It’s been a while since I did one of these (#125), and I previously always did my AI lore books’ “notes” for each individual volume. But times have changed, and that’s not how it worked out this time.
These are all the books I was able to publish since a few days I’ve been on this current trip:
- 126: Starship Pop-Up
- 127: Onboarding
- 128: Weak Password
- 129: Saint Mercy Gardens
- 130: Mysterious Data Centers
- 131: Wormwood
- 132: The Time-Emitting Animal
Briefly, I’ll run down each one in terms of what it’s about:
- Starship Pop-up – This was an idea I had right after #125 more than a year ago, but never got around to it til picking up my AI-sword again sometime last week. It’s basically Star Trek, but imagining that their on-board computers are as shitty and riddled with ads and subscriptions and other bullshit as the modern web is.
- Onboarding – This is basically bureaucratic dystopian job-hunt fiction, which, in my experience, is how almost all job hunting is now, give or take. It’s set in the “Conspiracy Dude-iverse” of the Topia series, but the tone is a bit off and strange relative to the more manually-written installments in that series.
- Weak Password – Basically a meta-reality take on how annoying it is when Zoom or whatever other app constantly tries to get you to log back in, authenticate, blah blah blah… the modern tech hell we live in.
- Saint Mercy Gardens – This one is based on two different premises that ChatGPT came up with in “Work” mode (more on that later), which I had it combine together to run through my pipeline, which is my main focus right now in getting back into all this.
- Mysterious Data Centers – I’ve seen tons of conspiracy-adjacent content opposing data centers on mainstream platforms like Facebook groups; it seems to really be fueling the public imagination. So I refined my pipeline to be able to create this “mode” (one of four identified in the entire series so far) of writing as well, which is highly expositional quite different from more character driven pieces.
- Wormwood – Wormwood (the “Changer”) is one of the most important Quatrian “powers” and I am happy that I was able to put together such an intriguing volume exploring this energy.
- The Time-Emitting Animal – This creature, more formally known as a mnemani is one that has been haunting my art and fiction for a few years now, and has taken a number of different forms. It appears in a couple other books in the series (Timehunters, and Order of Chronos, I believe) but this book, told as a recovered manual for keepers of this creature (who ostensibly use it to harvest its liquid time secretions), takes it in a whole new direction. There is something really intimate and interesting about the beast and the relationship portrayed here (textually and visually) that is hard to put into words, and strangely emotional and deep.
I got a fire lit under me again about these books due to an interview that might come out hopefully toward the end of the month. After reviewing the terms of the Gumroad storefront, I decided to split for hopefully greener pastures. I had to migrate all the old titles over to Payhip URLs. A bit dodgy without proper forwarding, but hopefully the trail of notes I left on Gumroad will be helpful to anyone looking.
In any event, I had to go through and manually update all the interior URLs linking out to other volumes in all 128 prior volumes. This got me thinking again about automation – which is something I would not trust to do all that fiddly updating I did by hand just yet. And by the time I was done, I was raring to go to try out some new volumes using a free offer for one month of ChatGPT Plus, which I hadn’t subscribed to (paid or free) in quite some time.
Turns out ChatGPT just recently released “Work” mode, which interfaces with Codex, and there are a host of new ways to use the tools that I’m still only just beginning to explore – though I’ve learned a lot in a short time…
My initial idea was to use “agents” to write blind experiments as modified surrealist exquisite corpse writing experiments. I had the idea initially of like okay, well, here’s a writer agent, an editor, and then a counter-editor. Then I was like, well what if between those roles, there are “wreckers” that come in and materially distort and “mess up” the work done prior to them. And to whatever degree, the wreckers would be secret, something something. It was a half-baked idea which, once I had ChatGPT save it all (quiet awkwardly I might add) into a config markdown file of some sort, yielded the first few books in this latest batch: at least Starship, and Onboarding. The project was also fed the text of previous lore books and another set of old articles from EC to pull from as background.
And those first couple manuscripts generated that way were for the most part pretty interesting, though they required some light-moderate editing. But then I realized somehow or other, when I saw some other little animations running in the ChatGPT UI that perhaps the system had been completely shitting me prior to that about having actually run processes separated by agentic magic. I think it just emulated the whole thing and did not tell me. And since I didn’t realize at first, and the results were basically within range of good, it seemed like it fooled me.
But then like I said, I saw other processes using like agent icons, and realized I’d been had. Probably. Maybe. The thing is, you kinda don’t know. It’s a *really* tricky and weird system, and you have to kind of wing it.
So over the course of the next volumes, I kept tinkering with all of that, and whether getting “agents” to run actually would even improve anything. And I saw how resource heavy the Work/Codex mode was, and how quickly it would blow through my 5h usage limits. I eventually changed the goal, scaling it down to fewer passes (each generation pass that it does on text is significantly slower than what you get out of normal Chat mode, where usage is not metered so strictly or in the same way – highly confusing, I know). And over the course of several books, I eventually settled back down to just having one single pass that covers all the different elements of what I want accurately enough to yield usable results. So, ultimately, emulating is more cost-effective.
The other big arm of this work, of course, was refining my config files, which I came up with memorable names for (ALPHA, ALPHA+1, ALPHA+2, WORMWOOD, PYRAMID, LEAN PYRAMID, WILD LEAN PYRAMID, PASTA SAUCE, MYSTERY SAUCE… etc). What I found in that process was that kooky names work and give you something to latch onto in a chaotic environment, and refer back to.
