The evolution of LLMs within the Enterprise will be different from that outside the enterprise.

Introduction

Last week, I spoke at? European Economic and Social Committee on behalf of the European Internet Forum .

In my talk, for an audience of policy makers, I mentioned that the evolution of LLMs is likely to be complex and different from what we see today.?

Here is a broad flow of thought I am thinking of?

Consider LLMs in the Enterprise

The implementation of LLMs within the Enterprise could be very disruptive - and different from that outside the Enterprise. Specifically, within the enterprise, LLMs could be used with knowledge graphs (KG) for the reasons we discuss below

Outside the enterprise, LLMs will grow in terms of number of parameters i.e. size.?To create these LLMs, you need large, well-funded companies.?OpenAI leads this pack.?

However, inside the enterprise the same LLM maybe implemented differently

Within the Enterprise, knowledge graphs help to create a semantic layer for the Enterprise that complements (and overcomes the limitations) of LLMs.

Today, knowledge graphs in the enterprise solve two big problems of LLMs:

a) lack of explainability and

b) reducing the potential for hallucination.?

Over time, knowledge graphs could help to create a reasoning engine for AI within the Enterprise.??

Since last year, I have been working with LLMs and KG. It's nice to see that Gartner also has knowledge graphs at the center of their impact radar this year

The Enterprise Semantic layer

In an Enterprise, the semantic layer could have a number of components - such as? Metadata, Taxonomy & Information Architecture, Business Glossary , Ontology and of course,? Knowledge Graphs.? Knowledge graphs represent factual data about real world entities in the form of networked nodes and edges. When this symbolic knowledge is linked to LLM data,? we are able to ground the LLM providing explainability, reliability (and in future), reasoning.?

LLMs and deep learning models in general, are excellent at learning patterns from large amounts of data. However, they are black boxes. Also, they need vast amounts of data to learn and might struggle with tasks that humans find easy. Knowledge graphs - or more broadly,? neurosymbolic AI systems provide a set of constraints / context that effectively narrows down the search space that the neural network needs to consider. This not only improves efficiency but can also enhance the system's ability to generalise from fewer examples, as it leverages structured human knowledge to guide learning.

Tomaz Bratanic provides one such implementation in? Enhancing Interaction between Language Models and Graph Databases via a Semantic Layer where we extract parameters from the user input and uses a template based on user intent to provide the LLM with context.?

In my teaching at the University Of Oxford, we are also exploring other models such as small LLMs like Phi and Otca and the chaining GPTs to create more complex software?

Many thanks? to Louise Grabo for hosting the event and for European Economic and Social Committee and European Internet Forum. for inviting me. My talk was 45 mins and the Q and A was almost an hour!. I will share more about the talk including the #AIact discussions.

You can meet us at the The Oxford AI summit. We also announced two of our well known courses: low code AI course at the university of oxford - open to non developers and digital twins for AI course

with Eusebiu Croitoru Maria Rosa Gibellini . also thanks to Anthony Alcaraz David Stevens Philip Rathle for their ongoing insights to our work at the #universityofoxford on #knowledgegraphs