I'm Bart, an independent mechanistic interpretability researcher. I aim to contribute to a brighter future for all by addressing technical issues in AI safety.
- Exploring the concepts learned by LLMs and the relationships between them with sparse autoencoders.
- MSc. Artificial Intelligence, University of Amsterdam (2016-2019)
- Bachelor's degree, University College Twente (2013-2016)
- Independent Alignment Researcher and Freelancer (2023 - Now)
- Machine Learning Researcher, IDLab, University of Antwerp and imec (2019-2022)
- Machine Learning Developer, Bit Amsterdam (2017-2019)
- Teaching Assistant, VU University Amsterdam (2017-2018)
- Learning Multi-Level Features with Matryoshka Sparse Autoencoders (2025)
- Sparse Autoencoders Do Not Find Canonical Units of Analysis (2025)
- BatchTopK Sparse Autoencoders (2024)
- Neural Additive Vector Autoregression Models for Causal Discovery in Time Series (2021)
- Towards Empathic Deep Q-Learning (2019)
- Trojan Detection Challenge Red Teaming Track - Best Black Box Method, NeurIPS (2023)
- PRINCE Out-of-distribution Generalization Challenge - Winner, ECML-PKDD (2022)
- Learning By Doing Competition (CHEM Track) - 3rd place, NeurIPS (2021)
- Causality for Climate Competition - 2nd place, NeurIPS (2019)