firehose> #llmops

Motivated Exposition

An explanation deliberately crafted by one individual to build motivation — organizing concepts in the right order and making clear why each next idea matters to the next problem — teaches better than a correctness-maximizing consensus reference, even though the motivated version may be "a little bit wrong" on the way and correct itself as it goes. Sanderson's contrast is concrete: a single-author reference like the Stanford Encyclopedia of Philosophy or the "Princeton Compendium of math" (his phrase) is written by one person who "tries to actually craft a motivation around it," whereas Wikipedia is a "local minimum" reached by crowd-sourced editing "where basically every sentence has to be correct." Optimizing every sentence for local correctness edits out the productive wrongness — the deliberately imprecise stepping-stone you correct later — that a good exposition uses to carry a reader forward. Dwarkesh converges on the same point from the other side: his most productive learning sessions are with a human artifact that "organizes the relevant concepts in the correct way and builds up the motivation of why building up the next idea would be relevant to solving the next problem," idea after idea.

The load-bearing tension for the vault: correctness and motivation are not the same objective, and a process that maximizes the first can suppress the second. Sanderson asserts current LLM explanations "feel a lot like Wikipedia" — committee-flavored, correctness-first — which makes this concept the diagnostic behind LLM as Resource Router (why you route out of the model to a human voice) and a live design tension for any distiller that must choose between a defensible, every-sentence-correct summary and a motivated through-line.

Claims


Linked from