Yuan Cao

Co-founder & CEO, Unreasonable Labs

About

Yuan Cao is co-founder and CEO of Unreasonable Labs, a Cambridge, Massachusetts startup building an AI discovery engine that composes new knowledge across chemistry, materials science, biology, and beyond. He was previously a Senior Staff Research Scientist at Google DeepMind, where he made core contributions to the Gemini model family and Google Search's AI Mode, and co-authored the influential reasoning frameworks ReAct and Tree of Thoughts. His earlier career spans Google, Baidu, and the speech-AI startup KITT.AI, following a PhD in Electrical and Computer Engineering from Johns Hopkins University. In 2025 he left DeepMind to co-found Unreasonable Labs with MIT professor Markus Buehler; the company emerged from stealth in March 2026 with a $13.5 million round led by Playground Global to build 'superintelligence for knowledge discovery.' He argues that the real bottlenecks of AI for Science are neither compute nor data but physical-world verification speed and the invention of new concepts, and advocates hybrid LLM-plus-symbolic approaches to push models beyond recombining existing knowledge.

Key Contributions

Co-founded Unreasonable Labs (2025) with MIT professor Markus Buehler, building an AI discovery engine that composes new knowledge across chemistry, materials science, and biology
Core contributor to Google's Gemini model family and Search AI Mode as Senior Staff Research Scientist at Google DeepMind
Co-author of Tree of Thoughts and ReAct, two influential frameworks for LLM reasoning and deliberate problem solving
Raised a $13.5 million round led by Playground Global (2026) to scale 'superintelligence for knowledge discovery'
Advocates hybrid LLM-plus-symbolic approaches to AI for Science, arguing the field's real bottlenecks are verification speed and concept invention

Videos & Interviews

对话前DeepMind曹原：AI for Science爆发，一个新时代到来了

Silicon Valley 101 (硅谷101) host Chen Xi talks with Cao Yuan the week Jeff Dean left Google to found Discovery Loop. Cao argues the AI for Science explosion is real but bottlenecked by neither compute nor data: physical-world verification can't keep pace with model iteration, and inventing genuinely new concepts — the last mile of AGI — may not even be a Turing-computable problem. Along the way: why math and code move faster than biology and materials, automated labs and research agents, and his bet on hybrid LLM-plus-symbolic systems at Unreasonable Labs.

View Details

Papers & Publications

Tree of Thoughts: Deliberate Problem Solving with Large Language Models

2023

Generalizes chain-of-thought into a search over trees of intermediate reasoning steps, letting language models explore, evaluate, and backtrack while solving problems

Read Paper

ReAct: Synergizing Reasoning and Acting in Language Models

2022

Interleaves reasoning traces with actions so language models can plan, consult tools, and adjust — a foundation of today's agent designs

Read Paper