From Stone Tags to AI: A Brief History of Ontology and Who Defines Reality

Brief History of Ontology

This piece explores how the concept of "ontology"—the study of what exists—has been repeatedly re‑drawn, from prehistoric stone markings to modern AI systems that decide what is real.

First Act – Philosophers' Ontology

Aristotle (c. 350 BC) created a ten‑category map ( Categories ) that split reality into entity, quantity, quality, relation, place, time, posture, condition, active, passive. He argued that labeling the world gave humans power to shape reality, a "knife" that could privilege some beings (e.g., slaves) over others.

Anselm of Canterbury (1078) used the same labeling trick to argue for God's existence: the "most perfect being" must exist, otherwise a greater being could be imagined, a classic ontological argument.

Descartes (1637) split existence into res extensa (extended things) and res cogitans (thinking things), establishing a dualism that later let machines treat the world as merely measurable objects.

These philosophical moves introduced the idea that categorizing reality is a form of power, not a neutral description.

Second Act – Machines' Ontology

Programmers independently reinvented the ancient trick: before a computer can process a domain, it must be told what kinds of things exist. Early databases reduced a person to three columns (ID, name, department), a violent "skinny‑down" of humanity.

Tim Berners‑Lee (2001) launched the Semantic Web, replacing string‑matching search with a knowledge graph of objects, properties, and links (RDF, OWL), making machines "understand" entities and relationships.

Palantir (2004‑present) added a third layer— Action —to the object‑link model, enabling systems to not only see but also act on the world (e.g., reroute a supply chain, fire a weapon). Their architecture (Data → Logic → Action → Security) provides a complete decision‑making pipeline with full auditability.

These systems illustrate that the same categorization power now resides in code and data pipelines, often in military or high‑risk contexts.

Third Act – Algorithmic Ontology

Large language models (LLMs) emerging around 2020 learn an implicit, statistical ontology from massive corpora. They can answer factual questions and propose actions, but their knowledge is "ghost‑like"—they never experience the world they describe.

Yao Shunyu , co‑author of the ReAct paper, emphasizes that an agent's success depends on a good environment and context . Without tools (e.g., a "order food" API), even the smartest model cannot act.

Palantir's object‑link‑action graph provides the concrete environment that ReAct agents need, turning implicit model knowledge into actionable, auditable operations.

The convergence of explicit ontologies (engineered by engineers) and implicit ontologies (emergent in LLM weights) creates a new power shift: machines now understand and can manipulate reality, but they still rely on human‑crafted maps for safe action.

Final Chapter – Who Owns Ontology?

The narrative ends with a warning: each era hands the "right to define reality" to a new actor—gods, philosophers, programmers, platforms, and now data‑driven algorithms. The ancient Blombos stone reminds us that reality resists being neatly boxed.

In a data‑centric worldview ("Dataism"), value equals information‑processing efficiency, turning humans into mere data nodes. The ultimate question is whether we retain agency when algorithms pre‑empt our judgments.

Practical Hook for B‑Side Readers

For enterprise AI projects, the key checklist is:

Identify the real objects in your business (customers, orders, assets).

Define their relationships.

Specify which actions an AI may perform on them.

Assign clear authority and audit trails (Security) for those actions.

Without a well‑crafted ontology, AI projects fail not because models are weak, but because the underlying map of reality is missing.

Thus, building a lightweight, composable "ontology environment" is the first step before scaling to heavyweight, security‑hardened solutions like Palantir.