Feijoo's Challenge

SoundtrackBanda sonora: Djrum – Meaning’s Edge
0:00 / 32:01

LLM-translated from the Spanish original. Read the original

TL;DR

Father Feijoo challenged his readers to find lines translated from French journals in his works. Almost three centuries later, I found about ten cases that meet his conditions.

Intro

Ever since I read AI labs need to start funding historical research a few days ago, I’ve been obsessed with the implications. In short, current frontier models have reached the point where they can produce original historical knowledge on their own, and the author’s conclusion is that labs should fund collaborations with historians and archivists to push the frontier.

If you look past the incentives (the person writing it is a historian by profession), what I take from it is that we’re at a unique moment where anyone sharp enough can make novel discoveries, before the labs start tying up all the remaining loose ends for a bit of promo.

I’m not the first to think this, it’s heavily inspired by the project of Daniel Bourdeau (forgive him the slop website…) who is deciphering WWII messages faster than they can be checked, and a new eyewitness account of the dodo among others.

Similar to what’s happening in mathematics, there’s some uncertainty when it comes to verifying that the discoveries are correct. Several of the first Erdős problems solved by LLMs already had solutions in the literature that hadn’t been linked to the problem, rather than a new idea behind them. This exact skill seems especially valuable for the study of history, because of the huge amount of digitized documents (and the many still missing). Just like in software, the bottleneck moves to verification.

To find topics to explore, I looked for options related to Galicia / Spain that I could reasonably verify on my own. I ended up with a list of a few that I’d like to tackle as soon as possible. In this post I present the results of the first one, to which I devoted a couple dozen hours of my own and several hundred hours of agents. This is a somewhat long post, since I try to talk about everything I find interesting. I encourage using the table of contents to skip anything you’re not interested in.

The challenge

In the Prólogo apologético of Teatro crítico universal III (1729), §5, Benito Jerónimo Feijoo throws down this challenge to all his readers:

«Lector mío, si estás en Madrid, y entiendes el Francés, ruégote que busques las Memorias de Trevoux, y el Journal des Sçavans, que no pueden faltar en la Biblioteca Real, y en otras; que unos, y otros libros vuelvas, y revuelvas bien; y cuando halles ni un párrafo sólo, ni aun cuatro líneas, que sean traslado, o traducción de ellos, o en este Tomo, o en alguno de los antecedentes, quiero que todos tres los des al fuego, y me obligo a restituirte el dinero que te han costado.»

Feijoo's challenge in the prologue of volume III, first edition of 1729, with the passage underlined
Source: Biblioteca Nacional de España, first edition of volume III (1729), image 18, CC BY 4.0..

If it catches your eye that the “s” looks like an “f”, it’s known as the long S and it has a fascinating history, related to the printing presses of the time, but if I stop at every detail like this I’d never finish. I know because I made that mistake in a previous post that I never got around to finishing.

The original idea is to compare volumes I–V (Feijoo himself extended the challenge in 1733 to his new works) of the Teatro crítico universal (TCU from here on) with the Memorias de Trévoux and the volume of the Journal des Sçavans that Feijoo said he had, and see if we find copied passages.

«Solo me resta un recurso; y es el que propondré ahora. Desafío al Anónimo Autor de la Carta, (sea el que se fuere) y a todos los demás que quieran conspirar con él, para que en una o muchas hojas volantes den al público señalados los lugares de las Memorias de Trevoux, de donde pretenden que haya sacado yo lo mejor que he empleado para el fondo de mi Obra. En vista de las citas ofrezco exhibir las Memorias de Trevoux, (ciento y veinte y cuatro tomos son los que tengo) ante dos Caballeros de los principales de esta Ciudad, y dos Eclesiásticos de la primera distinción, que unos y otros entienden bien el Francés, los cuales, leídos con exactitud los lugares señalados, darán certificación pública, firmada de sus nombres, de que es falsa la acusación, y fingido el robo que me imputan.»

Feijoo's second challenge, volume V, first edition of 1733, page 388, number 44
TCU V, discourse XVII, no. 44, p. 388. First edition, Universidade de Santiago de Compostela; transcription.

Historical context

Benito Jerónimo Feijoo was born in 1676 in Casdemiro, near Ourense, the firstborn of a well-off family of the Galician middle nobility. He entered the Benedictine monastery of Samos early, going against the natural path of the first son (within the marriage). He studied and taught in Galicia, Salamanca and León, and arrived at San Vicente de Oviedo in 1709, where he stayed for the rest of his life (despite multiple invitations to move to more important cities). There he combined his study and university teaching with his ecclesiastical career until his death at 87. His wake and burial drew huge crowds. His fame attracted many visitors and, many years after his death, people were still visiting the places where he had lived.

Portrait of Feijoo engraved by Juan Bernabé Palomino
Source: Biblioteca Virtual Miguel de Cervantes. Palomino engraved it from a portrait made around 1733–1734.

There’s quite a lot of information about his life, since he was one of the most celebrated people of the time. He wrote a short autobiography, at the request of a German baron, which along with the funeral honours and testimonies like Campomanes’ Noticia de la vida de Feijoo, as well as letters and news, helps us get a good idea of the life he lived1. Even so, a large part of his papers were lost with the disentailment, and many of the ones left in the monastery of Samos burned in a fire in 1951.

Monasteries and universities formed an unmatched network for passing knowledge around at the time, they lent books to each other and helped each other get hold of new releases. Feijoo read Latin and French. He often knew English or German ideas through their French version. In this period (late 17th century) a sort of Spanish enlightenment begins, physicians, mathematicians and philosophers (called “novatores”) who urged people to pay attention to observation and to the knowledge coming from Europe.2 Feijoo wasn’t one of the first to discuss these ideas, but he helped bring them to a much wider audience.

The first volume of TCU, his first big essay, came out in 1726, when he was 49 and Spain had had a quarter century of Bourbon dynasty. The title is misleading, it has nothing to do with theatre, they were «discursos varios en todo género de materias, para desengaño de errores comunes». Medicine, astrology, church music, languages… In Paralelo de las lenguas he argued that Galician and Portuguese were actually the same language. One of the most controversial was the Defensa de las mujeres.

Title pages of the first three volumes of the Teatro crítico universal
BNE first editions: volume I, 1726, II, 1728 and III, 1729.

The publication of these volumes was largely pushed by his ecclesiastical superiors, which is especially relevant since printed works needed approvals and licenses, which in turn are useful to reconstruct his character and standing at the time.3

approvals and permissions for volume I
Preliminaries of volume I, with the license from the Benedictine order
The order’s license authorized Feijoo to print the book after its examination.
Approval by Domingo de Lossada in the preliminaries of volume I
One of the approvals, written by Domingo de Lossada. Besides the judgement on the book, it contains praise for the author.
Preliminaries of volume I, with the summary of the Council's license
The summary of the license records the Council’s authorization to print it according to the approved original.

Source: Biblioteca Nacional de España, first edition of volume I, 1726.

Up to eight volumes were published between 1726 and 1739. After that he continued with the Cartas eruditas y curiosas, published in five volumes between 1742 and 1760. One estimate puts the circulation of his work at about 440,000 volumes.4 My description here probably falls short of getting across how important he was, he must have been the most recognized Spanish writer inside and outside Spain at the time, he wrote in Castilian when this kind of content was usually in Latin, and his ideas were extremely controversial.

Funnily enough, this year Oviedo hosted a conference for the three centuries of the first volume.

Controversy

Living in Oviedo, getting hold of foreign novelties was really hard. An important contact was Martín Sarmiento, also a Galician Benedictine, who supplied him with materials from Madrid, corrected his texts and defended his positions. Journals like the Memorias de Trévoux, which since 1701 had been collecting reviews and news on all kinds of subjects, were especially relevant.

A few weeks after TCU was published, writings by different authors responding to it were already going around, and several of his friends came out in his defense. There’s already the odd accusation, but it isn’t until 1728, in the Tertulia histórica y apologética, that a canon who claims to have studied in Paris says TCU is a translation of several French works, without many more details.

Accusation of translating French works in the Tertulia histórica y apologética of 1728
Tertulia histórica y apologética, 1728, pp. 8–9. The canon proposes checking the French sources in the Royal Library.

I find the way they accused each other especially elegant, in the case of Estrado crítico (1727), it’s a conversation between four ladies, in Tertulia histórica y apologética, one of the characters complains that, after his wife read Feijoo, the marriage was ruined and the woman started studying Latin and French and talking about Descartes.

Title page of the Estrado crítico en defensa de las mujeres contra el Teatro crítico universal, 1727
Estrado crítico, 1727. The title page is on the right of the scan. Biblioteca Nacional de España, copy 3/32988(11).

Feijoo answers this last one in the prologue of the third volume, in 1729, where he says he has a hundred volumes of Trévoux, but distinguishes between making use of a book and copying it. This is also where he sets out the public wager this post looks into.

That same year, Salvador José Mañer did exactly that, and in Anti-Teatro (1729) pointed to specific places. Feijoo answered him, and Mañer came back in 1731. The fight went on for decades, with new opponents and defenders. In 1750, a Royal Order from Fernando VI banned the publication of the third volume by one of Feijoo’s critics, as well as any further rebuttals. This is my attempt at reconstructing the accusations made up to that point:

Extended timeline of the controversy (1726–1750)
Date Author Position Work or intervention What it questions or contributes Digitization / text
3 Sep. 1726 Feijoo Author Teatro crítico universal, I Sixteen discourses, including the ones on medicine, astrology and the defense of women. Scan, BNE; transcription
5 Oct. 1726 Martín Martínez For Carta defensiva sobre el primer tomo del Teatro crítico Praises Feijoo’s work and discusses his skepticism about medicine. Transcription; BNE
22 Oct. 1726 Pedro Aquenza Against Breves apuntamientos en defensa de la medicina y de los médicos Considers that Feijoo insults physicians and a faculty created by God. Reproduction, BVMC/UCM
Oct. 1726, according to Caso y Cerra José Francisco de Isla, attributed For Blanda, suave y melosa respuesta a los ferinos y furiosos apuntamientos Burlesque reply to Aquenza. Facsimile, Biblioteca de Bizkaia
29 Oct. 1726 Francisco Suárez de Ribera Mixed Templador médico de la furia vulgar Defends Martínez and Feijoo against Aquenza, but disagrees with the discourse on medicine and defends the bezoar. Scan, Wellcome
Nov. 1726 Feijoo Author Respuesta a los doctores Martínez, Aquenza y Ribera Answers the first interventions on his medicine discourse. Reply to Martínez, transcription of a later edition
3 Dec. 1726 Eustaquio Cerbellón de la Vera Against Diálogo harmónico Defends church music against Feijoo’s criticism. —
17 Dec. 1726 «Laurencio Manco de Olivares», pseudonym Against Contradefensa crítica a favor de los hombres Contests the defense of women in the first volume. —
24 Dec. 1726 Francisco Suárez de Ribera Against Medicina cortesana satisfactoria Replies to Feijoo and sticks to his defense of the bezoar. —
late 1726 Diego de Torres Villarroel Against Posdatas de Torres a Martínez; Montante christiano y político Steps into the dispute over medicine and astrology. Posdatas, transcription
1726–1727 Jorge Irún y Adecha For Desengaño de delirios Replies to Manco de Olivares in defense of women. —
4 Jan. 1727, date of the text Agustín Castejón, attributed Against Dudas y reparos sobre que consulta un Escrupuloso Questions Feijoo’s medical opinions and his defense of women. —
4 Feb. 1727 Martín Martínez For Juicio final de la astrología Contests astrology and replies to Torres. Scan; Transcription
4 Feb. 1727 «Ernesto Frayer», Martín de Mendoza de Pina Specific disagreement Discurso philológico crítico sobre el corolario del Discurso XV Disputes the identification of Galician and Portuguese. Transcription
16 Feb. 1727 Juan Antonio Santareli Against Estrado crítico en defensa de las mugeres Dialogue defending the traditional ideal of woman. The reference to Bellegarde on p. 20 repeats a quotation Feijoo acknowledged; it doesn’t prove hidden copying. BNE
25 Feb. 1727 Jerónimo Zafra Against Antiteatro del Teatro crítico Contests the first volume, before Mañer’s Anti-Theatros. —
25 March 1727 Ignacio Ximénez Saforcada, as «Geminiano Zafra Ciscodexa» Against Antitheatro délfico judicial jocoserio Another rebuttal of the first volume. —
25 March 1727 Bernardo López de Araujo Against Residencia médico-cristiana Defends traditional medicine against Feijoo. —
6 May 1727 Ignacio García Ros Against Medicina vindicata Weighs in against Feijoo’s criticism of medicine. —
1 Jul. 1727 Anonymous Against Cátedra de desengaños médicos Continues the medical rebuttals of the first volume. —
1727 Anonymous For Papel de Marica la Tonta en defensa de su sexo Another reply to Manco de Olivares. —
1727 Diego de Torres Villarroel; Julián Salinero Against Entierro del Juicio final y vivificación de la astrología; Pragmática del tiempo Replies to Martínez’s Juicio final. Entierro, transcription
1727 Anonymous For Blanda, suave y melosa curación del Escrupuloso y de sus flatos espirituales Satirical reply to the Escrupuloso. Feijoo disapproves of its insults. Facsimile, Biblioteca de Bizkaia
1727 Feijoo, attributed Author Satisfacción al Escrupuloso Answers the objections and rejects the anonymous defender’s tone. Transcription
1727 «José Madaria», attributed to Feijoo Author Respuesta al señor Asiodoro Musical reply to the Diálogo harmónico. —
1727 Eustaquio Cerbellón de la Vera, attributed Specific disagreement Respuesta de Asiodoro a Madaria Continues the technical discussion on music. —
1727 Francisco Dorado; Feijoo Against / author Discurso fisiológico-médico; Respuesta al discurso fisiológico-médico Dorado criticizes Feijoo’s medical opinions; Feijoo replies. Feijoo’s reply, transcription
1727 Francisco García Cabero Spin-off dispute Templador veterinario de la furia vulgar Defends farriers against Suárez de Ribera. Google Books
6 Apr. 1728 Feijoo Author Teatro crítico universal, II Second volume, also targeted by Mañer’s Anti-Theatro. Scan, BNE; transcription
20 Apr. 1728 «Jaime Ardanaz y Centellas», unidentified author Against Tertulia histórica y apologética Accuses Feijoo of copying Naudé and of translating French journals. The accusation about Trévoux and the Journal des Sçavans on p. 9 doesn’t point to specific passages. BNE; Google Books
1728, licenses from May Felipe Brizeño y Zúñiga Against Juicio particular del Juicio Universal Contests the discourse on the antipathy between the French and the Spanish. Includes a censure by Torres. —
Nov. 1728 Mémoires de Trévoux News «De Madrid», p. 2140 Reports on how the controversy is growing and on its replies and defenses. —
1728 Manuel José de Medrano Against Vida de Santa Inés de Monte-Policiano Criticizes the passage on Savonarola. The Tertulia picks up his accusation of copying Naudé. Google Books
1728 Juan Martín de Lessaca Against Apología escolástica en defensa de la Universidad de Alcalá Defends scholastic medicine against Martínez and Feijoo. Google Books
31 May 1729 Feijoo Author Teatro crítico universal, III, «Prólogo apologético» Replies to the Tertulia and sets the four-line challenge. Scan, BNE; Transcription
7 Jun. 1729 Salvador José Mañer Against Anti-Theatro crítico, on volumes I and II Points out seventy oversights and specific passages supposedly taken from Trévoux, among them the Parent one from 1716. Scan; Bibliographic record
6 Sep. 1729 Francisco Antonio de Tejeda, attributed Against Apelación sobre la piedra filosofal Contests the discourse on the philosopher’s stone in volume III. —
1729 José Ortiz Barroso Against Reflexiones physico-curiosas Discusses, among other things, the discourse on the old rooster’s egg in volume II. —
1729 Antonio Heredia y Ampuero Against El estudiante preguntón Contests Feijoo, Martínez and the piscatores. Transcription
10 Jan. 1730 Feijoo Author Ilustración apologética Replies to Mañer and refutes his seventy oversights. The title page says 1729. Scan, Universidad de Oviedo
Sep. 1730 Letter from Zaragoza, attributed to Tejeda by Feijoo Against Letter in Trévoux, pp. 1693–1696 Claims Feijoo took the best of his work’s content from the journal. —
1730 Lucas Montoya y Rada For Rebeses al estudiante preguntón Defense of Feijoo against Heredia by way of an allegorical play. —
1730 Diego de Torres Villarroel Against Último sacudimiento de botarates y tontos; y si me vuelven a enfadar no será el último Attacks, among others, the defender Lucas Montoya y Rada. —
30 Jan. / March 1731 Carlos de Montoya y Uzueta; Martín Sarmiento Against / for Crítico y cortés castigo de pluma; Carta a don Carlos Montoya Montoya attacks Feijoo and Sarmiento. Sarmiento replies as «Sancho Revulgo y Cantalapiedra». —
Jun. 1731 Jean-Baptiste Boyer For «Lettre sur un ouvrage du R. P. Feijoo», Mercure de France Praises the reception and standing of the Teatro crítico. —
7 Aug. 1731 Salvador José Mañer Against Anti-Theatro sobre el tomo tercero and Réplica satisfactoria Announces 998 errors and responds explicitly to the challenge. Repeats the accusation about Parent and adds others. Scan, Universidad de Alicante
Apr. 1732 Jean-Baptiste Boyer For Review in the Mercure de France, pp. 743–752 Comments on volume III and the Ilustración apologética. —
23 Dec. 1732 Martín Sarmiento For Demostración crítico-apologética, two volumes Defends Feijoo, among other ways by explaining common sources. Talks about more than a hundred «papelones» against him. Volume I; volume II; transcription
7 Jul. 1733 Feijoo Author Teatro crítico universal, V, discourse 17 Replies to the Zaragoza letter and renews the challenge. States he owns 124 volumes of Trévoux out of the 128 published. Transcription
1733 Jacinto Segura Against Norte crítico Disputes the treatment of Savonarola in the prologue of volume III. Google Books
14 Sep. 1734 Manuel Mariano Ballester y de la Torre Against Combate intelectual Contests three discourses of the Teatro crítico. —
19 Oct. 1734 «Álvaro Menards», Salvador José Mañer Against El famoso hombre marino del P. M. Feijoo Attacks the account of the Liérganes man in volume VI. —
7 Dec. 1734 Manuel Marién y Rubio Against Primera parte de la singular vida de don Alonso Pérez de Saabedra, falso nuncio en Portugal Contradicts Feijoo’s account of how the Inquisition was introduced in Portugal. —
1734 Salvador José Mañer Against Crisol crítico, two parts Replies to Sarmiento’s Demostración. PDF scan; Google Books
1735–1737 Ignacio de Armesto y Ossorio Mixed Theatro anti-crítico universal, three books Presents himself as arbiter of the dispute between Feijoo, Sarmiento and Mañer. Scan, Universidad de Alicante; HathiTrust catalog
1735 Jacinto Segura Against Vindicias históricas por la inocencia de fray Gerónimo Savonarola Keeps up the defense of Savonarola against Feijoo. —
1741 Alonso Rubiños Against Theatro de la verdad Defends the exorcism of animals against volume VIII. —
1742 Nicasio de Zárate Against Bayles mal defendidos y Señeri sin razón impugnado Contests Feijoo’s defense of dances. Google Books
1743–1744, bibliographic dating Francisco Arango Against Carta apologética en favor del anual milagro de las flores Defends the supposed miracle of the flowers of San Luis. —
1744 Antonio Raymundo Pasqual Against El milagro de la sabiduría del B. Raymundo Lulio Defends Ramón Llull against Feijoo. Sermon preached in 1743. Scan, BNE
1744 Joaquín Javier de Aguirre Against El príncipe de los poetas Virgilio, mantenido en su soberanía Defends Virgil against Feijoo’s preference for Lucan. —
1744 Juan Ros Against Relación histórica de la portentosa anual maravilla de las flores de San Luis Another defense of the supposed miracle. —
1745 Feijoo Author Hecho y derecho en la famosa cuestión de las flores de San Luis Replies to the defenders of the miracle, in letter II.29. Transcription
1746 Bartolomé Fornés Against Liber apologeticus Artis Magnae B. Raymundi Lulli Defense of Ramón Llull, written in Latin. —
7 Nov. 1748 Fernando VI For Royal decree appointing Feijoo royal councillor Royal recognition of Feijoo. —
6 May 1749 Francisco de Soto y Marne Against Reflexiones crítico-apologéticas, two volumes Lists plagiarism accusations discourse by discourse, among them Parent and the iron in plants. The licenses are from 1748. Volume I; volume II
23 Sep. 1749 Feijoo Author Justa repulsa de inicuas acusaciones Replies to Soto y Marne and rejects his plagiarism accusations. Transcription
23 Jun. 1750 José de Carvajal y Lancáster, by order of Fernando VI For Royal order on the rebuttals of Feijoo Bans Soto y Marne’s third volume and any new rebuttals. Original cited: BNE, ms. 10.579, ff. 31v–32r. Transcription in Caso y Cerra, no. 283
Sep. 1750 Francisco de Soto y Marne Against Memorial a la Majestad Católica Reacts to the ban on his rebuttals. —
1750 Lucas Ramírez, attributed For La derrota de los alanos Reply to Soto y Marne’s Reflexiones. Scan, BNE

Sources: Caso González and Cerra Suárez, Bibliografía feijoniana (1981); Campomanes, Noticia de la vida y obras (1765).

This whole debate is pretty unfair to Feijoo, who made it clear that he read these texts and took inspiration from them, although he distinguished between making use of a book and translating it. Anyway, I’ll leave some of my incredible titles, like:

I wish more posts had titles like these!

Process

Four lines take up about 35–40 words in the first editions. At first I compared volumes I, II and III with the earlier issues of the Memorias de Trévoux and with the only volume of the Journal des Sçavans that Feijoo said he had. Then I extended the search to volumes IV and V and to Trévoux up to 1732 to check the 1733 challenge, which no longer mentioned the Journal or set a four-line limit. The search grew whenever I found promising cases, since I also went looking for the books mentioned in the corresponding journal article.

There are endless thematic overlaps, but for a passage to count it has to be a clear translation or a paraphrase with some concrete sign of dependence: same error, same unusual sequence of facts…

Corpus

Getting the documents with good OCR was not trivial at all. The Munich Digitization Center (MDZ) kindly gave me access to their files, but I ended up redoing the OCR of Internet Archive scans for better quality.

Downloaded documents and OCR
Document Downloaded OCR
TCU I–V Transcriptions; first editions I, II and III, 664 openings; PDF and images of IV–V BNE OCR for I–III; working transcriptions for I–V
Trévoux, Getty / Internet Archive, 1701–1732 129 volumes: images and ABBYY text Redone with Kraken + PP-OCRv6-medium: 66,964 leaves from 1701–1729 and about 7,034 from 1730–1732
Trévoux, BSB/MDZ, 1701–1732, and Amsterdam reprint 134 volumes, including nine from Amsterdam; occasional images. July–September 1727 is missing 78,908 hOCR pages downloaded
Journal des Sçavans of 1682, Amsterdam, 1683 482 images BSB OCR for 60; Tesseract for the other 422
Histoire de l’Académie royale des sciences, 1699–1728 Text of 30 volumes Existing OCR
Bayle, Dictionnaire historique et critique, 1702 Three volumes: I, II, III Existing OCR
Menagiana, 1729 Four volumes Existing OCR
Fontenelle, Œuvres diverses, 1728–1729 Three volumes Existing OCR
Moréri Four volumes from 1717 and two of the 1716 Supplément Existing OCR
Montfaucon, L’Antiquité expliquée, 1719 Ten volumes Existing text
Dictionnaire de Trévoux, 1721 Text Existing text
Writings from the controversy PDFs, images or texts of Estrado, Tertulia, Mañer 1729/1731/1734, Sarmiento 1732, Soto Marne 1749, Feijoo’s replies and Bibliografía feijoniana OCR or transcriptions, depending on the document
Sources for each case Reviewed books, scientific memoirs and other texts: PDFs, images or transcriptions Depending on availability

OCR

Most of the documents have online versions, but they usually carry corrections from later editions and changes to make them easier to read, so we need the first edition. I did most of the processing on my server with an RTX 3090. The OCR ran on the GPU and the CPU in parallel for about three days. At first I figured it would need more than a week, but several optimizations along the way brought the time down.

After trying a few I went with Kraken, a very flexible model specialized in historical texts, to locate the lines on each page, and PP-OCRv6-medium to recognize the text. The Kraken documentation explains the two stages.

It was important to take into account the quirks of the printing presses of the time that I mentioned before. The Internet Archive text already had OCR, but with quite a few problems (long s -> f, lots of split words…). I made a benchmark with a few pages from each volume where I had a reference transcription to compare against. Multimodal LLMs all worked really badly (gpt6, moondreamV2, Gemini 3.8) and the cost wasn’t viable.

Character error rate of the OCR engines on the Trévoux sample

Comparison of the underlined Trévoux passage with the old and the new OCR

Examples
Faded scan of Trévoux from 1703, page 202
1703, p. 202: faded scan and names in italics. Internet Archive original.
Trévoux from 1729, page 1410, with italics and tight spacing
August 1729, p. 1410: italics, ligatures and tight spacing. Internet Archive original.

Matching

With the texts available, we have to look for copied passages and verify them. The work by Hinderks, Ledins, Ginter and Tolonen on “translation mining” is very close to this. The idea is to compute embeddings of the fragments, which represent their content as vectors and let you compare texts in different languages. Two sentences that say something similar should end up close together regardless of the language. There are other precedents, like Roe, Olsen and Morrissey, who looked for translations of the Cyclopaedia in the Encyclopédie using machine translation and text alignment.

Applying this naively gives a ton of false positives. Feijoo and the French journals were constantly talking about the same topics, citing the same authors and works. I tried several models and the one that gave the best results by far was LaBSE. This model was trained specifically for translations. I considered retraining it with Feijoo-specific data, but that stays as possible future work, I don’t think it’s worth it.

A high score on its own said little, so I used margin scoring. It compared the similarity of each pair with the similarity both fragments had to their closest alternatives in the other language. If they were very similar to each other, but also to lots of others, the score went down, and vice versa. This helped pick which matches were worth reviewing. I also looked at proper names and numbers, which tend to survive a translation, to find regions of interest. Latin quotations weren’t useful since they tended to repeat in many places.

To complement this, I also machine-translated Feijoo’s text into French and compared rare words. The kind of false positives it found was different from the embeddings’, but it turned up some extra case like the planets one I go into later. A lot of duplicates had to be cleaned up, because I was doing the analysis at the same time as I was processing the data.

Finally, to check how well this process worked, I planted artificial cases in Feijoo’s text. In the hardest examples, it got up to 42 out of 50, the ones it missed being mostly paraphrases of the machine translation. I’m sure I haven’t detected every possible case.

Verifying

This is where most of the work went. The detectors turned up a lot of cases but each candidate was quite open to interpretation.

A first agent read the fragments and classified them as translation, close paraphrase, shared fact or noise. The most promising cases got an adversarial review, and any disagreement went back to the first step to be reclassified. If it passed this first filter, it was compared again against the first editions and the other sources mentioned. Trévoux was a review journal, the parallel usually came from the book being reviewed, and I wanted to identify those cases as such. Same with translations that had already been identified, Mañer, Sarmiento, Soto Marne and the later literature backed up many of the cases I found.

Finally, I checked by hand every one that made it through the whole process, which took a good while, but in exchange I come away with the ability to read 18th-century books fairly fluently.

Full pipeline: download, OCR, text preparation, two search paths, evaluation, source review and case classification

Working with agents

Unlike a lot of my friends, I’m still somewhat reluctant to throw swarms of agents at a task with little supervision. In my experience, they still don’t work as well as when you hold their hand a bit.5 The trend is clearly toward them needing less and less supervision and being able to take on longer tasks, I wouldn’t be surprised if in a few months this very blog could be replicated with a relatively simple prompt, but for now I had to be pretty involved in the process.

One thing I notice is that they fixate on details and waste huge amounts of time and tokens on them, details that fall apart with the slightest bit of big-picture view. But by far the biggest problem was dealing with context loss over time. The way I ended up working was a main agent orchestrating multiple subagents (usually around 10) á la cursor projects, which were the ones doing the actual work (reviewing promising cases, monitoring the OCR runs, reviewing literature…). I used both Opus5.5 and GPT-6Astra, with anecdotally better results with Opus.

The other big problem was documentation, since most of the subagents had a defined task and then disappeared, and I was trying to avoid filling the orchestrators’ context with irrelevant stuff, I documented almost every relevant fact in mds. The result was that I had hundreds of documents, written in prose that is extremely uncomfortable to digest, and a lot of them full of questions already answered, hypotheses already tested, data we had already found out was wrong…

The solution was a mix of agents tasked with routinely reviewing the documentation, and a lot more manual review than I’d like to admit. I don’t have enough hands to count the times a subagent jumped in with a “groundbreaking discovery” that sank under the slightest scrutiny. LLMs tend to overrate their work and underrate their capabilities at the same time.

Another thing to keep in mind in a set-up like this is that it’s going to overindex on your words massively, and it’s important to keep that in mind every time you talk to the orchestrator. More than once it started steering multiple subagents with instructions meant for one specific task. At the same time, the way you talk to it has a huge effect, the clearest example being Terence Tao’s chats, where, without really complex prompts, and with access to the same models as me, he gets incredible results.

They were pretty useless for writing this post, but they did help quite a bit with fixing errors and suggesting places where the prose could be better. As I wrote more, the suggestions got quite a bit sharper but never quite hit the mark. Having a corpus of your own writing is useful for keeping a similar voice, however I do find that sometimes my way of writing (even of talking!) starts to look like an LLM’s. I had already noticed this in colleagues before, but writing here I caught myself correcting myself several times. It worries me that this is happening at a large scale.

Results

So far I’ve found about 10 places where I’m fairly sure they settle the challenge, plus several more that could count depending on how you read the rules.

Search Candidates judged Strict (A) Dependent or short (B/C) Cited translations (D)
1729 challenge: TCU I–III against Trévoux and the 1682 volume of the Journal 633 9 + 2 borderline 4 B + 4 C 4
1733 extension: TCU IV–V against Trévoux up to 1732 237 1 1 B + 2 C 8
Table of challenge cases
Case Feijoo Trévoux Strict words Why it counts; what was known
C001, sunspots II.14, 1728 Parent, Feb. 1716 ≈128 Shares an error about the book of the Georgics and «Tum caput». Mañer already pointed out the article in 1729.
C002, iron in plants II.14, 1728 Lémery, Mar. 1707 ≈180 Follows the syntax of the review and details missing from the original memoir. Soto Marne pointed out part of the passage.
C003, abbess of Fontevrault I.16, 1726 Dec. 1704 ≈54 The combination of queen, king and burned verses is missing from Moréri. I found no earlier identification.
C004, Philippines II.2, 1728 Taillandier, Jul. 1715 ≈127 Calques and the same geographical error: Dapitan becomes Magallanes. Mañer had already denounced it.
C005, medical sects I.5, 1726 Barchusen, Nov. 1710 ≈120 Follows the order and mergers of the review, not of the book. Feijoo revealed in 1727 that he had that extract from Trévoux.
C009, musical modes I.14, 1726 Bonnet, Apr. 1716 37–51 Repeats a sequence of characterizations and «Subphrigio». I found no earlier identification.
C010, longevity I.12, 1726 Temple, Jul. 1702 43 Keeps the review’s «Nesmond» and the scope of «toute l’Angleterre». I found no earlier identification.
C017, planetary rings III.2, 1729 Feb. 1718 ≈82 Attributes the story to Camilo Leonardo, but his book doesn’t contain those details; it follows the French review. I found no earlier identification.
C020, giant’s hand I.12, 1726 Sep.–Oct. 1701 54 Feijoo cites the English Transactions, but follows the French excerpt, which leaves out the porpoise. I found no earlier identification.
C012, saints’ lives III.6, 1729 May–Jun. 1701 ≈48–51, borderline One sentence follows the reviewer’s condensation; the dependence on Trévoux is close to the threshold.
C038, fashions II.6, 1728 Henrion, Feb. 1702 38, minimum 30 Sentence translated clause by clause. It came out of a sample read by hand, not from the detectors; the Mercure galant still needs checking.
D2-C007, Behaim IV.8, 1730 Stuvenius, May 1716 ≈98, minimum 39 Follows the order and details of the review that aren’t in the Latin book. No earlier identification found. Searches in the Journal and the Acta, including the supplements checked, found no other equivalent review.

The sunspots

Mañer had already pointed out this case in 1729. In the Paradojas físicas, Feijoo talks about sunspots, gathering stories and observations of enormous spots. This same topic shows up in an article by Antoine Parent in Trévoux 1716, twelve years before the second volume of TCU.

When it comes to Virgil, both texts put a few lines in the second book of the Georgics, when they’re actually in the first, lines 466–468. Feijoo rearranges part of the explanation and fixes some other detail, but it’s one of the clearest examples of how he was probably working off the French text and adapting it as he went.

Passages and sources for case C001.

A ring to get rich and read minds

This case came up through the other search path, after translating Feijoo’s text into French. In Secretos de Naturaleza, Feijoo attributes to Camilo Leonardo a list of seven stones, seven metals and their corresponding planets. Then he explains how to make a ring that would bring wealth. The Trévoux review from February 1718 says the same thing, but it’s talking about three authors. The first is Camilo Leonardo; the list of correspondences comes from the second, Pierre d’Arleu; the ring example, from the third, Albinius. Feijoo keeps the name that shows up at the start and attributes the whole block to him.

Passages, original books and sources for case C017.

Examples

These are the clippings from the first editions. Open each case to compare them.

C001 · sunspots
Feijoo clipping: sunspots (C001)
Feijoo · II (1728), p. 250, n. 23 · BNE (CC BY 4.0)
Trévoux clipping: sunspots (C001)
Trévoux · February 1716, art. XXIV, p. 331 · BSB (NoC-NC 1.0)
C002 · iron in plants
Feijoo clipping: iron in plants (C002)
Feijoo · II (1728), p. 258, n. 39 · BNE (CC BY 4.0)
Trévoux clipping: iron in plants (C002)
Trévoux · March 1707, art. XXXIV, p. 479 · BSB (NoC-NC 1.0)
C003 · abbess of Fontevrault
Feijoo clipping: abbess of Fontevrault (C003)
Feijoo · I (1726), p. 361, n. 122 · BNE (CC BY 4.0)
Trévoux clipping: abbess of Fontevrault (C003)
Trévoux · December 1704, art. CLXXXIX, p. 2119 · BSB (NoC-NC 1.0)
C004 · Philippines
Feijoo clipping: Philippines (C004)
Feijoo · II (1728), p. 53, nn. 73–74 · BNE (CC BY 4.0)
Trévoux clipping: Philippines (C004)
Trévoux · July 1715, art. XCVII, pp. 1163–1164 · BSB 1; BSB 2 (NoC-NC 1.0)
C005 · medical sects
Feijoo clipping: medical sects (C005)
Feijoo · TCU I (1726), p. 112 · BNE (CC BY 4.0)
Trévoux clipping: medical sects (C005)
Trévoux · Nov. 1710, p. 1954 · Internet Archive
C009 · musical modes
Feijoo clipping: musical modes (C009)
Feijoo · TCU I (1726), p. 274 · BNE (CC BY 4.0)
Trévoux clipping: musical modes (C009)
Trévoux · Apr. 1716, p. 597; Apr. 1716, p. 598 · Internet Archive 1; Internet Archive 2
C010 · longevity
Feijoo clipping: longevity (C010)
Feijoo · TCU I (1726), p. 234 · BNE (CC BY 4.0)
Trévoux clipping: longevity (C010)
Trévoux · Jul. 1702 (title page misprinted «1701»), p. 78; Sep.–Oct. 1701, p. 299 · Internet Archive 1; Internet Archive 2
C017 · planetary rings
Feijoo clipping: planetary rings (C017)
Feijoo · TCU III (1729), p. 25 · BNE (CC BY 4.0)
Trévoux clipping: planetary rings (C017)
Trévoux · Feb. 1718, p. 334; Feb. 1718, p. 336 · Internet Archive 1; Internet Archive 2
C020 · giant's hand
Feijoo clipping: giant's hand (C020)
Feijoo · I.12, §VIII, n. 23, p. 243 · BNE (CC BY 4.0)
Trévoux clipping: giant's hand (C020)
Trévoux · September–October 1701, p. 291 · BSB (NoC-NC 1.0)
C012 · saints' lives (borderline)
Feijoo clipping: saints' lives (borderline) (C012)
Feijoo · III.6, §I, n. 5, p. 99 · BNE (CC BY 4.0)
Trévoux clipping: saints' lives (borderline) (C012)
Trévoux · May–June 1701, p. 62 · BSB (NoC-NC 1.0)
C038 · fashions (borderline)
Feijoo clipping: fashions (borderline) (C038)
Feijoo · II.6, §II, n. 6, p. 141 · BNE (CC BY 4.0)
Trévoux clipping: fashions (borderline) (C038)
Trévoux · February 1702, pp. 10–11 · BSB 1; BSB 2 (NoC-NC 1.0)
D2-C007 · Behaim (1733 challenge)
Feijoo clipping: Behaim (1733 challenge) (D2-C007)
Feijoo · IV.8, §XXXIX, n. 85, p. 210 (continues on p. 211) · USC Minerva
Trévoux clipping: Behaim (1733 challenge) (D2-C007)
Trévoux · May 1716, art. LVI, pp. 849–850 · BSB 1; BSB 2 (NoC-NC 1.0)

Copies from other books

The challenge only mentioned Trévoux and the Journal, but since I had all this set up I ran it over other French sources (the Histoire of the Académie des sciences, Bayle’s Dictionnaire, Moréri, the Menagiana and Fontenelle). About 17 passages come out, almost all from Moréri. Not surprising, Feijoo cites these works in several of his writings, but these are the places where he didn’t:

Table of examples from other books
Case Feijoo Book Uncited words What gives it away
O153, the cavalier Borri III.2, 1729 Bayle 1702 (Moréri reprints the main text) ≈800 Follows the article and its notes in the same order. In the whole discourse he only cites Moréri, and for something else.
O508, the learned Italian women I.16, 1726 Moréri, seven articles ≈585 Bucca, Nogarola, Cereti, Fidele, Cibo, Marchina and Cornaro. Moréri is the only text checked that has all seven.
O526, Apollonius of Tyana II.5, 1728 Du Pin 1705 ≈420 «Ciento y veinte años», like Du Pin’s «six vingt ans»; Trévoux says 110. He cites Philostratus and Lucian, not Du Pin.
O475, the Sibyls II.4, 1728 Moréri ≈330 Tarquin’s «trescientos escudos» and a counting error («Eliano cuatro») that comes from Moréri.
O471, Delphi II.4, 1728 Moréri ≈320 Echecrates and the maidens «consagradas a Diana», who aren’t in Van Dale or in Fontenelle.
O481, the women of Curzolari I.16, 1726 Moréri ≈75 Two of Moréri’s errors: it moves the Korčula story to the Lepanto islands, and the date «el año antecedente».
O474, Poverty I.3, 1726 Moréri, Paris edition ≈48 «Curio, y de Camila» for Camillus, which comes from the French «Camille».

The errors he copies also give away which copy he had. For example, I had downloaded the Amsterdam Moréri (1716–17), but «Camila» and the Aristophanes line about Poverty are only in the Paris editions (1707 and 1718).

Documents for the examples from other books
O153 · the cavalier Borri
Feijoo clipping: the cavalier Borri (O153)
Feijoo · III (1729), p. 37, n. 38 · BNE (CC BY 4.0)
Bayle clipping: the cavalier Borri (O153)
Bayle · Dictionaire, 1702, t. I, p. 654, BORRI · Internet Archive

Case sheet O153

O508 · the learned Italian women (Casandra Fidele)
Feijoo clipping: the learned Italian women (Casandra Fidele) (O508)
Feijoo · I (1726), p. 364, n. 128 · BNE (CC BY 4.0)
Moréri clipping: the learned Italian women (Casandra Fidele) (O508)
Moréri · 1717, t. II, p. 71, FIDELE · Internet Archive

Case sheet O508

O526 · Apollonius of Tyana
Feijoo clipping: Apollonius of Tyana (O526)
Feijoo · II (1728), p. 110, n. 13 · BNE (CC BY 4.0)
Du Pin clipping: Apollonius of Tyana (O526)
Du Pin · L'Histoire d'Apollone, 1705, pp. 6–7 · BSB

Case sheet O526

O475 · the Sibyls
Feijoo clipping: the Sibyls (O475)
Feijoo · II (1728), p. 78, n. 3 · BNE (CC BY 4.0)
Moréri clipping: the Sibyls (O475)
Moréri · 1717, t. IV, p. 386, SIBYLLES · Internet Archive

Case sheet O475

O471 · the oracle of Delphi
Feijoo clipping: the oracle of Delphi (O471)
Feijoo · II (1728), p. 82, n. 11 · BNE (CC BY 4.0)
Moréri clipping: the oracle of Delphi (O471)
Moréri · 1717, t. II, p. 335, DELPHES · Internet Archive

Case sheet O471

O481 · the women of Curzolari
Feijoo clipping: the women of Curzolari (O481)
Feijoo · I (1726), p. 332, n. 47 · BNE (CC BY 4.0)
Moréri clipping: the women of Curzolari (O481)
Moréri · 1717, t. II, p. 311, CURSOLAIRES · Internet Archive

Case sheet O481

O474 · Poverty
Feijoo clipping: Poverty (O474)
Feijoo · I (1726), p. 65, n. 38 · BNE (CC BY 4.0)
Moréri clipping: Poverty (O474)
Moréri · Paris, 1707, p. 181, PAUVRETÉ · Internet Archive

Case sheet O474

Full table of other books
Case Feijoo Book Uncited Only from that book Citation
O153, the cavalier Borri III.2 §XI, nn. 37–40, 1729 Bayle 1702, BORRI; Moréri Suppl. 1716 ≈800 ≈80 none
O508, the learned Italian women I.16 §XVIII, nn. 124–131, 1726 Moréri 1717, seven articles ≈585 ≈150 none
O526, Apollonius of Tyana II.5 §V, nn. 12–13, 1728 Du Pin 1705 ≈420 ≈60 inherited
C022, Elena Cornaro Piscopia I.16, n. 131, 1726 Moréri 1716, Trévoux Oct. 1713 and Leti ≈355 ≈35 Leti, only for the praise of her
O475, the Sibyls II.4 §I, nn. 3–5, 1728 Moréri 1717 and Suppl. 1716 ≈330 ≈60 inherited
O471, Delphi II.4 §II, nn. 9–11, 1728 Moréri 1717, two DELPHES articles ≈320 ≈80 inherited
C007 + C021, the women painters and Madame Le Hay I.16 §XXII, nn. 142–145, 1726 Trévoux 1706 and 1713, Moréri 1716, Carducho 1633, Palomino 1715, Leti ≈290 24 from Trévoux none
O466, Sitti Maani I.16 §XX, n. 135, 1726 Moréri 1717 (Thévenot 1663, Rocchi 1627) ≈280 ≈15 none
O467, Persian compliments II.15, n. 11, 1728 Moréri 1717, which abridges Olearius ≈116 ≈19 none
O510, the Anca bird and Chederles I.1 §VII, n. 20, 1726 Moréri Suppl. 1716 (Bochart, Busbecq via Bayle) ≈115 0–6 none
C006, what different peoples eat III.10, n. 10, 1729 Lémery 1702 or its extract in Trévoux ≈79 inherited
O481, the women of Curzolari I.16 §VII, n. 47, 1726 Moréri 1717 (Graziani 1624) ≈75 ≈25 none
O476, diamond cutting II.2, n. 66, 1728 R. de Berquen 1661, Moréri Suppl. 1716 ≈70 0–5 none
O478, Ami Perrin I.4, n. 41, 1726 Moréri, which abridges Maimbourg 1682 ≈58 ≈20 inherited (Maimbourg)
O474, Poverty I.3, n. 38, 1726 Moréri, Paris edition ≈48 ≈40 inherited
O472, the god Terminus I.4, n. 2, 1726 Moréri 44 ≈14 inherited
O017, Villemot’s comets I.10 §IV, n. 15, 1726 Fontenelle, Histoire of the Académie 1707 35 (≈130–155 with nn. 13–14) 35 none
O512, councils against magic II.5, nn. 63–64, 1728 Thiers 1697, Moréri ≈31 0 below the threshold
C034, the modern medical sects I.5, nn. 18–21, 1726 Barchusen 1710 ≈30 below the threshold
O469, the mottos of Saint Malachy II.4 §VI, nn. 37–40, 1728 Moréri Suppl. 1716 ≈455 cites the work in n. 41, to criticize it
O234, Jacques Aymar’s rod III.5, nn. 17–18, 1729 Bayle, ABARIS, and Mercure galant 1693 ≈118 from the Mercure cites Bayle
O490, the Isle of Pines I.12, nn. 20–22, 1726 Moréri Suppl. 1716 ≈230 cites the work
O157, Agrippa II.5, nn. 23–27, 1728 Bayle ≈250 cites the work
O496, Nicolas Flamel III.8, n. 30, 1729 Moréri ≈110 cites the work
O255, Gómez Pereira III.9, nn. 11–12, 1729 Bayle ≈110 cites the author
C032, Homberg’s gold II.14, nn. 4–5, 1728 Histoire of the Académie 1702 and 1707 ≈120 cites the author
C033, Hecquet’s trituration I.6, n. 11, 1726 Hecquet 1709 ≈110–130 cites the author
C018, Duncan and coffee I.6, n. 13, 1726 Duncan 1705 ≈50 cites the author

“Inherited” means Feijoo only names the authorities the book he’s copying already cited (Suidas, Maimbourg, Lucan…).

Conclusions

We’re at a unique moment in history. This is probably the closest a regular person is going to get to the frontier of models, and the gap will widen over time, a consequence of scaling laws, frontier labs as geopolitical actors and AI existential risks entering mainstream discourse.

This whole post is an arbitrage between what models can actually do and what most people think they can do. There are many more arbitrages like this with much greater financial returns, and contrary to what I thought in 2023 (I was an avid LessWrong reader lol), they’ll last much longer than a technical person would think 6.

The Diffusion Gap: AI capabilities, adoption and opportunity


Nothing I did here is technically complex or complicated. My biggest hope with this is to get more people to put some of their time and tokens into this, if you’re interested please get in touch:) That said, I do worry that swamping historians with low-quality slop could trigger a backlash, like already happened with mathematics. While working on this, I tried to get in touch with several experts on the subject, most of them didn’t reply, and with some interested ones I’m still in contact, but it’s still an open front, since I wanted to get the post out as soon as possible.

Another important point is that this work depends on being able to access scans of the relevant documents. There’s so much left to digitize! A 2017 European survey estimated that archives had digitized 10% of their holdings, and libraries 17%.

Google Books did a lot for digitization. The labs are in a race to get more data, with very strong incentives. Projects like the digitization of the Boston Public Library’s collections, or the fact that Anthropic bought millions of books, scanned them and threw away the originals, give an idea of where we’re headed. As I already argued in how AI might help save copyright 3 years ago, intellectual property laws urgently need reform, information deserves to be free and the second-order consequences are huge.

That’s all, here’s the code and relevant files, I’ll edit the post if I find anything else related to this challenge. Thanks for reading.


  1. From his cell he measured the heat in Oviedo with a thermometer on the balcony, experimented with preserving tobacco and chocolate, used a microscope, argued that a kitchen helper they were trying to exorcise had epilepsy. Vida del Padre Feijoo en la comunidad monástica de San Vicente de Oviedo, 2017, pp. 7–9 of the preprint. ↩︎

  2. Other relevant names include Martín Martínez and Tomás Vicente Tosca. For context, Antonio Mestre Sanchís, Los novatores como etapa histórica. ↩︎

  3. He didn’t always escape censorship. In 1739 the Inquisition ordered two paragraphs of volume VIII to be struck out. ↩︎

  4. Rodríguez Cepeda’s 2008 estimate, in Inmaculada Urzainqui’s biography of Feijoo. It refers to volumes, not readers or copies sold of a single title. ↩︎

  5. When I did this work, Opus 5.5 had just come out and for me it was a notable jump in capabilities. I’m not sure I could replicate this post with Opus 5, I really hated that model, 0 soul. ↩︎

  6. The reason is a mix of Reality has a surprising amount of detail and underestimating real-world friction. Something being technically possible isn’t the most relevant factor for having an impact. I’ve had a draft stuck on this for several months, talk to me if you’re interested in the topic! On that note, Del rigor en la ciencia, by Borges. ↩︎