Artificial general intelligence through recursive data compression and grounded reasoning: a position paper

This paper presents a tentative outline for the construction of an artificial, generally intelligent system (AGI). It is argued that building a general data compression algorithm solving all problems up to a complexity threshold should be the main thrust of research. A measure for partial progress in AGI is suggested. Although the details are far from being clear, some general properties for a general compression algorithm are fleshed out. Its inductive bias should be flexible and adapt to the input data while constantly searching for a simple, orthogonal and complete set of hypotheses explaining the data. It should recursively reduce the size of its representations thereby compressing the data increasingly at every iteration. Abstract Based on that fundamental ability, a grounded reasoning system is proposed. It is argued how grounding and flexible feature bases made of hypotheses allow for resourceful thinking. While the simulation of representation contents on the mental stage accounts for much of the power of propositional logic, compression leads to simple sets of hypotheses that allow the detection and verification of universally quantified statements. Abstract Together, it is highlighted how general compression and grounded reasoning could account for the birth and growth of first concepts about the world and the commonsense reasoning about them.

💡 Research Summary

Arthur Franz’s position paper argues that the central problem of building artificial general intelligence (AGI) can be reduced to constructing a universal data‑compression algorithm capable of solving all problems up to a certain complexity threshold. The author contends that if a system can compress sensory data efficiently, it automatically acquires a model of the world, because compression is equivalent to finding short programs (hypotheses) that generate the observed data. This “compression‑first” approach is presented as a more principled alternative to the current narrow‑AI paradigm, which focuses on solving specific tasks and suffers from the curse of dimensionality.

The paper outlines six essential properties that a general compressor must possess:

1. Data‑dependent search‑space expansion – the algorithm should dynamically enlarge its hypothesis space based on the structure of the incoming data rather than relying on a fixed inductive bias.
2. Feature‑hypothesis sequences – data are to be represented as a hierarchy of increasingly abstract features, each corresponding to a candidate program that explains a portion of the data.
3. Progress measurement via compression rate – a quantitative metric R(L) is defined. For a given complexity bound L, all programs of length ≤ L are enumerated, their outputs collected, and the compressor’s ability to recover the optimal (shortest) program for each output is measured. The average compression ratio across all strings provides a continuous gauge of partial AGI progress.
4. Recursiveness – once an initial compression yields a set of hypotheses, these hypotheses themselves become “data” that can be further compressed. This recursive process yields a hierarchy of models, each level more compact than the previous, mirroring the way humans build layered concepts.
5. Orthogonal feature basis – features (or hypotheses) must be mutually orthogonal to avoid redundancy and to ensure that new hypotheses do not conflict with existing ones. This property supports efficient inference and clear interpretation.
6. Interpretability – the final compressed representation should be human‑readable, allowing the system’s reasoning to be inspected and understood.

Building on this compression core, Franz proposes a grounded reasoning framework. Compressed hypotheses are “simulated” on a mental stage, providing a richer semantic substrate than classical propositional logic. Because the hypothesis set is simple and complete, the system can detect and verify universally quantified statements (∀‑type reasoning) and perform interventions to test hypotheses, thereby achieving a form of commonsense reasoning that is both resource‑efficient and flexible.

A novel contribution of the paper is the partial‑progress metric based on compression performance. Unlike the binary Turing test, R(L) offers a graded assessment: as the algorithm successfully compresses all strings generated by programs up to length L, R(L) approaches 1, indicating that the system has mastered all “simple” problems at that complexity level. The author suggests a bottom‑up research strategy: first solve all low‑complexity compression tasks, then gradually raise L, thereby scaling toward more sophisticated cognition.

The paper also discusses theoretical limits. Kolmogorov‑complexity theory tells us that most binary strings are incompressible; however, the subset of compressible strings corresponds precisely to the predictable portion of the world. Consequently, focusing on this small but meaningful subset is sufficient for intelligence. The author draws an analogy with human DNA, noting that the entire genetic code (~715 MB) is far smaller than modern software systems, implying that a general‑intelligence algorithm could be relatively compact.

In summary, Franz’s work reframes AGI as a problem of universal, recursive data compression coupled with grounded simulation of the resulting hypotheses. The proposed properties of the compressor, the compression‑based progress metric, and the grounding mechanism together outline a coherent research agenda. While conceptually compelling, the paper lacks concrete algorithmic designs, efficient hypothesis‑search strategies, and empirical validation, leaving substantial work for future investigations.