OpenAI's Yao Shunyu Joins Tencent as Chief AI Scientist

Yao Shunyu, a former OpenAI scientist, has joined Tencent, taking on dual roles as Chief AI Scientist and Head of the AI Infra Department and Large Language Model Department. His appointment follows earlier speculation within the AI community regarding his next career move.

Tencent has confirmed an internal architectural upgrade to its large model research and development system. This restructuring includes the creation of new departments: AI Infra, AI Data, and Data Computing Platform. These departments are designed to bolster capabilities across computing power, data, and platform infrastructure for large models.

Yao will report to Tencent President Martin Lau in his capacity as Chief AI Scientist within the CEO/President's Office. Concurrently, as Head of the AI Infra Department and Large Language Model Department, he will report to Dowson Tong, President of the Technology Engineering Group.

Academic Background and Research

Yao is an alumnus of Tsinghua University's Yao Class, where he earned his bachelor's degree in computer science. He received a silver medal in the National Olympiad in Informatics in 2014 and was admitted to Tsinghua with the third-highest science score in Anhui Province in 2015. After graduating in 2019, he pursued a Ph.D. at Princeton University, completing it in 2024 before joining OpenAI.

His primary research focus has been on "agents." At OpenAI, he investigated language agents for digital automation, contributing to projects such as WebShop, SWE-bench, and tau-bench. His work includes contributions to concepts like ReAct, Reflexion, Tree of Thought, SWE-agent, and CoALA. According to Google Scholar, his papers "ReAct" and "Tree of Thought" have garnered over 4,000 citations each, contributing to a total of nearly 16,000 citations for his work.

Yao's doctoral dissertation systematically summarized the value of language agents in transitioning from "next token prediction" to "digital automation," proposing new benchmarks, methodologies, and frameworks. He publicly shared his Ph.D. defense on Bilibili, where he also discussed his collaboration with his Ph.D. advisor, Karthik Narasimhan. Narasimhan, a co-author of the GPT paper, served as a visiting researcher at OpenAI from 2017 to 2018.

AI Development Philosophy

In April, Yao articulated his perspective on AI development trends, stating that reinforcement learning has become effective and that "evaluation" will surpass "training" in importance. He suggested that AI has entered a "second half," shifting focus from "problem-solving" to "problem-setting." He emphasized that "evaluation will be more important than training" is a key trend.

Yao highlighted the need to define what AI should accomplish. He believes that success in the new era of AI requires adapting mindsets and skill sets to align more closely with a product manager's role: defining problems, setting metrics, organizing iterations, and translating AI capabilities into measurable real-world value.

Recent research from OpenAI also supports this view, indicating that evaluation methods are crucial for addressing model hallucination and can unlock further potential in large models.