I am a Senior Research Scientist at DeepMind, where I work on frontier LLM post-training, with a focus on improving training data. Prior to that, I worked on pre-training data, and before that on protein folding and protein design, all at DeepMind. I completed my PhD in Machine Learning at the University of Oxford, supervised by Professor Ingmar Posner. Before that, I studied physics in Germany and the UK, with stops along the way in biophysics, astrophysics, nanophotonics, and management consulting.
These days, I’m mostly thinking about LLMs, agents (from both a research and deployment perspective), and model self-improvement. At some point I might go back into the natural sciences, leveraging the extraordinary advances in machine learning, but for now I am purely focussing on LLM research.
My PhD topic was learning invariant and equivariant representations, which was more theory-heavy than what I do today. Simply put: whereas most of deep learning is concerned with finding the important information in an input, I focussed on ignoring harmful or irrelevant parts of information. This is tantamount to leveraging symmetries and can be important to counteract biases or to better leverage structure in the data. While almost all machine learning tasks have some symmetries (which are often leveraged, e.g., by CNNs being translation equivariant), they become particularly prevalent on the length scales of molecules and below. I argue that, if we want to make machine learning enabled breakthroughs in fields like biochemistry and material science, we need to become good at leveraging symmetries.
I am originally from Southern Germany. Before switching to Machine Learning for my PhD, I studied physics at the Universities of Erlangen, Heidelberg, and Imperial College London. In my spare time, I enjoy rock climbing, tennis, playing the guitar, and the occasional carpentry project.