Spell Check is Global
AI is going to follow the path of all computer automation. Spell check is the completed example. Energy to run is the only true metric. What remains is joules.
Teaching figures for /papers/spellcheck/.
The curve below is a labeled illustration of cited endpoints. No coefficient is estimated.
Estimates are not package measured_j. board_synth_claimed=false.
The many (7B+) and the few (under 500M) are an access framing, not a census.
How to read this page
- Three prices. Bureau, then moat, then the separate price is gone.
- Two dots, not a fit. Nordhaus published the endpoints. The dashed line only connects them.
- The race is cited, not refit. Epoch's rates stay in Epoch's notes.
- Spell check finished the money curve. The joules did not go to zero.
- Seven hardware records. Each category has its own source and year. Do not add the rows.
- The installed generation is the reach. The latest die is a shipment. Software for the generation in hand has the largest impact.
sc-anim-01 Commoditization illustration
Labeled illustration Not a regression Cited endpoints only
The dashed connector draws from Nordhaus's manual-era price to his 2006 price. The motion joins two published numbers. No slope is fit. The callout is Epoch's summary of a recent inference-price window. Spell check is the bottom row: the separate money price of a correction is gone. The joule remainder stays lit. Commoditization leaves the physics in place.
Sources: Nordhaus, Journal of Economic History, 2007. Epoch AI, 12 March 2025. Emberson and Roodman, 22 September 2026. Hunspell README. Landauer 1961. Horowitz, ISSCC 2014. Study: /papers/spellcheck/.
sc-reg-01 Three price regimes
Teaching diagram
Read left to right. Moat rent is the signal that invites a rival to bundle the same job into hardware people already have. Spell check is that race after it ends. AI inference, on Epoch's notes, is still in the middle box.
sc-tok-01 One token, taught in order
Norvig's documented call
correction('speling') returns spelling in Norvig's essay. The steps are the companion laws on a finished path. Lookup, then a short edit. The suggestion is a proposal until commit. Further candidates stop when the token is kept. A frontier model that also says "spelling" has not produced a better word. It has produced the same word at a money price commoditization removes. His reported final test, 68 percent of 400 pairs at 35 words per second, is the toy's score, not an industrial claim.
| Step | What happens | Which study |
|---|---|---|
| 1 | The string is already on the machine. | Automation path |
| 2 | Is it in the local list? If not, try edit distance one. | Mixture of Limits |
| 3 | Show the suggestion. Write it only on acceptance. | Notational Intelligence |
| 4 | Once kept, more candidates do not change the predicate. | Satiation |
| 5 | The invoice can be zero. The probe still spends joules. | Metabolic Intelligence |
sc-tab-01 Cited anchors
Every number in this table is the source's number. The rightmost column states how this page uses it.
| Anchor | What it says | Use here |
|---|---|---|
| Nordhaus 2007 | Computation price from about $500 per MCPS to about 6×10⁻¹¹ by 2006, in 2006 prices. About seven trillion times. | Two endpoints. Connector is not a fit. |
| Epoch AI 2025 | Fixed-performance inference prices fell on the order of 9× to 900× per year across that note's benchmarks. | The race is fast and uneven. Not one slope. |
| Emberson and Roodman 2026 | About 47 percent per quarter, summarized as about 13× per year, since about 2023. | Their summary. Not extended past their window. |
| Damerau 1964; Kukich 1992 | Single-error correction, then the three-problem map. | The capability is a published procedure. |
| McIlroy 1982 | 30,000-word list squeezed to 26,000 16-bit words so the checker could run on minicomputers. | Paired to the machine people already had. |
| Norvig 2007 | 75% of 270; 68% of 400; 41 and 35 words per second. Written with no network. | Local gear. Not the industrial ceiling. |
| Hunspell README | Free library. Local dictionary. Named in ordinary editors. | Separate money price eliminated. |
| Hard et al. 2018 | 1.4 MB on-device next-word model. Raw typing stays on the phone. Their install claim is not our 7B+ framing. | Pairing, not a spelling score. |
| Landauer 1961; Horowitz 2014 | Erasure has a floor. Practical CMOS cost is mostly moving data. | Why joules remain. No new total computed. |
| Section 2.10 sources | Shipments, subscribers, revenue, and one survey. See sc-hw-01. | Seven records. Not a sum. Not 7B+. |
| Section 2.11 prices | Anita, HP-35, TI-2500, VisiCalc, Lotus, WordPerfect, Word. See sc-price-01. | Dated observations. Gaps stay blank. Disagreeing dates are both printed. |
Klere (klere.ai) is the product home of Metabolic Intelligence, not a prison. Spell check finished this path in editors that are not Klere. PDF: /pdfs/spellcheck.pdf.