Summary
Frank Coyle, a computer science educator at UC Berkeley with over 35 years of experience and a background in neuroscience, delivered a foundational talk at AI Engineer on why agentic AI systems require formal ontologies to function reliably. His central argument: large language models are inherently probabilistic, and hallucination is not a defect but a structural feature. Ontologies — formal, symbolic representations of entity relationships — provide the scaffolding that keeps LLMs operating within defined boundaries, giving rise to what researchers increasingly call neuro-symbolic AI.
Coyle traces dual intellectual lineages: agents back to John McCarthy, Marvin Minsky, and the 1956 Dartmouth conference; ontologies from Aristotle’s categories of being through philosopher Quine and Gruber’s 1993 formal definition — ‘a formal specification of a shared conceptualization.’ He explains how graph databases emerged from the rigidity of relational schemas, and how augmenting technologies like RDF Schema (RDFS) and OWL enable inference over graph structures. For example, RDFS domain and range declarations allow a system to infer from ‘Bob teaches Scooter’ that Bob is a teacher and Scooter is a student — deriving new facts without explicitly storing them. OWL’s transitive and functional properties enable further constraint and inference.
The talk surveys reusable public ontologies practitioners can build on today: schema.org for general web entities, FOAF (Friend of a Friend) for social network modeling, the Dublin Core for research metadata, and DBpedia — the ontology underlying Wikipedia’s search functionality. Coyle’s practical recommendation is to formalize organizational domain knowledge into explicit ontologies rather than relying on LLM probabilistic reasoning alone, giving agents a precise, auditable conceptual model of the domain they operate within.
📺 Source: AI Engineer · Published July 23, 2026
🏷️ Format: Deep Dive
