LLMs Are Data Compressors, Not Optimal Predictors—The Road to AI Singularity Needs Symbolic Reasoning

A new paper on arXiv from researchers in algorithmic information theory challenges a prevailing assumption about what large language models actually do, and what that means for the prospect of achieving artificial general intelligence (AGI).

How LLMs Actually Work#

The conventional training objective for language models—cross-entropy loss and next-token prediction—optimizes them to compress and reproduce patterns in data rather than to implement Solomonoff induction, the theoretical gold standard for optimal prediction under bounded computational resources. This distinction matters because Solomonoff induction weights candidate hypotheses by their program length, a principle that remains fundamentally different from fitting a model to observed patterns.

When researchers scale up compute within the same training objective, models improve their fit to data, but they do not automatically transition into better universal predictors in the algorithmic information theory sense. The data-processing inequality and fundamental conservation laws in information theory continue to hold—there is no violation of the underlying physics of computation. But for finite learners working with finite data, these theoretical boundaries become less constraining than commonly assumed.

What Gets Lost in the Standard Approach#

The paper argues that language models operating under conventional objectives cannot implement Solomonoff induction on their own. They are, in essence, sophisticated data compressors that latch onto statistical regularities rather than discovering the simplest underlying mechanisms that generate those patterns. This leaves a crucial gap: the ability to recognize and retain genuine improvements to a system's own reasoning and decision-making—the self-improvement feedback loop required for an intelligence explosion or singularity.

The Path Forward: Symbolic Model Synthesis#

Rather than predicting doom or breakthrough from scaling alone, the research points toward a different trajectory: the integration of symbolic computation with statistical learning. Advanced frontier models—including systems like Fable and Astra—are already moving in this direction, incorporating components that go beyond pure statistical pattern matching to include explicit model synthesis and symbolic reasoning.

This neurosymbolic approach represents a qualitative shift in how AI systems operate, allowing them to generate and test hypotheses about their own decision-making process, not merely reproduce learned patterns. Without this layer of symbolic reasoning, the authors suggest, computational power alone will hit a ceiling in terms of meaningful progress toward AGI.