Created on 2024-12-16 23:44
Published on 2024-12-17 05:00
I stumbled upon an insightful podcast by BG2 (Bill Gurley & Brad Gerstner) featuring Microsoft's Satya Nadella over the weekend. For anyone interested in the current AI-driven industry shift, this is a must-watch.
During the podcast, Satya discussed the historical challenges around Bill Gates' vision for what was then known as Windows Future Storage (WinFS), an abstract information layer for Windows based on People, Places, and other well-known entities informed by an overarching schema—essentially “Schematizing the world.” He viewed this as somewhat disconnected from an approach that leverages the power of natural language, which he now embraces, as demonstrated by Microsoft's investment in OpenAI.
Today, the "Schematizing the world" vision of WinFS, which also existed in broader open form as the Semantic Web concept, is finally being realized thanks to the emergence of Large Language Models (LLMs)—essentially the long-awaited killer app for this vision. Why? Because logic inherently functions as the overarching schema for language, and shared dictionaries (or ontologies) have long existed based on fundamental principles that can be expressed and serialized using various notations and formats.
The major hurdle in the past was the lack of tools, grounded in natural language, to translate between a myriad of notations and formats. This problem often led to protracted technical disputes over achieving an improbable single solution. LLMs solve these challenges with their ability to generate data in a variety of formats. They go a step further by addressing the even trickier task of incorporating hyperlinks into structured data representations, where hyperlinks act as magical words with unique properties for entity naming, disambiguation, and de-referencing (lookup)—just as the World Wide Web has already demonstrated, where the primary focus has been documents and document content types.
Naturally, this has inspired a new batch of live demonstrations showcasing the combined prowess of Generative AI (GenAI) and knowledge bases (or knowledge graphs) deployed using Linked Data principles (i.e., a Semantic Web constructed with hyperlinks for naming entities and their relationships).
Generating notes from the BG2 Podcast's YouTube page and uploading them to a Knowledge Graph.
Interacting with the generated Knowledge Graph using natural language questions. Notably, prompts and eventual queries are fine-tuned by similarity analysis performed by the LLM, based on information retrieved through SPARQL (SPARQL Protocol and RDF Query Language) queries.
These demonstrations highlight how the synergy between LLMs, Knowledge Graphs, Ontologies, and Linked Data principles is revolutionizing our approach to knowledge representation and interaction. Exciting times lie ahead, as what once seemed improbable is now much more achievable!