Leonard F. Bereska

4.125, LAB42

Science Park 900

1098XH Amsterdam

I’m Leonard, a PhD candidate at the University of Amsterdam working to make AI go well for humans. My research has migrated from the internals of neural networks toward bigger questions about what happens when increasingly capable AI systems interact with human institutions at scale.

research focus

I started in mechanistic interpretability, reverse-engineering how neural networks compress and represent information. That work led to a a review paper documenting foundations for the field. When I became interested in its connection to adversarial robustness, I had to formalize and measure superposition as a prerequisite. But (maybe prompted by recent developments in the world) recently I became interested in what happens outside the model: how do AI systems reshape the humans and institutions that deploy them? How can AI be a force for freedom?

I’m now developing theoretical foundations for understanding human agency in an AI agency-saturated world. The core intuition: genuine agency requires being predictable to yourself and unpredictable to others. I am excited about how this connects thermodynamics, active inference, and political philosophy. More on this soon.

I supervised Master’s students on projects ranging from bias detection and truthfulness in LLMs to interpretability in medical AI.

community building

In 2023, I co-founded the AI Safety Initiative Amsterdam (AISA) with colleagues who shared the sense that the “ChatGPT moment” demanded action. We’ve organized technical talks, panel discussions with OpenAI researchers and UvA professors, and facilitated reading groups bringing new researchers into the field. AISA is now run by the next generation of AI safety researchers and thriving.

I’m now working to establish an AI Safety Hub in Amsterdam, aiming to build continental Europe’s premier AI safety hub. This involves navigating the question of how safety research can benefit from proximity to industry without being captured by it.

beyond research

When I’m not thinking about neural networks or institutional dynamics:

Practicing yoga and meditation
Reading science fiction (Vernor Vinge for plausible singularity scenarios, Ted Chiang for everything else)
learning languages, (treating language acquisition as a window into cognition), currently learning Spanish
Playing with my two chihuahuas, Cicchetti and Pancetta
Building my digital fortress against surveillance capitalism

latest posts

Dec 15, 2025	Superposition as Lossy Compression – Measure with Sparse Autoencoders and Connect to Adversarial Vulnerability
Aug 19, 2024	Mechanistic Interpretability for Adversarial Robustness — A Proposal
Jul 10, 2024	Mechanistic Interpretability for AI Safety — A Review

selected publications

TMLR
Superposition as Lossy Compression — Measure with Sparse Autoencoders and Connect to Adversarial Vulnerability

Leonard F. Bereska, Zoe Tzifa-Kratira, Reza Samavi, and Stratis Gavves

TMLR, Aug 2025

Abs Bib HTML PDF

Neural networks achieve remarkable performance through superposition: encoding multiple features as overlapping directions in activation space rather than dedicating individual neurons to each feature. This phenomenon challenges interpretability; when neurons respond to multiple unrelated concepts, understanding network behavior becomes difficult. Yet despite its importance, we lack principled methods to measure superposition. We present an information-theoretic framework measuring a neural representation’s effective degrees of freedom. We apply Shannon entropy to sparse autoencoder activations to compute the number of effective features as the minimum number of neurons needed for interference-free encoding. Equivalently, this measures how many "virtual neurons" the network simulates through superposition. When networks encode more effective features than they have actual neurons, they must accept interference as the price of compression. Our metric strongly correlates with ground truth in toy models, detects minimal superposition in algorithmic tasks (effective features approximately equal neurons), and reveals systematic reduction under dropout. Layer-wise patterns of effective features mirror studies of intrinsic dimensionality on Pythia-70M. The metric also captures developmental dynamics, detecting sharp feature consolidation during the grokking phase transition. Surprisingly, adversarial training can increase effective features while improving robustness, contradicting the hypothesis that superposition causes vulnerability. Instead, the effect of adversarial training on superposition depends on task complexity and network capacity; simple tasks with ample capacity allow feature expansion (abundance regime), while complex tasks or limited capacity force feature reduction (scarcity regime). By defining superposition as lossy compression, this work enables principled, practical measurement of how neural networks organize information under computational constraints, in particular, connecting superposition to adversarial robustness.
@article{bereska_superposition_2025, title = {Superposition as Lossy Compression --- Measure with Sparse Autoencoders and Connect to Adversarial Vulnerability}, author = {Bereska, Leonard F. and Tzifa-Kratira, Zoe and Samavi, Reza and Gavves, Stratis}, year = {2025}, month = aug, journal = {TMLR}, }
TMLR
Mechanistic Interpretability for AI Safety — A Review

Leonard F. Bereska, and Efstratios Gavves

TMLR, Apr 2024

Abs Bib HTML PDF

Understanding AI systems’ inner workings is critical for ensuring value alignment and safety. This review explores mechanistic interpretability: reverse-engineering the computational mechanisms and representations learned by neural networks into human-understandable algorithms and concepts to provide a granular, causal understanding. We establish foundational concepts such as features encoding knowledge within neural activations and hypotheses about their representation and computation. We survey methodologies for causally dissecting model behaviors and assess the relevance of mechanistic interpretability to AI safety. We investigate challenges surrounding scalability, automation, and comprehensive interpretation. We advocate for clarifying concepts, setting standards, and scaling techniques to handle complex models and behaviors and expand to domains such as vision and reinforcement learning. Mechanistic interpretability could help prevent catastrophic outcomes as AI systems become more powerful and inscrutable.
@article{bereska_mechanistic_2024, title = {Mechanistic Interpretability for AI Safety — A Review}, author = {Bereska, Leonard F. and Gavves, Efstratios}, year = {2024}, month = apr, journal = {TMLR}, eprint = {2404.14082}, }