Ingest: Why AM and EURISKO Appear to Work
Type: kb/sources/types/ingest-report.md
Classification
This is a retrospective scientific paper: it explains the behavior and methodology of two implemented discovery systems, responds to a published critique of AM, and reports selected qualitative and aggregate observations rather than a controlled evaluation. Authors Douglas B. Lenat and John Seely Brown built or closely studied the systems and therefore offer first-hand design knowledge, while also defending the research program under criticism.
Summary
Lenat and Brown argue that AM's productive discovery of mathematical concepts did not follow from heuristic search alone. Its small LISP characteristic functions happened to make many simple syntactic mutations correspond to meaningful mathematical variations. The same mutation scheme failed on heuristics represented as large LISP programs because most local edits changed implementation details rather than domain meaning. EURISKO became more productive after people replaced those programs with a richer language of short, typed slots, making meaningful changes expressible and many meaningless code-level changes inexpressible. The paper therefore treats representation design as part of the effective hypothesis class of discovery. It also documents work outside the advertised learning loop: people chose starting concepts and slots, supplied domain-specific representations, encoded omitted heuristics, interpreted underconstrained outputs, recognized discoveries, and corrected some judgments.
Quotes
No source quotes have been retained yet.
Connections Found
The paper is a technical and historical anchor for the claim that representation affects productivity through the operations it makes locally expressible, extending the encoding-and-consumption axis in Representational form. It is evidence for A hand-crafted bootstrap fits the Bitter Lesson only if learning can outgrow it: AM and EURISKO searched productively inside vocabularies, mutation operators, and evaluation structures that people had substantially designed. It also supports A methodology governs its own extension only as far as it settles the meta-decisions it raises, because making heuristics a subject of discovery did not settle how heuristics should be decomposed; people had to redesign that representation before search improved.
Learning Claims (our opinion)
On the paper's account, AM and EURISKO learn by proposing syntactic variations of represented concepts or heuristics, evaluating their plausibility and interestingness, and retaining useful additions to an expanding vocabulary. Their effective update space includes the values of available slots, compositions permitted by their mutation operators, and, in a few reported cases, splits of existing slots. The history, agenda state, examples, and declarative relations exposed to heuristics condition which operations are attempted. The representation fixes the more consequential boundary: what distinctions are available, which edits count as small, and which semantic mappings the hypothesis class can express.
This is not theory refinement in Commonplace's established sense. The systems chiefly construct new concepts and heuristics rather than revise an explicit tentative theory against labelled cases while preserving prior knowledge. It does, however, sharpen the precondition for any refinement loop: separately editable parts help only when those parts and the allowed edits correspond to consequential distinctions. The contrast between two-page LISP heuristics and specialized short slots is strong mechanism evidence for Learning inside a fixed decomposition inherits its mistakes, but it is not an ablation. The paper reports a historical redesign alongside improved behavior, without isolating representation from accumulated heuristics, domains, implementation changes, or human practice.
The source also limits claims of autonomous self-extension. People selected much of the initial ontology and heuristic corpus, devised rich domain-specific slots, supplied slot semantics, interpreted ambiguous outputs, recognized familiar concepts the program did not value, found interest for reasons different from the program's, and corrected some misjudgments. EURISKO's few reported slot splits show that limited representation change entered the update space, while the larger decomposition and the criteria for a natural representation remained substantially external.
Extractable Value
- State representation productivity as operator-relative. The relevant property is not representational fidelity alone, but whether the system's available local transformations tend to produce meaningful variants and suppress implementation-only mutations. This sharpens the current representational-form account. [quick-win]
- Use heuristic learning as a fixed-decomposition case. Adding “heuristics” as another subject did not make AM able to learn useful heuristics because their large program representation gave its existing mutators the wrong units of change. [quick-win]
- Separate learned content from redesigned learning machinery. EURISKO learned values within a slot language and occasionally split slots, while people supplied much of the slot vocabulary, domain decomposition, and semantics that made this learning productive. [quick-win]
- Retain the human-machine system as the evaluation unit. Human observers supplied denotation, interpreted underconstrained descriptions, recognized rediscoveries, found alternative reasons for interest, and corrected judgments; attributing all resulting meaning or productivity to the program would erase causal work. [deep-dive]
- Treat the evidence as retrospective mechanism evidence. The paper offers implementation history, examples, and aggregate observations, but no controlled comparison that attributes improved heuristic discovery specifically to the new representation. [just-a-reference]
Limitations (our opinion)
The authors analyze their own systems while answering critics, so their access to design history is paired with an interest in defending the program. The paper does not present a controlled benchmark, matched runs, or an ablation of the old and new heuristic representations. Its examples show that particular syntactic edits can become semantically useful under a better decomposition, but do not measure how often this occurs or exclude accumulated heuristics and human intervention as alternative causes of improvement. Reported successes in several domains remain selected cases, and the three slot-splitting examples do not establish general autonomous representation learning. The account is unusually candid about omitted heuristics and human interpretation, yet it cannot reconstruct all volatile implementation details and therefore cannot fully separate system behavior from undocumented operator choices.
Recommended Next Action
Update Representational form with the operator-relative distinction that a representation constrains productivity by determining which semantically meaningful variations are reachable through the learner's available edits, using this ingest as evidence and preserving human redesign as an explicit boundary.