NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

From OpenAI to Running Tencent's AI — What Shunyu Yao Built in His First Months

Shunyu Yao (姚顺雨), the researcher behind ReAct and Tree of Thoughts, left OpenAI for Tencent — and within a year held its foundation models, compute and data.

2026-11-19 · 1132 words · NeuroAI
From OpenAI to Running Tencent's AI — What Shunyu Yao Built in His First Months

Tencent's founder and chief executive, Pony Ma (马化腾), has placed a researcher who joined the company less than a year earlier in charge of its core AI work.

That researcher is Shunyu Yao (姚顺雨), Tencent's chief AI scientist (首席AI科学家). His rapid rise says as much about how Chinese technology companies now recruit AI talent as it does about him.

Key takeaways

  • Yao, born in 1998, studied at Tsinghua's Yao Class (姚班) and earned a doctorate at Princeton. His ReAct and Tree of Thoughts papers are widely cited foundations of modern AI agent (智能体) work.
  • Former role: research scientist at OpenAI from 2024 to 2025, where he worked on agent products including Operator and Deep Research. Current role: Tencent's chief AI scientist, reporting to president Martin Lau (刘炽平), with responsibility for its Hunyuan (混元) models.
  • Tencent announced his appointment in December 2025. By July 2026 it had merged its language and multimodal model teams into one foundation-model department under him.
  • Reports say his first months included a rebuilt training stack, a Hy3 preview in April and a Hy4 release in November.
  • Tencent has publicly disputed reports of an unusually large salary for him. This article does not repeat those figures.

The researcher behind the methods

To understand why Tencent moved quickly, start with the papers. ReAct, published in 2022, proposed letting a large model (大模型) reason and act in the same loop instead of doing one then the other. Tree of Thoughts, a year later, let a model explore several lines of reasoning, backtrack and choose among them. Both became reference points for the agent systems built by companies around the world.

Yao did that work as a doctoral student at Princeton after graduating from Tsinghua's selective Yao Class. He joined OpenAI in 2024 and, according to Chinese financial media, contributed to Operator, Deep Research and computer-use agents. He has argued publicly that the next phase of AI is about applying capability rather than only producing it — a framing that later echoed in how he described his move to Tencent.

He left OpenAI in September 2025. In December, Tencent announced a restructuring of its model research and named him chief AI scientist in the CEO and president's office, reporting directly to president Martin Lau. He also took charge of AI infrastructure and the large-language-model department.

That is an unusual entry point. A newcomer reporting to the president, with control of both the team and the compute it needs, is not a standard research hire.

What changed after he arrived

The first months were spent on foundations rather than launches. Chinese reporting says that in early 2026 Yao led a rebuild of the pretraining and reinforcement-learning infrastructure behind Hunyuan, Tencent's family of foundation models. A Hy3 preview followed in April.

In July, Tencent merged its large-language-model department and its multimodal department into a single foundation-model department, placing it under Yao. By September his remit reportedly covered that department plus AI infrastructure and AI data — in effect the whole chain from raw data and compute to the model itself.

Chinese media reported in November that he led the release of Hunyuan Hy4, near the anniversary of his joining. I could not confirm the technical claims made around that release from independent benchmarks, so this article treats it as a reported event rather than a verified result.

The pattern is clear regardless: consolidate scattered teams, shorten reporting lines, and put one person in charge of the pieces that have to work together. Tencent was not alone in doing this. Other large Chinese technology groups reshaped their model organisations in the same period, merging divided teams into single units as the industry shifted from chasing parameter counts to controlling cost and deployment.

Why a young researcher got the keys

There is a practical logic. In foundation models, the technical ground moves faster than traditional seniority accumulates. A researcher who worked inside a frontier lab during the agent boom brings knowledge that is hard to build by promoting someone through a conventional ladder.

There is also a business logic. Tencent's strength is not publishing the most eye-catching model every quarter; it is putting AI into products that hundreds of millions of people already use — chat, gaming, office tools and video. A researcher who understands both the underlying methods and the value of real distribution is more useful there than a pure laboratory star.

Yao has said publicly that what is scarce now is not training technique but finding problems worth solving, and that such problems live inside operating businesses. That is a reasonable description of why a company like Tencent would hand him the keys. Whether the structure produces better models is a separate question, and one that needs more than a year of releases to answer.

A note on names: this is Shunyu Yao 姚顺雨 of Tencent. A different researcher, Shunyu Yao 姚顺宇, worked at Anthropic and Google DeepMind on theoretical topics. They are often confused in translation.

Honest limitations

The career facts here come from Chinese financial and news outlets — among them CNR, Securities Times, The Paper and 21st Century Business Herald — and from Tencent's own December 2025 announcement as reported by them. I have not interviewed Yao or Tencent, and I have no access to internal organisation charts. The exact scope of his authority is assembled from public reporting and may lag recent changes.

The ReAct and Tree of Thoughts attributions are well established in the research literature. The descriptions of OpenAI project work and of the Hunyuan releases rely on media reports, not on primary documents I have read directly. The Hy4 report in particular is single-source in what I could access, which is why it is flagged as reported.

Widely circulated salary figures for Yao are not reproduced here. Tencent has publicly disputed them, and I have no way to verify them. Political and governance dimensions of China's technology industry are outside this article's scope.

What readers can do now

  • Read the papers, not the profile. If you want to understand why this researcher matters, ReAct and Tree of Thoughts are public and more informative than any career story.
  • Watch team structure as a signal. When a company merges model teams and shortens reporting lines, it is usually telling you where it thinks the bottleneck is. That is often more revealing than a launch event.
  • Separate the person from the product. A star researcher can improve an organisation, but model quality depends on data, compute and product feedback too. Judge the releases, not the hire.

Yao's first months show how far a frontier-lab researcher can rise in China's technology industry when the company decides the old structure is too slow. The interesting test is not the appointment itself but what the next few model cycles look like.

Related coverage

Cite this article

NeuroAI (脑机网), "From OpenAI to Running Tencent's AI — What Shunyu Yao Built in His First Months," 19 November 2026, https://neuroai.site/a/na-people-shunyu-yao-tencent-ai

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Builders” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories