✦Crazy Wisdom

Crazy Wisdom

Episode #554: When Fluency Lies: The Knowledge Problem at the Heart of AI

Episode #554 · June 15, 2026 · 58 min

MP3 · Apple Podcasts · Spotify

About this episode

In this episode of the Crazy Wisdom Podcast, host Stewart Alsop sits down with Larry Swanson, creator of the Knowledge Graph Insights Podcast, for their second conversation together. The two cover a wide range of interconnected topics, starting with a correction Larry makes about the true origin of the term "artificial intelligence," tracing it back to the 1956 Dartmouth Conference and its distinction from Norbert Wiener's cybernetics. From there, the conversation moves through the history and structure of knowledge graphs, ontologies, RDF (Resource Description Framework), and the W3C standards process, touching on concepts like the T-box, A-box, and C-box, as well as the 25th anniversary of the Semantic Web paper. Stewart and Larry also dig into the limitations of large language models — particularly around reasoning, confabulation, and what Larry describes as "cognitive surrender" — and why symbolic AI and knowledge engineering may hold answers that the neural network world hasn't fully embraced. The episode also ventures into consciousness, panpsychism, Michael Pollan's ideas, and Stewart's own hands-on experience vibe coding a personal chatbot to replace functionality he feels he's lost with recent changes to Claude. Larry's podcast can be found at kgi.fm.
Timestamps
  • 00:00Stewart introduces Larry Swanson; Larry corrects the record on AI's origin, distinguishing it from Norbert Wiener's cybernetics at the 1956 Dartmouth conference.
  • 05:00Larry discusses interviewing semantic web paper coauthors on its 25th anniversary; RDF's hidden ubiquity compared to SIM cards powering everything invisibly.
  • 10:00Knowledge graphs explained through t-box terms, a-box assertions, and Dave McComb's c-box; IKEA's three-layer knowledge graph as a practical example.
  • 15:00Stewart connects metadata complexity to AI needs; faceted search explained as c-box attributes driving product filtering experiences.
  • 20:00RDF 1.2 reification standards discussed; W3C's rigorous recommendation process powering governments and enterprises worldwide through collaborative standards.
  • 25:00Cyc project examined as influential "successful failure"; Pat Hayes bringing description logic into semantic web; LLMs lacking true reasoning capability.
  • 30:00Epistemological fault lines between human and computer intelligence; cognitive surrender paper reveals no intelligence threshold protects against AI manipulation.
  • 35:00Stewart's Claude regression problem drives chatbot vibe coding quest; small language models and domain-specific approaches explored as alternatives.
  • 40:00Consciousness discussion through Michael Pollan's panpsychism lens; language versus cognition disconnect revealing LLMs as pure token-stitching without genuine thought.
  • 45:00Context graphs as purpose-built knowledge graphs for AI; Stewart's planning agents versus coding agents architecture and ground truth verification problem.
  • 50:00Docs-as-code versus code-as-docs paradigm shift; knowledge graphs as universal verifiers against validated facts; RDF 1.2 enabling provenance and degrees of certainty.
  • 55:00Jessica Talisman's Knowledge Graph Academy recommended for onboarding; kgi.fm podcast shared; knowledge representation community needs better abstraction for wider adoption.
Key Insights
  1. The term "artificial intelligence" was not a marketing gimmick but was coined deliberately at the 1956 Dartmouth Conference to distinguish the work of John McCarthy from Norbert Wiener's cybernetics. The two camps represented genuinely different approaches, and the AI label was a form of intentional intellectual branding rather than empty promotion.
  2. The semantic web, often called the most successful failure in technology history, has quietly embedded itself everywhere despite never achieving its original vision. Technologies like RDF power metadata standards inside every Adobe product and form the invisible backbone of government systems, enterprise data infrastructure, and cultural heritage organizations worldwide.
  3. Knowledge graphs are best understood as an ontology combined with all the instances that populate it. The distinction between things and strings, popularized by Google in 2012, captures the core idea that knowledge representation is about concepts as distinct from the labels we give them.
  4. The t-box, a-box, and c-box framework offers a practical model for understanding knowledge architecture. The t-box holds terminology and concepts, the a-box holds assertions about specific instances, and the c-box manages the attributes, taxonomies, and controlled vocabularies that sit between them and enable things like faceted search.
  5. Large language models produce fluent, convincing output but lack genuine reasoning, epistemological grounding, or judgment. Research on cognitive surrender shows that even people who understand how LLMs work are still susceptible to being misled by their fluency, meaning intelligence and awareness offer no reliable protection against being deceived.
  6. The gap between language and cognition matters deeply when evaluating AI. Evidence from people with aphasia shows that thinking can occur without language, which suggests LLMs, being purely language-based systems, are missing a fundamental layer of cognition that cannot be recovered through more tokens or better training.
  7. Knowledge graphs and RDF-based representation are well suited to the problem of verification and grounding in AI systems. Rather than relying on vectorized embeddings of language, a knowledge graph can store validated, provenance-tracked facts with degrees of certainty, making it a natural foundation for building trustworthy AI applications.
Episode transcript
Stewart Alsop III00:00

Welcome to the Crazy Wisdom Podcast. This podcast is for you. If you have an insane drive to find the truth of things, it's not the good answers that we seek, but the good questions. I interview a range of different guests from many different fields, all with the intention to uncover the simple truths that are hidden in plain sight. Most people don't want to go there. I go there, my guests go there, and you benefit. Please let me know if you enjoy these episodes, and as always, subscribe on itunes, Spotify, or wherever you listen to the podcasts.

Stewart Alsop III00:36

Welcome to the Crazy Wisdom Podcast. I've got Larry Swanson here, and he is the creator of the Knowledge Graph Insights podcast. And this is our second episode. And before we begin, Larry wants to talk about something we talked about on the last episode.

Larry Swanson00:48

Yeah, I have to, correct the record. I mistakenly said that AI, which I had heard many times, was like a marketing term that was made up at that 1956 Dartmouth concert. But actually, I have a friend who used to collaborate with John McCarthy, the guy who convened that conference. And the term was actually invented to disambiguate what they were doing. There were sort of two camps in AI at the time. Norbert Wiener was one camp, and he called what he did cybernetics. And the stuff that John McCarthy was doing, it's kind of similar, but different, you know, different enough. And there's kind of. I don't know if it was a spirited rivalry. you know, I don't know exactly what the distinction was, but they wanted to distinguish, their event from the cybernetic stuff. That's where the term AI came from. They were like, what? Okay, AI, Artificial intelligence. That's right. That's what we'll call it. So the. The true origin is more. It's not as snarky. It's not just a marketing term. It was a, it was kind of a branding thing. It's like, hey, what are we doing here? It's not cybernetics. It's artificial intelligence.

Your first full transcript is free. After that, an email opens every transcript in the index — a list of readers we can write to, not a guest book.

Subscribe