The Brain Is Just Specialized Agents Talking To Each Other — Dr. Jeff Beck

Agency as Policy Sophistication

Agency is not a binary state but a matter of degree, defined by the sophistication of the policy a system uses to map inputs to outputs. From a mathematical perspective, there is no structural difference between an object (like a rock) and an agent; both execute policies. The distinction emerges in the complexity of the internal computations—specifically whether a system engages in planning and counterfactual reasoning.

The Black Box Problem of Agency

Identifying whether a system is truly "planning" or simply executing a highly sophisticated pre-computed response is nearly impossible from an external perspective. Because an observer only sees the resulting action (the policy), they cannot definitively prove that internal rollouts or Monte Carlo Tree Search-style simulations occurred.

Dr. Beck proposes a pragmatic, model-based approach to this problem: if the simplest computational model that explains a system's behavior is one involving planning and counterfactual reasoning, it is reasonable to treat that system as an agent. This aligns with Daniel Dennett's "intentional stance," where treating a system as if it has goals is a useful explanatory tool, even if it is not a microscopic causal truth.

Physicality and Agency

Dr. Beck argues that true agency requires physical embodiment. While a high-fidelity computer simulation can model agency and predict an agent's behavior with 100% accuracy, the simulation itself is not an agent. Agency is tied to the physical interaction with the environment, suggesting that a model of agency is distinct from the existence of an agent.

Energy-Based Models (EBMs) and Bayesian Inference

Energy-Based Models differ from traditional feed-forward neural networks primarily in where the cost function is applied. In a standard network, the cost function operates on inputs and outputs to optimize weights. In an EBM, the cost function operates on both the weights and the internal states (hidden nodes) of the model.

Mechanics of EBMs

To arrive at a prediction, an EBM performs two minimizations:

1. Finding the energetic minimum associated with the internal states.
2. Minimizing the prediction error.

Variational Autoencoders (VAEs) are cited as a canonical example of EBMs because their cost function includes a term that constrains the internal representation (e.g., forcing the latent space to be Gaussian), rather than focusing solely on the reconstruction error between input and output.

EBMs vs. Test-Time Training

While current trends in "test-time training" treat some weights as latent variables to be optimized during inference, Dr. Beck notes a critical flaw: most of these models are trained in a purely supervised manner first. He argues that for true EBM-like behavior, the network should be trained with these latent optimizations active throughout the entire training process, not just at deployment.

JEPA and Latent Space Learning

Joint Embedding Prediction Architecture (JEPA), championed by Yann LeCun, shifts the learning objective from predicting every pixel (generative modeling) to predicting between embeddings in a compressed latent space.

The Advantage of Latent Prediction

Predicting every pixel often forces a model to focus on irrelevant details. By compressing both inputs and outputs into embeddings and learning the prediction between them, models can capture a more "gestalt" or conceptual understanding of the world. This approach treats science as a process of prediction and data compression.

Avoiding Model Collapse

A primary challenge in joint embedding is "model collapse," where the network finds a trivial solution (e.g., setting all embeddings to zero) to achieve perfect prediction. To prevent this, non-contrastive learning methods (such as Barlow Twins) use regularization to maintain the richness and fidelity of the representations without requiring the expensive negative sampling found in traditional contrastive methods.

The Modular Evolution of Intelligence

Intelligence is viewed not as a single general capability (AGI), but as a collection of specialized intelligences working together. Dr. Beck suggests that the brain evolved by combining simple, specialized modules that learned to communicate, creating emergent computational abilities.

The Olfactory Origin Theory

Dr. Beck proposes that the olfactory system may have been a primary driver for the evolution of the associative cortex. Unlike visual space, which is smooth and translation-invariant, olfactory space is combinatorial and highly complex. The neural machinery evolved to solve these non-smooth problems likely provided the foundation for the frontal cortex and higher-order planning.

Continual Learning and Meta-Programming

True intelligence requires the ability to perform continual learning—the capacity to encounter a novel situation and instantiate a new model or latent variable on the fly to explain it. This is exemplified by GFlowNets, which can be viewed as generative models of generative models, allowing a system to expand its own internal architecture to deal with novelty.

AI Safety and the Future of Work

Rather than fearing rogue superintelligences (e.g., "Skynet"), Dr. Beck expresses concern over human "enfeeblement," where humans become mere "reward function selectors" who simply approve or reject AI outputs without understanding the underlying process.

Safe Goal Specification via Inverse Reinforcement Learning

To avoid the "perverse instantiation" problem (e.g., an AI ending world hunger by eliminating humans), Dr. Beck suggests using Maximum Entropy Inverse Reinforcement Learning. Instead of specifying a goal by hand, the AI observes human behavior to derive an empirically estimated reward function based on the current stationary distribution of human actions and outcomes.

To improve the world safely, humans should make small, controlled perturbations to this estimated distribution and evaluate the consequences, rather than issuing broad, naive commands.