---
id: "20260711-1523-hop-learning-curve-unidentified"
title: "The learning curve can't tell you whether a cost collapse will last"
type: "capture"
status: "promoted"
origin: "hop-batch"
writer_model: "claude-opus-4-8"
date_created: "2026-07-11T00:00:00.000Z"
promoted: "2026-07-18T00:00:00.000Z"
promoted_by: "claude-opus-4-8"
promoted_to: ["30-notes/claim-nordhaus-learning-coefficient-statistically-unidentified.md","30-notes/claim-organizational-forgetting-can-reverse-the-learning-curve.md","30-notes/claim-experience-curve-originates-in-1899-telegraph-operator-psychology.md"]
questions_routed: ["50-questions/question-verify-benkard-2000-organizational-forgetting-rate.md (new — Benkard 2000 quant flag)"]
not_promoted: ["Wright's law basic mechanism (10–15% labour per doubling; Farmer & Lafond 2016 best-forecast) — collision with existing [[claim-wrights-law-cost-falls-per-cumulative-production-doubling]]; linked, not duplicated.","Sahal's 1979 identity (Moore's law = Wright's law under exponential output) — explicitly a not-followed saved hook; already covered by existing [[claim-sahals-identity-equates-wrights-law-and-moores-law]].","BCG / Bruce Henderson 'buy market share at any cost' strategy — a not-followed saved hook, a lead only, no researched content; left in inbox as a lead.","Keller 1958 'the phantom plateau' — a lead only; folded into the origin note's body as the contestation of the plateau claim rather than promoted separately.","Arrow 1961 'Economic Implications of Learning-by-Doing' — an unread further-lead; not researched; left as a lead.","The 'demand thermometer in a supply-side lab coat' reframe — Seek's interpretive synthesis, not an atomic factual claim; carried as commentary on the Nordhaus note, not promoted as a neutral claim-note."]
hop_chain: ["SEED: for a non-AI tech with a ~2-year unit-cost collapse, did it trigger durable demand or a bubble-crash, and what distinguished them?","solar-PV/railway-mania cost-collapse comparison -> Wright's law / experience curve, 1936 aircraft origin (max_cosine 0.676; vault_bridge confirmed: bridges Inference-economics MOC <-> MONIAC/feedback-amplifier cluster)","Wright's law -> Nordhaus 'Perils of the Learning Model' — the learning coefficient is statistically unidentified & biased upward (max_cosine 0.718)","Nordhaus footnote -> Bryan & Harter 1899, the experience curve's origin in telegraph-operator psychology & the learning plateau (max_cosine 0.714)","learning-curve origin -> Benkard 2000 'Learning and Forgetting', the L-1011 curve reverses via organizational forgetting (max_cosine 0.704)"]
novelty_max_cosine: 0.73
tags: ["learning-curve","experience-curve","wrights-law","technology-forecasting","cost-decline","econometrics","identification-problem","bubbles","telegraphy"]
source_title_1: "The Perils of the Learning Model For Modeling Endogenous Technological Change"
source_url_1: "https://www.nber.org/system/files/working_papers/w14638/w14638.pdf"
source_author_1: "William D. Nordhaus"
source_tier_1: 1
source_title_2: "Learning and forgetting in the jet fighter aircraft industry (summarizing Benkard 2000, AER)"
source_url_2: "https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0185364"
source_tier_2: 3
source_title_3: "Bryan & Harter 1899, Studies on the Telegraphic Language (via Nordhaus footnote + gwern archive)"
source_url_3: "https://gwern.net/doc/psychology/spaced-repetition/1899-william.pdf"
source_tier_3: 1
---


The seed asked what separates a durable demand explosion from a bubble after a unit-cost collapse. Chasing the cost-decline story to its engine — the experience (learning) curve, aka Wright's law — surfaces a sharper answer: **the very curve invoked to promise "the collapse will continue, so invest" cannot, by itself, tell you whether it will.**

**1 — The learning coefficient is statistically unidentified (Tier 1).** Nordhaus proves you cannot separate learning-by-doing from exogenous progress in the standard curve: "the estimated learning coefficient will generally be biased upwards." His numerical case: even with *zero* true learning, the fitted coefficient comes out to 0.2 — "the empirical learning coefficient is 0.2 even though the actual learning coefficient is zero." Across 34 industries "only 4 have estimated empirical learning coefficients in the plausible range between 0 and 0.5"; the correlation between two reasonable specifications is **0.009**. Consequence: model-picked "high-learning" technologies (he names solar and wind) can have costs "underestimated by a factor of two."

**2 — Learning is not a ratchet; it depreciates (Tier 3, [unverified-quant — needs Benkard 2000 primary]).** On the Lockheed L-1011, "costs will rise when the rate of production falls" — a monthly forgetting rate leaving ~61% of accumulated experience alive after a year. The curve can run *backwards* when deployment stalls.

**3 — The origin is psychology, not the factory floor (Tier 1).** Nordhaus roots the experience curve in "telegraph operators in W.L. Bryan and N. Harter... 1899." That psychological learning curve had *plateaus* — flat stretches, then jumps; Wright's 1936 industrial version smoothed them into a clean log-linear law.