There’s a period in the older series around Shelvin Parz, and Tree & Leaf Law where I tinkered with using ChatGPT to build a rudimentary interface to the API so that I could have structured outputs based on some form elements I would fill out. It kinda worked, but wasn’t exciting enough without adding more pizazz. But the technology on both the coding and the writing side has improved a lot since then (I started the earliest of these AI volumes way back in 2022), and what we have now with GPT6 Astra and Codex/Work, and all this other bullshit is qualitatively different…
And I made an important and interesting discovery: ChatGPT Work/Codex has something called App Server, which evidently means that you can use it as basically the backend to support a small simple web app that runs locally. Probably it works in other ways as well, but that’s how I use and understanding it (sort of). So, instead of hitting the conventional API and incurring costs per use, using App Server just uses your existing Codex usage limit.
But I should qualify all of that, and back up a little bit: seeing how easy it still is for Codex/etc to dupe me about agents, or about the various passes that it was claiming to have done, I realized that having the system be a bit more strongly regularized would be a good thing. It would result in a more repeatable process with some fixed inputs and outputs instead of a conversation that drifts and drifts and drifts, like it would otherwise.
So, I had Codex build out a webpage of app UI that runs locally in HTML/CSS/Javascript, and then through “magic” connects on the backend to Codex App Server. I can iteratively describe changes to Codex right there, and it updates and refreshes the page for me (or I can run it in a browser apart from that). And I can input my new book title, and premise, target length, and some other variables I’m experimenting with around whether to reference existing sources, and which type/mode/voice book this is going to be…
And then it basically… just works. It’s slow, and I have to frequently call for it to get the usage limits remaining. And I’ve blown through them easily many times in a row, but it gets the text generated in roughly the approximate shapes I want.
I’ve been able to devise other concrete parts of my book pipeline, which have not yet been integrated into the local web UI. The next big piece after that is exploring visuals for the book, which we (Codex and I) decided to call “Visual Ecology” where it sort of roves through visual subjects in the book across a series of different styles/media/etc. But again, you’re making images in Codex, not in Chat, which has different and more severe usage restrictions. What I usually do there is watch live while it’s producing a stream of new images, and I give it instructions to steer by as it continues.
Again, like the text generation stuff purely in a chat, without more formal structure, it becomes a pain to piece together after the session if you do multiple replacements, or extra visual requests. Which is why it would be nice to integrate that directly into my local app UI (the whole of which I’m calling “Pressworks”), to have there be a cleaner consistent presentation of all the associated pre-assembly book assets in one place, and I could accept/reject on a per-image basis in a more easily trackable spot.
I haven’t gotten there yet, but I will. Provided I don’t explode all my usage limit for the week(s) and month. I’ve not been using this that many days, and already had to use one of my two weekly resets; the usage restrictions are pretty tight for getting real substantial work done on a Plus-level account, and everything seems to be strongly pushing you towards moving up to a much more expensive ($100-200) Pro account, with elevated usage limits.
Anyway, after image production and selection (I’m just doing around 2 or so images per chapter on average for most of these, to keep volume low and reduce usage), I have the system do text mapping, where it decides locations to insert images into the text. Then it does document assembly, where it takes the generated first draft text, and the mapped images embedded into it, outputting a .docx file. I can then open this file up as an import in Vellum, which is where I do my ebooks. And I can update the front and back matter from other volumes in the series, and go directly to editing the text for its final version, and preparing the new product page on Payhip.
I don’t measure any of this anymore in “time saved” – at least not relative to any prior versions of how I’ve approached book production. Yes, overall time per title is probably greatly reduced in many cases with the new pipeline system and tooling I am carving out. But to figure out how to build it out, to run all the tests, and stumbling along through the tools still takes time enough on its own. So, okay, perhaps time-per-title is down, but time-per-tooling is way up. Though, that said, if I can solve that tooling stuff to a certain level, it will make the rest down the road (potentially) somewhat smoother sailing.
Toward the ambition of smoother sailing, I’ve started looking at how could I pull the generation tasks I’m currently using Codex App Server for out of there, and hand them off to other models. I haven’t had a successful run with it yet, but ChatGPT turned me onto LM Studio, which has a program you can run on desktop, called, Bionic, where you can download and run free open source models locally. And not only that, you can expose a local server to access those models from other apps. So I’m trying to have my local web UI (“Pressworks”) have a drop-down to select one of those locally-installed models, and pass at least my text generation tasks to those tools, and pull its responses into my local app to continue processing. Presumably there’s a way to do something like that with images too, though I’m not sure yet whether Bionic or a different tool would power it for my purposes.
One experiment I was trying to do was see if I could use Codex to automatically open up regular Chat mode, and pass Visual Ecology prompts into it, just using its Browser skills (where it emulates a human user using ChatGPT.com on its web browser). The idea being to make use of Chat mode’s greater seeming allowance for image generations relative to Work mode. I got it to create the test image over in Chat mode, but could not yet get it to bring that image back into my app. Maybe there’s a way, but if nothing else, it seems to mean I could at least have it create all the images for a book title in a single dedicated thread. I’m not yet sure that’s “better” and further testing will need to be done around consuming usage under that model. Likewise, it seems like I could do much the same for having it pass certain of its text generation functions over that way – if if if I can get the handoffs to work.
It’s been tricky and weird so far, and is absolutely a mindfuck to work this way, but there is a lot that is happening for me with it in a short time that is nothing short of revolutionary. And it’s making me strip away regularly my assumptions as I chase the White Whale here. Meanwhile, along the way, I’m at least getting some cool new book volumes out of it, whether or not I will ever quite arrive perfectly at the destination I am seeking.
PS. There is a lot of really impressive visual art coming out of this new image gen upgrade in ChatGPT as well. Some of the stuff in the new books is really amazing, and I’m really pleased with it. And lots of compelling stuff in the writing too, styles and emotions that I never hit in any prior models. Pretty wild process to see so up close and personal, let alone share anything from in a meaningful form.
