LLM as Resource Router
Use the model as a souped-up Google that zeroes in on the right human-written resource — ask it "who should I read?" and go read/watch that artifact — rather than treating its own explanation as the terminal answer. Sanderson's frame: the most useful part of a Wikipedia page is often just the references at the bottom; you go to them and read them, and that gives a better overview than the page. So he asks an LLM who to read (optionally specifying how he wants to learn — e.g. "a well-visualized video"), takes the pointer, and leaves. The value the model adds is routing to a human voice, not being the voice. This is the practical consequence of Motivated Exposition: while model output "feels like Wikipedia," the win is to use it to find the single-author artifact that Wikipedia-flavored text can't be.
Crucially, the router can be confidently wrong about provenance: Sanderson recounts being "gaslit" when Claude recommended a real, working video link but misattributed it to 3Blue1Brown (it was someone else's). The pointer was useful; the attribution was fabricated — and it was caught only by clicking through to the artifact. So the discipline is: take the pointer, but verify the source by going to it, don't trust the model's claim about the source.
Claims
- Use the LLM as a super-Google to find the right human-written resource ("who should I read?"), not as the terminal source; then go learn from that artifact. best practice — context: learning a topic that has good human expositors and where a motivated single-author treatment exists to be found; the pattern degrades when no such artifact exists (the router has nothing to point to) or on post-cutoff material the model can't reliably index.
- The highest-value part of a reference is often its pointers outward (a Wikipedia page's references), so treat the model's answer as a routing layer to primary/human sources. (principle — as asserted by Sanderson) — durable: for learning, a good pointer to a motivated source beats a mediocre direct summary, because the summary is the thing you were trying to improve on.
- Specifying the mode you want to learn in (e.g. "a well-visualized video") sharpens the routing. best practice — context: when a modality fits the material (visual/spatial topics); it is the operator supplying taste the model can't infer, akin to scoring novelty against a personal skill matrix rather than a global one.
- The source recounts an LLM confidently recommending a real, working video link but misattributing it to a specific creator (3Blue1Brown) — a provenance hallucination caught only by going to the artifact. observation — a specific anecdote ("got gaslit"); check-worthy and attributed, illustrating that the pointer can be right while the claim about the pointer is fabricated.
- Clicking through to the human artifact was a better learning experience than continuing to ask the model questions. (observation — the source's report) — one operator's comparison, offered as the reason to route out rather than iterate in-model.
Related
- Motivated Exposition — why you route out: the human artifact is single-voiced and motivated in a way current model output is not.
- Curatorial-Voice Learning — what the router is looking for: the right person to read, by voice and resonance.
- Recency-Grounded Research — the same "the model points you to real, checkable human sources rather than laundering its own recall" discipline, there as a productized skill with cited links; here as an ad-hoc learning move.
- Evidence-Gated Completion — the verify-by-going-to-the-artifact reflex: the misattribution was only caught because the pointer was checked against the real thing.
- Search-Then-Get — a machine analogue: a cheap step that returns pointers (previews + citations) and a targeted step that fetches the real content; resource-routing is the human-facing version of the same "route first, read second" split.
- Distillate: The Trick to Using LLMs to Learn — Grant Sanderson (3Blue1Brown) × Dwarkesh Patel