Ingest: Learning By Writing

Type: kb/sources/types/ingest-report.md

Source: learning-by-writing.md Captured: 2026-08-10 From: https://www.cold-takes.com/learning-by-writing/

Classification

Genre: conceptual-essay -- a first-person account of an investigation method, supported by one extended example and retrospective experience rather than a comparative evaluation. Domains: writing-as-thinking, directed-reading, belief-revision, human-agent-transfer Author: Holden Karnofsky writes as a practitioner explaining the method behind his own published investigations. That gives the procedure useful first-person detail but no independent authority for its efficiency, retention, or transferability claims.

Summary

Karnofsky organizes learning around a repeatedly rewritten bottom-line view rather than around comprehensive reading. After a small initial survey, he states and strongly defends a deliberately premature hypothesis, lists its weakest points, reads or consults others on the subquestion most likely to change it, and revises or reverses the claim as soon as his view moves. He repeats this cycle until his own criticism is exhausted, then brings in outside feedback as another source of weaknesses. The central mechanism is not merely that prose exposes confusion: an explicit current claim determines which evidence is decision-relevant, prevents endless undirected reading, and preserves the reasoning behind successive views.

Connections Found

This essay is the constructive investigation procedure in the existing writing-as-thinking cluster. Borretti and Grunewald emphasize what naive delegation loses, while Karlsson makes conjectures and their premises vulnerable to counterexamples. Karnofsky adds the missing inquiry-routing role: the current written hypothesis supplies the lens for directed reading, because it determines which source or subquestion could change the bottom-line view. His repeated articulation, attack, revision, and feedback cycle also supplies a detailed human comparator for the adversarial-loop account. It should not be treated as evidence that an automated KB learns in the same way: human analogies motivate functions without fixing component boundaries, and durable agent learning additionally requires retained artifacts, activation, and behavioral effect.

Extractable Value

  1. A written working hypothesis can serve as inquiry-control state -- The explicit current claim does more than expose error: it defines what could change the author's mind, which evidence is relevant, and which reading can be skipped. This sharpens directed reading from a supplied task lens into an evolving lens revised by the investigation itself. [deep-dive]

  2. Writing and evidence acquisition form a coupled learning loop -- Articulation exposes a weakness; the weakness selects a subquestion; reading changes the claim; the changed claim reallocates subsequent reading. This explains why “writing rather than reading” is not a choice between media but an ordering and feedback relation between two operations. [deep-dive]

  3. Subquestion choice is a value-of-information judgment -- Karnofsky prioritizes what is most likely to change the big-picture claim rather than what is easiest to research or locally interesting. An agent-operated analogue would need an explicit signal for expected decision impact; retrieval relevance alone does not supply it. [experiment]

  4. Deliberate stance reversal is a search operator, not a conclusion -- Arguing the opposing side before being convinced can escape a locally stable defense and expose different evidence needs. This is operationally distinct from neutral summarization, but the source offers no comparison showing when it improves rather than distorts inquiry. [experiment]

  5. The claimed retention target is a view plus its reasoning -- Karnofsky contrasts retaining scattered information with retaining an opinion and why he holds it. That is a useful target distinction for the proposed writing workshop, though the essay provides only self-report that the process improves retention. [just-a-reference]

  6. Outside feedback extends the same weakness-revision loop -- Publication is framed as escalation after self-critique stops finding problems, not as certification that the claim is correct. This supplies a simple functional placement for external review within the larger process. [quick-win]

Limitations (our opinion)

The essay is a selected retrospective from one unusually experienced investigator. It has no comparison with reading-first, outlining, discussion-first, or other workflows, and no independent measure of speed, accuracy, or retention. Because Karnofsky explicitly allows conversation in place of writing, the causal mechanism may be externalizing a definite hypothesis and subjecting it to criticism rather than writing itself.

The early-hypothesis strategy can also anchor the investigation, reward confirmation, or make evidence outside the initial frame invisible. “Defend aggressively,” notice the weakest point, choose the highest-value subquestion, and switch sides at the right moment all depend on tacit judgment that the procedure does not explain. Its example shows substantial belief revision but cannot reveal failed investigations or unnecessary deep dives; the author acknowledges both poor predictability and wasted work.

Most importantly for Commonplace, changed human understanding, changed artifact content, and changed agent behavior are different outcomes. The essay gives first-person evidence only for the human case. An automated analogue would need to specify which signal triggers revision, which operations can be composed, where the updated state is retained, how it is activated later, and whether use changes behavior. The workshop should compare functions across those boundaries without treating the human sequence as an agent architecture.

The written-artifacts-in-learning-loops workshop that examined this source has run and closed. Its distinctive contribution from Karnofsky — organizing source-reading around a provisional written claim that routes what to read next — was promoted as the conjecture claim-routed reading may beat reading-first for synthesis notes. Its open next action is that note's discriminating test: build a synthesis note both claim-routed and read-first, and compare warranted update against context cost.