h t t p s : / / p r o c e e d i n g s . m l r . p r e s s / v 1 3 1 / t h o r i s s o n 2 0 a / t h o r i s s o n 2 0 a . p d f

Ingest: Seed-Programmed Autonomous General Learning

Type: kb/sources/types/ingest-report.md

Classification

This is a scientific paper that presents a unified cognitive-architecture theory and uses previously implemented AERA agents as supporting examples rather than reporting a newly controlled experiment. Author: Kristinn R. Thórisson, a principal developer of the constructivist-AI methodology, AERA, and the causal-relational learning account being argued for, writing from Reykjavik University and the Icelandic Institute for Intelligent Machines.

Summary

Thórisson argues that an autonomous general learner must begin with a small seed that grounds drives in observable and manipulatable variables, then build corrigible knowledge by turning correlations into fine-grained causal-relational models. Deduction, abduction, induction, and analogy continually create, compare, apply, revise, unify, compact, and delete those models; shared variables and patterns connect new phenomena to existing knowledge, while self-reference makes internal mechanisms available to the same selective operations. The paper presents this combination as a route to cumulative learning in partially observable, novelty-rich worlds and cites AERA's S0 and S1 agents as initial evidence, but those demonstrations support the compound architecture in bounded interaction tasks rather than isolating the claimed necessity of causal organization, representational granularity, reflection, or the chosen seed.

Quotes

No source quotes have been retained yet.

Connections Found

The paper is a technical basis and bounded implementation case for a hand-crafted bootstrap fitting the Bitter Lesson only if learning can outgrow it: its seed supplies drives, observables, an ontology, and initial models while the learner constructs task knowledge that the seed does not enumerate. It supports reflection buying addressability through an architecture whose small causal-relational models can refer to one another and be selectively inspected and revised.

It is also a useful limitation case for learning inside a fixed decomposition inheriting its mistakes. S0 and S1 learn within fixed transducers, seed drives, initial ontology, causal-relational representation, comparison and attention machinery, and available primitive actions. Their task success shows that this compound setup admitted useful learning; it does not establish that those fixed choices were necessary, preferable, or sufficiently general. The paper's multidimensional treatment of autonomy and generality further compares with learning not being only about generality, because speed, reliability, resource control, and capacity for further learning remain distinct evaluation dimensions.

Learning Claims (our opinion)

On the paper's terms, cumulative learning is model unification under novelty: prior models and active goals direct attention to relevant similarities; correlation generates tentative causal-relational models; deduction predicts forward, abduction plans backward, induction adjusts scope, and analogy transfers structure across partially shared patterns. Models remain corrigible, and fine granularity is meant to let the system change a small part without rebuilding its knowledge. This is close to theory refinement when an explicit causal-relational model has consequences that experience can contradict, exposes candidate repair locations, and can be edited separately. The mapping is incomplete because the paper also calls model creation, application, retrieval, compression, and deletion learning, and it does not separately demonstrate a retained model being diagnosed, revised, and then changing a later learning episode.

The effective update space is broad but bounded. Behavior can condition on sensor histories, current observations, active goals, prior models, inferred patterns, prediction error, and the architecture's self-references. The learner can compose primitive actions and reasoning operations to predict, plan, compare, create, modify, unify, compact, or delete causal-relational models. Its hypothesis class expresses temporally ordered mappings among fine-grained variable patterns and hierarchies of those mappings. Outside that space sit the transducers and their accessible variables, primitive action basis, top-level drives, initial ontology and models, causal-relational representational form, model granularity, similarity computation, resource-control mechanisms, and much of the architecture that schedules reasoning. The S0 and S1 demonstrations therefore show improvement within a designed decomposition. They do not test the decomposition against alternatives, and no reported ablation attributes success to any one of its fixed choices.

Extractable Value

  1. The seed is a grounded bootstrap rather than a miniature task solution -- Drives must refer to observables and possible effects, while learned causal-relational models supply later task knowledge. This gives the existing bootstrap note a concrete architecture and a precise condition under which hand-designed structure can be outgrown. [quick-win]
  2. Fine-grained self-reference connects reflection to selective revision -- Because models can reference other models and internal operations, the same comparison and planning machinery can address parts of the learner's own knowledge use. This supports the addressability mechanism while leaving persistence and improvement to be established separately. [quick-win]
  3. Novelty handling depends on overlap with existing observables or inferred features -- The theory's advertised domain independence is conditional: a new phenomenon must share a bridge with the seed or acquired knowledge. This is a useful boundary on claims of autonomous generality. [quick-win]
  4. The demonstrations expose a fixed effective update space -- S0 and S1 reportedly acquire grammar, reference resolution, task models, and action strategies, but only through supplied sensors, actions, drives, ontology, representation, and architectural machinery. Comparing alternative seeds or representations would be required to test the fixed decomposition. [experiment]
  5. The four reasoning modes divide complementary learning jobs -- Deduction simulates consequences, abduction searches backward from goals, induction adjusts generality, and analogy finds relevant partial overlap. This is a useful functional vocabulary for comparing learning loops, though the paper does not ablate the division. [just-a-reference]
  6. Generality is explicitly coupled to autonomy but remains multidimensional -- The account usefully rejects task count as a sufficient measure and includes novelty handling and autonomous knowledge acquisition, while leaving reliability, speed, resource cost, and learning capacity as separate empirical questions. [just-a-reference]

Limitations (our opinion)

The central contribution is a broad theory assembled from the author's research program. Its strongest empirical examples are summaries of earlier AERA work, with no matched baseline, seed ablation, alternative representation, or controlled test of causal organization, analogy, fine granularity, and reflection. Claims that these elements are necessary for general learning therefore outrun the evidence; the demonstrations establish at most that the compound architecture learned the reported bounded tasks under its supplied interfaces and evaluation conditions.

The paper's signals and histories are constrained by fixed transducers and initially named observables; its responses are constrained by the primitive action basis; its expressible mappings are constrained by causal-relational models and the architecture's reasoning operators. If a response-relevant distinction is absent from those histories, an action cannot be composed, or a mapping cannot be represented, later optimization cannot repair the omission. The requirement that novel phenomena overlap with existing observables or inferred features acknowledges part of this boundary, but does not establish how often the bridge exists or how the learner recovers when it does not.

The S0 and S1 accounts report successful end behavior and numbers of acquired models, but this paper does not provide sufficient protocol, baseline, failure distribution, or independent replication to assess robustness or scalability. The claim that fine-grained models make large knowledge sets tractable also faces an unresolved cost: S1 required roughly 1,400 models for a bounded interview setting. Finally, self-reference makes internal processes addressable, but addressability alone does not demonstrate accurate diagnosis, warranted revision, retained improvement, or safe recursive self-modification.

Update A hand-crafted bootstrap fits the Bitter Lesson only if learning can outgrow it to use this ingest as its concrete seed-programming case and state the paper's observable-overlap condition and fixed-decomposition limit.