China AI talent poaching is no longer a slow bleed. It is a structural shift, and two of the most prominent recent moves show how deliberately Beijing-aligned companies are building out their research benches at American firms’ expense.
| Researcher | Former Role | New Role |
|---|---|---|
| Yao Shunyu | OpenAI researcher | Chief AI Scientist, Tencent |
| Wu Yonghui | VP of Research, Google DeepMind | Head of Foundational Research, ByteDance Seed |
| Hao Zhou | Google DeepMind researcher | Qwen AI team, Alibaba (reported) |
| Yang Zhilin | Meta AI / Google Brain | Founder, Moonshot (Kimi AI) |
What Tencent Is Building Around Yao
Yao Shunyu, speaking Friday at a Tencent event in Beijing co-organized with local authorities, said the path to AGI runs through smaller models with consistent performance on everyday tasks, not the heavyweight frontier systems dominating U.S. labs. He put the untapped commercial opportunity at trillions of dollars. “I don’t think ChatGPT or Claude will be the only super-app,” he said.
Tencent has given Yao a substantial platform to pursue that thesis. He reports directly to Tencent President Martin Lau and leads the newly created AI Infrastructure Department inside the company’s Technology Engineering Group, overseeing both the LLM and AI Infra teams. Tencent also built out three new organizational units around the same restructuring: AI Infrastructure, AI Data, and Data Computing Platform, all aimed at accelerating large model research and development.
That kind of organizational commitment is significant. Most AI hires at large tech companies land inside existing structures. Tencent is reorganizing around these people.
China AI Talent Poaching: The Structural Shift
Wu Yonghui’s move to ByteDance tells a similar story. Wu joined Google in 2008, spent nearly a decade at Google Brain working across machine learning, genomics, and natural language understanding, and was promoted to Google Fellow in September 2023. He left for ByteDance in February 2025.
At ByteDance Seed, Wu reports directly to CEO Liang Rubo. Several algorithm and technology leaders who previously reported to another executive shifted to report to Wu after his arrival, with the prior lead retaining oversight of model applications. Wu focuses on foundational research. His work there has already touched the Doubao 2.0 large language model and the Seedance 2.0 video-generation model.
China AI talent poaching is partly a function of U.S. immigration uncertainty. Chinese nationals who built careers at American labs now face a less predictable path to staying in the country. Lower salaries at home carry less weight when the alternative is visa limbo. China is simultaneously ramping up state investment in basic research as part of its five-year scientific push, which raises the floor for what domestic roles can offer.
Anthropic’s Warning Cuts the Other Way
The talent flow is happening as the U.S. AI establishment is fractured over what comes next. Anthropic said Thursday that frontier models are approaching the point where they can improve themselves without human oversight, and that “the human role is narrowing at each step in the AI development process.” The company argued a worldwide slowdown in cutting-edge AI development would “likely be a good thing,” but acknowledged the prisoner’s dilemma: if only one company stopped, rivals would simply race ahead.
That caveat matters enormously when the rivals include ByteDance and Tencent, now staffed with researchers who helped build the systems Anthropic is worried about.
Anthropic’s own 2028 AI leadership research paper argued that Chinese labs have stayed competitive partly through what it called distillation attacks, illicitly extracting innovations from American frontier models, and through accessing export-controlled U.S. chips via offshore data centers in Southeast Asia not currently covered by export law. That framing sits uncomfortably alongside a simultaneous call to slow down: a pause by U.S. labs that leaves chips and model knowledge already in circulation does not obviously reset the competitive position.
China AI talent poaching accelerates that problem. The knowledge walks out the door with the researcher.
Yao’s framing, smaller models and reliable performance rather than capability maximalism, suggests Chinese labs may not need to match U.S. frontier systems to win commercially. If that bet is right, the competitive question stops being who builds the most powerful model and starts being who deploys the most useful one at scale. Watch whether Tencent’s WeChat ecosystem gives Yao’s smaller-model thesis its first real test.
