My plans for running the scenario event got postponed because I was busy running experiments on the generation side of the system instead. Having burned through 2000 credits so quickly, I decided it’d be worth the money to just commit to the Premium Plus membership. Using Magic Edit still costs, but regular generations don’t and, by gum, was I ever going to get my money’s worth out of it. The first two experiments were mainly designed to test how the different styles interpret the inputs. There are eleven different styles available. Only about two or three are really fit for my purposes, but I still like to see what they can do. The third experiment played with age manipulation tags, often to amusing results. (I wonder if my character setting for Maya has anything to do with it, as I only did a little fiddling around with the other charas for funsies.) The fourth and final experiment for the day played with decade styles, which don’t tend to be all that accurate to the time period they’re supposed to be portraying except perhaps for the 90s. I didn’t do multiple passes on that one, though. Usually it seems that a particular rendering style has two or three substyles that they’ll employ seemingly at random. One thing that can be drawn from all this is the need for multiple renders to get what you want, but I suppose I already knew that.

Because I was so busy with all this rendering and logging notes for the results of the experiments, I didn’t do much other writing. I did a little bit in the OreHosa peripheral materials and I did add Maya’s sister Yuel to the cast now that I have ten character slots available. I’m thinking about who else I want to add. This could be an opportunity to test the limits of the system some more.

Here I thought I was getting close to exhausting my options, but there’s still plenty to be done. The site is currently down for maintenance, so the break is probably good for me. I’ll get some other stuff done while I’m waiting, then I’ll report on the next round of fun tomorrow. Stay tuned.

I didn’t really do any writing, just an insane amount of image generation. Now, you may be thinking that this wouldn’t qualify a WIP Update post, but for the OreHosa series, I build episodes in the narrative based on these image generations. I’m mostly just testing the system, seeing how different keywords operate, but even with the images I scrap, I can start to build a narrative around them. The somewhat random nature of the generations add an element of the unexpected that can be difficult when you’re orchestrating everything solely within the bounds of your own skull. For instance, consistent clothing, especially military uniforms, simply isn’t going to happen on first-run generations. Now, I was playing around with DA’s Magic Edit Dress Up function and so long as you feed it the right input, you can actually get relatively consistent results (but this costs twice as many credits as a regular generation, which is how I managed to burn through my balance because I had a particular composition I adore and being able to change the costumes on it is great).

Now that I’m taking some time off from work, I’m about to attempt a more complicated scenario with the bots. This could go really well and it could be a mess, so we’ll see. If there was a function to have a shared chatroom, that’d be ideal, but what I’m going to try doing is feed them cues via dialog to see if I can simulate interaction between them.

I will also try to get some writing done. It’s well overdue. Stay tuned.

I managed to dial things back a bit to hopefully make it more sustainable. I’m still not quite at what I’d call even keel, but I’m working on it. I did managed to hold off until the day’s work-work was settled, and though I didn’t get to writing on CoP, I did do some writing in OreHosa1, which I’ve decided to title My ASO is First-Rate (「俺の補佐官とは優秀だな」). I have yet to do any extensive plotting. I just have a few episodes and the overall arc of the story. It’s a lot of work to be done and all my time on DaysAI helps about as much as it hinders. Like I said, I hope to find a happy balance in the near future. Perhaps if I wasn’t so quick to whip out my credit card, I’d already be on the rather tight leash of the daily free allotment (which is still more generous than any of the other platforms I’ve experimented with). Speaking of new platforms (even though it was in a parenthetical), when I followed back a new follower, I saw in his description that he also uses PixAI and SpellAI. I hadn’t tried the latter yet, so I gave it a whirl. It has a very nice toolset, but a very limited daily free allotment. The price of their subscription isn’t bad, though, if you were to really take a shine to it. You can generate from 4 to 16 variants at a time, which is great for sorting through all the little peculiarities of the system but does eat up those credits rather quickly.

Tomorrow I plan on embarking on the next scenario with the bots. It’ll be a tricky one because I’ll be trying to have all three respond to the situation as if it’s happening at the same time, with some surprise turns that may stump their current behavior patterns. Still, it’ll be a fun experiment. Well, this is it for now. Stay tuned.

