LLMs Navigate Knowledge, They Do Not Create

Vishal Misra is a professor of computer science at Columbia, specialising in networking and information theory. He is not an AI researcher. That is precisely what makes his framework for understanding large language models worth serious attention. Outsiders see structure that insiders, immersed in the daily work of prompt engineering and benchmark chasing, systematically miss.

His core claim: LLMs navigate a pre-existing knowledge space. They do not create new knowledge. This distinction sounds pedantic. It is not. It has profound implications for what these systems can and cannot do, and for the trajectory of the entire field.

The Accidental Theorist

Misra came to this question sideways. His background in information theory and network routing gave him a set of mathematical tools that, applied to language models, produced insights the ML community had overlooked. He did not set out to build a theory of LLMs. He noticed that the mathematics of knowledge navigation already existed, and that LLMs were instantiating it.

The key insight from information theory: any system that processes information must operate within a space defined by the information it has received. A network router cannot send packets to addresses it has never seen. A search engine cannot return documents it has never indexed. And a language model cannot generate knowledge that was not, in some form, present in its training data.

This seems obvious. But its implications are radical.

The Matrix and the Manifold

Misra’s framework begins with a geometric metaphor that turns out to be more than metaphor. The training corpus defines a high-dimensional manifold, a surface in mathematical space. Every piece of text in the training data is a point on this manifold. The LLM learns the shape of the manifold.

When you prompt a language model, you are specifying a location on this manifold. The model’s response is a traversal: it moves along the manifold’s surface, generating text that is consistent with the local geometry. Good responses stay on the manifold. Hallucinations are points where the model falls off the manifold into regions where it has no data to guide it.

This geometric view explains several puzzling behaviours:

• Why LLMs are fluent but sometimes factually wrong. Fluency is a property of the manifold’s local structure. Factual accuracy requires being on the right region of the manifold. These are independent properties.
• Why larger models hallucinate less. More parameters allow a higher-resolution representation of the manifold, reducing the gaps where the model might fall off.
• Why fine-tuning works. It reshapes the manifold locally, adding detail in specific regions without distorting the global structure.

The manifold metaphor is not just a teaching aid. It is a mathematical framework that makes testable predictions.

Why Chain-of-Thought Works

The chain-of-thought technique, prompting a model to “think step by step,” has become ubiquitous. But why does it work? The standard explanation is vague: something about “giving the model more room to think.” Misra’s framework provides a precise answer.

Chain-of-thought works because it constrains the traversal path. Without chain-of-thought, the model must jump directly from the question to the answer, a long traversal across the manifold that may cross regions where the manifold is poorly defined. With chain-of-thought, the model takes many small steps, each staying in well-mapped territory.

Think of it as navigation. You can drive from Paris to Berlin in a straight line, but you will end up in a field. The roads follow the terrain. Chain-of-thought is the instruction to follow the roads.

This also explains why chain-of-thought fails on certain problems. If the knowledge required to solve a problem is not on the manifold at all, no amount of careful traversal will reach it. You cannot navigate to a destination that does not exist on your map. Chain-of-thought helps with problems that are on the manifold but require a complex path to reach. It does not help with problems that require genuine novelty.

The Impossibility of Self-Improvement

Misra’s most provocative claim: LLMs cannot fundamentally improve themselves. Self-improvement, in the strong sense, would require the model to generate knowledge that is not on its training manifold and then incorporate that knowledge. But the model can only navigate the manifold it has. It cannot extend the manifold by traversing it.

This is not a limitation of current architectures. It is a mathematical constraint. A function cannot produce outputs outside its range. If the training data defines the range, then no amount of inference-time computation can extend it.

The implications for recursive self-improvement narratives are stark. The scenario where an AI makes itself smarter, which then makes itself smarter still, in an accelerating loop, requires exactly the capability that Misra’s framework says LLMs cannot have. The model can get better at navigating its existing knowledge. It cannot create new knowledge to navigate.

This does not mean AI systems cannot improve. It means improvement must come from outside the system: new training data, new architectures, new objectives. The human in the loop is not a temporary scaffolding to be removed. It is a permanent necessity. The system needs external input to extend its manifold.

The Plateau and What Comes Next

If LLMs are navigating a fixed manifold, then there is a natural ceiling on their performance: perfect navigation of the existing knowledge space. Once the model can traverse any path on the manifold with perfect accuracy, further scaling provides no benefit. You have a perfect map. Making the map larger does not help if the territory it covers has not expanded.

Misra argues we are approaching this ceiling for many tasks. The rapid improvement in benchmark performance is not evidence of increasing intelligence. It is evidence of improving navigation. And navigation has a ceiling that intelligence does not.

What comes next, in Misra’s view, is not bigger models but different architectures that can extend the manifold. Systems that can run experiments, gather new data, and update their representations in real time. Systems that do not just navigate knowledge but actively create it. These systems will look very different from current LLMs, and the transition will require fundamental architectural innovation, not just scaling.

Conceptual Toolbox

Misra’s framework provides several tools for thinking about AI capabilities:

• Navigation vs. Creation. Before attributing intelligence to an AI system, ask: is it navigating existing knowledge or creating new knowledge? If it is navigating, then its capabilities are bounded by its training data, no matter how impressive the navigation appears.

• Manifold coverage. The quality of an LLM is a function of how well its training data covers the knowledge manifold. Gaps in coverage produce hallucinations. Redundancy in coverage produces robustness. Data curation is manifold engineering.

• Path complexity. Some problems require simple paths across the manifold. Some require complex paths. Some require paths that do not exist on the manifold. The difficulty of a problem for an LLM is determined by the complexity of the path, not the complexity of the answer.

• The boundary test. To determine whether an LLM is truly reasoning or merely navigating, give it a problem that requires knowledge outside its training distribution. Genuine reasoning would produce correct novel answers. Navigation produces confident wrong answers. This is exactly what we observe.

The Cultural Problem

Misra’s framework has been resisted by parts of the AI community, and the resistance is revealing. The economic incentives of the AI industry depend on the narrative that LLMs are creating intelligence, not merely navigating knowledge. Billions in investment are predicated on the assumption that scaling will produce artificial general intelligence. Misra’s framework suggests it will produce artificial general navigation, a valuable but fundamentally different capability.

This is not a reason to dismiss the framework. It is a reason to take it seriously. The history of technology is full of capabilities that were mislabelled in their early years. Early automobiles were called “horseless carriages,” a framing that obscured their true nature. LLMs may be “knowledge navigators” mislabelled as “artificial intelligence,” and the mislabelling may be causing us to invest in the wrong research directions.

The practical implications are significant:

• For product builders: Design for navigation, not creation. Your AI assistant can find and synthesise existing knowledge brilliantly. Do not expect it to generate genuine insights.
• For investors: The value is in the knowledge base and the quality of navigation, not in the promise of emergent intelligence. Invest in data moats and retrieval systems, not in the hope of spontaneous capability emergence.
• For researchers: The interesting frontier is not scaling. It is architecture. How do you build systems that extend the manifold rather than merely navigating it?

LLMs are the best knowledge navigation systems ever built. That is an extraordinary achievement. But navigation is not creation, and confusing the two will lead us to underinvest in the architectures that might actually create.