Meta Rational Pragmatics article
MRP Article 16 Macro-AI and Conceptual Discovery

Macro-AI: Automated Conceptualization and Recursive Improvement

From governing execution to governing the production of abstractions.

Author: Sînică Alboaie Series: Meta Rational Pragmatics Focus: Automated conceptualization
Permalink & citation

Beyond Orchestration: Automating Conceptualization

The strongest extension of the previous argument is this: if Meta-Rational Pragmatics is taken seriously, then its research horizon is not limited to orchestrating existing semantic components more reliably. Its deeper horizon is the automation of conceptualization itself. That means the partial automation of generalization, the invention of more useful intermediate concepts, the discovery of more compressive working theories, and the gradual replacement of fragile prompt-heavy procedures with more stable local competence.

In that sense, MRP is not only about governing execution. It is about governing the production of the very abstractions through which execution later becomes tractable. The MRP program states this direction quite explicitly when it describes MRP-VM as a runtime for governed interpretation, explicit routing, typed intermediates, validators, and the gradual conversion of broad neural competence into more structured and reusable forms, with research automation named as a central strategic track rather than a side effect. [MRP-INDEX-2026] [MRP-SITE-2026]

The Distinction: Micro-AI vs. Macro-AI

Intelligence debates often remain trapped at the wrong level of description, focusing on micro-AI: better representations, losses, policies, or architectures. While these matter, they may not be the most decisive layer. There is another level—macro-AI—where intelligence appears as the capacity to frame a problem, isolate its structure, invent useful distinctions, and compress a messy field into a tractable theory.

Micro-AI concerns local engines (embeddings, world models, solvers, retrievers) that operate within a given framing. Macro-AI concerns the organization of inquiry itself—judgment, method, and prudence. It asks what counts as a relevant decomposition, which variables deserve to become concepts, and when a solution pattern should become an explicit reusable skill. MRP belongs to this macro layer, aiming to make the organization of intelligence computationally explicit without pretending one fixed ontology can do all the work. [MRP-REGIME-2026] [MRP-FOUND-2026] [MRP-AGI-2026]

Recursive Self-Improvement Reconsidered

Recursive self-improvement (RSI) should be interpreted more soberly than the "singularity" rhetoric suggests. The more realistic route is architectural and epistemic. A system improves itself when it learns better decompositions, concept libraries, routes between regimes, and validation criteria.

Recent work examines test-time forms of self-improvement without external feedback, but for MRP, the gain comes from converting weakly structured behavior into compact, explicit, and reusable modules. Improvement is not a monolithic mind rewriting itself from first principles, but the progressive specialization of execution paths. [YAMPOLSKIY-2015] [WANG-2018] [TRT-2026]

Automating Scientific and Conceptual Discovery

If a system only reuses the concepts it was handed in advance, its improvement is bounded by the semantic prejudices of its designers. However, growing evidence shows that parts of scientific discovery can be automated. We are moving from simple equation discovery to autonomous discovery systems that learn and refine interpretable concept libraries.

Systems such as AI-Descartes and AI-Hilbert show that discovery becomes stronger when raw data are constrained by theory, complexity, and formal background knowledge. This suggests that one can automate not only solution search but also the invention of the conceptual tools that make search more effective. [ASD-2023] [LASR-2024] [AIDESCARTES-2023] [AIHILBERT-2024]

Compression as the Engine of Tractability

A useful concept is a compression device with operational value. It reduces descriptive burden, organizes variation, and makes inferences cheaper. Scientific laws, type systems, and reusable skill contracts all have this character. Meaningful scientific models must balance accuracy and complexity.

In MRP terms, conceptualization is not ornamental; it is how a runtime manufactures tractability out of complexity. By generating human-interpretable (or system-auditable) knowledge, the system reduces the computational cost of subsequent reasoning rounds. [AIDESCARTES-2023] [ASD-2023] [LLM-SCI-2025]

The Loop: Automating Generalization

Automating generalization does not mean deriving one final universal abstraction. It means building a disciplined loop where:

  • Problems are decomposed.
  • Tentative frames are induced.
  • Candidate concepts are proposed and tested.
  • Stable patterns are lifted into explicit skills or typed intermediates.

This progressive conversion of expensive cognition into cheaper organized competence is the core of the MRP self-improvement model. Stable skill chains are replaced by deterministic code, validators, smaller routers, and specialized interpreters. [MRP-REGIME-2026] [MRP-RELATED-2026]

Methodological Competence and Disciplined Judgment

The missing pieces in AI may lie less in a new low-level miracle and more in the better organization of what exists. MRP raises the possibility that existing pieces—world models, small models, retrievers—are sufficient once placed inside a runtime that can conceptualize, route, and validate.

Wisdom, in this context, is higher-order compression and governance: the capacity to see what matters and choose a framing that makes action possible. MRP moves from semantic competence toward methodological competence, and finally toward theory-building under explicit control. It supplies the higher-order runtime where micro-AI engines can be framed, compared, and gradually transformed into compact conceptual machinery. [CHOLLET-2019] [ASD-2023] [MRP-FOUND-2026] [MRP-SITE-2026]

References

  • [AIDESCARTES-2023] Cornelio, C., et al. (2023). Combining Data and Theory for Derivable Scientific Discovery with AI-Descartes. Nature Communications.
  • [AIHILBERT-2024] Cory-Wright, R., et al. (2024). Evolving Scientific Discovery by Unifying Data and Background Knowledge with AI-Hilbert. Nature Communications.
  • [ASD-2023] Kramer, S., et al. (2023). Automated Scientific Discovery: From Equation Discovery to Autonomous Discovery Systems.
  • [CHOLLET-2019] Chollet, F. (2019). On the Measure of Intelligence.
  • [LASR-2024] Grayeli, A., et al. (2024). Symbolic Regression with a Learned Concept Library.
  • [LLM-SCI-2025] Zhang, Y., et al. (2025). Exploring the Role of Large Language Models in the Scientific Method: From Hypothesis to Discovery. Nature.
  • [MRP-AGI-2026] Alboaie, S. (2026). AGI Will Not Be One Thing. Article 01.
  • [MRP-FOUND-2026] Alboaie, S. (2026). Meta-Rational Pragmatics. Article 02.
  • [MRP-INDEX-2026] Alboaie, S. (2026). MRP Index. Index.
  • [MRP-REGIME-2026] Alboaie, S. (2026). Regime Selection and Tractable Computation as Regime Induction. Article 09.
  • [MRP-RELATED-2026] Alboaie, S. (2026). Related Research for Meta-Rational and Executable Pragmatics. Article 11.
  • [MRP-SITE-2026] AGISystem2. (2026). System 2 Engineering for Reliable AI.
  • [TRT-2026] Zhuang, Y., et al. (2026). Test-time Recursive Thinking: Self-Improvement without External Feedback.
  • [WANG-2018] Wang, W., et al. (2018). A Formulation of Recursive Self-Improvement and Its Possible Efficiency.
  • [YAMPOLSKIY-2015] Yampolskiy, R. V. (2015). From Seed AI to Technological Singularity via Recursively Self-Improving Software.