Currently, the most valuable asset in the United States is Chinese AI talent: top students from Tsinghua University, Peking University, and the University of Science and Technology of China are "dominating" the AI circle in Silicon Valley.

In the past two weeks, the most talked-about thing in the AI industry has not been any product, but people. Often, when you wake up, your social media timeline is filled with news that's more of the same: yet another AI big shot has been poached.

Top AI talents are becoming the scarcest and most brand - effective assets in the AI race.

In the center of this storm of talent mobility, we've noticed a particularly prominent detail: among the core members who have led the R & D of large models such as ChatGPT, Gemini, and Claude, the proportion of Chinese scientists is surprisingly high.

This change didn't happen suddenly. In the AI wave that has emerged in recent years, the proportion of Chinese among top AI talents in the United States has been constantly increasing. According to the "Global Artificial Intelligence Talent Tracking Survey Report 2.0" released by MacroPolo, the proportion of top AI researchers from China increased from 29% to 47% between 2019 and 2022.

In the "Research Report on the Background of the ChatGPT Team" previously released by Zhipu Research, it was found that among the 87 - member core team of ChatGPT, 9 are Chinese, accounting for more than 10%.

Therefore, we've re - sorted out the profiles of Chinese AI researchers who have recently attracted wide attention in top Silicon Valley companies and tried to summarize some of their characteristics:

1️⃣ Graduated from top - tier universities with extremely strong academic abilities

Most of them graduated from top universities such as Tsinghua University, Peking University, University of Science and Technology of China, and Zhejiang University for their undergraduate studies, mostly with backgrounds in computer science or mathematics. At the postgraduate stage, they generally went to prestigious schools like MIT, Stanford University, University of California, Berkeley, Princeton University, and UIUC for further studies. Almost everyone has highly - cited papers in top conferences (such as NeurIPS, ICLR, SIGGRAPH, etc.).

2️⃣ Young and productive, with a concentrated burst period after 2020

Most of them are between 30 and 35 years old. During their master's and doctoral studies, they happened to be in the global explosion period of deep learning, with a solid academic foundation and familiarity with engineering systems and team collaboration. For many, their first career stop was to work on AI products or platforms of large companies or those serving a large - scale population, starting at a higher level and with a faster pace.

3️⃣ Strong multi - modal background, focusing on post - training of models

Their research directions generally focus on unified reasoning systems across modalities (text, voice, image, video, motion), including specific details such as RLHF, distillation, alignment, human preference modeling, and voice intonation evaluation.

4️⃣ Even with frequent mobility, they basically don't leave the ecosystem

Google, Meta, Microsoft, NVIDIA, Anthropic, OpenAI... Their mobility spans both AI startups and large - scale giants, but their research topics and technological accumulations often remain consistent, and they basically don't switch tracks.

OpenAI→Meta

Shuchao Bi

Shuchao Bi graduated from the Department of Mathematics at Zhejiang University for his undergraduate studies. Then he went to the University of California, Berkeley for further studies, obtaining a master's degree in statistics and pursuing a doctorate in mathematics.

From 2013 to 2019, he served as a technical leader at Google. His main contributions included building a multi - stage deep - learning recommendation system, which significantly increased Google's advertising revenue (in the order of billions of dollars).

From 2019 to 2024, he served as the head of exploration for YouTube Shorts. During this period, he jointly created and led the Shorts video recommendation and discovery system, and established and expanded a large - scale machine - learning team covering areas such as recommendation systems, scoring models, interaction discovery, trust, and security.

After joining OpenAI in 2024, he mainly led the multi - modal post - training organization and is a co - creator of the voice mode of GPT - 4o and o4 - mini.

During this period, he mainly promoted RLHF, image/voice/video/text reasoning, multi - modal agents, multi - modal voice - to - voice (VS2S), VLA, cross - modal evaluation systems, etc. He was also involved in multi - modal chained reasoning, voice intonation/naturalness scoring, multi - modal distillation, and self - supervised optimization. His core goal was to build a more general multi - modal AI Agent through post - training.

Huiwen Chang

In 2013, Huiwen Chang graduated from the Department of Computer Science at Tsinghua University (Yao Class). Then she went to Princeton University in the United States to pursue a doctorate in computer science. Her research focused on image style transfer, generative models, and image processing, and she won a scholarship from Microsoft Research.

Before joining OpenAI, she worked as a senior research scientist at Google for more than six years. She was engaged in research on generative models and computer vision for a long time and invented the MaskGIT and Muse text - to - image architectures at Google Research.

