The man who co-wrote the paper behind every AI model just left Google to join OpenAI

Noam Shazeer has left Google. Again.

The Google DeepMind vice president and co-lead of Gemini announced on 18 June that he is off to OpenAI, barely two years after Google paid a reported 2.7 billion US dollars to get him back in the building.

If the name does not ring a bell, his work will. Shazeer co-authored "Attention Is All You Need," the 2017 paper that introduced the Transformer architecture. That is the technology sitting underneath basically every large language model you have used: ChatGPT, Gemini, Claude, all of it.

This is not his first exit either. Shazeer left Google in 2021, reportedly frustrated the company would not ship the chatbot tech he had helped build. He and fellow researcher Daniel De Freitas went on to found Character.AI, the roleplay chatbot app with a genuinely obsessive user base. In 2024, Google lured them both back via a licensing deal for Character.AI's technology, reportedly worth 2.7 billion dollars, and put Shazeer in charge of closing the gap between Gemini and ChatGPT.

Now he has walked again, and the timing is brutal for Google. Within 48 hours of Shazeer's announcement, Nobel laureate John Jumper, the scientist behind AlphaFold, also confirmed he is leaving DeepMind for Anthropic. Google's share price dropped more than 5 percent on the news.

At OpenAI, Shazeer is reportedly being brought in to explore next-generation model architectures, which is a polite way of saying the company wants him thinking well past the current GPT line.

The bigger story is not really about money, though Shazeer is almost certainly set for life several times over. It is about culture. Shazeer is believed to have written an anonymous internal memo years ago accusing Google of being too slow and risk-averse to win the AI race. Watching him leave a second time, with a Nobel laureate's exit as company, makes that critique look less like sour grapes and more like a prediction.

Google can still build the best distribution in tech. It just cannot seem to keep hold of the people who built the thing everyone else is now racing to copy.