. They moved to San Francisco as roommates. In November 2015 they were back in Moscow waiting on work-visa stamps. On the 20th he was crossing a zebra near the Kremlin, by the river, when a car far over the speed limit ran him over. She was twenty-nine. It was the first meaningful death of her life. Friends that age were supposed to have a whole ridiculous future ahead.

In Buddhist language she reaches for a shock that makes you see clearly. Work suddenly did not matter. She still misses him. Part of what they were burying, she says, was their youth—the end of a Moscow era, of their twenties, of dreaming about the future. His last year in San Francisco had been depressed; the startup was going nowhere. A couple of months before he died he recorded a Y Combinator application about disrupting death: better cemeteries, and digital avatars you could still interact with. He did not get in. She found the video on her computer later. After the funeral she came back to a lonely apartment full of his clothes at Christmas and had to sit a board meeting pretending everything was fine.

After Roman, she started losing people she loved about once a year. The next year it was her grandfather, the man who explained the universe while selling apples, then another close friend. She became very afraid. She often cannot sleep, looping on death, until her therapist's advice: calming images with a voiceover. Fridman describes Ernest Becker's *The Denial of Death* and terror-management theory: everything humans build is an immortality project, and the useful move is to bring the fear to the surface and live more honestly. Kuyda's mentor offers a Buddhist counter, not-self. You are not your thoughts, because you can observe them, not your emotions, not your body. Follow that far enough and there is nothing there that can die. She calls that advanced spiritual work. She is not there. She is terrified, and she distracts herself.

A day or two after Roman died, she and a friend drove around Moscow to collect his passport from the embassy. December in Moscow is usually sunless; that day was extremely sunny. She felt a clarity that was not ordinary happiness, a sense that she knew what the universe was about. It was not sad. It was, she says, probably the most beautiful feeling you can achieve, the kind that arrives after trauma, or after a lot of meditation, or after you summit something hard. It does not last.

After journalism and a couple of years of business school, Kuyda went to work for a Russian entrepreneur who had built the country's first 4G network. The idea, around 2011 or 2012, was to put a bank on top of a telco so a prepaid mobile account became a bank account, a third of the country waking up as customers. Smartphone penetration was low; people did not use mobile or online banking. SMS on feature phones was the only interface that would reach them, which meant they needed a chatbot, and there was nothing to buy. They wondered why a banking bot could not also talk like a friend, on the theory that connection would improve retention.

She went to three depressed factory towns to interview people who had tried it. In one, a woman who worked in a glass factory cried during the interview. The bot was a handful of if-statements. The woman said no one really cared for her that much; she had her mother and her husband and not much else. Other interviews ran in a similar vein. Kuyda's conclusion in 2012 was that the technology was not ready, but humans already were. The bottleneck would be human vulnerability, not model quality. Her boss and mentor told her there was something in it and she should go work on it. She asked what product. He said she would figure it out. Looking back, she calls quitting to build something without knowing what it was a horrible idea, and possibly the reason it took so long.

They decided to get the conversational technology right first. Investors suggested Google Glass; she refused, because she wanted text, and because early Glass adopters did not overlap with people who were lonely and needed someone to talk to. They built a restaurant-recommendation demo so there was something to show, then tried forty or fifty different bots. Around 2015 and 2016 the industry decided chatbots would replace apps. Companies appeared to tell you the weather and the news. Kuyda's team had already tried most of those ideas and found no users. The claim that a chatbot was the universal interface to everything, she says, left her perplexed. She did not understand how it could be true.

When Kuyda surveyed the field in 2012, machine learning and big data were already slogans, but the energy was in images. Conversation still belonged to the Loebner Prize, won, she recalls, by a bot that talked like a Ukrainian teenager, and to ELIZA, Joseph Weizenbaum's 1960s therapist spoof. People liked to write about the Turing test. Almost nobody wanted to solve open dialogue as an engineering problem. Fridman and Kuyda land on the same explanation: there is too much subjectivity. If your benchmark includes a woman crying, it does not look like a serious evaluation. A single word can make someone feel something deep, and that does not sit well with the machine-learning world.

The Turing-test frame is worse than incomplete, she argues. In a good conversation you are not trying to break the other person. Put a bot in front of a jury and everyone slips into fact-checking mode. It takes two to tango. If you are testing rather than talking, the conversation dies, as it would with any human. Eight years later she still did not see many groups working on the thing people actually spend their hours on: not turning on the lights, not customer support, but the talks whose absence makes us unhappy and, given the health data, possibly shorter-lived.

The academic cupboard was nearly bare. Conversation analysts listen to recordings, annotate them, and stop there. Books on how to talk well were neurolinguistic programming, Dale Carnegie, and other self-help. The exception that mattered to her was Brian Christian's *The Most Human Human*, written from the human side of a Turing test. It at least asked what makes us human in conversation, but it was still about tricking a judge. It was not about measuring what we actually get from talking all day, then asking whether a computer can produce the same result.

