If You Can't See Inside, How Do You Know It's THINKING? [Dr. Jeff Beck]

Agency is a spectrum of computational sophistication

Beck posits that no structural line separates agents from objects; instead, agency exists on a continuum defined by the complexity of internal states, the time scales over which they persist, and the context-dependence of policies.

Planning and counterfactuals define true agency

The defining characteristic of an agent is not sophisticated behavior but the capacity for planning and counterfactual reasoning—internally simulating future consequences of actions—which remains invisible to external observers measuring only input-output mappings.

Physical embodiment is necessary for genuine agency

Beck asserts that a high-fidelity simulation of a brain lacks agency unless physically embedded in the world; only through physical interaction and causal disconnection from immediate environmental impulses can something be considered an agent rather than a sophisticated reflex machine.

Agency requires adopting the intentional stance

Following Dennett, we treat systems 'as if' they are agents when modeling them as executing planning yields the most parsimonious explanation, though we can never confirm agency from behavior alone without inspecting internal computations.

EBMs impose constraints on latent variables

Unlike standard neural networks that optimize only weights for input-output mapping, energy-based models apply cost functions to internal states, requiring dual minimization of both hidden state energy and prediction error.

Energy minimization approximates Bayesian inference

Energy functions correspond to log probabilities in Bayesian frameworks, where minimizing energy equals maximum a posteriori (MAP) estimation without expensive normalization; the Free Energy Principle adds an entropy regularization term to this approach.

Test-time training risks methodological inconsistency

Modern approaches that activate latent variable optimization only during deployment (test-time training) are problematic because the base network was trained without accounting for these dynamics, unlike traditional EBMs where latent optimization occurs throughout training.

JEPA prioritizes latent prediction over pixel reconstruction

Joint Embedding Predictive Architectures compress inputs and outputs into abstract representations before prediction, enabling conceptual understanding rather than forcing models to reconstruct irrelevant pixel-level details.

Non-contrastive methods avoid biological implausibility

By eliminating the need for negative sampling and contrastive divergence through regularization techniques (as in Barlow Twins or VICReg), these architectures avoid biologically implausible backpropagation requirements while preventing representational collapse to trivial solutions.

End-to-end embedding learning replaces pre-processing

Rather than using fixed pre-trained encoders (like VAEs) as separate pre-processing stages, modern approaches aim to learn optimal compression jointly with the prediction task, treating embedding discovery as part of the optimization problem.