Biography
My lifestyle is guided by analytical thinking, curiosity, and the development of coherent definitions. In addition to my research topics, my professional interests include the computational sciences, mathematics, the implementation of holistic learning systems, and systems thinking. I enjoy collaborating with businesses, the public sector, and the media.
Research interests
The goal of AI alignment is to guide AI systems toward the goals, preferences, or ethical principles set by an individual or group. Mechanistic interpretability is a subset of AI safety that aims to reverse-engineer neural networks by breaking them down (hopefully) into parts that are understandable to humans, such as features and circuits. Breaking down a neural network into its components fosters a deeper understanding of neural networks, which enables, for example, model steering even during the generation process itself.
In my current research, I focus on understanding mechanistic interpretability as a field of study, investigating how neurons and circuits can be interpreted directly without sparse autoencoders, developing an intuitive understanding of the risks of artificial intelligence, and how mechanistic interpretability can be combined with AI alignment to create safer and more stable AI models. Developing AI safety is particularly important so that we can prevent various AI risks from materializing!