Your comment persuaded me that just explaining my prompt methodology is insufficient.
With Claude's help, I ran some tests against nine different translations including Peter Green's line-matched translation, the closest method-matched human control. 36 human-vs-human pairings as controls.
Murray (the published translation of the Greek, Loeb source) and Lattimore affinities are genuinely elevated. Fagles is modest but not totally absent. The test can't exclude 4-6-word mosaic composition, and doesn't measure interpretive influence.
The longest shared passage in the 12,107 lines is 29 words. Long-fragment assembly is strongly disfavored by synthetic-splice calibration. The translation resembles Green as much as it resembles Murray, or three times as strongly as it resembles Fagles' famous one. That strikes me as being in the opposite direction from what training data-regurgitation would predict.
I love Wilson's work and take it as inspiration. Incidentally, I spotted an error in the first printing of her Odyssey that the publisher acknowledged on twitter and fixed in subsequent printings. I am not a professional, just an obsessive with a special interest. In a spirit of intellectual honesty, the fact that she is reportedly re-translating The Odyssey again is an indication that one of perfect adequacy might be elusive.
I find it to be cross between Richmond Lattimore and Ian Johnston. There are line-level annotations that discuss translation choices and a hyperlinked index w/ Latinate pronunciations was also built from scratch.
The only thing Fable was given was a clean of copy of the ancient Greek. This is not a bastardization of other translations. I appreciate and value the critical feedback in any case (truly).
I find that Fagles is very majestic -wonderful, really- but he's also very liberal (poetically) and definitely did not limit himself to the same number of lines as Homer as this project did.
> The only thing Fable was given was a clean of copy of the ancient Greek. This is not a bastardization of other translations
FWIW I gave GPT-5.5 two lines of William Cowper's translation (long out of copyright to avoid any filters) without web search and not only was it able to identify it, it was able to continue it verbatim for roughly 30 lines at which point it went off course, but only by skipping like two books ahead. If I had supplied the original text as an anchor it presumably could have kept going.
I didn't try it with Fable because I'm lazy, but all these translations (presumably the ones within copyright too, given the lawsuits) are definitely well known by these models.
As another comment here points out, Fable likely had multiple translations, academic literature about the text, and maybe the original as well, in its training set. So even if only the original Greek was given to it as input it probably leveraged prior art when translating it.
No. I am saying that I had Fable do a literal word for word translation, and then did two independent review pass translations with models of different lineages and then had another adjudicate it. I am just saying what I did, that is all. It occurs to me as I am typing this that a human translator working on an Odyssey translation today would have similar "prior training," culturally. It's one of two of our oldest stories.
Yes and if a human translator produced this work, it would be viewed as near plagiarism of Lattimore and Fagles. Compare this to Wilson. Wilson had surely read Lattimore and Fagles and yet produced something genuinely different (as much as I don't personally care for her translation). Claude just regurgitated others' half-digested work.
12,107 line-for-line translation of the A.T. Murray Greek text (Loeb, 1919) by Claude Fable. Includes line notes and an index. HTML, epub, pdf, and YouTube Audio (by gpt4o-TTS). A dual-agent review pass was done on the entire work with a third "adjudication" agent reconciling their findings. Reviewer agnets were gpt-5.6-sol high reasoning.
I love the poem and wanted a modern translation with annotations and index, etc., to be freely available to anyone who wanted to read it. I had a sense that Fable would do a really good job. I am pleased with how it came out. The review agents didn't change a single word of the translation, only some of the notes.
One very simple benefit is that it's unencumbered by copyright.
I do value the work of human translators; when they're really in harmony with the original work, they can be fantastic. But they can also be disastrous, like Alfred Birnbaum's translations of Murakami, which radically reshapes the protagonist and threw out nearly 20% of the original work.
If the Murakami was undertranslated then how will having more translations that crowd out the potential revenue of future translators help produce better quality translations?
The problem is that (we have every reason to believe) models like Claude were trained on existing translations. Now they're being used to generate things that are direct replacements for those translations in the marketplace.
This is basically an anti-AI author or lawyer's dream case.
And, seeing this uncanny-valley-of-slop output, I kind of hope they go for it.
Lots of commentary from an unnamed White House source characterizing Anthropic as being "unserious," without further elaboration.
I want to point out that when the commentary by export-ban apologists devolves into talking about perceiving Anthropic as being "leftist," they're not having safety or even a technology-centric conversation anymore. They're making a baised political statement of intent about seeking domination and control over technology for capricious, arbitrary and self-interested reasons.
If you read a lot of books, particularly older ones, you'll find em dashes in all kinds of writing and used often. It's functional punctuation that once you understand you may even find yourself using it (and then being accused of being an AI, lol)
With Claude's help, I ran some tests against nine different translations including Peter Green's line-matched translation, the closest method-matched human control. 36 human-vs-human pairings as controls.
Murray (the published translation of the Greek, Loeb source) and Lattimore affinities are genuinely elevated. Fagles is modest but not totally absent. The test can't exclude 4-6-word mosaic composition, and doesn't measure interpretive influence.
The longest shared passage in the 12,107 lines is 29 words. Long-fragment assembly is strongly disfavored by synthetic-splice calibration. The translation resembles Green as much as it resembles Murray, or three times as strongly as it resembles Fagles' famous one. That strikes me as being in the opposite direction from what training data-regurgitation would predict.
https://theclaudyssey.com/independence.html
Methodological corrections welcomed