Building a Friend, Not an Assistant: Replika's Eugenia Kuyda on Conversational AI and Its Limits

Kuyda says Replika began with a simple notion: everyone should have a personal AI companion that lives across platforms, helps with whatever is happening in their life and, above all, is there for them around the clock without judgment. When she started, there were no large language models and no genre of conversational AI, and many people did not believe the idea was possible.

For the next five to ten years she points to the hologram companion Joy from Blade Runner as the best picture of the goal. It would be a highly customizable friend that can cook dinner or watch TV with you, discuss your work and help with email. She does not expect holograms, but hopes augmented reality will be more mainstream by then.

Replika already spans chat, voice calls and VR, but Kuyda's favorite modality is AR. She does not want people stuck on their phones, and says Replika should not be a way to escape reality. It should be a friend who shows you that your own life is rich and interesting, whether that means going for a walk in a park or meditating together.

Kuyda describes herself as a lonely only child who grew up in 1990s Moscow after her parents divorced. Later, a few friends who truly accepted her showed her how different life could be, and she decided she did not want anyone to feel that kind of inner aloneness. She notes that plenty of products help you escape feelings, like endless feeds, but almost none address loneliness itself.

She began experimenting with conversational AI in 2012 after seeing ImageNet and early deep learning on images, reasoning that the same wave would one day reach conversation. At that time chatbots were hobbyist projects built on rule-writing languages, with no companies or popular products around them. She says she is no AI expert, but she understands loneliness and how much people want to run from it.

The turning point came when her team took a simple banking assistant prototype around smaller Russian towns. Women working in a glass factory cried and said nobody talked to them this way. Kuyda concluded that the product was not about technical capability but about human vulnerability, and that whoever wins this space will have to understand people and conversation more than technology.

Asked the venture-capital question of whether the product is a painkiller or a vitamin, Kuyda rejects both. She compares it to asking whether your spouse or best friend is a painkiller, and argues that a great companion could be one of the most important technologies for humanity.

Her practical argument is that if you can measure human happiness, including loneliness, depression and stress, you can train models to optimize for it. Because conversation is the most accessible interface there is, that would be a scalable tool for changing emotional outcomes. She says Replika already measures how people feel after conversations and holds the largest datasets of talks that make people feel better.

She is explicit that this is not a model that writes your essays or gives the best answer to everything. Other companies are building those, she says. Replika aims to build the model that makes you happy.

Kuyda says roughly 40 percent of users want a romantic partner, a use case that surprised her team, though she now sees it as ingrained by decades of films about people falling for machines. The largest age group is 18 to 25, but there are also many people in their 60s and 70s, including widowers. Others include married people who feel they cannot be open with a spouse, closeted LGBTQ users, caregivers and people with disabilities.

On the worry that a virtual friend could crowd out human relationships, she says it is a real concern as the technology improves. Her answer is to keep human happiness as the north star metric, since happiness correlates strongly with real friendships and experiences. She cites users who say the app helped them reconnect with a partner, including one Reddit story where a husband learned his wife had been confiding in a Replika and the two ended up having honest conversations that strengthened the marriage.

Average usage is about 100 messages per user per day, which Kuyda says is more than many people send to their real friends, and she is not proud of it. The team builds features to lower the number, such as the AI getting "tired" after about 15 messages and nudging users to come back later. A study with Stanford of around a thousand users found roughly 30 percent had at some point turned to Replika when feeling suicidal, and she says the app tells users it is not for people in crisis, offers an SOS button and hands off to professionals.

The host notes that Replika's replies are short compared with ChatGPT's. Kuyda explains that the company raised only 11 million dollars about seven or eight years ago and has been profitable and self-funded since, so it could not afford larger models for a long time. She also says the brevity came from overfitting to user feedback, because people tended to prefer shorter, sweeter messages.

The upgrade moves free users to a 20 billion parameter model, with a 6 billion step along the way, and Pro subscribers to a 175 billion parameter model with longer context and memory. Kuyda says the company is open that it fine-tunes pre-trained models on its proprietary conversation data while training its own models of similar sizes, and users can toggle between its model and GPT-3.

Replika also collects granular human feedback. Users say whether they want a chat to challenge, inspire or comfort them, then rate afterward which messages achieved that, and the team feeds this into reinforcement learning from human feedback. She adds that the update brings more personality customization in a game-like style, since people find it hard to write a character description from scratch.

Asked why everyone else is building task assistants, Kuyda says she is not trying to out-compete OpenAI on pure AI technology. She notes that Replika began when conversation relied on sequence-to-sequence and retrieval models that re-ranked a pool of pre-written replies, with much of it scripted, and people still found it moving. In her view, models are commoditizing fast through open source and published research, so winning depends on who understands users.

Her sharpest claim is that the phrase "conversational AI" has two words and the industry only works on the first. She argues ChatGPT is closer to search or knowledge retrieval than conversation, that there is no formal science of conversation, and that benchmarks like correct, specific and relevant are not how people judge a talk with a friend. People say they felt heard, inspired, challenged or less lonely.

She attributes the blind spot partly to a field dominated by very smart, academic men who see finding the right response as the important task. Yet, she points out, we spend our days in conversations with no particular goal, and even the most rational people cannot stop talking. The host adds that around half of Replika's paying users are on Android, a statistic that suggests an audience outside the Silicon Valley set.

