With advances in computing and the arrival of generative AI, we can now go through trial and error, turning ideas into something concrete and testing them, much faster than before. Anyone can now create programs, text, images, and videos to their liking, and the possibilities people encounter while creating have expanded greatly. Yet the more options there are, the harder it becomes to pick out the one that fits you best.

It is not realistic for people to compare every one of countless options by eye. That is why we need optimization techniques that efficiently find what fits each person. At the same time, people cannot always express clearly, in numbers or words, what they are looking for. That is why it is also important to design interfaces that let people naturally convey intentions that are hard to put into words, for example by comparing two candidates or sharing a rough image of what they want. I have combined optimization techniques and interfaces to develop systems that find what users are looking for, together with them. My work spans a wide range of domains, from programming and 3D modeling to photo editing and even exploring preferred tastes.

Recently, I have come to believe that it is important not only to reach the goal, but also to support the process of searching itself. This is because, while searching for what fits them, people encounter many possibilities and broaden their perspectives. In graduate school, I am continuing research that focuses on this process of searching.

My motto is "Just build it." When I come up with an interesting idea, I give it shape first. This goes back to elementary school, when I started programming with Scratch and discovered the joy of creating worlds exactly as I imagine them.

I also value building things from scratch whenever I can. As the physicist Richard Feynman put it, "What I cannot create, I do not understand." I believe that I truly understand how something works only after building it myself. For example, I have built my own programming languages, including one for education and one for quantum computers. To understand how large language models (LLMs) work, I also trained a Japanese LLM from scratch on a single GPU in our lab. During my internship at Preferred Networks, I trained an image generation model from scratch using only rights-cleared images. My experience with LLMs and image generation models in particular has been a great help in building new tools that use them.