Why Agentic Systems Need Ontologies — Frank Coyle, UC Berkeley

Summry

• Computer science degrees no longer guarantee employment due to AI's impact on the job market.
• Agents and ontologies are emerging as key AI components, blending probabilistic and formal representations.
• Hands-on learning is prioritized over traditional reading-based education.
• Ontologies and graph databases enable flexible knowledge representation compared to relational databases.
• Neuro-symbolic AI combines neural networks with symbolic systems to reduce LLM hallucinations.

Computer Science Education and AI Impact

• A computer science degree no longer ensures job security (Frank Coyle, 30-35 years of experience).
• AI and agents are the future, with 5,000 attendees noted at a related event.

Agents and Ontologies

• Agents and ontologies merge probabilistic (LLMs) and formal representations.
• Neuro-symbolic AI integrates neural networks with symbolic systems (e.g., rule-based systems, knowledge graphs).

Learning Philosophy and Methods

• John Cage’s philosophy: "Nothing is a mistake, there is no win or fail, only make."
• Hands-on engagement (writing, sensory activities) outperforms passive reading for learning.
• Writing engages the whole brain, accelerating learning compared to typing.

Ontologies and Knowledge Representation

• Ontologies trace back to Aristotle’s "philosophy of being" and were formalized by Von Quine and Gruber (1993).
• Graph databases enable flexible data modeling without restructuring (unlike relational databases).
• Existing taxonomies (e.g., schema.org, FOAF) can be reused to avoid redundancy.

Neuro-Symbolic AI and LLM Limitations

• Neuro-symbolic AI mitigates LLM hallucinations (an inherent LLM trait) through structured reasoning.
• LLMs are probabilistic and require external validation (e.g., Pydantic) for tool execution.

Practical Applications and Tools

• Ontologies detect errors like duplicate refunds or misdirected payouts using OWL disjoint properties.
• Pydantic adds type safety to Python and integrates with ontologies for validation.
• Agent design should avoid hardcoded type information to preserve flexibility.

Key Takeaways

• AI’s rise demands adaptability beyond traditional computer science education.
• Combining probabilistic (LLMs) and symbolic (ontologies) systems improves AI reliability.
• Hands-on learning methods surpass passive reading for skill acquisition.
• Ontologies and graph databases offer scalable, flexible knowledge representation.
• Neuro-symbolic AI addresses LLM limitations like hallucinations through structured reasoning.
• Tools like Pydantic and OWL enable robust validation and error detection in AI systems.