Generative AI can only answer questions or complete tasks based on what it's been trained on - unless it's given access to external knowledge. The"technology du jour" for providing that external knowledge is the embeddings or vector database, but anyone who knows anything about semantic knowledge graphsknows that a knowledge graph will give much better results.

In the last few days GPT-4 has been credited with "sparks of artificial general intelligence" so I thought I'd ask it which it would prefer. Here's what GPT-4had to say...

---

Steve:You're given two stores of external information to choose from, a semantic knowledge graph or a vector database. Think about the steps involved incompleting a question or task you (as an LLM) are given, which of these two store types would be more likely to give you the text you need?

GPT-4:
As an LLM, I would choose a semantic knowledge graph for completing a question or task. A semantic knowledge graph represents relationships betweenentities and concepts in a more structured manner, making it easier to identify the relevant information for a specific question or task. On the otherhand, a vector database represents information as numerical vectors in a high-dimensional space, which might be more difficult to interpret and extractthe required text directly. The structured nature of a semantic knowledge graph would allow me to easily navigate and find the information needed tocomplete the given task.

Steve:
Please give me five examples of questions or tasks that illustrate why you would make this choice.

GPT-4:
1. "What are the main causes of air pollution?"
A semantic knowledge graph would have structured relationships between entities and concepts, making it easier to identify the causes of air pollution.For example, it could have nodes for "air pollution" and "causes," and edges connecting different causes like "industrial emissions," "vehicleemissions," and "burning of fossil fuels."

2. "List the works of William Shakespeare."
In a semantic knowledge graph, there would likely be a node for "William Shakespeare" connected to nodes representing his individual works, such as"Romeo and Juliet," "Hamlet," and "Macbeth." This structure would make it simple to extract the required information.

3. "Describe the process of photosynthesis."
A semantic knowledge graph would have nodes and edges representing the process of photosynthesis, including concepts like "light," "chlorophyll,""carbon dioxide," "water," and "oxygen." The structured relationships between these concepts would make it easier to understand and describe theprocess

4. "What are the symptoms of the common cold?"
A semantic knowledge graph would have a node for "common cold" connected to nodes representing its various symptoms, such as "sneezing," "coughing,""sore throat," and "runny nose." This structured information would make it straightforward to extract and list the symptoms.

5. "Which countries are part of the European Union?"
In a semantic knowledge graph, there would be a node for the "European Union" connected to nodes representing its member countries, like "Germany,""France," "Italy," and "Spain." The connections between these nodes would make it easy to identify and list the countries that are part of the EuropeanUnion.

Steve:
Hallucination is a significant problem for LLMs such as yourself. If the information you need to perform a task is available to you in a semanticknowledge graph, do you think this problem could be eliminated?

GPT-4:
While using a semantic knowledge graph can significantly reduce the chances of hallucination by providing structured and accurate information,it may not completely eliminate the problem. LLMs, like any other machine learning models, are still prone to errors and might occasionallygenerate incorrect or irrelevant information. However, the structured nature of a semantic knowledge graph would undoubtedly help mitigate thehallucination problem by providing more accurate and reliable information for the LLM to work with.

---

So there it is - from the horse's mouth. Large Language Models (LLMs) need access to external knowledge and data if they're be generally useful and the semantic knowledge graph is the structure of choice. If you're a developer or provider of semantic graph database applications there will probably never be another opportunity like this one.

---

Stephen Young is CEO of FactNexus and a knowledge graph and complex data specialist. His lifelong passion for new knowledge technologies comes from his belief that our biggest problems are too complex to solve without artificial help. Steve re-engineered the famous ELIZA chatbot and built his first neural network as a CompSci and Psychology undergraduate and he published a web-based Knowledge graph, with web search and a conversational UI, a full two years before Google.