Early text - to - image generation mainly relied on diffusion models (such as DALL·E 2, Imagen). Although these models have high - quality generation, they have slow inference speed and high training costs. MaskGIT and Muse, on the other hand, use the method of "discretization + parallel generation", which greatly improves efficiency.

MaskGIT is a new starting point for non - autoregressive image generation, and Muse is a representative work that pushes this method to text - image generation. They are not as well - known as Stable Diffusion, but in the academic and engineering systems, they are very important technological cornerstones.

In addition, she is also one of the co - authors of the top - tier paper on diffusion models, "Palette: Image - to - image diffusion models".

This paper was published in SIGGRAPH 2022. It proposed a unified image - to - image translation framework and outperformed GAN and regression baselines in multiple tasks such as image inpainting, coloring, and completion. It has been cited more than 1700 times and has become one of the representative achievements in this field.

Since June 2023, she joined the multi - modal team at OpenAI, jointly developed the image - generation function of GPT - 4o, and continued to promote research and implementation in cutting - edge fields such as image generation and multi - modal modeling.

Ji Lin

Ji Lin is mainly engaged in research on multi - modal learning, reasoning systems, and synthetic data. He is a contributor to multiple core models, including GPT - 4o, GPT - 4.1, GPT - 4.5, o3/o4 - mini, Operator, and the 4o image - generation model.

He graduated from the Department of Electronic Engineering at Tsinghua University (2014–2018) for his undergraduate studies and obtained a doctorate in electrical engineering and computer science from the Massachusetts Institute of Technology. His supervisor was the well - known scholar Prof. Song Han.

During his doctoral studies, his research focused on key areas such as model compression, quantization, vision - language models, and sparse reasoning.

Before joining OpenAI in 2023, he worked as an intern researcher at NVIDIA, Adobe, and Google, and was engaged in research related to neural network compression and inference acceleration at MIT for a long time, accumulating a solid theoretical foundation and engineering practice experience.

Academically, he has multiple high - impact papers in areas such as model compression, quantization, and multi - modal pre - training. His total Google Scholar citations exceed 17800. Representative achievements include the video - understanding model TSM, the hardware - aware quantization method AWQ, SmoothQuant, and the vision - language model VILA.

He is also one of the core authors of the technical documentation for the GPT - 4o system (such as the GPT - 4o system card) and won the Best Paper Award at MLSys 2024 for his AWQ paper.

Hongyu Ren

Hongyu Ren obtained a bachelor's degree in computer science and technology from Peking University (2014–2018) and then a doctorate in computer science from Stanford University (2018–2023).

He has won multiple scholarships such as the PhD Fellowship from Apple, Baidu, and the SoftBank Masason Foundation. His research focuses on large - language models, knowledge - graph reasoning, multi - modal intelligence, and foundation - model evaluation.

Before joining OpenAI, he had multiple internship experiences at Google, Microsoft, and NVIDIA. For example, in 2021, when he worked as an intern researcher at Apple, he participated in the construction of the Siri Q&A system.

After joining OpenAI in July 2023, Hongyu Ren participated in the construction of multiple core models such as GPT - 4o, 4o - mini, o1 - mini, o3 - mini, o3, and o4 - mini, and led the post - training team.

In his own words: "I teach models to think faster, harder and sharper."

In the academic field, his total Google Scholar citations exceed 17742. Highly - cited papers include: "On the Opportunities and Risks of Foundation Models" (cited 6127 times); the "Open Graph Benchmark" (OGB) dataset (cited 3524 times), etc.

Jiahui Yu

Jiahui Yu graduated from the Juvenile Class of the University of Science and Technology of China, obtaining a bachelor's degree in computer science. Then he obtained a doctorate in computer science from the University of Illinois at Urbana - Champaign (UIUC).

His research focuses on deep learning, image generation, large - model architectures, multi - modal reasoning, and high - performance computing.

During his tenure at OpenAI, Jiahui Yu served as the head of the perception team, leading the development of important projects such as the GPT - 4o image - generation module, GPT - 4.1, and o3/o4 - mini, and proposed and implemented the "Thinking with Images" perception system.

Before that, he worked at Google DeepMind for nearly four years. During this period, he was one of the core contributors to the PaLM - 2 architecture and modeling, and co - led the development of the Gemini multi - modal model, being one of the most important technical backbones in Google's multi - modal strategy.