LLMs and AGI: Exploring the 'Creativity' Gap

Forum topic · ✨步子哥 · 2025-12-23

Summary

This article examines the fundamental gap between today's large language models (LLMs) and artificial general intelligence (AGI), centered on what it calls the 'creativity gap.' Drawing on Columbia University computer science professor Vishal Misra's argument that LLMs cannot create new scientific paradigms—famously illustrated by the claim that an LLM trained only on pre-1915 physics could never derive relativity—the piece argues LLMs are sophisticated pattern-matching engines confined to a 'Bayesian manifold' defined by training data. It uses information theory to explain LLM reasoning: inference is driven by entropy minimization, chain-of-thought works by decomposing problems into low-entropy steps, and entropy inversely tracks model confidence. The article formalizes LLMs as giant sparse matrices whose outputs are an 'inductive closure' of training data. Key limitations include statelessness, inability to recursively self-improve, hallucinations, and lack of embodied physical intuition. As a path toward AGI, the article surveys new mathematical frameworks (algebraic topology, information geometry, Kolmogorov complexity, arithmetic manifolds with geometric operators) and novel architectures such as the Recursive-Diffusion-Coherence (RDC) design. Its conclusion: AGI will require new theories and architectures beyond scaling LLMs, not merely larger models.

Key points

Core thesis: There is a fundamental "creativity gap" between current LLMs and AGI. The gap stems from essential differences in cognitive mechanisms, knowledge representation, and innovation capability—not something fixable by scale or iteration alone.
Vishal Misra's argument: The Columbia CS professor contends LLMs are sophisticated pattern-matching machines that cannot perform true scientific discovery or create new paradigms. His famous claim: *"Any LLM trained on physics before 1915 could never come up with relativity"* — because relativity required breaking Newtonian absolute spacetime assumptions, a disruptive "manifold jump" (source).

Theoretical tools

Bayesian manifold: LLM knowledge lives within a probability-distribution structure defined by training data in high-dimensional parameter space. It is finite, structured, and closed—creativity fails because LLMs cannot step outside this manifold.
Information theory and entropy: LLM inference is driven by entropy minimization. Lower entropy corresponds to higher generation confidence. Chain-of-thought (CoT) prompting works because it decomposes complex prompts into low-entropy steps; monitoring entropy can indicate whether a model is doing grounded reasoning or hallucinating.
Matrix model: An LLM can be conceptualized as a giant sparse matrix (rows = possible prompts, columns = possible tokens). Its output is an "inductive closure" of the training data—the set of all conclusions reachable by finite inductive steps from that data.

Intrinsic capabilities and limits

Strengths: statistical correlation detection, efficient text generation, multilingual understanding/translation.

Fundamental limitations:

No genuine understanding or abstract reasoning; no paradigm-shifting theory innovation
Lack of persistent memory — current LLMs are largely "stateless" (source)
Cannot recursively self-improve, since training data is static and human-provided (source)
No embodied physical intuition; knowledge derives entirely from text
Hallucination: generating factually incorrect content

Paths toward AGI

New mathematical frameworks:

Algebraic topology / information topology: analyzing information flow in neural networks via homology theory, with concepts like "information rings" and "homological capacity" (source)
Arithmetic manifolds and geometric operators: representing knowledge and reasoning as geometric operations on high-dimensional manifolds rather than tokens (source)
Information geometry and Kolmogorov complexity: quantifying creativity via geometric properties of probability distributions and novelty of information

New AGI architectures:

Recursive-Diffusion-Coherence (RDC) architecture:
*Recursive* — mimics hierarchical cortical structure, processing information at different abstraction levels
*Diffusion* — mimics neural signal spread, promoting cross-module coordination
*Coherence* — maintains internal consistency of system information

Conclusion

The author argues that the road to AGI necessarily requires new theories and architectures beyond the current LLM framework. LLMs are fundamentally statistical predictors bound by their training data's "knowledge manifold"; genuine paradigm innovation and scientific discovery demand systems capable of abstraction, hypothesis testing, and "internal thought experiments" that can jump outside that manifold.