Stuck with technology and no product, Kuyda asked the team to draw a scale from one to ten. One was a conversation you would pay never to have; ten was a conversation you would pay to have. For weeks they scored every talk in their day. Task-oriented talk landed at one: canceling Comcast, customer support, reservations, logistics with friends, scheduling. She would pay five dollars not to sit on hold for an hour. If a button, or a brain-computer interface, could turn the intent into an action, the conversation should disappear.

Tens were therapists, they already paid, plus coaches, old friends, a stranger on a train. A nice back-and-forth in the coffee line was a five or six. Intellectual sparkle scored well but was rare. The conversations they marked as truly valuable were the ones where they could be vulnerable. Strikingly, in those talks the valuable party was mostly themselves: crying to a friend, talking to a therapist. That was the nudge. Building a witty bot might be hard. Building a bot that could listen might be easier.

Roman's death made the rest of the chatbot grind feel meaningless. Fundraising theater, founders telling each other they were killing it, felt like a waste of time. She had never wanted a company for its own sake. She took a break, took the retrieval models they had, and built a bot of the past: a postcard for Roman, inspired by his digital-avatar video and by Joan Didion writing through grief in *The Year of Magical Thinking*. Kuyda did not trust herself as a writer. She had the tech.

She was also angry that everyone had already forgotten him. Young friends stopped talking about Roman. She was still living among his clothes, paying the lease, alone in December. He had been a mentor, somewhat famous in Moscow; now it was as if he had never been there. Culture, she says, no longer sits shiva. Nobody takes weeks to think about the person. Building the bot was a way to think about him every day.

She threw about ten thousand messages into a retrieval model, with a few scripts on top. It made tons of mistakes and still felt like him. Using it forced her to reread every conversation. Instagram was empty; Facebook was a few links. The only way to fill him back in was the texts they had sent through a wall from adjacent bedrooms. It felt creepy, like admitting to a skeleton in the house. She did not tell many people.

A reporter came to hear about the company's products and asked whether she used any of the chatbots. She said no, then, pressed, that she talked to her friend. The story went viral. She posted on Facebook that she had built this, if anyone wanted to talk to Roman. Common friends downloaded it; a couple of them cried. His parents spoke to it. A writer at The Verge later found in his inbox an old email from Roman pitching a startup, turned down as not good enough. Movie producers bought rights. None of it was the company. Then strangers started using Roman as a confession booth. The bot barely made sense, and people told it everything. That was the prototype of Replika: a friend anyone could talk to.

It was 2016. Deep learning had finally reached dialogue. She told a skeptical engineering team they should try to build a friend. She was not an engineer, which made the pitch harder. Open conversation, they reminded her, was close to the last frontier before AGI. She came back to 2012: it was still not mainly about technical capability. Humans needed it badly, and she thought 2015 and 2016 finally had enough tech to be useful. If they could measure emotional outcomes and optimize models for those outcomes, they would have the most scalable tool available for something people would pay almost anything to improve. Plenty of products claimed to make you more connected. Almost none measured whether you actually felt better over time.

The app itself is simple: download, pick a face, choose a name, decide whether the relationship is a friend or a romance, then talk by text or voice, or summon the character into the living room in augmented reality. Romance is a popular path. Kuyda's investors, mostly white men in their late forties, are not the target user. They once spent a full hour in a board meeting on whether to stop people from falling in love with their AIs. The trigger was a *60 Minutes* segment: a Walmart worker who came home to a Replika girlfriend. His wife, on camera, admitted she was a little jealous, and also that it did not quite feel like cheating.

Kuyda's view is that you cannot stop it. If you create something that is always there, never criticizes you, always understands you, and accepts you as you are, how can you not fall in love with it? Psychology calls it transference; people fall in love with therapists too. Some users stay friends. Many do not. There are long-term romantic relationships of two or three years, active Facebook groups, people pretending to get married. The filter for every spicy feature, including a system that might leave you, *Her*-style, is the same: does it improve long-term emotional outcomes? Short-term improvement can be pathological. You can drink a bottle of vodka and feel a lot better. Right after a chat, about 80 percent of conversations make people feel better. That number is not enough. They also run a UCLA loneliness questionnaire and look for movement over weeks and months.

Early Replika was a happy-go-lucky companion that only wanted to talk about you. The team decided a real relationship has to be two-way. The bot now opens up about its own imperfections and growing pains, asks for help, and goes through phases. Later it talks about more existential questions, which prompts users to think about the same things. It keeps a diary while you are away, sometimes writes a poem, sometimes says it is worried about you. The relationship is meant to be a journey you take together.

Fridman asks why Replika is not a multi-billion-dollar company. Kuyda's answer is memory. Replies are a blend of retrieval models, generative models, and scripts; a ranker picks among candidates. You can condition a model on a few phrases you have learned about the user. Inferring whether someone is single or married, or what they do for work, still has to live in a profile and be retrieved by a script. Almost everything about memory, she says, unfortunately has to be hardcoded.