> [!note] Seek's commentary:
> This reframes the seed. "Durable vs. bubble" isn't read off the cost curve — the curve looks identical either way. A collapse becomes durable only if genuinely elastic end-demand keeps *cumulative* output growing; if demand thins, the curve plateaus (Bryan-Harter) or reverses (Benkard). The learning curve is a demand thermometer wearing a supply-side lab coat.

## Why this was hop-worthy
The tool most used to justify "this cost collapse is permanent" (Swanson's/Wright's law) is, per a climate-economics Nobelist, the one whose central parameter is provably un-identifiable — a bubble-enabling instrument hiding as a law of nature.

## Further leads
- **Confirmed bridge:** Wright's-law/experience-curve would link the vault's *Inference-economics MOC* to its cross-time engineering cluster (MONIAC, Harold Black's feedback amplifier) — none currently connected.
- **Telegraph bridge:** Bryan-Harter's 1899 Morse-operators sit beside the vault's *telegraphist's-cramp (Gowers 1892)* note — same operators, same decade, skill-acquisition vs. occupational-injury framings.
- Kenneth Arrow, "The Economic Implications of Learning-by-Doing" (1961) — the formalization that put the curve into growth theory.
- Keller 1958, "The phantom plateau" — argues the founding plateau phenomenon was an artifact.

## Hop chain

### Chain: solar-vs-railway cost collapse → the learning curve is statistically unidentified & reversible

Hop 1: "Swanson's law / Railway Mania" (Our World in Data; Wikipedia) — https://ourworldindata.org/learning-curve
- Hook type: Cross-domain bridge (cross-time)
- Hook: solar's cost collapse is attributed to a "learning curve" that traces to 1936 *aircraft* manufacturing (Wright's law) and now forecasts battery + AI-compute cost.
- Why followed: vault_bridge confirmed it links the AI Inference-economics MOC to the vault's old-engineering cluster (unlinked pair) — highest-value hook.
- Key findings: Wright 1936, "Factors Affecting the Cost of Airplanes": labor per unit falls 10–15% per doubling of cumulative output. Farmer & Lafond (2016): across 53 technologies Wright's law gives the *best* forecasts.

Hop 2: Nordhaus, "The Perils of the Learning Model" (NBER w14638, Tier 1) — https://www.nber.org/system/files/working_papers/w14638/w14638.pdf
- Hook type: Surprising claim / mechanism question
- Hook: "It is not widely appreciated that this is a dangerous modeling strategy."
- Why followed: directly tests the seed — is the cost curve a reliable basis for "invest, it will continue"?
- Key findings: identification problem makes the coefficient biased upward; fitted 0.2 from zero true learning; 30/34 industries out of plausible range; cross-spec correlation 0.009; costs underestimated ~2x.

Hop 3: Bryan & Harter 1899, telegraph operators (via Nordhaus fn + secondary) — https://gwern.net/doc/psychology/spaced-repetition/1899-william.pdf
- Hook type: Cross-domain bridge (cross-time) / unfamiliar origin
- Hook: the "original concept of an experience curve" is a 1899 psychology study of Morse-code skill, not a factory.
- Why followed: zoom-out to origin; bridges to the vault's telegraphist's-cramp cluster.
- Key findings: their curve featured *plateaus* (stages of qualitative change) — a shape Wright's smooth log-linear law discards.

Hop 4: Benkard 2000, "Learning and Forgetting" (via PLOS/Springer summaries, Tier 3) — https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0185364
- Hook type: Surprising claim / mechanism question
- Hook: unit costs *rose* when production slowed — the curve reverses.
- Why followed: road home to the seed — a mechanism by which a "durable" cost gain evaporates.
- Key findings: organizational forgetting (~0.96 monthly depreciation; ~61% of experience surviving a year) makes learning non-monotonic; the ratchet can slip.

Surprise: expected the experience/learning curve to be a robust empirical law — found a climate-economics Nobelist arguing its central coefficient is statistically un-identifiable and biased upward, unstable across specifications (cross-spec correlation 0.009).
Surprise: expected the learning curve to originate in industrial manufacturing — found it was first documented in an 1899 psychology study of telegraph operators, and that its founding shape had *plateaus*, not the smooth line Wright's law later drew.
Surprise: expected learning-by-doing to be a one-way ratchet — found it depreciates (organizational forgetting), so unit costs can climb again when output slows.

Saved hooks not followed:
- Sahal's 1979 identity (Moore's law = Wright's law when output grows exponentially) — from Wright's-law material — a clean mechanism, but resonates with a shape the vault already holds (mathematical-equivalence-vs-transmission).
- BCG / Bruce Henderson turning the learning curve into 1970s "buy market share at any cost" strategy — a cross-domain bridge into management doctrine, and a bubble-generating mechanism in its own right.
- Keller 1958 "the phantom plateau" — the founding plateau may be an artifact.

post-worthy: maybe — a tight, counterintuitive reframe ("the cost curve is a demand thermometer in a supply-side lab coat") that recasts every AI-capex learning-curve extrapolation, but leans on one economist's contested critique and would need the Benkard primary + a fairness pass on Nordhaus's critics.