Sharing a personal essay about Standard Ebooks

[Weijia Cheng]

Dear fellow contributors,

I recently wrote a personal essay reflecting on my experience contributing to Standard Ebooks titled "Work can be joyous" and I thought that it might be interesting to many of my fellow contributors. I have been thinking about what it means now that AI coding assistants like Claude Code can potentially complete a lot (or even most) of our workflow, and in my essay I explain what contributing to Standard Ebooks means for me and why I think vibe coding SE projects would undermine my deeper purposes for producing ebooks.

I hope you'll forgive the self-promotion, but I'd like to think I've earned a bit of leeway after all of these years of answering questions and reviewing projects :)

Best,

Weijia

[Alex Cabal]

Thanks for sharing Weijia. I'm glad to hear our shared work has meant so
much to you.

I've also been thinking about how AI is going to impact our process.
People have reached out to me privately from time to time to offer AI
improvements to our process but none have impressed me so far.

In the long run, I think *some* of our work might be automated, but much
of it cannot. As you note, a lot of it depends on taste: selecting a
cover, deciding how to organize compilations/corpuses, deciding on some
edge case in the text.

And, a lot of literature doesn't fit neatly into the kind of "plain
prose fiction novel" that AI will have trained on. Last year someone
tried to pitch me an AI spelling/grammar checker for use at SE, and as
an example they sent me the its findings for _The Sound and the Fury_...
needless to say, both the person and his AI were hopelessly misguided.

Then, an AI's context window matters, at least for now. How can it be
expected to keep _War and Peace_ in its context window as it hunts for
typos, or whatever?

Finally, SE's process is oriented around precision and accuracy. AI
output is the opposite of that, and that doesn't seem to be changing much.

So, I don't see AI *significantly* impinging on our work, at least in
the mid term. Of course, with things progressing as fast as they have,
who knows that the 10 year term will look like!

On 4/15/26 12:32 PM, 'Weijia Cheng' via Standard Ebooks wrote:
> Dear fellow contributors,
>
> I recently wrote a personal essay reflecting on my experience

> contributing to Standard Ebooks titled "Work can be joyous" <https://
> www.aiandour.faith/p/work-can-be-joyous> and I thought that it might be

> interesting to many of my fellow contributors. I have been thinking
> about what it means now that AI coding assistants like Claude Code can
> potentially complete a lot (or even most) of our workflow, and in my
> essay I explain what contributing to Standard Ebooks means for me and
> why I think vibe coding SE projects would undermine my deeper purposes
> for producing ebooks.
>
> I hope you'll forgive the self-promotion, but I'd like to think I've
> earned a bit of leeway after all of these years of answering questions
> and reviewing projects :)
>
> Best,
> Weijia
>

> --
> You received this message because you are subscribed to the Google
> Groups "Standard Ebooks" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to standardebook...@googlegroups.com
> <mailto:standardebook...@googlegroups.com>.
> To view this discussion visit https://groups.google.com/d/msgid/
> standardebooks/e3268f02-5865-4f65-8434-665614ee27f5n%40googlegroups.com
> <https://groups.google.com/d/msgid/standardebooks/
> e3268f02-5865-4f65-8434-665614ee27f5n%40googlegroups.com?
> utm_medium=email&utm_source=footer>.

[David]

Nice post, Weijia! I enjoyed the read, and left a comment – not as thoughtful or apposite as Alex's here, though! :)

[Asher Smith]

This is right at the intersection of my interests, Weija, though I have historically come at it more from the angle of the philosophy of work. Thanks for sharing! Alex, I agree about your assessment of the technical state of play; I have run across entire missing sections in multiple productions, and cannot imagine the sort of context window with today's technology that would be able to hit something like that, figure out that what needed to be done was to find other transcriptions and compare them, and then be able to do so.

I have a few thoughts I'd add:

