Very interested to see, on Kurt Cagle’s Substack “The Ontologist” ontologist.substack.com… that he is outlining the equivalent of our “AI sandwich” (osf.io/8v5bd) as the model for how people are now using LLMs with knowledge graphs. The LLM translates the human’s question from English for the knowledge graph and then translates the knowledge graph’s answer into English for the human.

After describing how knowledge graphs were gaining traction till shiny new LLMs came along, but then how users tired of hallucinations, he says now

“there’s a shift taking place in organisations. The LLMs aren’t necessarily going away, but they are increasingly being relegated to the role of being transformers for back-end symbolic data stores (knowledge graphs). That is to say, the LLMs become a natural language bridge for querying the stores, and once the results are produced, they also become a similar bridge for converting the results of such queries into natural language output”