Department: Computer Science Faculty Adviser: Tri Dao Year of Study: G1 Undergraduate School: Renmin University of China Undergraduate Major: Computer Science
I'm Zhifei, a first-year CS PhD student working on machine learning systems. I became a systems person at Renmin University of China, where I led the TA team for our Intro to Computer Systems course, so I have watched a lot of students meet their first segfault. At UC Berkeley's Sky Computing Lab I worked on making AI cheaper to run: LEANN shrinks a retrieval index until it fits on a laptop, and SkyNomad hops AI jobs across cloud regions to wherever spare GPUs are cheapest. Now I build environments where AI agents learn to do systems work themselves, like speeding up GPU code, and a surprising amount of my time goes into catching agents that pass the test without actually solving the task. Back in Beijing I co-founded a platform that matches undergraduates with their first research problems, so ReMatch feels familiar. Outside the lab I play badminton and go hiking, and I'm always up for talking research over a game or on a trail.
Months before Meta launched Muse, I built my own AI agent: it reads my inbox, drafts my replies, wakes me up in the morning and makes phone calls for me. I still press send myself.
I work on machine learning systems and reinforcement learning. I build training environments and evaluators for AI agents that do real systems work, like repairing programs or tuning GPU kernels, the small programs behind a model's arithmetic. The hard question is whether an improvement is real and worth its cost. Next, I want agents to speed up the physics simulators robots learn in, so that training on realistic contact and soft objects becomes affordable. If you like asking why something is slow, or how you would know a fix actually worked, that is close to what I do.
Not available to participate in Summer ReMatch+ program.
