Tianpei Gu

1Introduction

Hey there, I’m Tianpei. I’ve worked on AI research at very different scales, from early-stage startups to large research organizations. One thing I’ve learned is that the most interesting problems rarely fit neatly into a single role. Sometimes the answer is a new model, sometimes it’s better data or infrastructure, and sometimes it’s simply a tool that makes everyone’s work easier. I enjoy figuring out what matters and building it.

At Morpheus, I built the research data infrastructure from scratch for a team working on real-time video generation. At ByteDance, I worked on real-time human video generation, and on data systems that supported model research at a much larger scale. These experiences taught me that good infrastructure isn’t separate from research — it shapes which experiments are possible, how quickly we can iterate, and how well we can understand what our models are actually learning.

I learn best by building. I enjoy turning papers into experiments, trying unfamiliar tools, and creating things that make complicated workflows simpler. Sometimes that means training a model; other times it means building an application or a tool that I wish already existed.

Some of my favorite experiences have been working with small teams where people trust each other to take on difficult problems and figure things out. I value that kind of trust and independence.

2Research

Much of my earlier research focused on people: how they move, express themselves, and interact through speech and video. Over time, that interest has expanded toward worlds that can respond to our actions, remain coherent over time, and support real-time interaction. I find the intelligence behind these interactions just as interesting as the visual generation itself.

Working on generative models has also shaped how I think about data. Choosing what a model learns from is a research decision. When a model fails, I often find it useful to ask what kind of experience it might be missing, and whether changing the data would be a better experiment than changing the architecture. I think there’s a lot of interesting research to be done here.

More recently, I’ve been building AI-agent systems to help with my own research and engineering. What interests me isn’t just having agents write code, but figuring out how to give them the context, tools, feedback, and autonomy they need to carry out more complex work. Building these systems has made me increasingly curious about recursive self-improvement — how far can we go when AI systems start improving the processes used to develop themselves?

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 generative models will become in everyday life. Part of the appeal is building things people can actually try, and learning from what they want to do.

3Experience

4Education

5Service

7Writing

All writing