LLMs and AGI: Exploring the 'Creativity' Gap

*English translation/summary of a zhichai.net forum post.*

Executive Summary

A fundamental "creativity gap" separates today's large language models (LLMs) from artificial general intelligence (AGI). This gap cannot be closed by simple technical iteration or scale expansion; it stems from essential differences in cognitive mechanism, knowledge representation, and innovation capability.

LLMs are essentially complex pattern-matching machines performing statistical prediction over massive data. Their capabilities are strictly confined to the "knowledge manifold" defined by training data, and they cannot perform genuine paradigm innovation or scientific discovery. Therefore, the path to AGI requires theoretical and architectural innovation beyond the current LLM framework—not merely scaling.

1. Core Argument: The Fundamental "Creativity" Gap

Columbia University computer science professor Vishal Misra argues that although LLMs excel at pattern matching and text generation, they are inherently unable to perform genuine scientific discovery or theoretical innovation like humans:

> "LLMs cannot create new paradigms. They can only reason and combine within existing knowledge frameworks; they cannot step outside those frameworks to create entirely new scientific paradigms or theories."

The Einstein case study: Misra asserts that "any LLM trained on physics before 1915 could never propose relativity theory"—relativity required breaking Newtonian assumptions of absolute space-time, a disruptive "manifold jump."

| LLM limitations | True AGI characteristics | |---|---| | Cannot perform paradigm shifts or theoretical innovation | Can perform scientific discovery and theoretical innovation | | Confined to the knowledge manifold of training data | Capable of abstract thinking and hypothesis validation | | Lacks genuine scientific discovery ability | Can conduct "internal thought experiments" |

2. Theoretical Tools: Information Theory and Geometric Manifolds

Bayesian Manifolds: The Knowledge Boundary of LLMs

The "Bayesian manifold" is the core concept in Misra's framework: the probability-distribution structure an LLM learns in its high-dimensional parameter space, defined by training data. It is the model's "knowledge world" or cognitive boundary, characterized by:

• Finitude: fully determined by training data
• Structure: geometric structure in high-dimensional space
• Closure: cannot exceed the scope of training data

Information Theory and Entropy

• Entropy minimization drives LLM inference: the lower the entropy, the higher the model's confidence when generating content.
• Chain-of-thought (CoT) works because it decomposes complex prompts into low-entropy steps, reducing uncertainty.
• There is a direct inverse relationship between entropy and LLM confidence; monitoring entropy changes can indicate whether the model is "hallucinating" or reasoning soundly.

Matrix Model: Interpolation vs. Creation

Misra conceptualizes an LLM as a giant sparse matrix where rows represent possible prompts and columns represent possible tokens. Under the inductive closure principle, an LLM's output is the inductive closure of its training data—the set of all conclusions derivable from training data via finitely many inductive inferences.

3. Intrinsic Mechanisms and Capability Boundaries

Strengths: strong statistical correlation identification, efficient text generation/processing, multilingual understanding and translation.

Fundamental limitations:

1. No persistent memory — current LLMs are mostly "stateless," lacking durable memory of past interactions and long-term goal setting. 2. No self-improvement — LLMs fundamentally cannot perform recursive self-improvement because training data is static and human-provided. 3. No physical intuition — knowledge comes entirely from text data, lacking direct perception of the physical world and "embodied cognition." 4. Lack of genuine understanding and abstract reasoning; limited knowledge transfer; hallucination problems.

4. The Path to AGI: New Theories and Architectures Beyond LLMs

New Mathematical Frameworks

• Algebraic topology and information topology: analyzing information flow in neural networks via homology theory, with new concepts like "information loops" and "homological capacity."
• Geometric operators and arithmetic manifolds: thinking in "geometric operators" rather than tokens, representing knowledge and reasoning as geometric operations on high-dimensional manifolds.
• Information geometry and Kolmogorov complexity: quantifying creativity via geometric properties of probability distributions and measuring novelty with Kolmogorov complexity.

New AGI Architecture: Recursive–Diffusion–Coherence (RDC)

• Recursive: emulates the hierarchical structure of the cerebral cortex, recursively processing information at different abstraction levels.
• Diffusion: emulates neural signal diffusion, promoting coordination across modules.
• Coherence: ensures internal consistency and coherence of information.

5. Conclusion and Outlook

The post concludes that scaling alone will not bridge the creativity gap. Reaching AGI requires fundamentally new theories—drawing on topology, information geometry, and complexity measures—and new architectures capable of leaving the training-data-defined manifold, performing genuine abstraction, hypothesis testing, and paradigm-level innovation.