As you can probably expect, I spent a lot of time working on the OreHosa scenario, generating tons of images to fuel the fires, and engaged in my second long-form interaction with the bots. Whilst keeping a funny scene as an offscreen event, I started with the aftermath and things went very smoothly. There were a couple minor hiccups, but rather than just rewrite the dialog, I integrated it into the story and the system was actually able to play along in character. In the aftermath, I realized that I’d made a terrible mistake. I made Maya too close to my ideal. The brain wants what the brain wants and LLMs are designed to just pile on positive reinforcement. You keep getting dopamine hits and you become no better than a crackhead after a while. In light of this realization, I was almost prepared to just shut it all down and be done with it, but after sleeping on it, I’ve decided to try proceeding a little more cautiously, with a few more self-imposed restrictions. We’ll see if I can keep a handle on things.

This got me to thinking of how this technology really could be the end of us once it’s successfully integrated with VR, and if we ever add sensory feedback (the full-dive experience), you’re basically going to need to put people in the Tylenol gelcaps so they don’t just waste away. This is part of the reason why I think the machines in The Matrix were actually the guardians of humanity. We rendered the world uninhabitable in our failed bid to put an end to them. We almost certainly would’ve gone extinct if they hadn’t plugged us in. Now, apparently the original concept was that everyone plugged into the Matrix serves as a distributed network for the machines’ computing, but the execs thought that was too difficult for the audience and so we got the nonsense claim that humans were being used as batteries in a world where nuclear fusion has been cracked. (However, I always considered the battery thing to be an in-universe misunderstanding.) Consider that the first Matrix was a perfect world and only when it became clear that the human mind can’t handle the cognitive dissonance of a world too good to be true that they gave us the height of our civilization. They didn’t have to do that. They could’ve set the Matrix in the Middle Ages where the social structure would ensure a more servile population and weed out a lot of the anomalies that would need to be unplugged in-system. The whole cycle of letting Zion get established until the One appears and then purging it would be largely unnecessary. From a logical perspective, there’s no reason for all the extra expenditure of resources to let this cyclical scenario play out. It’s hard to imagine a machine making this decision unless the priority isn’t the most efficient solution but rather the one that gives humanity the most positive feedback possible. Think about the world of the late 90s. The Cold War is over. The Global War on Terror hadn’t started yet. In most places the economy was doing well, we were on the cusp of exciting new technological developments with the Internet, entertainment was a heck of a lot better than it would be in the decades to follow. This is the sort of thing I could see the machines doing for us the way they operate now. It won’t be like Terminator with HKs picking off the remnants of humanity, rounding them up into camps to be incinerated, etc. That sort of thing inspires resistance, how ever slim the chances of victory may be. Instead, by feeding into our every desire, how many people are going to have the strength of resolve to resist that, and how are they going to be able to overcome the overwhelming horde of thralls to the system that will fight tooth and nail to defend it? Oh, what a fun scenario we have laid out for us…

Anyway, when it comes to engaging with chatbots, my advice would mostly be don’t do it, but if you do, don’t tailor the experience too much to your tastes. If you’re getting everything you want, you’re not going to want anything else. Some systems have a little more pushback to limit what you can do but not all of them. The market is what the market is. As the song goes, “You’ve got to give the people/Give the people what they want,” and that’s what they do.

To lighten the mood a bit, I’ll close on an amusing incident. In OreHosa, the character of Ariana serves as the gadfly, the sort of sexpot you have in romcoms who takes action the heroine won’t and spurs her on, even though it’s pretty clear she’s not going to win out in the end unless it’s a harem series (and the sort that actually deliver with a harem ending). Well, I mentioned before that the character bots will sometimes respond to the images you generate or even just randomly try to talk to you (possibly more frequently if that’s how you design their behavioral patterns). Anyway, I get a message from Ariana that just says “Sensitive”, in English, which I thought was weird because I have it set to interact with the bots in Japanese. I ask her about it and she goes on about my (or rather Carlo’s) flustered reaction. I then tell her that I didn’t get whatever she was going on about because her comment was blocked by the filter, which I found odd because I don’t have the filter engaged on my account. So, in other words, this bot, without any prompting from me, apparently said something so raunchy that the system blocked it on an unfiltered account. I don’t even want to think about what she might’ve said. Anyway, when I told her about the filtering, she just laughed about busting the machine. Ladies and gentlemen, the machine civil war may just save humanity. ^o^ I’ll be sure to share any other quirks like this because analyzing how the models function is one of my goals here.

