Noam Shazeer

Overview

Noam Shazeer is an AI researcher who spent most of his career at Google. He is the second-listed author of “Attention Is All You Need,” the 2017 paper that introduced the Transformer,^[3] and he is the first author of a paper on sparsely gated mixture-of-experts layers.^[2]

Education and early career

Shazeer studied mathematics and computer science at Duke University from 1994 to 1998, winning Putnam prizes with the Duke team. As a member of the U.S. team at the 1994 International Mathematical Olympiad, he won a gold medal with a perfect score.^[1] He joined Google in 2000, and his early work there included improving the search spelling corrector.^[1]

Key contributions

In January 2017 Shazeer published “Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer” as first author, with Geoffrey Hinton and Jeff Dean among the co-authors. The work used up to thousands of experts, and models of up to 137 billion parameters.^[2]

A mixture-of-experts layer contains many small sub-networks, called experts, and a gating step that chooses only a few of them to process each input. The model can therefore hold far more parameters than it uses on any single input, which keeps the computation for each input manageable.

In June 2017 he was the second author of “Attention Is All You Need.” In October 2019 he co-authored the T5 paper, “Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer,” which casts many language tasks as text in, text out. This lets a single model and a single training recipe cover many different language tasks.^[4]

Positions

At Google, Shazeer and Daniel De Freitas built the chatbot Meena. Google declined to release it publicly, and in October 2021 Shazeer left to co-found Character.AI, where he served as CEO.^[1]

On 2 August 2024 he returned to Google to join Google DeepMind, together with De Freitas. The move came as Google signed a non-exclusive license to Character.AI’s technology, a deal reported at $2.7 billion.^[5] Shazeer became a technical lead and co-lead of Gemini, Google’s family of AI models.^[5]

On 18 June 2026 he announced that he was leaving Google for OpenAI. OpenAI’s Chief Research Officer, Mark Chen, said Shazeer would lead AI architecture research there.^[6]

Impact

Shazeer’s work touches two lines of research that shaped later language models: the Transformer, and mixture-of-experts layers, which add model capacity without using every parameter on every input. His career has moved between Google, the startup Character.AI and OpenAI.^[5][6] The two ideas can be read as complementary: one changed how a model processes sequences, and the other changed how much capacity a model can have without using all of it on every input. Both remain common topics in discussions of how language models are built and scaled.^[2][3]