Jiepeng Jin [金杰鹏]

Jiepeng Jin [金杰鹏]

Direct Ph.D. student @ ZJU CS

👋 Hi there! I’m Jiepeng Jin

I’m Jiepeng Jin (金杰鹏), a direct-entry Ph.D. student in Computer Science at Zhejiang University (ZJU).

I studied Computer Science & Technology at ZJU from 2022 to 2026, with a minor in ACEE (Advanced Class of Engineering Education) at Chu Kochen Honors College.

I’m currently rebuilding my understanding of language models from the fundamentals. I’m working through Stanford’s CS336: Language Modeling from Scratch, implementing core components and running small-scale training experiments to better understand how the pieces fit together.

My research interests are still taking shape. I’m particularly interested in understanding how intelligent systems work, how hardware and systems constraints shape their capabilities, and how we might move from using existing models toward constructing better ones.

Current Interests

  • Language Models from First Principles: Understanding tokenization, Transformer architectures, optimization, and training dynamics through implementation and experiments.
  • AI Systems and Performance: Learning how GPU execution, memory, computation, and system design affect the efficiency of modern AI systems. I’m still building my foundations in this area.
  • Autonomous Intelligence: In the longer term, I’m curious about how AI systems could identify worthwhile problems, learn from real feedback, and improve beyond tasks whose objectives are fully specified in advance.

These are directions I’m exploring rather than a settled research agenda. For now, I want to understand the mechanisms behind existing systems well enough to reason about how they could be improved.

The “Restart”

For a long time, I followed a fairly conventional academic path. I was good at learning established material and achieving good results, but I gradually became less satisfied with research when I could describe what a method did without really understanding why it worked.

The question that stayed with me was: Can understanding a system help us construct a better one?

Reading Xu Yang’s writing about moving from understanding intelligence toward constructing intelligence helped me articulate this interest. I decided to press the “Restart” button and rebuild my foundations, starting with language models.

That is why I began working through CS336, studying the components of a language model and trying to connect mathematical definitions, implementations, and experimental results.

I’m still at an early stage. I have a lot to learn, and getting code to run is only the beginning. I want to develop the ability to explain mechanisms, design meaningful experiments, question my own assumptions, and gradually find problems worth pursuing over the long term.

For me, Restart is an ongoing change in how I learn and do research.

Selected Work & Notes

CS336: Language Modeling from Scratch

I’m working through the CS336 assignments to understand how a Transformer language model is implemented and trained. My current work focuses on the basics, including tokenization, model components, optimization, and small-scale training experiments.

I use the implementation and write-ups to connect theory with code and to record what I have learned and what I still don’t understand. The work is ongoing.

LLM Notes

A personal knowledge base on large language models, covering tokenization, model architectures, GPU programming, distributed training, inference systems, scaling laws, and related topics.

The notes synthesize public learning resources, including Datawhale’s DIY-LLM materials and Stanford CS336, alongside papers, technical reports, and engineering documentation. I maintain source references and update the notes as I verify details and deepen my understanding.

Personal Learning Notes

I built and maintained a public website for course notes and learning experiences during my undergraduate studies. The idea was simple: organize what I had learned so that the notes could remain useful beyond my own coursework and help other students learn more efficiently.

How I Try to Work

  • Understand the mechanism: Go beyond observing that something works; try to understand why.
  • Make claims checkable: Keep track of sources, assumptions, experiments, and results.
  • Correct mistakes early: Treat failed experiments and revised beliefs as part of the learning process.
  • Share what I learn: Turn useful notes and experience into resources that others can inspect, reuse, and improve.

Let’s Connect

I’m open to thoughtful conversations, technical collaboration, and meeting people who enjoy exploring difficult problems.

Education

  • Ph.D. in Computer Science, Zhejiang University (ZJU CS), 2026–2031 (expected)
  • B.E. in Computer Science & Technology, Zhejiang University (ZJU CS), 2022–2026
    • Minor in ACEE, Chu Kochen Honors College