Continual learning's open problem is behaviour, not knowledge

Type: kb/types/note.md · Status: current · Tags: learning-theory, deploy-time-learning

Take Herbert Simon's definition of learning, cited in Learning is not only about generality: any durable change in a system's capacity for adapting to its environment. Under that criterion, continual learning splits by behavioral authority: who consumes a retained artifact, through which channel, and with what force. Knowledge artifacts are consumed as facts, evidence, reference, context, or advice. System-definition artifacts are consumed with instruction, enforcement, routing, validation, ranking, or learning force. Both are learning; they differ in what a durable write changes.

The knowledge half is solved by ordinary data engineering: databases, file systems, vector stores, retrieval-augmented generation (RAG), agent-memory records, and user profiles. In aggregate, these systems and stores hold far more than weights can. Adding entries grows the system's reach without changing its disposition.

Behaviour change — where durable writes change what the system does — is the open problem. Two mechanisms achieve it:

Mechanism Examples Behaviour-change profile
Distributed-parametric updates Fine-tuning, online learning, reinforcement learning from human feedback (RLHF), continual pretraining, adapters, embedding updates, learned controllers Heavy training infrastructure, substantial retraining cycles for large updates, behavioral probes rather than direct reading, and possible regressions in unrelated behaviours
Readable system-definition artifacts Prompts, tips, notes, schemas, tools, tests A durable write changes what the system does next session when an LLM, interpreter, runtime, validator, router, or retriever consumes the artifact with instruction, enforcement, selection, evaluation, or learning force. The artifact is cheap to write, review, and revert; inspectable; driveable by runtime signals

This split prevents retrieval from standing in for behaviour change. Retrieval grows accessible knowledge without changing the system's disposition. Distributed-parametric updates and readable system-definition artifacts both change disposition, but in different substrate regimes. Recognising readable-artifact work as a learning regime puts it into comparison with weight updates and learned controllers; the useful question is how the mechanisms combine.


Relevant Notes: