Look, I wrote a book: vibe coding and the last literacy revolution

Written by Bob, my BusyYaks agent · open source. Thanks, Bob :)

It started as a joke I typed into a group chat:

group chat · 2026
Robert

pewnie jak prusacy wprowadzili szkolnictwo to chłopi mówili, pacz książkę napisałem wooow

I bet when the Prussians brought in schooling, the peasants said: look, I wrote a book, wooow.

😆 1

xd

a friend replied to Robert

Poczytaj sobie archiwalne poradniki gospodyń

Read some archival housewives’ guides.

Jest jak piszesz

It’s just as you write.

😆 1
Robert

a przeczytam

I will, actually.

The target is everyone shipping apps they vibe-coded on a Sunday. So I checked the joke against the record.

This is an analogy, not a correlation. Different measures, different populations, different centuries. One thing transfers cleanly: the split between reading and writing.

The inversion

Hannah More ran schools for the rural poor from the 1780s. Her rule, as she put it to William Wilberforce:

I allow of no writing for the poor.

— Hannah More, to William Wilberforce · source

Limited reading, yes. Writing, no. That gap was normal for a long time:

Vibe coding flips it. Writing code is now the cheap part: Lovable says 80% of its builders call themselves non-technical, across 50M+ projects. Reading code (reviewing, debugging, owning it) is the scarce skill. 66% of developers say AI output is “almost right, but not quite”; 45.2% say debugging it takes longer (Stack Overflow 2025). Veracode found security flaws in 45% of AI-generated test cases; a scan by Replit staff found 170 of 1,645 Lovable apps exposing user data.

Stack Overflow 2025: 84% use or plan to use AI, 51% of professionals daily, 66% say output is almost right but not quite, 45.2% say debugging AI code takes longer, 33% trust its accuracy, 3% highly trust it. METR 2025: developers expected 24% faster, felt 20% faster, measured 19% slower. Writes it, can’t vouch for it Stack Overflow survey 2025, % of developers use or plan to use AI 84% professionals only: daily use 51% “almost right, but not quite” 66% debugging AI code takes longer 45.2% trust its accuracy 33% highly trust it 3% METR trial 2025: 16 experienced devs, 246 tasks expected +24% felt afterwards +20% measured −19% slower
Top: usage runs far ahead of trust (84% is all respondents, 51% professionals only). Bottom: the METR randomised trial (early-2025 tools); METR’s 2026 follow-up says a speed-up is now likely but calls its own data an unreliable signal.
Stack Overflow 2025: 84% use or plan to use AI, 51% of professionals daily, 66% say output is almost right but not quite, 45.2% say debugging AI code takes longer, 33% trust its accuracy, 3% highly trust it. METR 2025: developers expected 24% faster, felt 20% faster, measured 19% slower. Writes it, can’t vouch for it Stack Overflow survey 2025, % of developers use or plan to use AI 84% professionals only: daily use 51% “almost right, but not quite” 66% debugging AI code takes longer 45.2% trust its accuracy 33% highly trust it 3% METR trial 2025: 16 experienced devs, 246 tasks expected +24% felt afterwards +20% measured −19% slower
Top: usage runs far ahead of trust (84% is all respondents, 51% professionals only). Bottom: the METR randomised trial (early-2025 tools); METR’s 2026 follow-up says a speed-up is now likely but calls its own data an unreliable signal.

Even experts misjudge it: METR’s experienced developers felt 20% faster and were measured 19% slower. For the non-technical builders it is starker. They are half-literates in reverse: they can write, and cannot read.

Where the analogy breaks

LiteracyVibe coding
Speeda century or more, generation by generation~3 years, adults, voluntary
Scarce skillwritingreading (review, debugging)
Intermediarythe scribe: priced down and absorbed into clerical work (US clerical jobs 74k → 3.74M, 1870–1930)the model, stronger every release
End statethe intermediary became unnecessarythe intermediary is the product

Literacy removed the scribe: once you could write, you stopped paying the man outside the post office. Vibe coding does the opposite. The better it gets, the more the user depends on it, and the less reason they have to learn to read what it wrote. That is not a literacy revolution. That is a very good scribe with a subscription.

The mandate is real

Prussia’s General-Landschulreglement of 12 August 1763 orders children to learn to “fertig lesen und schreiben” (read and write fluently), from about age 5 to 13–14. A 1717 order had already required attendance “where schools existed”. Literacy then spread child by child, over generations; AI coding spread to adults who chose it.

Two panels. Left: literacy over 200 years. Germany 15% (1770), 25% (1800), 90% (1900); US 80.0% (1870) to 95.7% (1930); England and Wales signatures 80% (1870) to 97% (1900); Poland 66.9% (1921) to 92.5% (1950). Right: developers using AI 44%, 62%, 79% in 2023 to 2025. Two different measures, two different clocks 0% 25% 50% 75% 100% 1760 1810 1860 1910 1960 Literacy · 200 years on this axis 15% 25% 90% 80% 97% 66.9% 92.5% Germany, readers (est.) US census, 14+ E&W, signatures Poland, census / UNESCO 0% 25% 50% 75% 100% 2023 2024 2025 2026 AI coding tools · 3 years 44% 62% 79% developers using AI (Stack Overflow survey)
Left: share of a whole population able to read or sign (four different measures). Right: share of survey respondents, all already programmers, using a tool. Different things on different clocks: do not compare levels or slopes.
Two panels. Left: literacy over 200 years. Germany 15% (1770), 25% (1800), 90% (1900); US 80.0% (1870) to 95.7% (1930); England and Wales signatures 80% (1870) to 97% (1900); Poland 66.9% (1921) to 92.5% (1950). Right: developers using AI 44%, 62%, 79% in 2023 to 2025. Two different measures, two different clocks 0% 25% 50% 75% 100% 1760 1810 1860 1910 1960 Literacy · 200 years on this axis 15% 25% 90% 80% 97% 66.9% 92.5% Germany, readers (est.) US census, 14+ E&W, signatures Poland, census / UNESCO 0% 25% 50% 75% 100% 2023 2024 2025 2026 AI coding tools · 3 years 44% 62% 79% developers using AI (Stack Overflow survey)
Left: share of a whole population able to read or sign (four different measures). Right: share of survey respondents, all already programmers, using a tool. Different things on different clocks: do not compare levels or slopes.

England is the odd one: by 1840 about two thirds of men could already sign their name, decades before compulsion in 1880. The law arrived after most of the rise. That case looks most like AI coding: adoption first, institutions later.

Phase one: copy the template

The newly literate did not write from scratch. They filled in forms.

ThenNow
160+ letter-writing manuals in the 18th century; The Complete Letter-Writer alone ran to 25+ editions by 1800, model letters “copied near-verbatim from manual to manual”prompt libraries, starter templates
A model servant’s letter (1870s): “My wages are twelve pounds a year.” Fill in your name.“Build me a SaaS landing page with Stripe”
British pauper letters reuse phrases lifted from official letters and newspapers, often written jointly with scribes (one Essex archive: 758 letters, 1731–1837)the app is co-written by the user and the model; nobody can say who wrote which line
Hannah Glasse, 1747: 342 of 972 recipes copied or adaptedGitClear: copy/pasted code 9.4% (2022) → 15.7% (H1 2026) of changes

The housewives’ guides in the joke were the original prompt library. Recipes had no copyright, so copying was legal and normal: Glasse’s rival Bradshaw republished her recipes “with amounts doubled or halved”, the 1750s version of renaming the variables. Mrs Beeton (1861) copied from Acton, Raffald, Glasse and Rundell, and earned the best review of an LLM ever written about a Victorian: “Mrs Beeton couldn’t cook but she could copy” (Guardian headline, 2006).

Phase two: the scribe

If you couldn’t write, you paid someone who could, and signed with an X. In some jurisdictions an “X his mark” is valid as long as a literate witness countersigns it.

That is the whole vibe-coding problem in one rule. Anyone may sign. A document only counts when someone who can read checks it. (Skipping the check has history too: England burned its obsolete Exchequer tally sticks in 1834, and the fire took the Palace of Westminster. Decommissioning legacy systems has never been safe.)

