Hey there, I’m Tianpei, a Research Scientist at Roblox, based in San Mateo, California. I joined through the acquisition of Morpheus AI, and previously worked at ByteDance and Krea, training generative models and building the systems around them.
I like writing code and building things, and I enjoy taking an idea through the whole process of making it work. A lot of my work has involved end-to-end data systems, from raw images and videos through processing and curation to model training. At Morpheus, I built that infrastructure from scratch. I also build tools for inspecting data, comparing experiments, and understanding results, so I can follow a problem through the different parts of the system.
I’m quick to pick up new research, tools, and frameworks, and I learn best by putting them to use. I like turning a paper into an experiment or bringing a new tool into a project, then figuring out what it makes possible and where it fits. That habit shapes both what I build and what I become curious about next.
Much of my earlier research focused on people: how they move, express themselves, and interact through speech and video. That interest has expanded toward worlds that can respond to our actions, remember what happened, and stay coherent over time. I think the intelligence behind that interaction will be as interesting to work on as the images themselves.
Working on these models has also shaped how I think about data. Choosing what a model learns from is a research decision. A failure can suggest that some kind of experience is missing, and changing the data becomes a way to test that explanation. I think this part of research deserves more attention.
Lately, I’ve been building systems around AI agents to make them more useful in my projects. I enjoy figuring out how to give them the tools, context, and feedback they need to carry work forward with less manual coordination. As models become more capable, I want these systems to let them take on more of the process.
That has also drawn me toward recursive self-improvement: the possibility of systems improving their own methods, then using those improvements to become better at improving themselves. My experience with automation gives me a practical starting point, and I’m interested in understanding how far that process can go.
Alongside these interests, I keep coming back to my personal “mom benchmark”: could someone like my mom find a reason to use this? I want to build things people find useful, enjoyable, or worth coming back to. I don’t have a fixed picture of what video models will become in everyday life. Part of the appeal is building things people can try, and learning from what they actually want to do.
Outside work, I enjoy cooking, playing Dota 2, and trying out new technologies. I also have a cat named 贝壳, or Bayker in English. Below are a few photos of him; you can find more on Instagram and Xiaohongshu.
