I look for the right way to understand things.
Then the solution is usually simple.
Shuxin Zheng 郑书新
Vice President, Zhongguancun Academy of AI（中关村人工智能研究院副院长）
Previously Principal Researcher at Microsoft Research
I start from mathematics, not trends. When you see the structure clearly, ideas transfer across domains that seem unrelated — from distributed optimization to molecular science to language models.
I believe in less structure, more intelligence. Strip away the scaffolding and let the model learn what matters. If a method needs too many tricks to work, it's probably wrong.
I pick problems half a step ahead — early enough to shape the direction, not so early that nobody cares. If the timing is wrong, I'd rather not do it.
The domains change. The taste doesn't.
Compressed protein conformational dynamics simulation from years to hours — making the Boltzmann distribution of biomolecules computationally accessible for the first time.
A Transformer that natively understands graph structure. Won KDD Cup 2021 and the Open Catalyst Challenge — proving that general architectures can beat domain-specific ones.
A simple reordering of normalization layers that stabilized deep Transformer training. Quietly became a default in OpenAI's GPT series and most large language models since.
The first mathematically rigorous delay compensation for asynchronous SGD — solving a problem that had been patched with heuristics for years.
Now building at the intersection of research and industry — leading AI research at 中关村人工智能研究院, teaching large-model foundations at 中关村学院, and incubating the next wave of AI-native companies.