[lordgrenville]

While it's really cool that this is possible now, this project seems to be a thin wrapper around running Claude. Presumably, someone who wanted to read one of these books could just run it through Claude themselves with comparable results.

Sharing some artifact of your own added value, or things you learned in the process, would be more interesting to me than the output.

[sweisman]

I can assure you this is not a thin wrapper. Even the public repo of the first workable version of it is not a thin wrapper.

You are welcome to try. It is not rocket science, but what I developed would take much effort to reproduce.

I wrote an origin story about how I got started: https://aicentral.substack.com/p/pipeline-to-the-past

It describes some of the added value and lessons learned.

[mmooss]

Also, people can look at the Github repo of your prior version and see it's clearly not just 'running it through Claude'.

https://github.com/sweisman/translation-pipeline

What I haven't found in either the OP or Github repo is what your background is - software development? astrophysics? paleography? two or three of those? [edit: history of science?] - and what human review is involved. (My apologies if I overlooked anything; I don't have time to read it all!)

[sweisman]

Software development. I did work in an astrophysics lab decades ago, as one of my public repos documents.

[us-merul]

I think there's more to it than that. It's a combination of extracting the text from the PDFs, resolving persistently tricky issues (especially numbers), and flagging issues for human review at scale. You can view lessons from the author on the GitHub repo: https://github.com/sweisman/translation-pipeline

[lordgrenville]

Great, then rather post that! Although that repo looks like 100% AI artifact, not human intervention. (What does "flagging issues for human review at scale" mean in this context?)

[us-merul]

I was referring to this bullet under Lessons Learned: "Manuscript translation is feasible but rough. A 14th-century Sephardic manuscript hand is readable at ~70–80% accuracy. The output is a useful rough draft, not a polished translation. Plan for a longer discrepancy report and more editorial review." Some of the samples I looked at had question marks or original Hebrew text throughout, which I took to mean marked uncertainties.

[sweisman]

Manuscripts are doable but tougher than printed works. Hebrew likewise is tougher than Latin.

[sweisman]

This is not what I use to translate. It was the first version to create usable output, and intentionally implemented in a code-free manner (today I have an extensive body of code in addition to the Claude Code harness instructions). Read my origin story, linked to in another comment, for more background.

[lordgrenville]

Reiterating what I wrote above, sharing these would have been more interesting imo than sharing the pipeline output.