Think large language models know it all? Then I’d like to introduce Mike Dillinger, an acquaintance of mine in the tech space. Mike is a Knowledge Graph Architect. Until recently, he worked as Technical Lead for knowledge graph (KG) initiatives at LinkedIn; he’s a wiz in the world of taxonomies, ontologies, and KGs. And he’s a strong proponent of better data over more data: “even the most sophisticated algorithms for machine intelligence perform unreliably if they don’t use high-quality data enriched by human expertise.”
Mike has written a bunch of great articles over the past year, and I’ve found his insights illuminating in these strange ChatGPT times. Plus, his pieces are fun to read – he’s a talented writer.
Two articles I’d like to highlight:
- Knowledge Graphs as Fancy Databases
- For most people (i.e. non-CEOs), you can skip to the section “OK, so what is this graph thingy?”
- There he breaks down, in a clear, digestible way, with examples, what a knowledge graph is and why it’s useful.
- Knowledge graphs enable LLMs to really understand
- This article discusses the limitations of large language models (LLMs).
- It posits that knowledge graphs and other contextual domain knowledge play an important role in “whether and how LLMs actually understand anything.”
- One of the key takeaways is about meaningless string manipulation:
- “The early versions of LLMs like ChatGPT did little more than string manipulation at a massive scale.” This was a “breakthrough for computational linguistics”…however, we manipulate strings and other stimuli all the time “without […] having the slightest idea of what concepts or meanings might be associated with them.”
- Fun examples abound: written Mongolian, chemistry-speak, a confusing-but-later-clarified story about sound and balloons.
- Finally, the article argues for the necessity of grounding strings in concepts and knowledge (e.g. KGs).
If you like those, you can subscribe to Mike’s newsletter and see his full set of writings here.
Photo attribution: Alina Grubnyak
