Abstract
Research starts with a plan and then wanders off. Mine did too. I set out to climb an animal action recognition leaderboard and ended up somewhere more interesting: from plain video-level labels, the model learned which animal in the clip is doing what. The main finding of this work was not even the main goal. I'll show how we got there and how distillation and transfer carried it across four animal datasets. But the real story is how it got done. As a computer vision researcher, I live in chaos, with too many open questions at once. So I started treating research like exploration. Every question is a map. Everything we don't know yet is fog. The frontier is whatever we can attack today. Behind it sits everything we have already uncovered, ready to build on. Two months in, that sky holds 9 maps and more than 130 tickets, and I can watch it light up as a star-map of planets, explored and unexplored. Filled not only with the questions I asked at the start but, more importantly, with the ones we discovered along the way. None of this would have been possible without the right tools and skills. Literally: SKILL.md files. They turned a swarm of Claude worker bees into actual research collaborators. Part research update, part field guide to ClaudeMaxxing your research, with the wins and the pain points. Expect science, engineering, and a lot of AI magic.
About the speaker
Ulas Bingol is a doctoral researcher in the CVIA group at the University of Konstanz, working on computer vision, action recognition, foundation models and 3D reconstruction. He comes from robotics, with several years of building robots in industry, and is still getting used to a field where the state of the art changes faster than a robot can finish its calibration. His main focus is action recognition and representation learning, which mostly means poking at foundation models to find out what they really understand about how we interact with the world, and how to teach them more