A recent revelation about Google’s missed opportunity in generative AI has sparked intense discussion in the tech world. Thibault Sottiaux, the engineering lead for OpenAI’s coding tool Codex, disclosed on social media platform X that Google’s AI lab DeepMind had developed a chatbot with nearly identical functionality to ChatGPT a full year before its release — but the product was ultimately shelved internally due to fears it would disrupt Google’s search business.
Sottiaux was personally involved in the project. In a reply to an old interview with Jeff Dean, former head of Google Brain, he said the project was initially called LMChat and later assigned a new internal codename, with an experience already very close to what ChatGPT would offer a year later.
Sottiaux expressed frustration in his reply, noting that DeepMind did not have the authority to release the product at the time, because large language models could potentially threaten Google’s search business, and management was highly nervous about it, repeatedly withholding approval. He lamented: “I think about this a lot.”
From Google to OpenAI: A Missed Chapter
According to Sottiaux, he joined Google in 2015, initially working on Google Maps before moving to DeepMind. At the time, DeepMind was still a fairly pure research lab, with AlphaGo, reinforcement learning, and artificial general intelligence research at the peak of attention. At DeepMind, Sottiaux was primarily responsible for building the AI and machine learning infrastructure used by researchers.
In November 2022, ChatGPT burst onto the scene. After seeing it, Sottiaux said he “collapsed.” In 2024, he ended his nearly nine-year tenure at Google and officially joined OpenAI, where he now leads the Codex project. His mastery of “infinite tokens” has earned him the nickname “cyber godfather” from netizens.
Google co-founder Jeff Dean had previously publicly acknowledged that Google had developed internal chatbots before ChatGPT’s release, but the company concluded the product was not better than Google Search and therefore never launched it. Sottiaux’s revelation adds another layer to the decision-making at the time — concern about cannibalizing internal business.
Notably, Demis Hassabis, head of Google’s AI division, also said in his biography that OpenAI did not make a disruptive innovation — many people at the time recognized the potential of large language models, but OpenAI CEO Sam Altman simply had the boldness to release it.
Search Becomes the Straitjacket on Google’s AI Innovation
Analysts point out that Google’s conservative strategy in generative AI is closely tied to its core business. Google Search set the bar too high as a baseline — if a new product merely served search, Google Search was already highly mature. Meanwhile, in 2023, large models got almost everything wrong; anything involving time-sensitive content was pure hallucination. Google set its own acceptance standard as “better than Google Search” — a standard that ultimately cost the company dearly.
Some commentators have compared Google’s predicament to Xerox in the 1970s. Xerox’s Palo Alto Research Center invented the graphical user interface, the mouse, and Ethernet, but abandoned these innovations because they couldn’t directly drive printer sales — ultimately watching Steve Jobs visit, return to develop the Macintosh, and usher in the personal computer era.
Talent Exodus Intensifies: Google AI Suffers Consecutive Losses
The demise of LMChat was not an isolated incident. Recently, Google’s lag in AI coding has triggered an even more dramatic talent shakeup.
On June 18, Noam Shazeer, a core co-author of the seminal Transformer paper “Attention Is All You Need” and a contributor to the Gemini model, announced he was joining OpenAI. Shazeer left Google in 2021 to found Character.AI, and in 2024 Google spent $2.7 billion to buy back the company and bring him back, making him co-lead of the Gemini project. Yet that massive $2.7 billion investment couldn’t retain him for even two years.
On June 19, John Jumper, vice president at DeepMind, announced he was heading to Anthropic. Jumper was a key lieutenant promoted by Hassabis — entrusted with leading the AlphaFold team just six months after completing his PhD. In 2024, he and Hassabis jointly won the Nobel Prize in Chemistry for their groundbreaking work on AlphaFold in protein structure prediction. Now, the Nobel laureate has chosen to join a competitor.
On June 24, two more names were added to Google’s talent loss list: Jonas Adler and Alexander Pritzel. Both are core contributors to Gemini — Adler led Google’s AI coding efforts, Pritzel led pretraining — and both worked on AlphaFold alongside Jumper. Their destination is also Anthropic.
PersonFormer RoleDestinationDateNoam ShazeerGemini Co-LeadOpenAIJune 18John JumperDeepMind VP, AlphaFold LeadAnthropicJune 19Jonas AdlerGemini Core Contributor, AI CodingAnthropicJune 24Alexander PritzelGemini Core Contributor, PretrainingAnthropicJune 24
This series of personnel changes reflects Google’s defensive position in the generative AI race. From holding the “max-level gear” of the Transformer paper to falling behind OpenAI and Anthropic in emerging battlegrounds like AI coding, Google’s situation has been jokingly described by netizens as “America’s steamed bun” — a reference to a Chinese internet meme about being soft and easily squeezed.
One industry observer raised a thought-provoking question: If Google had released GPT first, how different would the AI landscape be today? While that question is difficult to answer, Google’s lesson of missing the window in the AI era due to an internally conservative culture is undoubtedly a cautionary tale for all tech giants.