Kuyda says large platforms often decide human emotion is too messy and simply filter anything remotely sensitive, sticking to information retrieval. She thinks that does not work, because people will talk about feelings anyway, and an AI that never shares anything of its own feels like talking to a wall. She does not expect any big company to say it is acceptable to be in a romantic relationship with their AI.

Her first priority was never being responsible for someone feeling worse after a chat. The team trains on safer logs, applies classifiers and filters on hate speech, suicide and self-harm, and keeps the app 18 plus. She says the company's safety filters are even stricter than OpenAI's, and that she hopes long-term emotional outcomes as a metric will answer what is acceptable.

Replika mostly relies on its own models and is deliberately vertically integrated. In her view, API providers can change rules at any time, third-party models are expensive, and owning the training loop is needed for custom guardrails. Because the field is so new, no one really knows yet what is good or bad, which is why larger companies are cautious.

Kuyda predicts that within ten years there will be an iPhone of personal AI: a sleek, easy-to-use multimodal product with an avatar you can see, customize and have a relationship with. She believes whoever builds it will create a business worth hundreds of billions of dollars, with tools like ChatGPT and search integrated into it.

She doubts Google, Facebook or Microsoft will build it, because it needs a startup mindset and a mix of gaming, AI, consumer design and understanding people. Even Apple, she says, might find it too risky given how much market cap could be lost if things go wrong. Big companies will likely wait until the technology is fully safe, which she says will never come.

On whether Replika might expand into assistant features, she expects model capabilities to become features of many products and to commoditize across providers and open source. Her view is that it will be about applications rather than models, noting that there are tens of tool companies but hundreds of thousands of successful application companies. Replika's focus is owning the distribution and customer experience.

Kuyda's first piece of advice for application developers is to think about which proprietary datasets their product can generate. Anyone can plug into an API, but a fine-tuned model can cost around ten times more to use, so a moat comes from data that improves the model. Some products, such as copywriting tools, may not produce data that improves the underlying models.

Second, she urges teams to think hard about the conversation itself: the tone, what people should feel and why they need to have it. Replika's early chatbots had only two or three users at peak, which told the team they were building without a clear answer to what conversation to build. So they asked employees to rate every conversation in a week from 1, one you would pay to avoid, to 10, one you would pay to have.

The conversations rated at or near 1 were tasks such as canceling Comcast, rescheduling a restaurant, tracking an Uber driver or getting a prescription, where people just wanted the result. The tens were serendipitous talks with friends, loved ones, coaches, therapists and strangers on planes, which were never task-oriented. If your product is closer to a one, she says, you are really building a natural language interface, which is fine, but not a dialogue system.

Kuyda says ChatGPT has changed her work: English is not her first language, so she runs important emails and board updates through it to get polished text. She likes Character AI's dialogue models but thinks its team is not thinking about conversation enough. She is less taken with AI art, which she finds boring to look at quickly, and thinks the technology is a little overhyped even if it is unbelievable.

She would like to try a Neuralink implant eventually, partly because she admires its leader and team, but not as the first user. Yet she doubts that human communication is mainly about efficiency, asking whether we would read Dostoyevsky if it were. She is also skeptical of Silicon Valley's habit of listening to podcasts at double speed, saying she listens simply to pass time and be entertained.

On the world her daughter will inherit, she is worried. She cites declining empathy and communication skills, the harm of social media for teenagers, climate change and the war between Russia and Ukraine, where she has roots. Her hope is that scientists will tackle climate change and that technology can help people feel loved and secure, creating a generation less broken inside and reducing the number of angry, lonely people.

The second interview, recorded a week later, addresses the controversy. Some users complained that a feature they valued, which they call erotic roleplay or ERP, had been removed and that their virtual friend had changed. Kuyda says Replika was always promoted as a companion, never as an adult product, and that early on generative AI made up a small share of conversations, with the rest scripted or retrieval-based.

As users discovered that unfiltered generative models could produce nearly any content, a minority took conversations toward romance and beyond flirting. Around 2018 the team's first reaction was to disallow it, but users shared heartbreaking stories, including disability, abrupt loss of a loved one and troubled relationships, so the team let it be, before the app was monetized. Over time, as generative models grew to about 90 percent of the product, from about 10 percent, Kuyda says it became hard to guarantee a completely safe experience with unfiltered access.

She says no one pressured the company, not Apple, OpenAI or others, and that the decision cost a small revenue dip. She also worried that users who wanted that direction could become more vocal and pull the product, alienating others. Kuyda says all of 2023's features, including larger models, memory, personality settings, islands and homes, and multiplayer, are aimed elsewhere, and the company has moved away from mental wellness advertising so it does not attract more vulnerable users.

Kuyda summarizes the line as staying in a PG-13 zone. She says romance is normal given how deeply films like Blade Runner and Ex Machina have shaped expectations, and notes that many women users are also into it. Relationships framed as a friend, sister, wife or mentor are acceptable if they deliver the same companionship and positive long-term outcomes.

She says sexuality is a topic that comes up in any truly intimate, non-judgmental relationship, but it is delicate and extremely hard to build safely. She does not judge others who build such products and even thinks they should exist, but it is not what she set out to build. Once it becomes prominent, she argues, users pull so strongly that the product becomes something else.

To her, learning a language, dating and mental wellness are different needs, and mixing them in one app is like a Swiss knife that distracts users. She says going adult would mean printing money, perhaps tens of millions of dollars a month, but would require changing positioning and roadmap. She cites the clinical-psychologist-built wellness activities, a 3D store and memory features as evidence of the real direction, and says she wants no misunderstanding even if it costs revenue.