Frank Coyle emphasizes the critical role of ontologies in agentic AI systems to provide structured knowledge that grounds probabilistic models like large language models, enhancing reasoning, error detection, and control. He advocates for a neurosymbolic approach combining symbolic ontologies with AI, encouraging hands-on experimentation and the use of existing resources to build robust, consistent, and reliable intelligent agents.
Frank Coyle, an educator at UC Berkeley with over 30 years of experience in computer science, opens his talk by reflecting on the evolving landscape of AI and computer science education. He emphasizes the importance of hands-on creation and learning through making, rather than just reading or typing, advocating for engaging multiple senses through writing and drawing to enhance understanding. This philosophy sets the stage for his discussion on the intersection of agent-based AI systems and ontologies.
Coyle traces the origins of agents back to the early days of AI, highlighting foundational figures like John McCarthy and Marvin Minsky, who conceptualized agents as entities that perceive, decide, and act. Ontologies, on the other hand, have philosophical roots dating back to Aristotle, who categorized being. In modern computing, ontologies represent formal specifications of shared conceptualizations, essentially structured knowledge about entities, their properties, and relationships, often implemented through graph databases. This formal knowledge is crucial for providing context and structure to AI agents.
He explains the convergence of probabilistic AI models, such as large language models (LLMs), with symbolic AI approaches embodied by ontologies, a fusion known as neurosymbolic AI. This hybrid approach helps keep LLMs “on guardrails,” mitigating issues like hallucinations by grounding their outputs in formal knowledge structures. Ontologies enable richer inference capabilities through technologies like RDFS and OWL, which define properties such as domain, range, transitivity, and functional constraints, allowing systems to derive new knowledge and validate data consistency.
Coyle also discusses practical methods for building ontologies, either top-down by domain experts or bottom-up by incorporating real-world data and interactions. He highlights existing taxonomies and ontologies like schema.org, FOAF, and DBPedia as valuable resources to avoid reinventing the wheel. Furthermore, he illustrates how ontologies can be integrated with agent loops in programming, using Python examples to show how LLMs can suggest tool usage but require external validation through ontologies to ensure correctness and prevent errors or infinite loops.
In conclusion, Coyle advocates for the use of ontologies as essential guardrails in agentic AI systems, enabling better reasoning, error detection, and control over probabilistic models. He encourages developers to leverage supporting technologies like Pydantic for type checking and to design agents with minimal side effects to maintain logical consistency. His overarching message is optimistic and creative: in AI development, “there is no win, no fail, only make,” urging continuous experimentation and learning.