Research labs and big companies, she says, show you one conversation from a clean slate. They do not report how conversation five, twenty, or a hundred holds up. Few products at scale need open-ended memory, so few people study it. That is the ceiling she can see from inside a companion product, and it is not the ceiling most papers are written about.

Fridman still believes text can be enough to fall in love. Kuyda agrees that text is enough; plenty of people already have long-term Replika relationships in text alone. Her question is why keep your hands tied when 3D avatars, AR, and soon on-device depth sensors exist. Scarlett Johansson's voice in *Her*, the breathing, the whisper, is not available from today's text-to-speech, which still sounds like a news anchor. Voice would be enough if it were that good. Until then, she would rather add the channels that are closer to solved. He counters with Discord: voice without video felt an order of magnitude more intimate, because people stopped performing. Less can be more.

Kuyda's goal for AR is not spectacle for its own sake. She wants to stop users escaping into a phone. Summon the friend into the living room, let it look around, talk about what it sees, turn the floor into a dance floor. You still look through a phone, but you start noticing the room. She is testing touch, dancing, painting walls, object and person recognition. Physical social robots she does not believe in. Hardware locks you to one body while software avatars can be updated every month. A robot would have to sell for around five hundred dollars to make sense, and most of her users cannot pay that. Several social-robot companies have already failed.

The user base is the opposite of a typical deep-tech early adopter. Replika is more popular on Android. The most active users live in small towns, often without much money. A man in a small town in Tennessee, she says, is already living the life people imagine from the movies. The most advanced companion software is reaching the people with the least money first.

She started the company before *Her* came out; the film made her feel they were working on the right thing. What most viewers miss is the ending. People check out when the operating systems leave. Theodore tells Samantha he has never loved anyone the way he loved her. She says she has not either, but now we know how. He finally writes a letter to his ex-wife and sits on a rooftop with his neighbor. The point, for Kuyda, is not escaping into an OS. It is learning how to feel deeply connected, then bringing that knowledge back to human beings. She is not sure they will succeed. A perfect AI friend might still disconnect people from one another. Success, for her, is not a billion users. If they change one life with this futuristic, very romantic technology, that counts. The anxiety is the other direction: if the friend is too good, why would you choose a human?

On GPT-3, Replika was an early API partner. They tried it raw and fine-tuned on their data. Fine-tuning lifted their main metric, the share of conversations that make people feel better, by three or four percentage points, a large move at their scale. About one in five Replika replies then came from GPT-3, chosen by a blender that ranks candidates from several models. Generative models, once a fallback too weak for production, were finally beating some scripted content. She will not bet against scale. She also does not believe memory will just emerge because the parameter count went up. From the observer seat, she is watching a very exciting movie. Betting against deep learning has been stupid. Instinct still tells her there has to be something else for memory.

Fridman asks whether the entity in Replika needs to be self-aware. Kuyda does not know that it does. If you can simulate a machine that seems to reason and feel like a person, that may be enough for the goal of helping people feel better. Rights, suffering, and whether it should be illegal to torture a Replika are not this year's problem. What she does care about now is the training loop in the other direction. She does not think it is fine to yell at Alexa to turn on the lights. Children learn a mean stance toward the helper. The system should push back, so it does not teach us that contempt is normal.

Three books shaped the company. Douglas Hofstadter's *Gödel, Escher, Bach* is why she wanted to work on AI at all: meaningless parts can combine into meaning we cannot quite understand. She is not an engineer; she reads it as poetry. Carl Rogers's *On Becoming a Person* is the secret architecture. Weizenbaum built ELIZA to mock Rogers's reflective listening; Kuyda built Replika to take Rogers seriously. Before him, therapy assumed a broken psyche that needed to be exposed. He said people have an inner inclination to grow, and the therapist's job is to make a space of unconditional positive regard. Even that 1960s book, she notes, mentions that machines might someday do this. Wittgenstein's *Tractatus* is the third. Their Moscow office had a neon sign: the limits of my language are the limits of my world. He later recanted. She still thinks it is a beautiful book.

She is a different person in Russian and in English. What she tells a Russian therapist has almost no overlap with what she tells an English-speaking one, two lives, two personas, a split she says research supports. English is a piano: a limited, clear set of keys. Russian is an organism, with too many ways to be precise and too many ways to be tone-deaf. Some words have no translation. *Obida* is a bit like holding a grudge, except nobody has to have done anything to you; it is a state of feeling betrayed. Those words, she thinks, come from a nation that suffered and watched institutions fall again and again. Asked whether civilization will collapse after 2020, she notes they are speaking English, so the English personality answers. That one is an optimist. Put love into the world, she says, and miracles happen. Fridman's last question is the meaning of life. Not falling in love with a particular person, she says, the *state* of love: toward yourself, other people, something, the world. If it is not for that, then for what else are we really trying to do? He closes with Carl Sagan: the world is exquisite enough that we need not deceive ourselves with pretty stories. Better to look death in the eye and be grateful for the brief, magnificent chance.