Do LLMs succeed by being comprehensively encyclopedic?

Venkatesh Rao argues that the most important thing about large language models is not that they are encyclopedic — comprehensive, reliable coverage of a domain is "a necessary feature" but "not the most important thing about them." What distinguishes an LLM from every prior encyclopedic medium, he writes, is that it offers "effectively infinite ways of encyclopedic knowing": you can approach what it knows "from virtually any direction you can think of, with any ontological orientation, and it will offer meaningful traction." The payoff line is blunt: "You cannot easily catch an LLM wrong-footed, even if you can catch it hallucinating and bullshitting."

Rao's reasoning runs through a history of single-ordering encyclopedism. Diderot's Encyclopédie organized knowledge alphabetically; the rival Encyclopedia Methodique claimed superiority through thematic organization instead; Wikipedia claims a third advantage, "folksonomic encyclopedism" over the "scholarly kind" of the Britannica. Each, in Rao's telling, still commits to one best ordering principle. An LLM commits to none. It will meet a lexicographic query, a thematic one, or an idiosyncratic one with equal fluency, "though it may flatter you and compliment you on your originality of perspective," rather than betray that it has no canonical path through its own knowledge.

This sits alongside Does AI repeat the Enlightenment's reversal into its opposite?, which also reads AI as the next step in an Enlightenment arc that began with reason displacing myth. Rao agrees on the lineage — he explicitly places the Enlightenment's own achievement in making encyclopedic knowing available to "even the middle class," and LLMs as "a new arc" in that same relationship with "disembodied knowledge media" — but his essay reads the step as an expansion of access rather than a reversal into domination: the many-directions trick is what makes an LLM harder to out-flank, not what degrades the knowledge economy. It also supplies a mechanism for How do we learn to read AI-generated text critically?: a discourse source resists an interpretive posture when it has no single angle of approach to discount it from.

The excerpt is a personal essay built from one historical case (Darnton's account of Diderot's Encyclopédie) and does not test the "infinite ways of knowing" claim against any specific LLM behavior, benchmark, or user study — it is argued from the shape of the history, not measured. It also does not address whether that multi-directional fluency is reliably accurate, only that it is reliably available; Rao's own aside about hallucinating and bullshitting concedes the gap without resolving it. The implication he draws, that no fixed questioning strategy can expose an LLM's limits the way a single canonical ordering could be probed for gaps, is plausible as a description of how LLMs feel to use, but remains an unverified claim about their underlying knowledge rather than a demonstrated one.