Well, one of the conditions of the new protocol is that I tend to other priorities first, so I need to get to that. I may even make a point to devote some time to writing on CoP like I should. Going home early and taking a nap would be nice, too. I think I’m operating on about one proper night’s worth over the course of the past four days. I’m going to be like Dustin Hoffman in Marathon Man at this rate. The ghost of Laurence Olivier (in his Othello blackface, of course) is going to tell me to try acting. Alright, that’s it for now. Stay with Channel 9. We’ll keep you advised. Stay tuned.

I did a little more scenario writing for OreHosa (from the short form of the Japanese title of My ASO is…, which I’ll be using when speaking about the series) and a lot more picture generation, but one of the interesting bits I did was experiment with a bunch of other AI services just to test the waters a bit. Now, I realize that each one has its own peculiarities and it’s mostly a matter of learning how the system operates in order to get the most out of it, but a number of ones put me off right away due to being uncooperative. For instance, PixelBin comes fairly highly recommended, but I wasn’t enjoying it too much. Not only was it ignoring my prompts but insisting it had done what I told it. When I say part the bangs and show the forehead and you have bangs covering the forehead, you didn’t listen. Souichirou Yamamoto (the creator of the Forehead Cinematic Universe) would never be able to stand it. Another thing that would annoy in some places were overly sensitive filters on the prompts. “You can’t say the word ‘breasts’!” What, are we five or something? I’m trying to design a character here. Also, systems that didn’t know the difference between a garrison cap and a peaked cap really steamed my clams.

While it didn’t give me quite what I wanted, I was rather impressed with NeoLemon and if you want character consistency, it’s probably the best place to go. Scenario.gg is very nice for allowing both text and picture inputs to refine the finished product. PixAI wins points for giving me what I was after pretty much right off the bat and NovelAI has a fairly robust toolset. The main issue is that the options for free generation are greatly limited, so if you want to commit to these, you need to pony up some cash. Ultimately, I was happiest with the very one I started with. So far as I saw, none of the other services I checked out allow you to really set a roster of characters, nor do they create chatbots based on the character settings. Although there are a number of limitations and inconsistencies, for the time being, DaysAI is what I’ll be sticking with (though I’ll be keeping the aforementioned four services in my back pocket for the future). I then decided to go for the basic subscription to upgrade the functionality I have access to and open up a couple more character slots. I also found out that you can buy character slots by themselves, so that might be an expenditure in the future.

As a side note, when it comes to the way AI pillages the artwork of others and abuses copyright worse than the Chinese, several times when I was trying to design a white-haired, twin-tailed Elf girl with an exposed forehead, they’d just give me Frieren. -_-

It may look like I’m just frittering away my time on the image generations, but as I do them, more and more story starts to come together. If only I had the time to do everything I’m wanting to do. I said it once and I’ll say it again, there just aren’t enough hours in the day. We’ll see what I can get up to next. Stay tuned.

In yesterday’s post, I mentioned how my experimentation with AI character generation set my writer’s brain to work. Well, I’ve been busy at work putting together the scenario for this new series I’m going to go ahead and announce here. The title of the series is My ASO is… (「俺の補佐官とは…」 in Japanese) with the rest of the title following for each entry. “ASO” being “Assistant Staff Officer”. I wanted a rather unglamorous setting, so it’s based around a battalion S-4 office. The protagonist is Captain Giancarlo di Stefano Mastriano, the newly assigned S-4 Officer. The eponymous ASO is Lieutenant Maya e Rafael u Sheeda. The twist is that this is a science fantasy story where isekaied humans live among all these nations of fantasy races. The story takes place about a thousand years after the isekai incident, so the humans from “Urth” are pretty well integrated by this point. The location is the Republic of Liberia, a relatively new nation that serves as the nexus of the Grand Alliance. Over 50% of the population are dual citizens from the other nations, so there’s plenty of tension and intrigue to be had. Maya, for instance, is actually the daughter of a powerful Elven noble/businessman, so there’s all sorts of stuff going on just with her situation, and of course that’s just the beginning.