• The existence of Standard Ebooks and similar volunteer-driven, open-source projects gives lie to the notion that people will only do labour for monetary gain, and that technical problems can only be efficiently solved by people in competition with each other. You quote Kropotkin in your essay, and I think that the anarchist position is incredibly apt; in conversations about the possibility of anarchist societies (or even things like the effects of UBI), I've seen it argued that coercion by threat of poverty is necessary to motivate creation, but SE demonstrates that people will do large amounts of painstaking labour by which they will only enrich themselves in their position as just one more person who benefits from the public domain.
• With the rise of LLMs and the threat of automation of knowledge work, I'd argue that it's important to distinguish what knowledge work is meaningful in and of itself. Those places are where the performance of that work becomes virtue ethics, as you touch on. Proofreading, tedious though it may be, falls under that banner in my estimation because it is a way of engaging directly with a piece of writing; I'm reading Pygmalion at the moment, and have found that I am reading it much less deeply than Shaw's other plays which I read in the context of producing for SE. I'm happy to automate plenty of the initial formatting,  but I want to automate the act of "engaging with art" not even one bit. The "work" of that frequently is joyous.
• I have over the past several years been continuously improving the scripts and tools that I use when I refresh the SE catalogue on Apple Books; at the start, I did a lot of work on it and developed a deep understanding of the technical standards involved, and I found that the act of developing rather than merely using tools improved my technical abilities. Yesterday, I plugged one of my scripts into an LLM and had it modify it to be better parallelised. I do not have the understanding of parallelisation tools that I would have if I had done the dev work myself, but also I wouldn't have had the time to do that work yesterday and so that trade-off - my knowledge in exchange for my convenience - felt like a worthwhile compromise yesterday. I'm not sure if I will always feel that it was the correct decision that I made yesterday.
• My understanding of the SE mission to produce "lovingly formatted" ebooks is that the "lovingly" part is important, and that implies a person doing the loving. I have seen unaffiliated people talking about the project online note that the quality of the collection, not just the size of it, is a strength, and I suspect that simply turning an LLM loose on the PG database would dampen that feeling. My general opinion of LLM-generated novels is that if you couldn't be bothered to write it, why should I be bothered to read it? The corollary of LLM-produced ebooks is that if you couldn't be bothered to read and edit it, why do you think it's worth preserving at all?

[C T]

I had some success doing a few specific tasks with Claude and GPT 5.4 (via MS Copilot harness) when I did The Cruise of the Nona. 

For each task, I included in its context the appropriate page from the style manual, and I had it iterate over groups of lines at a time. It was fairly useful when identifying things that need semantic tagging like book titles and ship names (it's pretty good at looking at the surrounding context and suggesting changes). It also cataloged all of the various possible changes I'd need to make regarding switching italic tags to quotes, etc.

Of course I did a read, but it was useful to have that as a backstop, because even after my read it found one ship name and a couple book names that I had missed in the reading.

It also identified a printer's error through its internal "compressed wikipedia" where it identified that the name of a river was misspelled, and was useful (again, by iterating through 50 or 100 lines at a time) in checking any possible spelling concerns.

I would say it's not a "game changer" in a deep way, at least yet. In a limited fashion, as another tool in the toolkit, it helped a bit. You need to be explicit with it, and even then it requires multiple passes, and likely even more care with prompt design, to have to capture everything. 

Even after all of that, Emma still caught a missed song name semantication.

CT

To unsubscribe from this group and stop receiving emails from it, send an email to standardebook...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/standardebooks/aca2448d-1999-4848-a0ad-d97e0fb62a5e%40standardebooks.org.

[C T]

Also, a very nice Essay, Weijia. I too have been reading A. Mac and Ellul recently.

--
You received this message because you are subscribed to the Google Groups "Standard Ebooks" group.
To unsubscribe from this group and stop receiving emails from it, send an email to standardebook...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/standardebooks/e3268f02-5865-4f65-8434-665614ee27f5n%40googlegroups.com.

[Weijia Cheng]

Thank you all for the comments! I'm glad that this resonated with people and has sparked some discussion.

One of the areas where I think it could be fruitful to explore the use of generative AI more is the transcription stage. There are some powerful OCR models these days which are capable of directly converting image scans to Markdown format, preserving some of the structure and semantics that are lost with older OCR tools like tessaract. From there, an LLM could be used as a basic proofreader to iron out basic transcription errors. My general experience as a producer is that as I've become more experienced, I've become less and less reliant on the quality of the original transcription (although obviously it is easier to work with a good PG transcription when possible). There are lot of interesting books out there that, for various reasons, no one has cared enough to transcribe, and generative AI could short-circuit a lot of that work. Let's say that on a 1 to 10 scale, a good PG transcription is 10/10 and raw tessaract OCR is 1/10. If the OCR model + LLM pipeline allowed me to generate transcriptions that start at 6/10, that would give me a lot more options about what ebooks I could feasibly produce.

[C T]

Weijia, if you're interested in OCR and transcription, I am actively working on a little project to use PGDP's "ground truth" text (available by page via API) to train recognition, detection, and structure ML models for OCR (to help PGDP, but also can help SE). The DocTR guys have built an open source library that has provided an excellent start and I've had pretty great success getting *much* better results than Tesseract after training on a few hundred labeled pages.

https://github.com/ConcaveTrillion/ocr-labeler (GUI mostly for Labeling and Tagging)

https://github.com/ConcaveTrillion/pd-book-tools (Python Library)

I intend to also build a GUI primarily for "transcription generation" based mostly on the labeler GUI.

It's heavily vibe coded, especially the UI (I started manually with a python notebook but the UI got far too complex, and I'm a data/backend engineer mostly).

To view this discussion visit https://groups.google.com/d/msgid/standardebooks/192248d9-2d2d-47db-a554-cd470da145b6n%40googlegroups.com.