Dr. Jeff Beck is a computational neuroscientist focused on the mathematical and physical foundations of intelligence, and has recently challenged the traditional philosophical boundaries that define agency in artificial systems. In an interview with an interviewer, Beck dissected the fundamental concepts of agency, intelligence, and architectural changes that drive modern AI research, with a particular focus on predictive models, energy-based systems, and the risk of human cognitive decline in an automated future.
Beck argues that the structural distinction often drawn between agents and mere objects is mathematically weak. When viewed through the lens of policy implementation, where the system maps inputs to outputs, a gravity-responsive rock is fundamentally no different from a complex algorithm. The only real difference is complexity. Beck says that from a purely mathematical point of view, “there is no structural difference between agents and objects. It’s really just a matter of degree.” Agents are simply objects capable of very sophisticated computation, allowing things like planning and counterfactual reasoning over vast timescales.
This leads directly to the agency “black box problem.” In other words, how do we know if the system is really thinking or just running a precomputed lookup table? If we can only observe input-output policies, we cannot conclusively conclude that the agent is engaged in deliberation, counterfactual reasoning, or goal-directed behavior. The only way to infer true agency is to look at internal states and calculations, that is, the means as well as the ends. This challenge has led researchers toward modeling approaches that make internal states tractable and interpretable. According to Beck, the science is fundamentally about data compression. We employ the simplest computational model that adequately describes the observed behavior. If the simplest models require internal states representing plans, then we pragmatically conclude that the system has agency, even if the absolute truth remains unknown.
The search for robust and generalizable AI architectures deeply reflects the need to model complex internal dynamics. Beck emphasizes the fundamental relevance of energy-based models (EBM), pioneered by figures like Jan LeCun. Unlike traditional neural networks, which simply optimize weights to minimize an output cost function, EBM incorporates a cost function across all states, including latent variables. This architectural choice forces the model to treat its internal representation as a trainable latent variable, allowing optimization at inference time. This is the essence of training-at-test, a key technique gaining traction across deep learning, and one that aligns neatly with Bayesian inference, where the energy function is similar to log-probability. The ability to optimize internal states during inference gives EBM powerful flexibility not found in standard feedforward networks, effectively building flexible inductive priors into the system.
A powerful representation of this change is Yann LeCun’s Joint Embedding Predictive Architecture (JEPA). JEPA goes beyond predicting every pixel in an image (a task that often results in models wasting capacity by learning high-frequency pixel-level artifacts) to predicting compressed abstract representations in latent space. This non-contrastive approach attempts to maintain semantic richness and fidelity while avoiding the “modal collapse” problem where embeddings become trivially identical. By making predictions in this compressed latent space, JEPA encourages networks to learn robust, high-level features essential to understanding the world. Focusing on learning abstract representations rather than brute force pixel prediction is seen as a necessary step to building truly generalizable models.
Beck sees true artificial general intelligence not as a monolithic, centralized entity, but as a system that can autonomously build novel and specialized models. He draws a convincing parallel to biological evolution, suggesting that intelligence advances through the combination of specialized and modular components. He argues that the highly complex and non-smooth nature of olfactory space, which requires associative processing rather than the simple translational symmetries found in visual space, may have been a key driver in the evolution of our association cortex. Therefore, the next major milestone in AI is to create systems that autonomously generate and test hypotheses, rather than simply achieving high performance on fixed tasks. This ability is what Beck calls “active reasoning.”
A commitment to autonomous, continuously learning systems introduces complex safety considerations. Beck is less concerned with the catastrophic and rogue superintelligence scenarios that are often portrayed in popular media and more focused on the risks of human cognitive decline. If AI automates all critical thinking, humans risk becoming mere “choosers of reward functions”, passively approving or rejecting the output of algorithms without deep understanding. Beck argues, “We really don’t want a situation where humans like value function selectors. They just basically say, oh, no, I don’t like the result, please do this instead.” He suggests that the solution lies not in automation per se, but in building AI that improves human understanding and facilitates scientific discovery, always employing inverse reinforcement learning (IRL) to derive goals from observed human behavior, and ensuring that suggested actions include small, controllable perturbations rather than broad, simple commands.
Ultimately, philosophical and technical arguments converge on the idea that intelligence is a measurable degree of computational sophistication that is best modeled through architectures that prioritize robust internal representations and the ability to dynamically combine expertise modules.