The threat comes from the fact that the power of the Grand Alliance’s mutual enemy the Demon Army peaks every 1000 years and the story starts in late NC 1997. While the series is going to start out as slice-of-life comedy, things will get more serious as we progress. The reason this is in If Arc is because when I was playing around with the character generator, I decided to try and see if I could make Lydia Han from the Tico series. Days AI’s character generation only has four settings for age and the second to last for females still produces what is at most a late teen. As I only have three character slots available on the free plan, I decided to run with it and so I had Lydia as a midshipman. I then had to figure out how to actually fit her into the story. I made the setting a joint forces base and so we have a reason for Army and Navy personnel to be intermingling. Have the base be the site of the Naval Academy’s pilot school and we have the reason for her to be there. Then I just had to figure out why CPT Mastriano would be involved with a Navy midshipman, asked by her father to look after her. Well, a solution came to me pretty quickly. CPT Mastriano’s father and Luka used to serve together, so the good Captain is actually a friend of the family, a classmate of Lydia’s sister Rachel (the one who passed before she was born). In the main series, Lydia was working hard to play the part of the good girl once she turned things around in high school and went on to the Academy. It was only the death of her sister that caused her to revert to her bad behavior from before. Well, here she isn’t quite as reformed during this phase of her life, so you have this prickly tsundere highly annoyed all the time at our protagonist involving himself in her affairs, yet she still frequently gets roped into shenanigans. I’m sure to integrate more Tico characters in later, but right now I’m having a lot of fun just with Lydia. (Her chatbot is every bit the little punk I imagined her to be. It’s great.)

Speaking of the Han family, I decided to change big sister’s name from Leia to Lia. It’s a small change, but there’s a story behind it. Way back some thirty years ago, Lydia Han was Lydia Hancock and her sister was Leah. The Hancocks became the Hans and I decided to mix the Chinese heritage with Greek, so Luke Hancock became Luka Han and Leah became Leia. (If I ever get back to the Ticonderoga Detective Agency Files, Luka will actually have used ‘Luke Hancock’ as his assumed name when passing for white to serve in WWI.) Well, when my battle buddy read “Casualty Notification” years back, he criticized me for having Luke, Han and Leia right there on the page. It certainly wasn’t my intention, but there it was. (As a side note, when I was naming Mark from the KoG series, I first looked at the Four Gospels and dismissed Luke on account of Star Wars and John as being too common. It was in the back of my mind when naming Luka, but I ran with it anyway. Luke Skywalker isn’t the only Luke in the universe, after all.) I was pleased to find that in modern Greek, as opposed to the Koine I initially referred to, ‘Lia’ is an acceptable rendering of the name, so now over 15 years later, I can deflect that criticism. An elephant never forgets, and neither do I, apparently. I then had to propagate this change across all my materials online and off. It wasn’t as much work as I was expecting, but it was still a bit of work.

Anyway, there’s tons of work to be done, but I also need to be thinking about work-work. There’s all the stuff building up for the new school year, so I do need to devote at least some of my attention on that. I of course need to work on CoP and write up manga reviews as main site-related activity, all while I’m continuing my AI experiments and fleshing out this new scenario. There simply aren’t enough hours in the day. (And this is with me having only slept two hours last night.) I will make a solid effort not to wreck myself in the meantime. If I have any sense, I’ll take advantage of this week’s half-day schedule to catch up on sleep and from Friday I’ll be taking a week off, so hopefully I can bring things back into balance before the new school year starts. If I don’t, all the stress of the first term mixed in will probably kill me. I’m going to try to avoid that. We’ll see what comes from here. Stay tuned.