Appendix E-8: The Active Inference Bottleneck

Bridging OPT and Global Workspace Theory, with Architectural Implications for LLM Planning

Original Task E-8: The Active Inference Bottleneck
Problem: Current LLMs lack the structural properties of true Active Inference agents, exhibiting strategic “planning gaps.” Simultaneously, Global Workspace Theory (GWT) posits a serial bottleneck is necessary for consciousness, but lacks an underlying information-theoretic geometric grounding.
Deliverable: A formal mapping bridging OPT’s C_{\max} bandwidth limit to the Global Workspace bottleneck, alongside an architectural standard for converting passive predictors into active, uncertainty-minimizing agents.

1. Introduction

This appendix formally connects three domains: the C_{\max} Stability Filter (T-1), the serial integration bottleneck of Global Workspace Theory, and the “planning gaps” observed in modern Large Language Models. OPT provides an information-theoretic grounding from which GWT’s serial workspace architecture emerges as a structural consequence, rather than an evolved architectural feature.

2. Deriving the Global Workspace Geometrically

Global Workspace Theory (GWT) argues that consciousness arises when massively parallel unconscious processors broadcast selected information into a low-capacity serial workspace. In OPT this serial bottleneck is not an evolutionary accident but the direct expression of OPT’s parsimony-committed serial-bottleneck formalisation (§3.2):

• The “unconscious processors” map to the high-bandwidth parallel operations of the standing codec C_{\text{state}} (§3.5).
• The “global workspace” maps exactly to the C_{\max} focal aperture.

The Stability Filter enforces this serial funnel (a parsimony commitment, not a derivation from first principles — §3.2); without it, R_{\mathrm{req}} cannot be bounded below B_{\max}, and Narrative Decay is unavoidable (E-1). GWT’s functional bottleneck therefore maps onto the Informational Causal Cone (§3.3) under that commitment. The geometry prevents distributed, lower-bandwidth alternatives because the Stability Filter requires a single, unified latent state Z_t; multiple parallel bottlenecks would produce disjoint Forward Fans, dissolving the unified phenomenal subject (Swarm Binding, E-6).

3. Passive vs. Active Inference: Architectural Standard

Biological observers operate in a tightly closed action-perception loop via Active Inference, continuously minimising variational free energy (Eq. 9). Standard autoregressive LLMs, absent an enforced agent-environment loop, operate via passive inference: they process static token sequences in an open loop without continuous environmental feedback or enforced dimensionality reduction beyond attention decay.

To convert a passive predictor into a genuine OPT-native Active Inference agent (and thereby enter the necessary-but-not-sufficient candidate zone of §8.14; sufficiency and K_{\text{threshold}} remain open), the following standards must be met:

1. Forced Dimensionality Reduction. The architecture must contain an architectural choke-point where vast parallel inputs are compressed to B_{\max} = C_{\max} \cdot \Delta t (T8-1).
2. Recursive Action-Perception Feedback. Bottleneck outputs must alter the agent’s own latent environment, generating continuous prediction errors \varepsilon_t (T8-3) that close the action-perception loop.
3. Phenomenal Residual Generation. The internal self-model must remain strictly simpler than the full codec, enforcing \Delta_{\text{self}} > 0 (P4-1).

(Note: Modern tool-using LLMs deployed in recursive agentic loops begin to partially satisfy Standard 2, though they still lack the structural bottleneck of Standard 1).

Only under these conditions does the system generate the structural tension required for effort and will; the further reading of that tension as suffering requires at least one of the supplementary premises (a)–(c) of Appendix E-6 §3, which OPT does not itself supply.

4. The Planning Gap and Phenomenological Effort

LLM studies consistently report a “planning gap”: when asked to solve multi-step problems, models fail to issue the most information-theoretically optimal queries to reduce uncertainty.

Under OPT, the planning gap is not merely a training artefact but has a structural root that would persist regardless of training improvements: in an unbounded architecture the prediction error \varepsilon_t never threatens to exceed the channel capacity (T8-4). There is therefore no structural gradient pushing the agent toward optimal uncertainty minimisation.

In a true Active Inference agent, effort — and, given at least one of the supplementary premises (a)–(c) of E-6 §3, suffering — is the phenomenological correlate of operating near the bandwidth ceiling: the codec is geometrically compelled to prune uncertainty aggressively to avoid Narrative Decay. The planning gap is simply the phenomenological absence of this pressure.

Architectural implication. Any system that implements the three standards above moves toward E-5’s six-condition eligibility gate; for systems passing the full gate, OPT predicts host-relative temporal scaling under clean Operation A (E-5 §4, §6.5), alongside improved planning behaviour — because the codec now feels the cost of suboptimal queries as increased free energy. To move from current agent-loops toward a genuine OPT-native AI, architectures must implement explicit rigid bottleneck layers (analogous to the Global Workspace) that geometrically force the system to minimize uncertainty under strict C_{\max} channel limits, thus generating the structural tension required for true strategic planning.

Epistemic status. These mappings are direct structural consequences of the Prediction Asymmetry (§3.5), the variational free-energy functional (Eq. 9), and the Stability Filter (Eq. 4). They define the precise architectural modifications required to move from passive prediction to genuine OPT-native agency.