Jennifer Prendki

Artificial intelligence and theoretical physics have historically evolved as largely independent disciplines. Yet both seek to answer fundamentally similar questions, such as how information is represented, how internal states evolve, and how interactions among those states give rise to complex behavior. Recent advances in AI have been driven primarily by large-scale language modeling (and specifically, Large Language Models). These systems have demonstrated that language contains a remarkable amount of information about the world and that many cognitive tasks can be approximated through statistical modeling of linguistic patterns. However, the success of this paradigm has also reinforced an implicit assumption: that reasoning itself can be represented and manipulated within linguistic space. We believe that this assumption may be fundamentally limiting. Because language is only a lossy observable manifestation of cognition, it is unlikely to constitute the substrate in which reasoning actually occurs. Instead, reasoning may emerge from stateful operations performed over latent internal representations that continuously evolve as an agent interacts with the world. From this perspective, intelligence becomes fundamentally a problem of dynamics rather than sequence generation. Many of the mathematical tools developed in theoretical physics (including state spaces, dynamical systems, field theory, symmetry, invariance, and coupled processes) provide a natural formalism for describing the evolution of intelligent systems. This talk presents a mathematical model of intelligence in which intelligent behavior emerges from the interaction between two coupled dynamical systems: cognitive dynamics, which govern the formation and evolution of internal world models, and volitional dynamics, which govern attention, goals, resource allocation and action selection. Internal representations are further examined through a field-theoretic formulation that pairs latent state spaces with their corresponding conjugate spaces and connects successive states through gauge fields. This gauge-invariant structure provides a formal mechanism for distinguishing genuine changes in an internal world model from arbitrary changes in its internal representation. Although the complete mathematical formalism and computational implementation are beyond the scope of this presentation, this framework provides a unified perspective on common-sense reasoning, active perception, and the relationship between intelligence, cognition, and consciousness. More broadly, it suggests that many observations emerging independently from physics, neuroscience, consciousness research, and artificial intelligence may ultimately be manifestations of the same underlying dynamical principles.

Jennifer Prendki is a physicist, AI researcher, and technology executive whose work focuses on developing a physics-based framework for modeling intelligence and cognition.

Originally trained in particle physics, she earned her PhD from the Sorbonne, where she chose experimental physics over a purely theoretical path because of a strong conviction that even the most elegant theories must ultimately remain anchored in observation and the physical world. Her research included work at SLAC's BaBar experiment, where she studied the fundamental properties of matter through large-scale experimental data.

Over the past two decades, she has held leadership positions across the data science and artificial intelligence industry, including serving as Chief Scientist at Atlassian, leading machine learning and data science organizations at WalmartLabs and Figure Eight, and heading AI Data at DeepMind during the period preceding the release of the first generation of Gemini models.

Today, she is the Founder and CEO of Dyssonance, where her research focuses on developing a foundational theory of thought that integrates ideas from physics, artificial intelligence, neuroscience, and consciousness research to model reasoning as a dynamic, stateful physical process.