Learning is not only about generality
Type: kb/types/note.md · Tags: learning-theory, deploy-time-learning
People often equate learning with generality: knowing more, handling more cases, and covering wider scope. But making something more reliable, faster, or cheaper is equally learning. A system that can now multiply without hallucinating has learned, even though it handles no new cases. Capacity is not a single axis.
Herbert Simon's definition grounds this: "learning is any change in a system that produces a more or less permanent change in its capacity for adapting to its environment." By this definition, almost every KB improvement is learning. But not all capacity changes are equal: a typo fix and a design principle discovery both increase capacity, but on different dimensions and at different scales.
Generality — how widely does the capacity apply?
| Change | Scope | Example |
|---|---|---|
| Fix a typo | One retrieval | System can now match a query it would have missed |
| Sharpen a description | One note's findability | All queries that might match this note work better |
| Add a connection | Two notes' mutual discoverability | Navigation between these ideas now exists |
| Define structured sections for a type | All future notes of that type | Every related-system note gets consistent structure |
| Discover a design principle | All future decisions in that area | "Types and directories are orthogonal" applies broadly |
| Improve methodology | All future KB operations | The verifiability gradient changes how everything constrains |
The scope axis contains a qualitative distinction, not just a quantitative one. Accumulation — adding knowledge to the store — is itself a learning operation, and the most basic one. But what you accumulate varies in explanatory-reach. At the narrow end are facts — "the key is on the table," a specific claim, a particular observation. At the broad end are rules and theories — "types and directories are orthogonal," a design principle, an abstraction. Both are genuine learning through accumulation. Explanatory-reach is the property that distinguishes them: facts are adaptive knowledge (useful for the immediate context but don't transfer), while theories are explanatory knowledge (they apply in contexts they weren't designed for, because they capture structure rather than circumstance).
Fact accumulation is real and valuable — AgeMem's 23-49% improvement on task completion comes entirely from storing and retrieving the right facts. But facts alone don't compound into deeper understanding. Positing an abstraction and recognizing particulars as instances of it is the operation that produces theories — and recognizing what the particulars share is its expensive step. Theories are the highest-explanatory-reach items accumulation can store.
Argyris's single- and double-loop distinction (snapshot required) separates error correction that leaves underlying objectives and policies unquestioned from inquiry that questions governing values. This note interprets that distinction as rough regions on the generality axis: single-loop changes tend to be narrower because they retain the governing commitments, while double-loop changes can be wider because those commitments enter the update space.
The compound — reliability, speed, cost
Capacity has a second cluster of dimensions that tend to move together: reliability (how consistently it works), speed (how fast), and cost (how cheaply). These form a compound because they often improve simultaneously — codification is the clearest example, where moving from LLM to deterministic code improves all three at once. But the compound isn't exclusive to codification: conventions improve reliability, caching improves speed, derived skills reduce cost.
The generality trade-off is the primary dynamic: constraining and extraction can trade generality for reliability, speed, or cost. By this note's analogy, learning can occupy either rough region: codifying one check into a script resembles single-loop change, while deciding that claim notes should use Toulmin-derived sections resembles double-loop change.
Other dimensions
Generality and the compound don't exhaust what matters. A verified claim is more useful as a premise than an unverified one — not because it improves the system's current capacity, but because it improves its capacity to learn further, making the artifact load-bearing for future reasoning. Composability is a meta-capacity: the ability to build on what the system already knows.
Sources
- Herbert Simon: "Learning is any change in a system that produces a more or less permanent change in its capacity for adapting to its environment."
- Chris Argyris: Organizational Learning and Management Information Systems (snapshot required) — primary source for the single-/double-loop distinction; its mapping to generality is this note's interpretation.
- Knowledge acquisition — extracting and structuring knowledge from sources; one region on the learning spectrum, not a separate activity.
Relevant Notes:
- constraining and extraction can trade generality for reliability, speed, or cost — extends: the two mechanisms that operate on the generality trade-off defined here
- constraining — one mechanism: constrains the interpretation space, trading generality for reliability, speed, and cost
- theory and methodology form a two-layer execution system — the other mechanism: working use-shaped artifacts out from reasoning under context budget constraints
- codification — the far end of constraining where the compound gain is largest
- the verifiability gradient — structures the compound dimension
- deploy-time learning — frames where the gradient-moving happens on the system-adaptation timescale
- first-principles reasoning selects for explanatory-reach — grounds: Deutsch's reach criterion (registered here as explanatory-reach) distinguishes facts (adaptive, no explanatory-reach) from rules (explanatory, with explanatory-reach)
- recognition, not linking, is the hard problem in knowledge systems — mechanism: converting facts into rules turns on recognizing what they share, which is where the cost falls
- memory management policy is learnable but oracle-dependent — exemplifies: AgeMem learns a policy for managing facts (when to store, retrieve, summarize) but operates entirely on facts, never producing rules or explanatory-reach
- A method's ceiling bounds the method, not the transfer it already made — extends: reliability, speed, and cost gains inside a fixed automation envelope transfer no responsibility away from people