Dr. Jeff Beck discusses the challenges of understanding agency and intelligence in artificial systems, emphasizing the importance of geometric deep learning for modeling the physical world. He explains that incorporating physical symmetries, such as translation and rotation invariance, into models leads to more accurate representations of reality. Beck notes that while brute-force discovery of these symmetries is possible, explicitly building them into models is more mathematically satisfying and efficient. He also highlights the recent advancements in tools and techniques that allow for such modeling.
The conversation delves into the philosophical and practical definitions of agency. Beck argues that, from a modeling perspective, there is no strict structural difference between agents and objects—agents are simply more sophisticated objects with complex, context-dependent policies and internal states that represent information over long timescales. He points out that agency is often attributed based on observed behavior, but without access to internal computations (such as planning or counterfactual reasoning), it is difficult to definitively determine whether something is truly an agent or just executing a complex policy. The distinction, he suggests, is often a matter of degree rather than a clear-cut boundary.
Beck further explores the idea that agency and intelligence are best measured by the sophistication of computation and policy, referencing concepts like transfer entropy and the intentional stance. He discusses the limitations of functionalist views, which equate any input-output mapping with agency, and argues that physical embodiment and the ability to perform planning and counterfactual reasoning are key features of true agents. However, he acknowledges that, in practice, we often rely on the simplest explanatory model that fits observed behavior, even if it means treating something as if it were an agent.
The discussion transitions to technical aspects of machine learning, such as energy-based models, variational autoencoders (VAEs), and the importance of learning representations that balance compression and fidelity. Beck explains the advantages of energy-based models over traditional function approximation, particularly in terms of inductive priors and regularization. He also touches on self-supervised and non-contrastive learning methods, emphasizing the need for models that retain rich, generalizable information rather than discarding potentially useful data for the sake of specific tasks.
Looking to the future, Beck expresses optimism about the role of AI in scientific discovery and society. He envisions autonomous agents capable of continual learning, experimental design, and modular combination of knowledge—traits he sees as hallmarks of true intelligence. While acknowledging concerns about AI-induced “infeeblement” or loss of human agency, Beck believes that humans will adapt and that AI will ultimately serve as a partner in human progress. He stresses the importance of careful reward specification and empirical evaluation to ensure safe and beneficial AI development, concluding with a techno-optimistic outlook on the co-evolution of humans and intelligent machines.
