// talks
Presentations, lectures, and appearances.
2026
No talks yet this year.
Past
LLMs: The perfect psychopaths – a case for JEPA
Budapest Data+ML Forum 2025 — Budapest, 2025
LLMs are hyper-intelligent, articulate, and eerily persuasive. Yet, beneath the surface, they are nothing more than cold, probabilistic machines, devoid of understanding, common sense, or genuine reasoning—perfect digital psychopaths.
In contrast, JEPA (Joint Embedding Predictive Architecture) represents a radically different paradigm: an AI model that learns like a real organism, constructs hierarchical world models, and predicts reality rather than blindly imitating past data.
This session will dismantle the myth that bigger LLMs lead to true intelligence and explore why JEPA-like architectures are the real path toward AGI, capable of reasoning, adapting, and even developing something akin to curiosity.
Formal verification of neural networks
Budapest Data+ML Forum 2024 — Budapest, 2024
Large neural networks, as solid backbones of many modern ML techniques are appearing inside more and more secure software environments. It is now more than ever important to shine light into their black box behavior and to ensure that their properties stay within certain boundaries. The topic of formal verification of neural networks intends to aim to formally guarantee properties of neural networks such as monotonicity, robustness and correctness. During the session we will briefly cover the primary theoretical background of bound propagation and zonotopes while also giving a short insight to the practical implementations such as alpha-beta-CROWN and nnenum.
Interested in having me speak? Get in touch.