Building Knowledge Clusters: Toward Grounded AI

Designing Hybrid Systems That Check What AI Says

Hi everyone,

I went from writing book reviews for the Los Angeles Review of Books to sketching out architectures for knowledge clusters in hybrid LLM systems. Strange, fascinating times we live in. But I digress…

Conspiracy theories, conflicting claims disguised as value judgments (think of public disputes over the Israel–Hamas war), ambiguity (“which John Smith?”), and the inescapable fuzziness of language all bedevil any attempt to engineer a definitive storehouse of truth. The longing for certainty is not new. Greek philosophers like Plato sought to anchor knowledge in what Kant would later call analytic truths—statements true by definition, like “all triangles have three sides” or “all bachelors are unmarried males.” Leibniz dreamed of a universal calculus of reason, a perfect language in which every statement could be encoded and errors rendered impossible. Gödel shattered this dream by proving that truth cannot be fully captured by formal rules: what can be proved and what is true will never perfectly coincide.

And yet, the oracle fantasy persists. We still search for a machine that will give us final, unassailable answers. Large language models seemed, for a moment, to deliver that oracle—an all-knowing voice on demand. But the reality is closer to a glib salesman or a drunk in the corner pub: fluent, confident, and often wrong. These models are prolific generators of text but have no internal mechanism for recognizing truth independent of statistical likelihood. What “sounds right” in their decision space often wins over what is right. Entire research programs now attempt to bolt scaffolding onto these systems to check, retrieve, and filter their claims. But despite billions invested, there is still no reliable, end-to-end solution for truth in machine-generated language—and perhaps there never will be. Still, we can try.

Building systems with free-text processing used to mean hacking together SpaCy, NLTK, or custom Python scripts to parse, tag, and chunk text. Now, LLMs do all that with a single prompt—and sometimes better. But they come with a deeper problem: they speak without grounding their words in truth. This post argues that we can’t fix this with more scaling. We need hybrid systems that verify what LLMs say against real-world sources of knowledge.

I’ve been thinking about this problem since before LLMs went mainstream. Back in the Python-and-SpaCy era, I sketched versions of this “truth-grounding” idea but never pursued it. Now, with hallucinations everywhere and high-stakes use cases multiplying, I’m curious whether others see the value in building hybrid systems that actually check what they say. Let’s work on this problem….