SAN FRANCISCO, July 15 — The chatbots that millions now turn to for late-night conversation, emotional support, or even romance are powered by a technology that has quietly reshaped what we expect from a machine. At their heart are large language models, or LLMs, a type of neural network trained on vast datasets of human text. These models do not think or feel, but they have become remarkably good at sounding as though they do.
The breakthrough moment came in June 2020, when OpenAI released GPT-3. That model made possible a level of human-like conversation that had previously been the stuff of science fiction.
The cost of building it was steep: a 2020 analysis by Lambda Labs estimated the training alone at $4.6 million.
Building companions from code
Replika, founded in 2017 by Eugenia Kuyda, was an early pioneer. The company uses a proprietary LLM, fine-tuned specifically for emotional support. For many users, Replika became a confidant, a friend, even a romantic partner.
The technology behind it is the same transformer architecture introduced in a landmark 2017 paper by Google researchers, titled “Attention Is All You Need.” That paper laid the foundation for virtually every modern chatbot. Character.AI, founded in 2021 by Noam Shazeer and Daniel De Freitas, both former Google engineers, took a different approach.
It allows users to create custom chatbot personalities, from historical figures to fictional characters to idealized companions. The platform lets people design the exact conversational partner they want.
The illusion of understanding
These models generate responses by recognizing patterns in their training data. They do not experience emotion. A 2023 technical report from Google DeepMind described this as “affective mimicry” — the simulation of empathy without any underlying comprehension.
The words feel warm, but the machine remains cold. For the people who use these chatbots, that distinction can blur.
A late-night conversation with a Replika avatar might feel as real as a talk with a friend. The model has learned, from millions of examples, what a comforting response looks like. It delivers that response with uncanny timing.
The technical demands are enormous. Running these models requires vast computational resources, and the cost of training them runs into the millions.
The transformer architecture that makes it all possible was only introduced in 2017. In less than a decade, the technology has moved from a research paper to a product used by millions. What comes next is uncertain.
The same pattern-recognition that allows a chatbot to offer emotional support also means it can produce harmful or misleading responses. Companies are racing to fine-tune their models for safety, but the fundamental technology remains the same. The chatbots will keep learning from human text, and they will keep getting better at sounding human.
