Cross-Stack Research Directions

This section covers cross-stack research directions that shape how Hyperon's modules learn, transfer, and cohere over time. These are currently explored theoretical approaches to governing the flow of learning, inference, memory, and self-revision across the system.

Predictive and Causal Coding

Responsible: Faezeh Habibi, Bezawit Abebaw (iCog), Daniel MacDonald, Willis Ferenbaugh, Charlie Derr, Matthew Behrend, Amir Mohammadi

Predictive coding is a neural learning framework in which hierarchical layers continually generate predictions and update themselves through local prediction-error dynamics. Within this picture, learning is not treated as a single monolithic end-to-end adjustment, but as an iterative inferential process in which latent states and parameters are refined through structured exchanges of top-down prediction and bottom-up error.

Predictive and causal coding work with information-geometric principles and commutator relationships to shape how learning propagates through the system: local influence estimates, mixed-curvature structure, and small-commutator dynamics help determine where updates should go, where they should not go, and how modular competence can be preserved under continual adaptation.

Causal coding extends this framework by introducing interventional influence into the learning process, so that updates are directed preferentially toward the modules that are actually causally implicated in a given context, while clarity and pruning pressures suppress redundant or merely correlational pathways. The result is a more modular learning dynamic intended to reduce interference, bound catastrophic forgetting, and support continual adaptation under distribution shift.

Recent formulations develop this further through a two-level architecture in which Bayesian routing governs which columns or modules should be active, reused, or forked, while predictive-coding microstructures within those modules are kept internally coherent through pruning, inhibition, and shell-based consolidation. In this form, predictive and causal coding are not merely neural techniques but a broader strategy for scalable neural-symbolic integration, offering a principled account of how sub-symbolic learning dynamics, modular reuse, transfer, and emergent higher-level structure may be coordinated across the Hyperon stack.

TransWeave

Responsible: Ben Goertzel

TransWeave is a framework for carrying useful structure forward when a system moves from one task, environment, or regime into another. The core idea is that an intelligent system should not have to either cling rigidly to an old solution or start over from scratch. It should preserve what is still true, adapt what has changed, and do so in a disciplined way.

In formal terms, TransWeave studies transfer maps that preserve the deep organization of a task while allowing local adaptation. In reinforcement learning, this is framed through Bellman–Darboux intertwining: a transfer is good when "transfer then learn" comes out nearly the same as "learn in the new setting," which gives a principled way to judge when prior structure remains valid, when adaptation can stay local, and when more substantial revision is needed.

Within Hyperon, TransWeave is best understood as a cross-stack strategy for continuity of intelligence. It complements predictive coding, causal coding, and modular memory by helping explain how learned structure can persist across changing contexts, so that modules still suited to the new reality remain largely intact while learning effort concentrates where the shift actually occurred.

This matters for practical scalability. Systems that transfer structure well should adapt faster, preserve more prior competence, and use less compute. In that sense, TransWeave is not just a narrow transfer-learning theorem family, but part of a broader AGI strategy for making intelligence cumulative rather than repeatedly rebuilt.

WILLIAM

Responsible: Ben Goertzel

WILLIAM is a compression-based cognitive pattern detector whose design makes it inherently cross-stack. Its core principle: the patterns worth retaining are those that compress experience most effectively. By embedding directly into MORK's trie infrastructure, WILLIAM becomes a real-time signal source for multiple Hyperon subsystems simultaneously — rather than a batch tool operating within a single module.

Its compressed pattern statistics feed PLN (prioritizing high-value subgraphs for inference), backward chaining (following statistically heavy edges), schedulers (compression-adjusted priorities), ECAN (compression-driven attention signals), and neural components (heavy-hitter features for efficient forward passes). This broad consumer pattern means a single data structure produces cross-system coherence without explicit inter-module messaging.

In the cross-stack picture, WILLIAM plays a complementary role to predictive and causal coding: where those frameworks govern how learning propagates through the system, WILLIAM governs what is worth learning by tracking statistical weight across the full knowledge graph. See WILLIAM for technical implementation details.