I really enjoyed the Semantics 2024 conference in Amsterdam this week. Some highlights, take-homes, and insights from the week.
Trust is crucial
Trust is paramount in AI. In his opening keynote, Ruben Verborgh pointed out the inanity of trying to establish trust in the BEOT (Brief Explosion of Trust) moment when you accept a websites cookies or agree to its terms of service. “Trust is a human relationship,” he says, “We cannot trust a machine, because it can’t trust us back. We build systems to support human relationships at scale. Only trust itself is immaterial, not the actions to build and sustain it.” So, HCD FTW here as elsewhere.
AI architectures are maturing
It’s apparent that we’re early in some sort of an AI-architecture transition/maturity curve. It started with basic RAG, is currently focused on “graph RAG” (which can mean anywhere from three to eight different things, depending on which expert you talk to), and is moving soon to hybrid AI that integrates LLMs, machine learning, NLP, knowledge representation, computer vision, and other tech. Fun times. And a particularly fun time to be in the semantics biz.
Long live the semantic web
The semantic web – the meaningful data/metadata part of the web built on the W3C RDF standard – may finally be coming into its own. It’s been here all along, hidden behind firewalls, in enterprise knowledge graphs, in Google’s knowledge graph, in Amazon’s product graph, in LinkedIn’s economic graph, and exposed to marketers and SEOs as they have used schema.org to promote their content to Google with linked RDF data. Vendors at the show reported strong enterprise interest in RDF-based knowledge tech (triple stores, ontology tooling, etc.). One interesting data point in that story is Samsung’s recent acquisition of Oxford Semantic Technologies and their RDFox knowledge graph and reasoning engine.
Data objects may be accounted for in multiple places
Lots of people in the semantics and AI worlds are thinking about how to use in AI architectures multiple representations of the same data. For example, accounting for the same data object as both a vector embedding and as an element in an RDF triple so you can account for both mathematical and semantic proximity.
Knowledge graphs provide true reasoning capabilities
There’s “reasoning” in the LLM sense, which is just more of the same plausible-sounding-but-fact-deficient genAI output, and then there’s the true kind of fact-grounded reasoning that RDF-based knowledge graphs enable.
LLMs and computational power alone can’t do everything
I saw at Semantics the same exasperation that I’ve seen in other similar communities with the propensity of developers and engineers to just throw LLMs and other computationally intensive methods at problems that could be solved more efficiently and economically with semantic tech and other well-established tech and methods.
This was Semantics’ 20th annual gathering but my first time attending. I’ll definitely be back.
