A nine-year-old in Brazil set out to build a chatbot. His parents had encouraged his interest in computing, and the young Daniel De Freitas began assembling a database from scratch. He could get the program to seem conversational for a while. Then he saw the limit: the technology available in the late 1990s could not scale into the sort of exchange he wanted. He put the project down. The question remained.
That childhood experiment is unusually useful for understanding his career. It contains both sides of De Freitas's work: a fascination with the tiny drama of one person talking to a machine, and the engineering question of how the machine might handle conversations it has never seen. Years later, he would bring that problem into a major research lab, then into a startup where anyone could try the answer.
De Freitas grew up in São Paulo and studied computer engineering at the University of São Paulo. After graduating, he moved to the United States in 2012 and worked at Microsoft. Part of that work involved software for Bing to produce responses to search queries. It was a practical setting for an engineer interested in the moment between a question and an answer. Search gives a user a destination; conversation asks the software to stay for another turn.
The side project that kept coming back
In 2016, De Freitas joined Google and worked on machine learning for YouTube. He had read research from Google Brain about systems trained on movie subtitles. It gave him a reason to revisit his old idea with newer tools. Through Google's allowance for engineers to pursue relevant projects outside their daily assignments, he began building a chatbot called Meena. He later said the name had come to him in a dream. Research programs do not often arrive with that kind of naming story.
The technical ambition was straightforward to say and difficult to deliver: keep a conversation going when the subject could be almost anything. A scripted bot can guide a customer through a narrow task. An open-domain bot has to meet whatever the other person brings to the exchange. It has to remember the thread, offer a response that belongs in it, and make enough sense that a second question is worth asking.
De Freitas and his Google colleagues described Meena publicly in 2020. The model had 2.6 billion parameters and was trained on 341 gigabytes of filtered public-domain social-media conversations. The numbers mattered because this was an attempt to find patterns across an enormous range of ordinary talk. Yet the research also showed a concern with qualities that anyone can recognize at a dinner table: did the reply make sense, and was it specific to what had just been said?
Their measure, called Sensibleness and Specificity Average, forced a distinction that chatbot demos can glide past. “That's nice” is a reasonable reply to many things. It is also a way of saying almost nothing. A particular response, one that fits the actual exchange, is the beginning of something more engaging. In the published evaluation, the full Meena system reached a score of 79 percent, compared with 86 percent for an average human in the same setup. Those figures describe one research test, not a claim that software had become a person.
Meena even produced a joke about horses going to “Hayvard.” Asked how a system trained to predict words might reach such a line, De Freitas explained the stages by which it learned to move beyond simple repetition toward more plausible combinations. He described the process without turning the pun into magic. That combination of curiosity and mechanical explanation runs through the way he spoke about the work.
“The more time I spent in it, the more it became kind of humanlike.”Daniel De Freitas, recalling his childhood chatbot
From a lab conversation to a public one
At Google Brain, De Freitas connected with Noam Shazeer, whose work on language models and the Transformer architecture helped shape the field. Together they developed LaMDA, another effort focused on dialogue. By then the young coder's private question had become a large team project. The work looked at whether a language model could move through open-ended exchanges rather than deliver a single isolated answer.
The pair wanted more people to use the technology. Google was cautious about releasing a public version, and its own Meena paper acknowledged questions around harmful outputs and bias while explaining why it was withholding an external demo. De Freitas and Shazeer left Google in 2021 and formed Character.AI. Their decision took the research problem into a very different environment: a product where the public would decide what conversations they wanted to have.
Character.AI opened a public beta in September 2022. Users could create and talk to AI Characters, defining a persona and letting an exchange unfold. A character might serve a creative writing scene, a lesson, a bit of improvisation or a conversation for its own sake. The model beneath it was important, but the interface made an equally strong bet: people wanted to shape the voice on the other side of the screen.
When the company introduced itself that December, it said the two-month-old beta was generating one billion words a day and that users had made more than 350,000 Characters. By March 2023, the company reported more than two billion messages sent in five months, with the second billion arriving in the final month of that period. Investors led by Andreessen Horowitz put $150 million into the company at a reported $1 billion valuation. The numbers made clear that open-ended conversation was an activity people would return to, not merely a demonstration they would sample once.
De Freitas's explanation of the product was practical. He wanted people to use and customize AI for what they needed. The company trained its own language model and built the interface around it, giving the team a way to tune both the engine and the experience. To a user, the result appeared simple: choose or make a Character, type a line, wait for another. Behind that apparent ease sat the older problem of sensible, specific replies.
A joke with a point
He could be dryly funny about the rivalry his company had created. Asked in 2023 about Google's Bard chatbot, De Freitas replied, “We're confident Google will never do anything fun. Because we worked there.” The line worked because Character.AI had chosen a playful form. It made room for roleplay and invention in a field where products often introduced themselves as assistants for tasks.
The joke also marked a serious difference in product instinct. A general assistant is usually asked to be useful and correct. A character can be interesting, idiosyncratic and responsive to the scene a user makes. That openness gives people creative control, while putting real pressure on the people who build the system to make its nature clear. Character.AI's conversations carried a reminder that Characters' statements were made up. The product invited imaginative use and needed users to understand the frame.
It is easy to describe De Freitas as a founder who spotted a market. His timeline is stranger and more patient than that. He tried a chatbot before large language models existed, studied engineering, worked on search responses, built a Google side project, published dialogue research and then founded a company. Each stage made the next one more plausible. The pursuit was continuous even when the machines, organizations and audiences changed.
A first chatbot experiment in Brazil reaches the limits of its era.
Google publishes the Meena work, with De Freitas as a lead author.
He co-founds Character.AI; its public beta opens the next year.
He joins Google DeepMind after a licensing agreement with Character.AI.
The return trip
In August 2024, Google and Character.AI reached a non-exclusive licensing agreement covering the startup's technology. De Freitas, Shazeer and some colleagues went to Google DeepMind. Character.AI continued as a separate company. For De Freitas, the move closed a loop on a résumé: Google researcher, founder, Google researcher again. The work that tied those positions together was still conversation.
The circumstances were different on the return. When he had built Meena, a public dialogue model was a research proposal whose release raised difficult questions. By 2024, people had spent years making and using AI Characters. They had shown that conversation could be a creative medium, a study aid, a way of playing with language, and an everyday product habit. They had also made the obligations of its builders more concrete. The lab could no longer imagine public use only in the abstract.
De Freitas's documented story offers no tidy ending. The nine-year-old's original program ran into a wall; Meena made measurable progress; Character.AI put a more open experience in users' hands; Google brought him back. What stays constant is the idea behind the first attempt: a computer response ought to meet the person who made the opening move. Making that happen takes more than a clever sentence. It takes memory, judgment, engineering and, as his career suggests, an uncommon willingness to try the conversation again.