NUPIAO
On July 8, NUPIAO learned from sources close to Tencent that former OpenAI researcher Yonglong Tian has recently joined Tencent’s large language model department, where he will work on visual language model (VLM) research and development.
Public records show that Tian earned his bachelor’s degree from Tsinghua University, then deepened his expertise in fundamental vision algorithms under the guidance of Tang Xiaoou and Wang Xiaogang at the Multimedia Laboratory (MMLab) of the Chinese University of Hong Kong for his master’s. He later pursued a Ph.D. at MIT EECS, advised by Phillip Isola, a leading authority in contrastive learning. During his doctoral studies, Tian led influential open-source projects such as SupContrast, building a solid academic reputation in the visual generation space.
On the industry side, after earning his Ph.D., Tian joined Google Research Cambridge as a senior research scientist. In May 2024, he moved to Google DeepMind to work on visual perception models. That October, he joined OpenAI as a technical researcher, focusing on computer vision and multimodal representation learning, and was deeply involved in the iteration of OpenAI’s multimodal generation technology until his recent departure and return to China to join Tencent’s Hunyuan team.
Tian and Tencent’s chief AI scientist Yao Shunyu were not only undergraduate classmates at Tsinghua but also core researchers working together at OpenAI. In September 2025, Yao left OpenAI for Tencent, and in December Tencent announced a major organizational restructuring, appointing Yao as chief AI scientist in the “CEO/President’s Office,” reporting to Tencent President Martin Lau. Yao also concurrently serves as head of the AI Infrastructure Department and the Large Language Model Department, reporting to Lu Shan, president of the Technology and Engineering Group.
When asked why he joined Tencent, Yao previously explained that in the second half of the AI race, methodologies have become very mature, but finding the right problems has become harder. A key factor for him was that Tencent is full of good problems and good products. He also noted that Tencent operates on a culture of trust rather than metrics, which he believes is vital for building a long-term-oriented AI organization.
The successive departures of Yao Shunyu and Yonglong Tian are part of a nearly year-long exodus of core talent from OpenAI. Most of those leaving have come from three critical divisions: AI safety, multimodal, and reasoning models, with strategic disagreements being the main driving factor.
Just in early July, Joshua Achiam, a nine-year veteran at OpenAI and its chief futurist, announced he would leave at the end of the month. He had been responsible for long-term AI safety and mission alignment, but the company disbanded that team in February, folding safety functions into commercial product lines.
Earlier, Jerry Tworek, the VP of research in charge of model post-training and reasoning, left at the beginning of this year after repeatedly failing to get approval for long-cycle fundamental research compute resources, leading to serious technical disagreements with management. Caitlin Kalin, head of robotics, resigned in protest against the company’s defense AI cooperation policies. Several economists and policy researchers also chose to leave as the company fully commercialized and de-emphasized long-term safety research.
Some analysts argue that OpenAI’s rapid shift from a nonprofit safety lab to a for-profit enterprise has squeezed basic research space, which is the root cause of the continued outflow of top researchers.
In contrast, Tencent has been doubling down on the large model race through organizational overhauls, heavy investment, and global talent poaching, steadily accelerating its business implementation. In April, Yao’s team launched the new Hunyuan Hy3Preview base model, which surpassed 3.66 trillion token calls in its first week, placing its performance in code generation, tool use, and complex reasoning among the top tier in China. Leveraging the massive ecosystems of WeChat, gaming, and advertising, Tencent also rolled out WorkBuddy, an all-in-one AI agent for the public, AI generation tools for games, and a multimodal creative generation system for advertising, tightly integrating model technology into its core businesses.
Beyond Yao and Tian, in February, Pang Tianyu, a former chief research scientist at Singapore’s Sea AI and a Tsinghua Ph.D., also joined the Tencent Hunyuan team as a chief research scientist, focusing on large model training optimization.
To attract more top-tier talent, Tencent is now heavily promoting its “Qingyun Plan,” recruiting bachelor’s, master’s, and Ph.D. graduates globally in fields like AI large models and infrastructure, offering highly competitive compensation and customized development programs.
Tencent’s Q1 2026 financial report shows that total capital expenditure for the quarter reached 31.94 billion yuan, up 16% year-on-year and a staggering 63% sequentially. The vast majority of those funds went toward AI computing chips, data centers, and large model training infrastructure. R&D spending for the quarter hit 22.54 billion yuan, a 19% increase year-on-year, with all incremental investment directed at Hunyuan model iteration, multimodal algorithm R&D, and compensation for top AI talent.
At Tencent’s shareholder meeting in May this year, Chairman and CEO Pony Ma commented on whether Tencent was lagging in AI, saying, “A year ago we thought we were on the boat, but later we discovered the boat was leaking. Now we feel like we’re standing on it, but we still can’t sit down, and we still hope the boat can go a bit faster.”
Ma admitted that Tencent’s early AI capabilities were not outstanding, but through talent building, team management, and internal training in recent years, the company has been shoring up its weaknesses and is now gradually getting on track.
He stressed that Tencent may not always be the fastest to seize the opportunity, but it insists on taking the right path, combining its unique strengths and advancing steadily. “We can’t just cross over and grab other people’s territory just because we see them doing something there. We’ve tried that before and mostly failed.”
In January this year, Ma also mentioned that 2025 is a big year for AI, and among Tencent’s business lines, AI is the only one that requires a lot of spending and is still worth investing in heavily. He added that Tencent has its own considerations and pace, with the core focus being the long-term competitiveness of products and user experience. The company is now carefully thinking through each business segment and platform, continuously lighting up new skills.
