In the race toward AGI, a long-standing idea has been that general intelligence would eventually emerge from language (or media) alone. The assumption was that if we trained sufficiently large models on the entirety of the internet, scaling would do the rest.

Yet even after training terabyte-scale models on vast corpora, they can still be outperformed by a six-year-old in basic reasoning and common-sense understanding.

The current phase of AI research, arguably underway for the past five years, has shifted focus. Instead of relying purely on next-token prediction in data space, the emphasis is now on learning structured world models and operating in latent space. The goal is to move beyond surface-level pattern matching toward models that internalize causal structure, physical intuition, and abstract representations of the world.

In this talk we'll deep dive to these.

Gergely's focus areas include formal verification of ML models, neo-architectures like entropy-based models, and the mathematics that ties it all together. So hyper-graphs and finite systems.
He’ve worn different hats over the years, CEO, CTO, researcher. These days he's deep in the research side of things.