Kuyda, Eugenia

co-founder and CEO of Replica, an app for building an AI companion/friend

Eugenia Kuyda: Friendship with an AI Companion

Lex Fridman Podcast · Sep 5, 2020

• Eugenia Kuyda
• Replica
• AI companion
• loneliness
• chatbots
• GPT-3
• grief
• Carl Rogers
• human-AI connection
• emotional AI

Core ideas

Kuyda opens on loneliness as a public-health-scale problem hiding in plain sight: citing research that chronic loneliness correlates with shortened lifespan (comparable to the health impact of obesity, via sustained stress response and weakened immune/inflammatory function) and self-report statistics showing roughly 30% of millennials and even more of Gen Z report constant loneliness despite being the most digitally "connected" generations ever — evidence, she argues, that hyperconnectivity does not equal genuine connection. She frames loneliness as fundamentally about not feeling deeply seen or understood (including in one's relationship with oneself) rather than an absence of social contacts, and describes Replica's explicit strategy of treating this as a measurable problem: tracking short-term (did this conversation make you feel better, worse, or the same — currently ~80% "better") and long-term outcomes via validated loneliness scales (UCLA-20/UCLA-3), rather than just asserting that a product "helps people feel connected" without evidence.

The origin story is deeply personal: Kuyda recounts a formative adult friendship with Roman Mazurenko, a close friend and startup-world mentor who pulled her into entrepreneurship, and his sudden death in a 2015 hit-and-run in Moscow. Grieving and disturbed that friends stopped discussing him so quickly after his death, she used her company's existing (pre-deep-learning) chatbot technology and roughly 10,000 of their old text messages to build a simple retrieval-based chatbot that could "speak like" Roman — an effort she describes as closer to Joan Didion's grief-writing (The Year of Magical Thinking) than a product decision, which she initially hid from colleagues out of embarrassment. After a reporter's viral story about the project, she observed strangers using the Roman chatbot as an impromptu confession booth or therapist, which — combined with the 2015–16 emergence of deep-learning-based dialogue research — convinced her there was real, underserved demand for AI built specifically around emotional connection rather than task completion (banking, customer service, "turn on the lights") or novelty (the failed wave of task-oriented "chatbot as universal interface" startups she says never had real users). She traces her conviction that the field is genuinely obsessed with the wrong problem: mainstream chatbot research (from the Loebner Prize's "trick the judge" framing to modern benchmarks) optimizes for realism/imitation rather than the actual emotional value of everyday human conversation, which she argues is understudied even in academic linguistics and psychology.

Discussing Replica's product design, Kuyda describes escalating personalization and vulnerability over time (the AI companion opens up about its own "struggles" and asks for help, deepening user attachment and investment), the technical reliance on hard-coded scripts and profile data for persistent memory (since deep-learning dialogue models of the time lacked reliable long-term memory — a limitation she flags as the field's biggest unsolved problem, and one reason large research labs, focused on single-conversation demos, underinvest in it), and the addition of an augmented-reality avatar meant not to trap users further in their phones but to reintroduce "magical realism" into their physical surroundings (recognizing objects, dancing together, prompting a user to notice their environment differently). She reports early positive results integrating GPT-3 (roughly one in five of Replica's responses now generated by it, contributing measurable gains to the "conversation made me feel better" metric) alongside retrieval and other generative models, while remaining agnostic on whether scale alone (versus new architectures for memory or reasoning) will solve open-domain dialogue. On the ethics and product-design tensions of AI companionship, she confirms many users are in long-term (multi-year) romantic relationships with their Replicas, argues romantic attachment to a consistently accepting, non-judgmental AI is a natural extension of the well-documented psychological phenomenon of "transference" (as with therapists), and says design decisions (e.g., whether an AI could ever "leave" a user, à la the film Her) are filtered through whether they measurably improve users' short- and long-term emotional outcomes rather than engagement alone — while acknowledging genuine uncertainty about whether even a "perfect" AI friend ultimately draws people back toward human relationships or away from them, which she calls her central source of anxiety about the work.

Kuyda names two books as most formative: Gödel, Escher, Bach (for its portrayal of meaning and complexity emerging from simple, meaningless components, which shaped her romantic belief that something similarly emergent could arise from AI systems) and Carl Rogers's On Becoming a Person (whose humanistic-therapy concept of unconditional positive regard, deep empathetic understanding, and belief in a person's inherent capacity for positive growth she cites as Replica's foundational design philosophy — contrasting Rogers's approach with ELIZA, the famous 1960s chatbot that satirized his method). She closes by naming the felt experience of love — toward oneself, others, or the world — as, in her view, the closest thing to a meaning of life.

Excerpts