Zhejiang University

I studied Information Management and Information Systems at Zhejiang University, in the School of Management. My degree is a B.S. in Management — back then, "CIO" was still a new word, and the program emphasized how organizations handle information.

But I was drawn to the technical side. I joined a small lab in the Computer Science department and picked up real software engineering skills alongside the management curriculum — data structures, OOP, databases, networking.

Into the Real World

After graduating, I discovered that information management isn't usually the most important function in a company. Marketing, sales, finance, R&D — those tend to matter more.

I caught the wave of e-commerce early and founded a company in 2009. Through persistent effort, it survived its early challenges and continues to operate today.

A New Curiosity

After I got married and had a kid, I started trying to think about everything in a new way. Intelligence and life became the two most interesting things I cared about. So I taught myself math and other foundations, working through online courses — Stanford's Machine Learning (2016), deeplearning.ai's Neural Networks and Deep Learning (2017), and Geoffrey Hinton's Neural Networks for Machine Learning from the University of Toronto (2017).

I started writing code to experiment — implementing neural networks from scratch with Python and NumPy, playing with TensorFlow, and building interactive demos to understand what I was learning.

University of Vermont

In 2019, I went to the University of Vermont to study Complex Systems and Data Science. Deep learning was one of the hottest topics, and the program also valued visualization as a tool for understanding complex systems. I studied both.

My advisor was Josh Bongard, who studies evolutionary robotics and is a great educator. Sam Kriegman, then a PhD candidate under Josh, was the one who introduced me to the lab and showed me what scientific research actually looks like. I consider both of them my mentors.

My thesis explored the Multi-Robot Continuous Control problem in deep reinforcement learning — how one controller can manage robots with different body shapes by arranging their sensory inputs and outputs in specific ways. During that time I also rewrote a CPU-based voxel physics engine into a GPU-accelerated one (Voxcraft), and contributed to the Xenobots project — living robots designed by computers and built from frog cells.

I also miss the peaceful, friendly life in Vermont.

Exploring

After graduating in 2021, I felt I wasn't ready for focused research. I was too curious about too many things, and I didn't have a clear perspective yet. So I paused, read widely — both books and papers — and worked on small projects to explore possibilities.

For a while I advised an investment company in Shanghai on AI trends, helping simplify complex AI concepts and evaluate emerging startups. But mostly, I was building my own understanding — reading about cognition, philosophy, mathematics, consciousness, and the nature of intelligence.

Residual Stream

Recently, I discovered that the residual stream inside large language models is very interesting. I started to focus on this — exploring mechanistic interpretability and beyond, trying to understand how these models actually work under the hood.

Now I live in Singapore and do independent research, using a quick prototyping platform to run experiments and sharing discoveries along the way.