Benjamin Berczi
Hi! I'm Benji, an AI safety researcher, machine learning scientist and former theoretical physicist. I'm currently a research scholar at MATS, mentored by Cozmin Ududec (UK AISI), studying how personas are induced in language models, how deeply models adopt them, and how that drives misalignment.
Before MATS I was an ML scientist at Depop, building large-scale recommender systems, and worked on interpreting goal-conditioned RL agents with a researcher at Google DeepMind. Earlier I did a PhD in theoretical physics at the University of Nottingham, writing simulations of quantum matter in curved spacetime to study how tiny black holes form.