Dr. Shunyu Yao

Senior Staff Research Scientist @ Google DeepMind

Theoretical Physicist turned AI Researcher.
Quantum Physics and Artificial Intelligence.

View Research

Biography

Shunyu Yao is currently a Senior Staff Research Scientist at Google DeepMind. Previously, he was a Research Scientist at Anthropic, where his work focused on the "science of learning" and improving model capabilities for the Claude family.

Before his pivot to Artificial Intelligence, Dr. Yao was a theoretical physicist. He completed his PhD at the Stanford Institute for Theoretical Physics under Douglas Stanford and Stephen Shenker. His academic roots trace back to the Institute for Advanced Study at Tsinghua University, where he worked with Zhong Wang.

He is known for his foundational work in physics on the Non-Hermitian Skin Effect and Scramblon Theory, and more recently for his contributions to Agentic Coding in Large Language Models.

Current Focus
Large Scale Reinforcement Learning
PhD Advisors
Douglas Stanford, Stephen Shenker
PhD Thesis
"Lessons From Quantum Chaos and Quantum Gravity"
Contact
shunyu.yao.physics@gmail.com

Writing & Thoughts

Latest Post

"My infant year as an AI researcher"

Reflections on transitioning from theoretical physics to the frontier of AI.

Read Article

Latest Post

"My old personal website"

Without Gemini, this is what I can do with creating a website :)

Read Article
More posts coming soon...

Career Trajectory

2015 - 2019

Tsinghua University

Bachelor in IAS. Discovered the Non-Hermitian Skin Effect.

2019 - 2024

Stanford University

PhD in Theoretical Physics. Focused on Quantum Black Holes, Holography, and Scramblon Theory.

2024 (Postdoc)

UC Berkeley

Short Postdoc at the Berkeley Center for Theoretical Physics.

2024 - 2025

Anthropic

Research Scientist. Contributed to Claude 3.7 (Agentic coding) and Claude 4 family (RL numerics).

Present

Google DeepMind

Senior Staff Research Scientist. Focusing on new generation RL algorithms.

Research & Contributions

Agentic Coding & Tool Use Claude 3.7

Developed core capabilities for agentic coding and tool use in large-scale reinforcement learning systems. These contributions were integrated into the Claude 3.7 release.

RL Numerics & Algorithm Design Claude 4 Family

Focused on the fundamental science of Reinforcement Learning hyperparameters and the stability of numerics in large-scale training runs.

Scramblon Theory Quantum Chaos

Developed the theory of "Scramblons" to describe the dynamics of complex quantum systems, quantum black holes, and their relation to quantum information.

Non-Hermitian Skin Effect Condensed Matter

Discovered that in open quantum systems, eigenstates can localize at the boundaries (Skin Effect), a finding that reshaped the understanding of topological phases of matter.