Über Nacht ging die Nachricht viral: 27-jähriger Yao Shunyu verlässt OpenAI. Wird der Genie aus Tsinghuas Yao-Class zum Produktmanager?

Tencent has just refuted the rumor of "Yao Shunyu joining the company", but the industry hasn't breathed a sigh of relief:

Where the top intelligent agent talents go may determine the direction of the second half of the AI era. At the age of 27, graduated from the Yao Class of Tsinghua University, with a doctorate from Princeton University, and having worked on the core intelligent agent projects at OpenAI, his works have been cited over ten thousand times... He proposed that "evaluation will be more important than training", steering the research towards products and measurable values.

Yao Shunyu was rumored to have joined Tencent yesterday, and this morning, "The 37th Hand" and Tencent's "E Factory Blackboard" officially refuted the rumor!

Why has Yao Shunyu attracted so much attention from the industry? Why is there a rumor of a sky - high annual salary of 100 million RMB?

Naturally, it's due to Meta's sky - high poaching in Silicon Valley: Mark Zuckerberg offered 100 million US dollars to core researchers at OpenAI!

In this regard, 100 million RMB is definitely not a high price. Similarly, if he joins a domestic company, who can afford a salary package of 100 million US dollars?

He is definitely worth the 100 - million - dollar offer from Meta - he is a core contributor to Deep Research.

You should know that Hyung Won Chuang, a core contributor to OpenAI's Deep Research, has joined Meta's Super Intelligence Laboratory.

Since joining OpenAI in August 2024, Yao Shunyu quickly joined the core team and was deeply involved in the R & D of the first batch of intelligent agent products launched in 2025: Operator and Deep Research, as well as the development of Computer - using Agent.

This is a crucial step for OpenAI towards AGI Level 3, representing that OpenAI is translating the capabilities of large - scale models into practical products that can independently execute complex multi - step tasks, marking a key leap from "chatting" to "getting things done".

This refutation only shows that Yao Shunyu didn't join Tencent. Most likely, he has started a new career stage.

Currently, Yao Shunyu hasn't updated his personal position.

At only 27 years old, Yao Shunyu's departure from OpenAI is enough to indicate that a deeper migration and reorganization of the global AI talent ecosystem is taking place.

In the second half of the AI era, will he become a "product manager"?

Existing data shows that among the Chinese talents working at OpenAI, the top three undergraduate institutions are Tsinghua University, Peking University, and Shanghai Jiao Tong University, with double - digit numbers of students from each school.

Yao Shunyu graduated from Tsinghua University for his undergraduate studies.

Like other alumni from the Yao Class, he has been on a "straight - A student path" in his studies.

In junior high school, Yao Shunyu studied at Hefei No. 45 Middle School, and after graduating from junior high school, he was promoted to Hefei No. 1 High School.

In 2014, he won a silver medal at the National Olympiad in Informatics (NOI); in the college entrance examination the following year, he ranked third in science in Anhui Province and entered the Yao Class at Tsinghua University, majoring in computer science. He also served as the chairman of the student union of the Yao Class.

After graduating from his undergraduate program in 2019, he directly entered Princeton University to pursue a doctorate.

He also posted the entire process of his doctoral defense on Bilibili.

After graduating in 2024, he joined OpenAI until the latest news came out.

Academically, he mainly studies intelligent agents. At OpenAI, he has worked on language agents (ReAct, Reflexion, Thought Trees, SWE - agent, CoALA) for digital automation (WebShop, SWE - bench, tau - bench).

Google Scholar shows that ReAct and Thought Trees have been cited over 4,000 times, and his total personal citation count has exceeded 15,000 times.

As early as his doctoral studies, he started focusing on the research of intelligent agents.

In his doctoral dissertation, he summarized the core value of language intelligent agents: from predicting the next token to digital automation, he proposed a series of new benchmarks, new methodologies, and new principled frameworks for language intelligent agents.

Link: https://ysymyth.github.io/papers/Dissertation - finalized.pdf

In his doctoral dissertation, he also recalled his harmonious relationship with his doctoral supervisor.

In 2019, although he chose the Princeton Graduate School, he still wasn't sure what to do or where to go.

At that time, he contacted Karthik R. Narasimhan and said:

This language model thing (GPT - 2) seems very promising. Can it directly solve text - based games?

Karthik readily agreed.

After that, Yao Shunyu started five wonderful years of research and life. Karthik is also called "the best supervisor he's ever had" by him, and Karthik was even the best man at Yao Shunyu's wedding.

Additionally, it's worth mentioning that as early as 2017 - 2018 , Karthik spent a year as a visiting researcher at OpenAI and is also the author of the groundbreaking GPT paper!

In April this year, Yao Shunyu also put forward a unique observation on AI development:

Reinforcement learning has finally worked, and from now on, evaluation will surpass training.

From now on, AI has entered the second half: it will shift from solving problems to posing problems.

He reached a startling conclusion: evaluation will be more important than training.

For a long time, although evaluation (Evaluation) and training (Optimization) are both among the three major elements of learning algorithms, no one has ever attached as much importance to evaluation as Yao Shunyu!

In Yao Shunyu's view, the current problem isn't "whether we can train a model" but "what exactly we want AI to do".

He also pointed out that to win in the second half, we need to shift gears in time:

Our mindset and skills should be more like those of a product manager - defining problems, setting indicators, organizing iterations, and transforming capabilities into measurable values in the real world.

OpenAI's latest paper also points out that evaluation is the root cause of model hallucinations, and simply changing the evaluation method can unlock more capabilities of large - language models.

Perhaps, this time Yao Shunyu will truly put his ideas into practice: getting closer to product management, using evaluation to define real AI.

References:  

https://www.linkedin.com/in/shunyu - yao - 099a41115/ 

https://ysymyth.github.io/papers/from_language_models_to_language_agents.pdf 

https://x.com/ShunyuYao12/with_replies%20 

https://www.ctol.digital/news/tencent - recruits - openai - scientist - 13 - 8 - million - chinese - tech - talent/ 

This article is from the WeChat official account “New Intelligence Yuan”. Author: New Intelligence Yuan. Republished by 36Kr with permission.