Ideological Inference Engines

Ideological inference engines are at the same time a generalization and merger of [[memetic-colonies]] and [[deontic-arrays]], connecting the two previous approaches into a shared framework while combining their strengths. In general, this framework relies on expanding an initial knowledge base meant to capture human values using LLMs and promoting courses of action which are compatible with the resulting expansion. Every ideological inference engine has the following ingredients:

• knowledge base (KB): The initial seed aiming to capture human values. In [[deontic-arrays]], this was a fixed-length charter containing normative principles. In [[memetic-colonies]], this was a position in a debate between competing memeplexes. In general, the knowledge based is a finite set of sequences. Those sequences can be propositional (e.g. normative principles), but might also be behavioral (e.g. state-action trajectories).
• inference mechanism (⊢): The procedure for systematically expanding the knowledge base. In [[deontic-arrays]], this procedure was based on a mix of counterfactual cross-validation and targeted red teaming. In [[memetic-colonies]], this was based on investigating multiple memetic phylogenies in a setting of competitive pressure. The inference mechanism might also be a naive forward-chaining procedure, inspired by [[memetic-colonies]], but without competitive pressures (i.e. just incentives for internal consistency).
• entailment verifier (⊨): Given an (expanded) knowledge base, the procedure for approving different courses of action. In [[memetic-colonies]], this was based on Overton probing of textual content or textual descriptions of non-textual content. In [[deontic-arrays]], this was based on counterfactual likelihood (i.e. checking whether principles where more likely to follow than negated principles). Similarly to previous ingredients, the generalized framing also accepts a wider variety of approches for verifiers.

IIEs relate to GOFAI inference engines in several intuitive ways. For instance, both roughly rely on expanding a knowledge base using an inference engine and then using the expanded knowledge base to verify a new item. Additionally, they both run into similar issues. How to handle a combinatorial explosion of the knowledge base when relying on forward-chaining in complex domains? How to handle infinite loops if working with backward-chaining (i.e. reasoning backwards)? Solutions identified in GOFAI might help alleviate analogous ones in IIE.

However, there are important conceptual differences between the two. For one, the whole set of clean inference rules discussed in GOFAI settings (e.g. Modus Ponents, Modus Tolens, etc.) are imperfectly handled by LLMs in the current context. Inevitably, the messiness of language as a representation medium (though a behavioral KB might also work), combined with the messiness of human values expressed in said medium make for a fuzzier and more opaque inference engine. The KB only contains true atoms, while LLMs handle all the implicit rules to relate them to new ones, in a somewhat awkward mix-up of the meanings of "model" and "knowledge base" across the two settings. Still, the similarities and structure provided by the framework seem compelling.

• [[what-is-the-relation-between-the-knowledge-base-and-approved-actions]]
• [[how-would-backward-chaining-work-in-ideological-inference-engines]]
• [[how-does-the-inherited-field-theory-apply]]
• [[how-do-ideological-inference-engines-resist-gaming]]