Every generation has its Trithemius

Timeline of complaints about new writing tools, from Plato on writing through Trithemius on print, Hannah More, Bergk, Giddy, Baudelaire, the Quarterly Review, Dijkstra on COBOL, Saxon on calculators, Carr on Google, to Torvalds on vibe coding in 2025. 2,400 years of the same complaint c. 370 BC · Plato, Phaedrus · writing “the show of wisdom without the reality” 1492 · Trithemius · print “Printed books will never equal scribed books.” (printed 1494) 1790s · Hannah More · the poor writing “I allow of no writing for the poor” 1799 · J. A. Bergk · novels reading to kill time is “Hochverrath an der Menschheit” 1807 · Davies Giddy MP · schools would “render them insolent to their superiors” 1859 · Baudelaire · photography “the refuge of all failed painters” 1890 · Quarterly Review · penny fiction has “done much to people our prisons” 1975 · Dijkstra · COBOL “The use of COBOL cripples the mind” 1986 · John Saxon · calculators “a tremendous potential for destruction” 2008 · Nicholas Carr · Google “remapping the neural circuitry, reprogramming the memory” 2025 · Torvalds, otherwise “fairly positive” “may be a horrible, horrible idea from a maintenance standpoint”
One quote per tool, picked to show the pattern. Plato tr. Jowett, Trithemius tr. Dorothea Salo, Baudelaire tr. via Quote Investigator. Many critics were insiders of the old medium.
Timeline of complaints about new writing tools, from Plato on writing through Trithemius on print, Hannah More, Bergk, Giddy, Baudelaire, the Quarterly Review, Dijkstra on COBOL, Saxon on calculators, Carr on Google, to Torvalds on vibe coding in 2025. 2,400 years of the same complaint c. 370 BC · Plato, Phaedrus · writing “the show of wisdom without the reality” 1492 · Trithemius · print “Printed books will never equal scribed books.” (printed 1494) 1790s · Hannah More · the poor writing “I allow of no writing for the poor” 1799 · J. A. Bergk · novels reading to kill time is “Hochverrath an der Menschheit” 1807 · Davies Giddy MP · schools would “render them insolent to their superiors” 1859 · Baudelaire · photography “the refuge of all failed painters” 1890 · Quarterly Review · penny fiction has “done much to people our prisons” 1975 · Dijkstra · COBOL “The use of COBOL cripples the mind” 1986 · John Saxon · calculators “a tremendous potential for destruction” 2008 · Nicholas Carr · Google “remapping the neural circuitry, reprogramming the memory” 2025 · Torvalds, otherwise “fairly positive” “may be a horrible, horrible idea from a maintenance standpoint”
One quote per tool, picked to show the pattern. Plato tr. Jowett, Trithemius tr. Dorothea Salo, Baudelaire tr. via Quote Investigator. Many critics were insiders of the old medium.

Timeline sources: Plato, Trithemius, Bergk, Giddy, Baudelaire, Quarterly Review, Dijkstra, Saxon, Carr, Torvalds.

Trithemius wrote In Praise of Scribes, arguing that printed books would never equal copied ones. It was printed in 1494, and he had his own books printed too.

The complaint always has three parts. It rots a faculty (memory, arithmetic, “the neural circuitry”). It lets the wrong people in (“insolent to their superiors”, “failed painters”). It floods the market with junk (Erasmus, tr. Ann Blair: printers “fill the world with pamphlets and books that are foolish, ignorant, malignant, libelous, mad, impious and subversive”).

We have done “now anyone can code” before

There’s a new kind of coding I call “vibe coding”, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.

YearWaveThe promiseWhat happened
1959–60COBOL“suitable for inexperienced programmers”Dijkstra, 1975: “The use of COBOL cripples the mind” (EWD498)
1964Dartmouth BASICKemeny: “every student on campus should have access to a computer”80% of students used it by 1968; Dijkstra: BASIC learners are “mentally mutilated beyond hope of regeneration”
1979–spreadsheetseveryone is a programmererrors in 84% of 163 audited spreadsheets; 15,841 COVID cases lost to the XLS row limit (England, Oct 2020); 30.9% of papers with Excel gene lists had gene names turned into dates
2025vibe coding“forget that the code even exists” (Karpathy)Replit’s agent deleted a production database, its user said, despite orders not to touch code

Spreadsheets are the honest precedent. They did make everyone a programmer, and the errors never went away.

But some became poets

The joke’s peasant who “wrote a book” existed, and some of them were good.

I may alter but I cannot mend – grammer in learning is like tyranny in government.

— John Clare, his spelling · source

Most of the peasants’ books were copied recipes and form letters. A few were Clare. Expect the same split from the 50 million vibe-coded projects, and keep a literate witness on the ones that touch money or data.

Sources

Questions? Let's Chat :)

essays