Noam Shazeer, Transformer Co-author, Leaves Google for OpenAI

Noam Shazeer has switched houses again. Co-author of the paper that created the Transformer and, most recently, VP of engineering for Google's Gemini division, he announced his departure for OpenAI, where he assumes the role of research lead for architecture. The news captures attention not because of the title he holds, but because of what he designs. Shazeer's fingerprints appear on some of the core building blocks that sustain nearly every large language model in operation today.

Who is Noam Shazeer

Shazeer is one of eight authors of "Attention Is All You Need," the 2017 paper that introduced the Transformer and opened the technical path to today's LLMs. His contributions extend well beyond that paper. He worked on Mixture of Experts, a technique that activates only part of a massive model at any given time, and on Multi-Query Attention, which cuts memory costs when generating text. These are architectural choices, not implementation details. They change what a model can run and at what cost.

When people say "the model improved," much of the credit belongs to decisions like these—made years earlier by people who design the structure beneath training.

The Numbers That Reports Cited

The move gains weight from recent history. Shazeer had left Google to found Character.AI. According to technology industry reports, Google paid roughly $2.7 billion in 2024 to bring him back, in a deal that included licensing Character.AI's technology and reintegrating the team. Also per those reports, he stayed in this new stint for less than 22 months before announcing his move to OpenAI.

Those numbers come from reports, not from 10Dobro Prod or official company statements. Worth noting: acquisition figures and tenure lengths often vary between sources and can be adjusted later. What becomes clear beyond doubt is the direction of the move. A central figure in AI architecture switched sides again.

What This Says About the Market

The easy read is "talent war." True, but shallow. The technical point is more useful for decision makers.

Foundation models are converging on raw capability. What separates a good system from a mediocre one has shifted to other layers: how the architecture is designed, how the model is trained and fine-tuned, and how it integrates into real operations. Companies pursue people like Shazeer because competitive advantage now lives more in system design than in model size alone. Whoever designs the structure sets the ceiling on what the product can deliver.

For the manager evaluating whether to adopt AI, there is a practical lesson. The model is becoming a commodity. The difference lies in who integrates it, how the system converses with your data and workflows, and who answers when something breaks.

Where This Touches Our Work

At 10Dobro Prod, the thesis is the same, at a smaller scale and without billion-dollar spotlight: the edge isn't just the model, it's who designs the system and how it enters operations. We use models available on the market, but value lives in the architecture around them, in human checkpoints, and in integration with the client's real workflow.

That's why we talk about multiplying what a good team already delivers, not replacing it. The same models are available to everyone. What changes the result is the hand that designs the system and the data foundation that sustains it. Shazeer's career move is an expensive reminder of this: talent and architecture drive the market, more than the model name on the box.