LLMs as Approximate Knowledge Bases

A useful mental model: LLMs function as approximate natural language knowledge bases that can be approximately queried.

The “approximate” qualifier matters in both directions. The knowledge is approximate, captured implicitly in weights rather than explicitly in retrievable facts, subject to hallucination, and unevenly distributed based on training data prevalence. The querying is also approximate, natural language prompts don’t have the precision of database queries, and responses vary based on phrasing.

This framing helps calibrate expectations. LLMs are not authoritative sources, but they do capture and can surface a wide swath of human knowledge (both commonsense and specialized) provided it appears sufficiently in training data. The practical implication: LLM outputs can dramatically accelerate work when paired with human expert verification.

Related:, 01-atom—human-in-the-loop