Why LLMs need knowledge graphs

Knowledge graphs have gone from a niche data architecture concern to the thing every boardroom is suddenly asking about. Tony Seale, known across LinkedIn as the Knowledge Graph Guy, has been making the case for over a decade, first at Deutsche Bank and then as the architect of the UBS knowledge graph.

In this episode, he talks to David Jones, co-founder and CTO of ipushpull, about why relational databases lose the nuance that matters, what an ontology actually is once you strip away the mystique, and why large language models swing from expert to useless the moment they step outside their training distribution.

Tony explains the neural symbolic loop: LLMs now remove roughly 80% of the effort of building a knowledge graph, and the graph in turn gives the model an internal, verifiable structure to reason over. He is blunt about the limits. There is no AI fairy. The work is still hard, and the organisations that put it off are the ones that will struggle in what he calls phase three, the agentic web, where data moves across organisational boundaries on open standards.

Also covered: why your out-of-distribution data is the only thing left that is worth anything, why distributed identifiers and open standards are unavoidable, and why "context graph" and "agent memory" are mostly rebrands of work that has 30 years of research behind it.