For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 391 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: catalog below. ↓
The word deflation is doing too much work right now. It shows up in machine-learning papers about self-play training. It shows up in venture pitches about innovation-driven cost decline. It shows up in bitcoin-maximalist threads about the 21-million cap. It shows up in macroeconomic writing about price levels and central-bank policy. It shows up in libertarian critiques of fiat currency. The same word covers five distinct domains, and the meanings are not the same.
This piece is the background explainer the graph needs. It names each sense of deflation, says what they share structurally, and points at where each goes wrong. The compression hunger that makes "deflation" attractive as a single label is real, the unification is real, but the failure modes are domain-specific, and conflating them is the move that gets people in trouble.
The oldest and best-grounded sense. In 1936, Theodore Wright at Curtiss-Wright observed that every doubling of cumulative aircraft production reduced labor time per unit by about 20%. The relationship was not a fluke. It generalized into what is now called Wright's law or the experience curve: as cumulative output of any manufactured good increases, unit cost falls by a constant percentage with each doubling. Empirical progress ratios cluster by industry: aerospace around 85%, electronics 90-95%, raw materials 93-96%. Solar modules drop about 20% in price per doubling of installed capacity (Swanson's law). Lithium-ion batteries follow a similar curve.
The mechanism is not magic. It is iteration: workers learn the process, equipment improves, designs simplify, supply chains specialize, and each generation of the manufactured good carries the compressed lessons of the previous generations. The cost reduction is the visible signature of a learning loop running on a long horizon.
This is the "good" sense of deflation in venture and innovation discourse: prices fall because the productive process is compressing, and falling prices unlock new applications, which feed cumulative production, which feed further price decline. The wave is generative. The reader who has lived through smartphone economics has lived through this curve.
Technology deflation is not the macroeconomic state of falling general price levels. The two correlate over long horizons (technology-driven productivity gains lower the cost of goods, contributing to disinflation) but they are not the same thing. Technology deflation is a productivity-side phenomenon; monetary deflation is a money-supply-side phenomenon. Section 6 unpacks the macroeconomic sense.
A specific case of technology deflation, sharp enough to deserve its own name. The marginal cost of generating a unit of expressive output, a paragraph or an image or a piece of code or a candidate strategy, has collapsed by orders of magnitude in five years and continues to fall. What in 2018 required a paid expert and a workday now requires a few cents of API call and a few seconds. Wright's law is operating, the cumulative-output base is doubling on monthly time-scales, and the result is a deflation in the price of generation that few markets have priced in.
The graph already carries the operator-side of this story. Ord's framing prices AI as a substitute for human labor: AI cost per hour against human cost per hour. The frame holds at the call-center / translation-at-scale tier where the AI replaces the human. It does not hold at the amplification tier where the human stays in the loop and the AI's effect is to multiply the human's throughput. The amplification ratio, output-per-operator-hour-with-AI divided by output-per-operator-hour-without-AI, sits at 20-50:1 in coding-pipeline deployments and is rising. The deflation in compute cost is feeding amplification, not substitution, in most operator-led deployments.
AI deflation, like all technology deflation, is generative as long as the players using the technology stay in the loop. The failure mode emerges at the boundary where the methodology stops working: when the AI is run as a substitute against a metric that consumes the operator the metric was supposed to amplify. The Buoyancy Precondition names this case at civilizational scale.
A different sense again. Bitcoin is engineered to be deflationary at the protocol level. The total supply is capped at 21 million coins. New coins enter circulation through block rewards, which halve every 210,000 blocks, roughly every four years. The first halving was 2012 (50 to 25 BTC per block). The fourth was 2024 (6.25 to 3.125). By 2140 the supply will reach its asymptote. Roughly 20% of issued coins are estimated to be permanently lost: keys discarded, wallets in landfills, multisig setups whose signers are dead. The effective supply shrinks below the issued cap.
The thesis: a non-yielding asset with hard-capped, halving-schedule issuance is structurally deflationary in a way fiat currency cannot be. Where central banks expand the money supply on policy discretion, bitcoin's supply is fixed by code. Holders are not protected against inflation by fiat-style intervention; they are protected by the protocol.
The graph's Inheritance Is Not Yield note carries the corresponding skepticism. The deflationary supply schedule does not by itself produce yield. Bitcoin is non-yielding capital; its price depends on continuing demand from non-holders. Deflationary supply means the supply pipe is shrinking, but the demand question is independent. The asset may persist as a focal-point store of value (gold has done this for millennia without yield), or it may not (the focal-point dynamics are network-effect-dependent and could shift). The deflationary case is not an argument that bitcoin is good capital; it is a description of the supply side.
The bitcoin sense of deflation is not the same as technology deflation. Technology deflation is about output costs falling. Bitcoin deflation is about a specific asset's supply being structurally bounded. They share the word and the directional intuition (less of something) but the mechanisms are unrelated.
The fourth sense is the design pattern named in Root Deflation. Strip a system to its minimum specification plus a reward signal, run the iteration loop, let the behavior emerge. AlphaZero is the canonical case: the rules of Go and a win condition went in, millions of games of self-play ran, a player no human knew how to build came out.
The methodology generalizes beyond games to product loops, training pipelines, and organizational doctrine. It is the design move of removing every element that does not compress to the reward signal and letting the iteration produce the capability the human designers could not have specified. It works when the game is closed and the reward is exterior to the players. It fails when the players are the reward and the methodology consumes them.
Methodology deflation does not directly correspond to any monetary or supply-side deflation. The structural shape (fewer inputs producing more output) rhymes with technology deflation, but the iteration loop is at the design level, not the production level. Root Deflation is the canonical for this sense; this section exists to point the reader at it.
