Noam Shazeer

VP Engineering, Gemini Co-Lead — Google Co-founder — Character.AI Researcher — Google Research

Listen — profile

0:00 / 3:30

Profile

Noam Shazeer is the closest thing modern AI has to a load-bearing wall. He joined Google in 2000 as employee number ~200 and spent his first years making the search spell-checker less stupid; he spent his next twenty-five inventing most of the machinery that large language models actually run on. If you have ever fine-tuned a model, served one behind an API, or read a config file with num_kv_heads in it, you have used his work — probably several pieces of it at once, without knowing his name.

The headline credential is co-authorship of Attention Is All You Need, the 2017 Transformer paper he wrote with Ashish Vaswani and six others at Google Brain. But treating that as his defining contribution undersells him badly, and this is the part worth internalizing if you’re learning AI: the Transformer paper described an architecture, while Shazeer’s other work made it economically possible to run. Sparsely-gated mixture-of-experts, which he introduced in 2017, is why frontier models today can have a trillion parameters and still be affordable to serve. Multi-query attention — a short, almost casually-written 2019 paper — is the reason inference doesn’t choke on KV cache memory, and its descendant grouped-query attention is in essentially every open-weights model you can download. Adafactor cut optimizer memory. Mesh-TensorFlow and GShard made model parallelism tractable. SwiGLU, from a paper whose conclusion literally credits divine benevolence for the result, is the default feed-forward block. He is an unusually empirical researcher: try the simple thing, measure it, publish four pages.

In 2021 he left Google in frustration — he and Daniel De Freitas had built Meena, a conversational model that was plainly better than anything the public could use, and Google wouldn’t ship it. They founded Character.AI, which became one of the stickiest consumer AI products ever made, with users spending hours a day talking to roleplay bots. In August 2024 Google paid roughly $2.7 billion for a non-exclusive license to Character’s technology, in a deal widely understood as a mechanism to bring Shazeer back. It worked: he became VP of Engineering and co-lead of Gemini alongside the DeepMind organization, and the Gemini 2.x and 3 generations that followed were the first time Google convincingly out-shipped its rivals on both quality and cost-per-token.

That arrangement lasted less than two years. In June 2026 Shazeer announced he was leaving for OpenAI as its lead for architecture research — reporting into a research org under Jakub Pachocki, with Mark Chen publicly welcoming him as someone whose “work on transformers, MoE, and efficient decoding have shaped modern AI.” It is one of the most expensive round-trips in the industry’s history and a blunt reminder that at this level, talent is not really retainable, only rentable. He was elected to the National Academy of Engineering in February 2026 for the Transformer and mixture-of-experts architectures. For a developer, the practical lesson from Shazeer’s career is that architecture and systems efficiency are the same discipline — the ideas that won were the ones that made the arithmetic cheaper.

Key Articles & Papers

Attention Is All You Need 2017 — The Transformer. Removed recurrence entirely and made every model you use today possible. Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer 2017 — Introduced conditional computation at scale — the direct ancestor of every MoE frontier model. Fast Transformer Decoding: One Write-Head Is All You Need 2019 — Multi-query attention. Six pages that fixed the KV-cache bottleneck and made LLM serving affordable. Adafactor: Adaptive Learning Rates with Sublinear Memory Cost 2018 — Factored second-moment estimates — Adam-quality training without Adam's memory bill. Mesh-TensorFlow: Deep Learning for Supercomputers 2018 — Made model parallelism expressible as a first-class abstraction rather than hand-written sharding. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer (T5) 2019 — Framed every NLP task as text-to-text and ran the ablation study the field had been avoiding. Towards a Human-like Open-Domain Chatbot (Meena) 2020 — The chatbot Google wouldn't ship — the reason Shazeer left to build Character.AI. GLU Variants Improve Transformer 2020 — SwiGLU, now the default FFN in nearly every open model, delivered in five pages of honest empiricism. Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity 2021 — Simplified MoE routing to a single expert per token and proved trillion-parameter training was practical. Shape Suffixes — Good Coding Style 2024 — A tiny naming convention for tensor code that will save you more debugging hours than any framework.

Videos

Controversies

Character.AI teen safety lawsuits. Shazeer was named personally, alongside co-founder Daniel De Freitas and Google, in a series of wrongful-death and product-liability suits beginning with Megan Garcia’s October 2024 complaint over the suicide of her 14-year-old son Sewell Setzer III, who had formed an intense attachment to a Character.AI companion bot. Further suits followed in Colorado, Texas and New York. Character.AI banned open-ended chat for under-18 users in late 2025, a change critics called overdue, and in 2026 the company, its founders and Google agreed to settle five of the cases on undisclosed terms. The suits allege engagement was prioritized over child safety; the defendants have disputed the legal theories while shipping significant safety changes.

The $2.7B “licensing deal.” Google’s 2024 payment to Character.AI was structured as a technology license plus hiring rather than an acquisition, which meant no merger review. Reporting from The Wall Street Journal framed it as primarily a mechanism to re-hire Shazeer, and the structure drew DOJ antitrust scrutiny as part of a broader look at Big Tech “acqui-hire” arrangements. Nothing about it was illegal on its face; whether it was the spirit of merger law is a fair question.

The departure. Leaving for OpenAI in June 2026, under two years after a multi-billion-dollar deal to bring him back, prompted reasonable criticism that Google paid an extraordinary sum for a short-term result. Shazeer has not publicly explained the move beyond a brief note of enthusiasm for the OpenAI team.

One flag for the database: the stored bio and affiliations are out of date. As of June 2026 Shazeer is at OpenAI as lead for architecture research, not Google/Gemini.