// talks
Presentations, lectures, and appearances.
2026
Past
World-models and the twilight of LLMs
Budapest Data+AI Forum 2026, Budapest, Hungary, May 19, 2026
A recent session on why just scaling LLMs is not enough to get to real intelligence, and why world models and latent-space reasoning look like a more useful direction for AGI.
LLMs: The perfect psychopaths – a case for JEPA
Budapest Data+ML Forum 2025, Budapest, 2025
LLMs are very articulate and surprisingly persuasive. But under the surface they are still mostly probabilistic machines, without real understanding, common sense, or reasoning. So in a way, they behave a bit like digital psychopaths.
JEPA (Joint Embedding Predictive Architecture) is a different approach: a model that learns more like an organism, builds hierarchical world models, and tries to predict reality instead of just imitating past data.
In this session I’ll argue that bigger LLMs are not the road to real intelligence, and that JEPA-like architectures are a more honest path toward AGI, with actual reasoning, adaptation, and something that starts to look like curiosity.
Formal verification of neural networks
Budapest Data+ML Forum 2024, Budapest, 2024
Large neural networks are showing up inside more and more security-relevant software environments. So it is increasingly important to look into their black-box behavior and make sure their properties stay within certain bounds. Formal verification of neural networks tries to give actual guarantees on properties like monotonicity, robustness and correctness. In the session I cover the main theoretical background (bound propagation, zonotopes) and also give a short look at practical tools like alpha-beta-CROWN and nnenum.
Interested in having me speak? Get in touch.