{"slug": "google-built-chatgpt-s-ancestor-in-2018-and-watched-its-creator-leave-twice", "title": "Google Built ChatGPT's Ancestor in 2018 and Watched Its Creator Leave Twice", "summary": "Google developed a ChatGPT-like chatbot called Meena in 2018, but executives blocked its release, leading creators Daniel De Freitas and Noam Shazeer to leave and found Character.AI, which grew to 20 million monthly active users by August 2025. Google later paid $2.7 billion to license Character.AI's technology and rehire the founders, but Shazeer left again on June 18, 2026, to join OpenAI, where he will lead AI architecture research.", "body_md": "*Google had a ChatGPT-like bot years before ChatGPT became the product everyone copied. The problem wasn't that Google lacked the technology. It was that the people who built it kept leaving.*\n\nAround 2018, Daniel De Freitas, then a research engineer at Google, started working on a side project with a simple aim: build a chatbot that didn't sound like a chatbot. It talked back. Noam Shazeer joined him soon after, bringing the credibility of a Google Brain researcher who had co-authored *Attention Is All You Need*, the 2017 paper that introduced the Transformer architecture. If you use a modern large language model, you're living with the consequences of that paper. The bot became Meena, a 2.6 billion parameter conversational model trained on 341GB of text from public domain social media conversations, according to Google's 2020 research post. It worked. Google said Meena was designed to give replies that were not only sensible, but specific to the conversation. Business Insider, summarizing Wall Street Journal reporting, later described the internal version more vividly: it could argue about philosophy, talk about television, and make puns about horses and cows. That detail matters. This wasn't a vague research ambition. It was a working demo of the product category that later made OpenAI famous.\n\nGoogle didn't ship it. According to the Journal account carried by Business Insider, executives blocked attempts by De Freitas and Shazeer to send the bot to outside researchers, add it to Google Assistant, or release a public demo. The stated concerns were safety and fairness, linked to the AI principles Google adopted in 2018 after employee protests over the company's military contracting work. The project later became LaMDA. It stayed inside Google while its creators kept pushing to release it.\n\nThey left in 2021. Sundar Pichai personally asked them to stay, according to the same Journal reporting, but De Freitas and Shazeer went anyway and founded Character.AI. Shazeer later explained the decision in an interview with Aarthi Ramamurthy and Sriram Krishnan: *It caused a bit of a stir inside of Google. But eventually we decided we'd probably have more luck launching stuff as a startup.*\n\nThey were right.\n\nCharacter.AI turned the idea loose. Users could talk to bots styled after Elon Musk, Mario, fictional characters, historical figures, you name it. By August 2025, Wired reported that the company had 20 million monthly active users spending an average of 75 minutes a day with its bots. That's a strange number to sit with. It means Google's shelved chatbot line became one of the stickiest consumer AI products on the internet once it escaped the building.\n\n## Google buys it back\n\nGoogle noticed, and the return was expensive. Reuters reported in August 2024 that Google signed a non-exclusive licensing deal for Character.AI's large language model technology and brought Shazeer, De Freitas, and some of the startup's researchers back to Google. The Information later reported, citing a regulatory filing, that Google paid Character.AI $2.7 billion in cash as part of the licensing and hiring deal. That is a large price for a lesson Google had already paid for once.\n\nLess than two years later, Shazeer left again. On June 18, 2026, he announced on X that he was joining OpenAI. Reuters identified him as Google's vice president of engineering and a co-lead of Gemini. Bloomberg reported that OpenAI's chief research officer Mark Chen said Shazeer would lead AI architecture research, while The Information reported that he would work on new model architectures and the evolution of the Transformer architecture. That is close to the work that made him important in the first place.\n\nSam Altman made the recruitment sound personal. In a public X post, he wrote that Noam was one of the people he had most wanted to work with since OpenAI began, adding that it *only took 10 years* and would be worth the wait. Google's response, given to Reuters, was shorter: the company said it was grateful for Shazeer's *meaningful contributions to Google over the years*. De Freitas stayed. His DeepMind profile still lists him as a research scientist.\n\n## What it costs Google\n\nHere's the part that should sting inside Google. The company had the chatbot before ChatGPT. It had one of the engineers behind the Transformer paper. It had the distribution - Search, Assistant, Android, all of it - the research staff, the Transformer co-author himself, and years of runway before OpenAI was anywhere near the mainstream. It chose caution. Some of that caution was defensible, because a public chatbot can say damaging things at scale. But caution also has a cost, and Google is still paying it.\n\nOpenAI shipped ChatGPT in November 2022 and changed the market. Google answered with Bard, then Gemini, and spent the next few years trying to turn its research strength into consumer momentum. Now the same researcher Google lost to Character.AI, then paid heavily to bring back, has crossed the road to OpenAI. Frankly, this was never just a story about one chatbot. It's a story about a company that kept building the future early and releasing it late.\n\n**Also read:** [DeepSeek's V4-Flash Undercuts OpenAI and Anthropic on Price Again](https://startupfortune.com/deepseeks-v4-flash-undercuts-openai-and-anthropic-on-price-again/), [Orchid's viral anniversary ad backfires into a debate over AI and intimacy](https://startupfortune.com/orchids-viral-anniversary-ad-backfires-into-a-debate-over-ai-and-intimacy/), and [Nvidia's AI Chip Demand Is Outpacing Supply 12 to 1, Says Dan Ives](https://startupfortune.com/nvidias-ai-chip-demand-is-outpacing-supply-12-to-1-says-dan-ives/)", "url": "https://wpnews.pro/news/google-built-chatgpt-s-ancestor-in-2018-and-watched-its-creator-leave-twice", "canonical_source": "https://startupfortune.com/google-built-chatgpts-ancestor-in-2018-and-watched-its-creator-leave-twice/", "published_at": "2026-08-02 12:36:52+00:00", "updated_at": "2026-08-02 12:54:16.510779+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-products", "ai-startups"], "entities": ["Google", "Daniel De Freitas", "Noam Shazeer", "Character.AI", "OpenAI", "Sundar Pichai", "Sam Altman", "Mark Chen"], "alternates": {"html": "https://wpnews.pro/news/google-built-chatgpt-s-ancestor-in-2018-and-watched-its-creator-leave-twice", "markdown": "https://wpnews.pro/news/google-built-chatgpt-s-ancestor-in-2018-and-watched-its-creator-leave-twice.md", "text": "https://wpnews.pro/news/google-built-chatgpt-s-ancestor-in-2018-and-watched-its-creator-leave-twice.txt", "jsonld": "https://wpnews.pro/news/google-built-chatgpt-s-ancestor-in-2018-and-watched-its-creator-leave-twice.jsonld"}}