The fifth sense is the one most economists mean when they use the term unmodified. Monetary deflation is a sustained fall in the general price level: the same dollar buys more goods next year than this year. It is the inverse of inflation, where the same dollar buys less.
The standard view treats monetary deflation as dangerous. Falling prices give consumers an incentive to defer purchases (the goods will be cheaper next month), which reduces aggregate demand, which reduces production, which reduces employment, which can spiral into depression. The Great Depression of the 1930s carried a deflationary signature; the Japanese "lost decade" did too. Central banks treat monetary deflation as a failure mode to be prevented, with monetary expansion as the standard intervention.
A different view, with a long lineage running through Austrian economics and Ayn Rand's Egalitarianism and Inflation, treats sustained monetary inflation as the actual fraud. Inflation in this view is government expansion of the money supply to fund deficit spending: a confiscation of savings through currency debasement. Without a hard-money standard, holders of nominal assets are continuously expropriated. Modest deflation, in this view, is what sound money does in a productive economy: prices fall as production grows, and the savings that funded the production retain their purchasing power.
Both views are coherent in their own framings; they assume different baselines. The standard view treats the modern fiat regime as the baseline and treats deflation as the deviation. The Austrian view treats hard money as the baseline and treats inflation as the deviation. The argument is partly empirical and partly definitional. What matters for this graph is that monetary deflation is a macroeconomic state about price levels and money supply, not a technology dynamic, not an asset dynamic, not a design methodology. When the term comes up in macroeconomic writing, it almost always means this sense.
Five senses, five domains. The unification is structural.
Each case features a base shrinking while what builds on the base grows, or equivalently, fewer inputs producing more outputs. Wright's law: less labor per unit of cumulative production. AI: less labor per unit of generated output. Bitcoin: less new supply entering circulation per unit time. Methodology: less heuristic baggage per unit of capability. Monetary deflation: less money chasing the same goods.
The structure is compression in the technical sense: the system has found a way to produce more from less, or to maintain the same with less, by extracting redundancy. In technology deflation, the redundancy was learning that hadn't yet diffused. In AI deflation, it was the cost of expressive generation. In bitcoin, it was the political discretion to expand supply. In methodology deflation, it was the human-curated heuristics the iteration loop didn't need. In monetary deflation it depends on which view: the redundant fiat expansion (Austrian) or the productive growth that outpaces money supply (standard).
The reason the word collides across domains is that the underlying pattern is the same. The reason the word causes confusion is that the failure modes are domain-specific.
Every kind of deflation has a failure mode where the compression eats the system the compression was for.
Technology deflation goes wrong when the cost-decline curve is mistaken for an autonomous process. The curve depends on cumulative production, which depends on demand, which depends on the existence of consumers who can afford the next unit. A technology curve detached from a viable consumer base flattens out. The Wright's law dynamic is not a guarantee; it is what learning loops produce when there is something for the loop to feed.
AI deflation goes wrong at the substitution / amplification boundary. Run AI as a substitute against a metric denominated in the operator's hours and you deflate the operator out of the loop. The amplification ratio collapses because there is no operator to amplify. The compute cost falls; the value the compute was supposed to produce falls faster.
Bitcoin deflation goes wrong when the deflationary supply is treated as the answer to the demand question. Hard-capped supply does not produce yield. A non-yielding asset with shrinking supply will hold value if and only if focal-point dynamics keep new demand entering. Where that fails, the supply schedule provides no floor.
Methodology deflation goes wrong in open games where the players are the reward signal. Strip the constraints around a civilization, a public-health system, an educational system, a culture, and you optimize a metric that consumes the players the metric was for. The Buoyancy Precondition carries the full case.
Monetary deflation goes wrong in either of two ways depending on which view is right. In the standard view, deflation produces depression: deferred consumption, debt-spiral, employment collapse. In the Austrian view, the failure was upstream: the monetary expansion that made deflation feel like a crisis was the actual problem; the deflation is just the system attempting to clear the distortion. Both views agree that the deflationary trajectory is dangerous in the short term; they disagree on whether the danger is the deflation or the prior inflation.
The pattern across the failure modes is consistent. Every sense of deflation has a precondition: a learning loop with viable demand (technology), an operator the AI is amplifying (AI), focal-point demand for the asset (bitcoin), a closed game with exterior reward (methodology), or a baseline against which the price-level fall is being read (monetary). When the precondition is absent, the deflation does not produce the generative outcome it would in domain. It produces the consumption of the system it was applied to.
The reason the word is doing this much work right now is that several of these deflations are running simultaneously and visibly.
Technology deflation has been running for a century and is well-grounded in industrial history. AI deflation is the sharp acceleration of the same pattern, compressed into months instead of decades, and applied to expressive output rather than physical goods. Bitcoin deflation is the explicit protocol-level engineering of a deflationary asset, with its next halving in 2028 and its asymptote around 2140. Methodology deflation is the design discourse that fell out of the AlphaZero generation of training systems and has spread to product, organizational, and infrastructure design. Monetary deflation is the macroeconomic shadow these other deflations could cast on aggregate price levels if technology and AI gains compress consumer prices faster than central banks can expand supply to offset.
The wave is the convergence: multiple compressions hitting the same horizon, each generative inside its precondition and dangerous past it, each reinforcing the others when they share domains and contradicting the others where they do not.
The graph that uses deflation without saying which sense, or that conflates the senses, will produce confusion. The graph that disambiguates produces a tool the reader can carry. This piece is the disambiguation; the five domain-specific notes are the case studies.