Chinese AI scientist Yao Shunyu Resigns from Google to Start a Business? His Personal Response

Recently, market rumors claimed that Shunyu Yao, a Chinese AI scientist, left Google DeepMind to launch a startup, establishing an AI company focused on Recursive Self-Improvement (RSI), with the project valued at as high as 2 billion US dollars. The rumors even indicate that its co-founding team has been basically formed, one of the co-founders is a senior employee of DeepMind, and the team is also attracting research and engineering talents from leading AI labs including DeepMind, Anthropic, and OpenAI.

On November 3, according to The Paper, Shunyu Yao responded: "I have not founded any company, and I am still an employee of Google DeepMind at present." He added, "I have no idea at all where this fake news came from."

This is not the first time that Shunyu Yao has been "arranged to start a business" by the market. As early as October 12, similar financing rumors spread in domestic investment communities. At that time, he posted on social platforms to clarify: "I have no plans to raise financing in China at present. I am also very confused about some reports, please do not waste time on me. I am just an ordinary researcher."

The two rumor-refuting statements are only more than one month apart, which shows that with the rising popularity of the RSI and agent tracks, capital's pursuit of top reinforcement learning talents is becoming increasingly fierce.

Shunyu Yao's academic resume is an important reason why the outside world is willing to believe the rumors. Born in 1997, Shunyu Yao completed his undergraduate study in the Department of Physics (Institute for Advanced Study) of Tsinghua University, where he won the Special Scholarship for Undergraduates of Tsinghua University and the Ye Qisun Physics Award. He once cooperated with his mentor Wang Zhong, published a paper in *Physical Review Letters* as the first author, proposed the non-Hermitian skin effect, established the non-Hermitian topological band theory, which is a foundational achievement in this field and opened up a new branch of non-Hermitian topological physics.

After graduating from Tsinghua University, he entered the Institute for Theoretical Physics of Stanford University to pursue his doctoral degree, with mentors Douglas Stanford and Stephen Shenker. His research directions cover quantum gravity, quantum chaos and holographic principle, and he developed the Scramblon theory. After completing his doctorate, he did short-term postdoctoral research at the University of California, Berkeley.

In October 2024, Shunyu Yao officially crossed into the AI industry, joined Anthropic as a research scientist, and deeply participated in R&D work related to Claude 3.7 Sonnet agent coding, large-scale reinforcement learning of the Claude 4 series, and training numerical stability, focusing on the direction of "learning science".

In September 2025, he announced that he left Anthropic and joined Google DeepMind, focusing on the new generation of reinforcement learning and Recursive Self-Improvement (RSI), which is the direction of autonomous model iteration that is very popular in the venture capital circle at present. He mentioned in his blog that among the reasons for leaving Anthropic, 40% were value differences, and the remaining 60% were internal information that cannot be disclosed.

Talking about the transition from theoretical physics to AI, he explained on his personal blog: "The main reason is that I want to find a direction that can provide more opportunities for young people." In his view, theoretical physics is an amazing training field full of intellectual challenges, but this field has lacked experiments for many years. In a field lacking experiments, it is very difficult to objectively judge the importance of a theoretical work, and it is also difficult to eliminate disagreements through systematic experiments.

"I also considered quantum computing, but the bottleneck of quantum computing at this stage lies in the experimental platform. After comprehensive trade-off, I finally chose artificial intelligence." He also made a widely circulated metaphor: the current large model research is just like the human research on thermodynamics in the 17th century, there is no mature and complete theory, but continuous experiments can continuously bring new insights.

The core goal of the reinforcement learning and RSI recursive self-improvement that Shunyu Yao is currently focusing on is to enable AI models to learn to optimize themselves autonomously. Most of the current mainstream large models complete training relying on human-labeled data and manually set reward signals. Recursive self-improvement expects the model to autonomously generate better training tasks, self-evaluate outputs, iteratively update its own capabilities during operation, reduce dependence on manual labeling, and move closer to the direction closer to general intelligence.

The reinforcement learning and RSI recursive track is still in the early exploration stage, with no mature commercialized products, but it is regarded by many institutions as the core breakthrough point for the next generation of large models. Globally, OpenAI, Anthropic, and DeepMind are all investing resources in layout, and many domestic startups and large tech company labs are also following up. The high expectation for the track has also pushed up the value of relevant core researchers.

Behind this oolong rumor is the multi-year large-scale flow of Google AI talents. Many Chinese researchers from the DeepMind and Google Gemini teams have chosen to join domestic large enterprises, bringing overseas cutting-edge technical experience in reinforcement learning, pre-training and multi-modality to domestic self-developed systems.

Zhou Hao, a former senior Google research scientist, quietly joined Alibaba in January this year, and currently serves as the person in charge of the post-training direction of Tongyi Qianwen, reporting directly to Zhou Jingren, CTO of Alibaba Cloud and head of Tongyi Lab. Tian Yonglong, who once worked at Google DeepMind, joined Tencent and is responsible for the multi-modal R&D of Hunyuan large model. Tian Yonglong graduated from Tsinghua University and also worked at OpenAI. His main research directions include computer vision, visual representation learning and generative models.

Some other Google-affiliated AI talents choose to start their own businesses independently, targeting directions such as agents, world models, and large model infrastructure. Such projects are generally favored by top VCs, and some projects have obtained high valuations even at the stage without commercialized products. For example, Chen Yongchao, a Chinese researcher with Google DeepMind background, founded Chaoyan Intelligence, focusing on AI autonomous scientific research and recursive iterative models. Less than half a year after its establishment, before mature commercialized products were formed, the total financing of its angel round and angel+ round has reached nearly 400 million RMB. Chaoyan Intelligence is also one of the few startups in China that has obtained financing and released RSI models.

This article is from the WeChat Official Account "Jiemian News", author: Song Jianan, published with authorization from 36Kr.