How Powerful Is Karpathy’s LLM Wiki? An Ontology and VSM Analysis

Part 1 – Ontology: What Exists?

In the AI era, building systems is so easy that the real question becomes what we are actually constructing and how components relate. Ontology is divided into three layers, with the first layer asking "What exists?" This layer contains five competing perspectives that each answer "What is real?" differently.

[Perspective 1] Aristotle : Reality consists of things with attributes. In AI this maps to files, rows, or documents stored in a database.

[Perspective 2] Whitehead : Reality is events, not static things. Understanding changes over time reveals that conversations are processes, not merely transcribed text.

[Perspective 3] Leibniz : Isolated things are meaningless; meaning arises from relationships. A note without links is dead, and a concept without neighbors lacks purpose, highlighting a graph‑oriented view.

[Perspective 4] Wheeler : Reality is information itself, not data. Schemas and API contracts are ontological assertions that most engineers overlook.

[Perspective 5] Heidegger : Reality is what is perceived by a particular person in a specific context. This phenomenological view brings humans into the ontology as the ultimate reason for a system’s existence.

Part 2 – VSM: How Does a System Remain Viable?

The fourth question—how a system stays viable—leads to the Viable System Model (VSM), originally created by Stafford Beer for any system that must survive in a changing environment.

Using a restaurant kitchen as an analogy, VSM defines six essential functions:

S1 Operations – the chefs actually cooking.

S2 Coordination – a scheduler preventing work‑station conflicts.

S3 Regulation – the head chef optimizing the current service.

S3* Audit – a food‑safety inspector verifying compliance.

S4 Scanning – the general manager observing trends, weather, reservations.

S5 Identity – the owner’s vision defining what the restaurant is and isn’t.

Removing any of these functions causes the system to degrade; for example, without S2 the kitchen becomes chaotic, and without S5 the restaurant drifts toward mediocrity.

Part 3 – When Ontology Meets VSM

Ontology provides diagnosis; VSM offers a prescription. Ontology tells which realities a system ignores, while VSM shows how to organize to avoid missing core elements. Their intersection can be visualized as follows:

S1‑S3 (Operations, Coordination, Regulation) sit on top of entity and information ontologies, handling concrete things and structured rules—areas where most AI systems excel.

S4 (Scanning) requires process and relationship ontologies to detect change and track connections; this is where systems become "blind".

S5 (Identity) is a phenomenological concern: the system’s purpose and what it deliberately excludes. Without it, a system may have capability but no direction.

Part 4 – Diagnosis: Karpathy’s LLM Wiki

Karpathy’s LLM Wiki is currently the most influential personal AI knowledge system blueprint. Its key insight is that maintenance, the traditional bottleneck of personal wikis, is solved by LLMs, reducing upkeep cost to near zero.

The model consists of three layers (raw source → Wiki → Schema) and three operations (ingest, query, check), with two special files: index.md (directory) and log.md (audit trail).

The diagnosis yields the same conclusion for both checks: the Wiki is strong in entity, relationship, and information dimensions (S1‑S3) but weak in process and phenomenology (S4‑S5). The ontology correctly predicted the architectural gaps, and the VSM confirmed them.

This is not criticism; rather, it shows that Karpathy’s design is an optimal core blueprint for personal AI knowledge accumulation, with a deliberately bounded scope.

Part 5 – Bridging the Gaps

The diagnosis suggests three extension directions:

Time Awareness (Process, Perspective 2) : The Wiki should distinguish current understanding from historical understanding, beyond what Git version history provides, so queries can retrieve the appropriate knowledge version.

External Scanning (S4) : In addition to internal linting, the system should autonomously detect external changes—new papers, deprecated libraries, or progress on concepts—and surface them without user prompting.

Explicit Identity (Phenomenology, Perspective 5 / S5) : An identity.md file should encode the system’s purpose and what it deliberately excludes, providing guiding principles rather than just schema conventions.

These extensions do not require rebuilding the Wiki; they augment the already robust S1‑S3 foundation. The gap lies at the top of the stack, not the base.

Both ontology and VSM are historic frameworks—ontology traces back to ancient philosophy, and control theory (VSM) emerged in the mid‑20th century. Their longevity, rather than being tailored for AI hype, makes them reliable foundations for building enduring systems.